China AI Monthly

AI competition is intensifying as Meta, Xiaomi and StepFun leapfrog in rankings, while the US-CN intelligence gap has widened to 20% from 12%. The API price gap reached 70%, while OAI cut lighter GPT-6 prices by 50-56%, putting more pressure on Chinese models. Anthropic also priced Opus 5.5 24% below Opus 5.0. World models could be the next big thing, but their heavy compute and memory needs pose further ROI challenges.

LLM competition intensifies as laggards leapfrog in rankings and new players enter. In Sep, US models occupied the top four spots in Artificial Analysis' (AA) intelligence ranking, with Anthropic retaining the lead with Claude Opus 5.5, followed by OAI's GPT-6 Astra, while Meta jumped to No.3 with Muse Spark 1.3. Chinese models occupied No.5-9, led by Xiaomi's MiMoV2.6 Pro, Qwen3.8 Max, and GLM-5.3. Importantly, previous laggards such as Meta (No.3), Xiaomi (No.5), and StepFun (No.8) largely jumped in the rankings, highlighting how difficult it is for LLM vendors to sustain technological leadership. Four new models also entered the ranking in Sep (here), pointing to intensifying competition and lower-than-expected entry barriers.

US-CN intelligence gap widens; aggressive RL can boost Chinese models but has tradeoffs. The US-China frontier intelligence gap widened to 20% in Sep from 12% in Aug on consistent criteria. We believe it could widen further, as US labs benefit from earlier access to next-generation GPUs such as Rubin and stronger anti-distillation measures. Chinese AI players can partly offset compute constraints through more intensive reinforcement learning (RL), boosting intelligence without drastically increasing model size, but this has drawbacks. Aggressive RL can weaken performance in areas without standardized answers, such as creative writing and business analysis, increase hallucinations due to overconfidence, raise token consumption and costs, and lengthen inference time. We recently hosted AI expert William Fong, who observed that Western enterprises rely more heavily on frontier US models for complex task planning and result verification given their lower hallucination rates (here).

The API price war is intensifying, putting more pressure on Chinese AI labs. The average API price gap between US and Chinese frontier models widened to 70% in Sep from 60% in Aug, partly because OAI launched GPT-6 Astra at a 150% premium to GPT-5.6 Sol. But competitive pressure is rising even in the US: Anthropic priced Opus 5.5 24% below Opus 5.0, while OAI cut GPT-6 Sol/Luna prices by 50%/56% versus their GPT-5.6 equivalents. These lighter models target less complex workloads, with GPT-6 Luna now 74% cheaper than DS V4.1 Flash, suggesting OAI is pushing into the mid- to low-end segments dominated by Chinese models. Sustained high capex, falling API prices, and more competitors are negative for AI-lab economics, particularly in China. We continue to prefer platform players, especially CSPs, given their multiple monetization channels and ability to cross-sell AI services into sticky enterprise and consumer user bases.

World models (WM) could be the next big thing for AI, but are still at very early stage. World models aim to understand and simulate characteristics of the physical world, with potential applications in autonomous driving, robotics, gaming, and entertainment. They would require much more compute and memory, making ROI an even bigger challenge. This is Part 1 of our deep dive into this emerging field. Notably, AMD announced yesterday its US$8.2bn acquisition of prominent world-model startup World Labs.

World models are still at an early stage, with large tech companies and start-ups betting on multiple technological routes. A world model is an AI system’s internal simulation of how the world works, enabling it to predict future states and plan actions by “imagining” outcomes before acting, much like humans mentally rehearsing an action before taking it. WMs matter because they can model how the physical world may change in response to an action, a capability that LLMs are not designed to provide. This could extend AI into physical-world applications, including autonomous driving, robotics, and robotaxis. We believe WMs will complement rather than replace LLMs, and large tech companies are investing in both. The technology is still nascent, with seven different routes emerging (Table 2). Among the key players, NVDA (Cosmos 3), World Labs (Atlas), Waymo, and Tesla are combining multiple approaches within individual products, while Google DeepMind, Meta, Alibaba, ByteDance, and Tencent are pursuing multiple technological routes in parallel.

WM use cases are emerging across multiple verticals, but commercialization remains highly uneven. 1) Already commercializing: The clearest ROI today is in autonomous-vehicle simulation and validation, where WMs can generate rare or dangerous corner cases that are costly or impractical to capture through real-world testing, lowering validation costs while potentially improving safety. 2) Monetizing, but not pure WM plays: Video/creative content and industrial digital twins are already generating revenue, but the former faces uncertain unit economics and copyright risks, while the latter represents more of an AI-driven upgrade to an existing revenue pool than a new WM market. 3) Next likely areas to scale: architecture, real estate, and design could benefit from explorable 3D environments integrated into CAD workflows, while defense/synthetic environments could benefit from large autonomy budgets and government interest, although WM-specific revenue remains difficult to isolate. 4) Commercial but unprofitable: Robotaxis and warehouse/factory robots are commercially deployed but remain structurally unprofitable, constrained by high hardware costs, long payback periods and limited autonomy in complex realworld environments. 5) Still largely at the demo stage: Interactive gaming and general-purpose humanoids remain largely experimental, characterized by short-duration demonstrations, limited business models, and particularly for humanoids, continued reliance on teleoperation, rather than true autonomy.

WMs could be very positive for AI compute and memory demand, but the upside is routedependent and more importantly, driven by the use cases' profitability. The most compute- and memory-intensive routes are generative video and interactive neural engines, which continuously generate high-dimensional visual outputs and often rely on diffusion-based architectures, making inference substantially more resource-intensive than typical LLM workloads. Memory could become an even greater bottleneck as longer, higher-resolution, and more interactive environments require models to maintain increasingly large amounts of contextual and state information. This supports a positive LT read-through for data center GPUs and HBM suppliers if generative WMs scale across content creation, gaming, autonomous-vehicle simulation, and synthetic-data generation. However, the thesis carries meaningful risk: robotics-focused approaches, such as JEPA, could be substantially lighter and potentially run on-device using LPDDR rather than HBM. If such non-generative architectures ultimately dominate physical AI, the incremental HBM opportunity would be more limited. The potentially much higher compute and memory cost also mean the use cases that WM helps to commercialize would need to be profitable enough to cover the much higher hardware costs.

The 3-Minute Guide

LLM Competition has Intensified, as Laggers (Meta, Xiaomi, Stepfun) Leapfrogged and New Non-US/CN Players have Entered.

Laggers were Able to Enter the Frontier Cohort, Again Proving Our View that LLMs do NOT have Moats.

As Expected, in 2H26, US-CN Frontier Intelligence Gap Widened to 20% in Sep from 12% in Aug.

Overall, US Models still Lead on Agentic Capability, which is a Crucial Pre-requisite for Model Recursive Self-Improving.

US-CN Model Price Gap Increased to 70% in Sep from 60% in Aug as OAI Hiked GPT-6 Astra Price by 150% (US$7.7/mn tokens).

Meanwhile, OAI Lowered Mid-/Low-Tier Models' API Prices by ~50%, Posing Strong Competition to Chinese Open-Weight Models.

横向滑动查看完整图表

Claude Opus 5.5, GPT 6 Sol, Mimo V2.6 Pro and GPT 6 Luna Now Sit at the Pareto Frontier.

Key Traits of a World Model: 1) Representation of the Environment, 2) Prediction of What Happens to an Object, and 3) When Condition Changes, it Will Plan and Act Accordingly.

Required Property 1. Representation. WM recognizes the object and knows its location and relationship with other objects. For example, it recognizes there is a glass on the edge of the table.

Required Property 2. Prediction. It can simulate and predict what will happen to an object. For example, it knows a glass would fall and break rather than float when it got pushed off the table.

Required Property 3. Plan and Action. When condition changes, it will plan and take actions accordingly. For example, it will move the glass to center of the table to avoid it from falling. If the glass is replaced with a cat, it will not take action.

World Model is A Crucial Component of Physical AI that Helps to Predict What Happens Next.

Currently, There are Seven Technological Routes to Developing World Models.

横向滑动查看完整图表

The Seven Tech Routes can be Categorized into Four Quadrants. Vertically, the Generative vs. Predictive Camp, Horizontally, Implicit Understanding vs. Explicit Structure.

Among All the World Model Use Cases, Autonomous Vehicle Simulation and Validation is the Most Commercially Ready One.

横向滑动查看完整图表

World Model Development Still at Very Early Stage, But there are some Indicators Worth Following...

横向滑动查看完整图表

China’s AI in a Global Context – Sep 2026

Global LLM Price War; Deep Dive into World Models (Part 1)

Executive Summary

Four US Models Led the AA Intelligence Benchmark in Sep; Anthropic’s Opus 5.5 Now Ranks No.1

• In Sep, Artificial Analysis (AA) released Intelligence Benchmark V4.3, adding AutomationBench-AA, a new agentic business-workflow benchmark, and removing τ³-Banking and GPQA Diamond. The updated index places greater weight on agentic and general capabilities, each at 30%, while reducing coding and scientific reasoning to 20% each.

• Anthropic’s Opus 5.5, OpenAI’s GPT 6 Astra, Meta’s Muse Spark 1.3 and SpaceXAI’s Grok 4.7 ranked in the top 4 globally on the latest updated AA model intelligence ranking in Sep. Meta’s ranking jumped from 6th in Aug to 3rd, while Kimi K3 fell from 4th to 9th.

• Xiaomi’s Mimo V2.6 pro overtook Kimi K3 (based on our Aug report) to be ranked 1st among Chinese frontier models and 5th globally, followed by Alibaba’s Qwen 3.8 Max 0902, Zhipu’s GLM 5.3 Max, StepFun’s Step 5 Preview (new) and Moonshot’s Kimi K3.

Anthropic and OAI Widened the Lead; More New Players Emerged

• In the US, Anthropic and OpenAI remain the world’s leading AI labs with state-of-the-art (SOTA) models. The two frontier labs released Claude Opus 5.5 and GPT 6 Astra in Sep. Meanwhile, Muse Spark 1.3 and Grok 4.7 are now ranked no.3/4 for model intelligence by AA.

• In China, the leader keeps changing. Xiaomi’s Mimo V2.6 pro and Qwen 3.8 Max 0902, GLM-5.3 are the top 3 Chinese models, while StepFun’s Step 5 Preview is now ranked the 4th best model in China. The fact that Xiaomi and StepFun’s new models were able to enter the frontier cohort signals intensifying competition in the Chinese model market.

• Four new models entered the ranking, three of which are non-US/China, indicating an even more crowded industry.

US-CN Frontier Performance Gap Widened from 12% to 20%, But Could Widen Further in Next 12

Months

• Currently, the leading Chinese frontier model (Mimo V2.6 pro) has reached ~80% of the intelligence level of the top U.S. frontier model (Claude Opus 5.5)

• The newly added AutomationBench-AA benchmark tests whether AI agents can handle real, multi-step work tasks in business software. AA sets a high bar: an agent only scores if it finishes the entire task without breaking any guardrails.

• Based on consistent evaluation criteria, the model intelligence gap between the US and CN widened from 12% in Aug to 20% in Sep.

• U.S. AI labs’ growing advantages in access to cutting-edge GPUs (NVDA Rubin) and growing anti-distillation measures could widen the performance gap between Chinese and U.S. frontier models in 2H26.

Claude Opus 5.5, Grok 4.7, and Qwen 3.8 Max have the Best Agentic Capability

• As Artificial Analysis stopped publishing its Agentic Index in Sep, we estimate frontier models’ agentic capability by blending their AA-Briefcase, GDPval-AA, and AutomationBench scores at a 15:10:5 ratio (matching their weights in the AA Intelligence Index).

• US models lead the pack: Claude Opus 5.5 ranks first, followed by Grok 4.7, Muse Spark 1.3, and GPT-6 Astra. Muse Spark 1.3 was the biggest climber in Sep.

• Among CN models, Qwen3.8 Max scores the highest in terms of agentic capability and ranks 3rd globally, followed by GLM-5.3, MiMo V2.6 Pro, Kimi K3, Step 5 Preview, and DeepSeek V4 Flash Vision. Models from non-US/CN regions lag in terms of agentic capability.

API Pricing Gap Between US and CN Frontier Models Widened Again in Sep

US – CN frontier model pricing gap widened from 60% in Aug to 70% in Sep, mainly because OpenAI’s GPT-6 Astra

• launched at premium pricing, while China’s average was dragged down slightly by DeepSeek’s 74% price cut on DS V4.1 Flash.

The average price of Chinese models has doubled since Jan 2026, driven by both rising inference cost (high-end GPU

• and memory price hikes) and Chinese models’ improving performance.

We introduce another version of US API pricing by including lighter versions of US frontier models – we call it US all-

• in API pricing. China’s API pricing gap vs US all-in API pricing expanded from 47% in Aug to 52% in Sep.

The Most Powerful US Model is Cutting API Price to Gain Market Share

• Anthropic released Claude Opus 5.5, its latest flagship and best-performing model, on September 22. It achieved an AA intelligence score of 58 — a 9% improvement over the previous leading model Fable 5.1, while cutting API pricing by 24% (vs Opus 5). The release reflects Anthropic’s cross-series development strategy: more cost-efficient Opus models surpass the preceding Fable generation, narrowing the gap until the next Fable-class model pushes the performance frontier higher.

• Claude Opus 5.5 leads globally, and is ranked No. 1 on the AA Intelligence leaderboard. It delivered a 9% gain over Anthropic’s previous flagship, Fable 5.1, and outperformed GPT 6 Astra by 10%, Muse Spark 1.3 by 21%, Grok 4.7, and Mimo V2.6 Pro by 25%, and Kimi K3, Qwen 3.8, and GLM 5.3 by an average of 30%.

• Stronger performance but lower API pricing suggests Anthropic is targeting greater market share and higher ARR. Claude Opus is priced at $0.2/$4/$20 per million tokens for cache/input/output, respectively, resulting in a blended price of $3.2 per million tokens based on a 7:2:1 cache-to-input-to-output ratio. This blended price is 24% lower than the previous Opus 5 model, and 56% lower than Fable 5.1. We believe this pricing reflects lower inference costs from infrastructure optimization and a strategy to capture market share as Anthropic seeks higher ARR and valuation ahead of its postponed IPO.

• Claude Opus 5.5’s lower API pricing is a clear sign that competition among AI model vendors is escalating. By moving away from premium pricing and setting blended pricing at $2.9 per million tokens—an all-time low—Anthropic has reduced its premium over the most expensive Chinese model, Kimi K3, to just 43%, from 226% previously, and is likely to force other rivals to respond.

US AI Labs Introduced Light Versions of their Flagship Models to Compete with Chinese Models

• The intelligence level of the US lighter model versions is still very competitive against the Chinese models. For example, Claude Opus 5 and GPT 6 Sol beat all Chinese models in intelligence score.

US AI Labs Cut Prices by 24%-80% for their Lighter Models to Compete with China

• US frontier models retain stronger pricing power than their Chinese peers, given the current US-CN model performance gap.

• In the US, OpenAI launched its new frontier GPT 6 Astra at premium pricing in Sep, raising blended prices by 150%, but at the same time, it lowered prices for GPT 6-Sol and GPT 6-Luna by 50%/56% compared to GPT 5.6 Sol/Luna. Meanwhile, Anthropic lowered Claude Opus 5.5 pricing by 24%.

• In China, DS cut the price of DS V4.1 Flash by 74% compared to DS V4 Pro in Sep, while StepFun increased the price of Step 5 Preview by 187%.

横向滑动查看完整图表

Light Versions of US Frontier Models have Become Much More Price-Competitive

• AA’s cost-per-task metric measures frontier models’ weighted average cost for answering a single question (task) when going through all the benchmark tests. Cost-per-task depends on both API pricing and inference efficiency.

• OpenAI released GPT-6 Sol and GPT-6 Luna on September 22 at headline prices that are 50% and 56% below those of their predecessors, respectively, citing caching and inference efficiencies.

• In particular, GPT-6 Luna is priced at US$0.07 per task, 74% lower than DS V4.1 Flash’s US$0.27. This means OAI now not only aims to provide SOTA models, but also competes directly with Chinese vendors in value-for-performance models.

Chinese Models are Still Cheaper than US Models Despite Tendency to Overthink

• Although most Chinese models have low blended API prices, their actual cost per task is not as low as it appears.

• The cost per task is a function of API price and the number of inference tokens. Since Chinese frontier models have adopted heavy RL to improve their intelligence, they tend to “overthink” (think longer and multiple rounds) in inference to ensure accuracy. This increases the model’s cost per task.

• But even so, leading Chinese models are still substantially cheaper than full-sized, frontier US models.

横向滑动查看完整图表

Claude Opus 5.5, GPT 6 Astra/Sol, Mimo V2.6 Pro, and GPT 6 Luna Now Sit at the Pareto Frontier

Claude Opus 5.5, GPT 6 Astra/Sol, Mimo V2.6 Pro, and GPT 6 Luna now sit at the Pareto frontier as they offer the

• most intelligence per dollar.

Claude Opus 5.5 sits at the premium end, offering the highest intelligence scores, but also has the highest cost per

• task ($5.98/Task).

GPT-6 Sol now directly competes with frontier Chinese models like GLM-5.3, Qwen3.8 Max, and Kimi K3, while GPT-6

• Luna offers the most value-for-performance and competes with DS V4.1 Flash.

A Deep Dive into World Model – Part 1

AMD just acquired the world model startup World Labs for US$8.2bn

A World Model Answers “What Happens If I Do This” — Not Just “What Happens Next”

• A world model is an AI system’s internal simulation of how the world works, which allows it to predict what will happen next based on an action, and plan actions by “imaging” outcomes before acting, much like how people mentally rehearse a move before making it.

• Unlike systems that simply classify inputs or generate outputs, world models aim to simulate dynamic processes, including physics, object interactions, and causal relationships.

• Yann LeCun (ex-Chief AI Scientist of Meta) argues that language carries very little information compared with what a child takes in by watching the world. A system trained only on text, therefore, has no grounding in physical reality, no lasting memory, and no real ability to plan. Thus, he believes LLMs are showing real progress, but in the wrong direction. He even predicted that in five years, nobody would use GPT-style LLMs.

• World models will be a hugely important new development area because they give AI something LLMs lack, which is the ability to predict what will happen in a physical world if an action is taken.

REQUIRED PROPERTY 1 Representation WM recognizes the object and knows its location and relationship with other objects. For example, it recognizes there is a glass on the edge of the table.

REQUIRED PROPERTY 2 Prediction It can simulate and predict what will happen to an object. For example, it knows a glass would fall and break rather than float when it got pushed off the table.

REQUIRED PROPERTY 3 Plan and Action When condition changes, it will plan and take actions accordingly. For example, it will move the glass to center of the table to avoid it from falling. If the glass is replaced with a cat, it will not take action. Source: Jefferies

World Model is A Crucial Component of Physical AI that Predicts What Happens Next

• Physical AI is artificial intelligence that perceives, reasons, and then acts on the physical world. It includes robots, autonomous vehicles, drones, and automated factories, etc.

• World models are the predictive component of physical AI: the part that works out what will happen if the machine acts, before it commits.

• Learning without breaking things. A robot’s training data does not exist until it moves, and every bad move damages something. The world model lets most learning occur via simulation, with only the last fraction on real hardware.

• A training-time tool, not yet a runtime one. In most deployments, the world model generates scenarios and validates policies offline; the robot runs a compact vision-language-action (VLA) model that makes the decision.

Role of a World Model and VLA Model in Autonomous Driving Scenario

How World Models Differ from Language Models

• A language model is a machine that has read thousands of texts on a glass breaking when it falls from a table, but does not really understand the physics of it.

• A world model is a machine that has watched the glass breaks, understands the physics of it, and can predict it (e.g. when it sees a cat push a glass off a table).

• Both turn what they are given – words or camera image –into a compact set of numbers, then use the same kind of engine, a transformer, to look at everything so far, and predict what comes next. Both learn that simply by practicing on vast amounts of ordinary material, with nobody labeling anything or teaching them the rules.

• World models, VLA models, and LLMs all require compute, memory, and electrical power for hardware infrastructure. They also share the same pool of AI research and engineering talents.

How a World Model Differs from an LLM

横向滑动查看完整图表

Will World Models Make LLMs Obsolete?

• World models are unlikely to replace LLMs; they are more likely to complement each other.

• Language is the interface: People control world models with words, and an LLM inside turns those words into instructions the world model can follow. World models don’t replace LLMs; they depend on them to understand what users want.

• Most valuable work isn’t physical. Code, contracts, research, knowledge, and customer service all work in language and symbols, where a model of physics adds little. Demand for LLMs would not shrink as robots get better.

• The two are merging, not competing. NVDA’s Cosmo 3 combines reasoning, world generation, and action prediction in a single model, and Google builds its robotics models on Gemini.

• LLMs already form a partial picture of the world. Harvard researchers showed that a model trained only on moves from the board game Othello has built an internal map of the board. But those internal pictures are patchy and fragile, which is exactly why connecting LLMs to world models helps.

• The economics favor LLMs for most tasks. A second of 720p video takes 4,000x more tokens than a second of text. OpenAI shut Sora down in Mar 2026, partly to free up compute for its core LLM products. For anything that can be done in words, an LLM will remain far cheaper than a world model.

• Large AI players have invested in both LLMs and world models.

• What would change our view: 1) A WM-first system beating LLMs at language and reasoning; 2) AMI Labs or similar lab turning JEPA-style models into a real product; 3) LLM revenue stalling while physical-AI revenue scales.

The Seven Technological Routes to World Models (1)

• Currently, there are seven technological routes to develop world models.

• Each route is trading against the same constraint — how much of the world you keep track of versus how cheaply you can run.

• The most crowded route is generative video, where Chinese labs are most competitive.

• The likely end-state is a combination of multiple routes: NVIDIA’s Cosmos 3 explicitly combines vision reasoning, world generation, and action prediction in a single mixture-of-transformers model; World Labs’ Atlas combines generation with explicit geometry; Waymo’s world model is a generative video system post-trained to emit LiDAR.

横向滑动查看完整图表

The Seven Technological Routes to World Models (2)

• We classify the seven technical routes using a four-quadrant chart.

• The horizontal axis runs from implicit understanding of the physical world hidden inside the WM, to explicit WM structure you can inspect, such as geometry, equations, or words.

• The vertical axis runs from generating a visible world (generative route), to predicting an action with no visual output (the predictive route).

Source: Jefferies

World Models vs. LLMs: Compute and Memory Intensity (1)

• Route 1 (generative video) and Route 3 (interactive neural engines) clearly cost more in terms of compute and memory than LLMs. Both routes use diffusion models to generate video pixels frame by frame.

• Route 4 (3D splatting) and route 5 (physics engine hybrids) are heavier only in one generative step each, but that step could also be compute-intensive.

• Compute: generative world models require far heavier compute per second of output than LLMs:

o Training is not where the difference lies: building a large world model takes a training budget roughly comparable to a frontier LLM. (Cosmos: 20m hours of video, reportedly ~10,000 H100 GPUs over ~3 months; Meta Movie Gen: 1.25m GPU-hours across 6,144 GPUs). The gap opens up in inference: the everyday cost of running the model each time a user or robot asks it to generate something.

o An LLM writes text one small piece (a token, roughly a word) at a time, and each piece is cheap to produce. A generative world model generates video, and a single frame is equal to thousands of these pieces (~16,000 per second of 720p, against ~4 for text).

o Most world models also use diffusion: they start from noise and refine the whole block of frames over dozens of passes, so each second of output consumes compute required to generate many more seconds of content.

o The cost gap shows up in real workloads. Generating a 5-second clip with the open-source HunyuanVideo model takes about an hour on a single A100 GPU. It is estimated a 10-second Sora clip would require about 40 GPU-minutes, and costs roughly US$1.30.

World Models vs. LLMs: Compute and Memory Intensity (2)

• Memory capacity: the clearest pressure point

o To keep a scene consistent (a room that stays the same room, an object that doesn't vanish when the camera turns), a world model has to keep a record of everything it has already generated. This short-term memory is the KV cache, and it grows steadily the longer the model runs.

o The numbers get big quickly. The WorldKV paper (May 2026) found that for a mid-sized 14-billion-parameter world model, the required memory exceeds 200GB after just one minute of interactive output. That exceeds the HBM on even the most advanced single GPU available today, so the workload has to be split across several chips.

o As the memory fills up, performance drops. In the same study, generation speed fell from about 8.9 to 3.6 frames per second over that one minute. In practice, memory capacity is what limits how long a world model can stay coherent, which is the central technical challenge the industry is trying to solve.

• Implications

o Generative world model routes need more compute and even more memory capacity, which is positive to GPU and memory players.

o HBM demand sits in the datacenters, where generative models train and produce simulated environments. The routes built for robot decisions (JEPA, Dreamer) are light enough to run on the robot’s local chip, which uses cheaper LPDDR memory, not HBM.

o Route selection therefore directly impacts compute and memory demand. If non-generative routes win in robotics, less of the physicalAI buildout turns into HBM and compute demand. Generative demand for content, gaming and synthetic training data would remain either way.

o Suppliers haven't confirmed the link yet: SK hynix's 2026 outlook does not cite world models as an HBM driver.

o Near term, LLMs are likely to remain the dominant source of HBM demand, given their far larger user base.

Data: Video is Abundant; Video with Actions Attached is Scarce

• Raw video is not scarce. Epoch AI puts the world’s usable public text at ~300 trillion tokens; Cosmos’s single 20m-hour training set implies several times that entire stock. There is no video wall in any meaningful sense.

• Action-labeled data is desperately scarce. AgiBot World holds over a million trajectories from 100+ robots; DROID is 76,000. Against a billion hours of internet video, the ratio is roughly six orders of magnitude.

• The economics is stark: real demonstration clips cost $50–500 each; synthetic equivalents $1–1.50. That 30–50× gap is why every serious player has a synthetic-data strategy. Unresolved: how much of a policy’s training can be synthetic before it degrades.

• Three bridges to watch: egocentric video from smart glasses; fleet data (Waymo ~200m autonomous miles, Huawei ADS in 1.9m vehicles); and state-organized data collection in China, with a Beijing target of 10,000 humanoids in commercial use by year-end.

• Legal risk is materially higher than for video. Video carries identifiable people, recognizable characters, and studio copyright. Any world-model business trained on scraped video carries an unquantified liability that a text model does not.

Unsolved Problems on the Algorithm Side

• Long-horizon consistency is the central blocker. The best public interactive world model, Google's Genie 3, keeps a scene coherent for a few minutes and remembers only about one minute back. A robot task or a game session needs hours. Small errors pile up frame after frame until the world drifts apart. At least five research papers in 2025–26 tackled this, but none have solved it.

• The world stops when you look away. In a real room, a kettle keeps boiling while your back is turned. In generated worlds, hidden events usually freeze. A June 2026 benchmark of 23 models found this across the board. Scaling one model family from 1.3B to 14B parameters made it worse, which suggests the fix is a design change, not more compute.

• Looking real is not the same as being physically correct. Leading image-to-video models score only about 25–40 out of 100 on Physics-IQ, the main benchmark for whether generated video obeys physics. The benchmark's authors say bigger models and better prompts are not enough to close the gap, and scores need to roughly double before generated worlds are safe to train robots in.

• No agreed way to measure progress. At least eight competing benchmarks (Physics-IQ, WorldModelBench, VBench-2.0, RoboArena and others) launched within about 18 months, each arguing the others are inadequate. Without a commonly accepted benchmark, buyers can't compare vendors, and vendor claims can't be checked.

• Learning actions from passive video is difficult. Models learn from vast amounts of ordinary video, but a robot needs to know what caused each change: which movement, which force. Linking actions to outcomes is still weak, which is why action-labeled data remains the bottleneck (thousands of hours of it, against a billion hours of internet video).

• The non-generative route has its own problems. JEPA's training goal is hard to define: which parts of the future should the model keep, and which should it ignore? Its understanding is also invisible, stored in abstract codes with no picture to check against, so progress is hard to verify or demonstrate.

• The sim-to-real gap persists. Even when a robot learns well inside a generated world, some of that skill is lost on real hardware. How much of a robot's training can be synthetic before performance drops is still unpublished.

Among All Use-Cases, Autonomous Vehicle Simulation and Validation is the Most Mature One

横向滑动查看完整图表

World Model Still at Early Development Stage, But these Indicators are Worth Following

横向滑动查看完整图表

China vs US AI Models: A China-Driven Price War May Have Started

Chinese Models Continue to Offer Superior Value for Performance vs. US Peers

Four US Models Led the AA Intelligence Benchmark in Sep; Anthropic’s Opus 5.5 Now Ranks No.1

• In Sep, Artificial Analysis (AA) released Intelligence Benchmark V4.3, adding AutomationBench-AA, a new agentic business-workflow benchmark, and removing τ³-Banking and GPQA Diamond. The updated index places greater weight on agentic and general capabilities, each at 30%, while reducing coding and scientific reasoning to 20% each.

• Anthropic’s Opus 5.5, OpenAI’s GPT 6 Astra, Meta’s Muse Spark 1.3, and SpaceXAI’s Grok 4.7 ranked in the top 4 globally on the latest updated AA model intelligence ranking in Sep.

• Xiaomi’s Mimo V2.6 pro is ranking at 1st among Chinese frontier models and 5th globally, followed by Alibaba’s Qwen 3.8 Max 0902, Zhipu’s GLM 5.3 Max, StepFun’s Step 5 Preview, and Moonshot’s Kimi K3.

AA Updated Intelligence Score Methodology in Sep Focusing More on Agentic Capabilities

• Since our last-published report, Artificial Analysis has updated its methodology twice, in August (V4.2) and September (V4.3). The latest methodology for measuring LLMs’ overall intelligence is weighted mainly toward general and agentic capabilities, which accounted for 60% combined.

• Newly-added tests, including AA-Briefcase, GDP.pdf, and AutomationBench, indicate that the new benchmark version focuses more on real-world, long-context, multi-step, and professional tasks.

横向滑动查看完整图表

US-CN Frontier Performance Gap Widened from 12% in Aug to 20% in Sep; New Players Emerged

• In the US, Anthropic and OpenAI remain the world’s leading AI labs with state-of-the-art (SOTA) models. The two frontier labs released Claude Opus 5.5 and GPT 6 Astra in Sep. Meanwhile, Muse Spark 1.3, and Grok 4.7 are now ranked no.3/4 for model intelligence by AA.

• In China, the leader keeps changing. Xiaomi’s Mimo V2.6 pro and Qwen 3.8 Max 0902, GLM-5.3 are the top 3 Chinese models, while StepFun’s Step 5 Preview now ranked as the 4th best model in China. The fact that Xiaomi and StepFun’s new models were able to enter the frontier cohort signals intensifying competition in the Chinese model market.

• Four new models entered the ranking; three are non-US and non-China, indicating an overly-crowded LLM market.

US AI Labs Introduced Light Versions of their Flagship Models to Compete with Chinese Models

• The intelligence level of the US lighter model versions is still very competitive against the Chinese models. For example, Claude Opus 5 and GPT 6 Sol beat all Chinese models in intelligence score.

Claude 5.5 Opus: The SOTA Model is Cutting API Price to Gain More Market Share

• Anthropic released Claude Opus 5.5, its latest flagship and best-performing model, on September 22. It achieved an AA intelligence score of 58 — a 9% improvement over the previous leading model Fable 5.1, while cutting API pricing by 24% (vs Opus 5). The release reflects Anthropic’s cross-series development strategy: more cost-efficient Opus models surpass the preceding Fable generation, narrowing the gap until the next Fable-class model pushes the performance frontier higher.

• Claude Opus 5.5 leads global peers, ranking No. 1 on the AA Intelligence Ranking and setting a new state of the art (SOTA). It delivers a 9% gain over Anthropic’s previous flagship, Fable 5.1, and outperforms GPT 6 Astra by 10%, Muse Spark 1.3 by 21%, Grok 4.7, and Mimo V2.6 Pro by 25%, and Kimi K3, Qwen 3.8, and GLM 5.3 by an average of 30%.

• Stronger performance but lower API pricing suggests Anthropic is targeting greater market share and higher ARR. Claude Opus is priced at $0.2/$4/$20 per million tokens for cache/input/output, resulting in a blended price of $3.2 per million tokens based on a 7:2:1 cache-to-input-to-output ratio. This blended price is 24% lower than the previous Opus 5 model, and 56% lower than Fable 5.1. We believe this pricing reflects both lower inference costs from infrastructure optimization and a strategy to capture market share as Anthropic seeks higher ARR and valuation ahead of its postponed IPO.

• Claude Opus 5.5’s lower API pricing is a clear sign that competition among AI model vendors is escalating. By moving away from premium pricing and setting blended pricing at $2.9 per million tokens—an all-time low—Anthropic has reduced its premium over the most expensive Chinese model, Kimi K3, to just 43%, from 226% previously, and is likely to force other rivals to respond. This price cut will likely increase pressure across the market.

GPT 6 Astra: Strong Performance, but Not Yet AGI

• OpenAI released its new flagship model GPT 6 Astra on Sep 3, 2026. GPT 6 Astra ties Claude Fable 5.1 in AA’s Intelligence Index at 53 (+12% vs GPT 5.6 Sol) and ranks 2nd globally. Its agentic capability outperforms Fable 5.1. On AA’s AutomationBench, GPT 6 Astra scores 15% higher than Fable 5.1.

• Computer use is the standout upgrade: OpenAI reports 72.6% for GPT 6 Astra on OSWorld 2.0 (long-horizon computer use benchmark), versus 65.7% for GPT 5.6 Sol (+6.9ppt). GPT 6 Astra achieved at ~40 minutes per task vs ~75 minutes for GPT 5.6 Sol, which is 46% faster. Three product innovations underpin the improvement: asynchronous function calls, mid-turn steering and reasoning-effort adjustments that preserve the prompt cache. Together, these strengthen execution control and improve the economics of long-running agentic workflows.

• The looped-transformer architecture remains unconfirmed: GPT 6 Astra is widely reported to adopt a looped-transformer, or recurrent-depth design. It applies a weight-shared layer stack repeatedly, increasing effective reasoning depth without a proportional increase in parameters and potentially reducing token consumption per task. The approach remains a maturing research direction, but no confirmed adoption in any frontier lab’s flagship model to date. OpenAI has not officially confirmed that GPT 6 Astra adopts such an architecture.

• GPT 6 Astra has not yet reached AGI. At the launch briefing, OpenAI president Greg Brockman declared “Welcome to the AGI era,” citing a 99.9% score on ARC-AGI-3. However, the headline figure is harness-dependent rather than a property of the model alone. ARC Prize, an independent 3rd party that maintains the benchmark, published a result of 62.7% on launch day under its own Standard harness. ARC Prize itself declined to characterize the result as AGI. The gap indicates that frontier scores are not directly comparable across models when it comes to agentic workflows.

DeepSeek V4.1 Flash: Key Upgrades and Core Innovations

• DeepSeek released DeepSeek V4.1 Flash, its fastest flash model to date, on 10th September. The model has 552B total parameters + 196B Engram conditional memory, with 8B/16B asymmetric active parameters for prefilling/decoding. It sustains a higher token rate and greater throughput than DeepSeek V4 Flash, while exhibiting better performance at long context lengths. It has already been integrated into WorkBuddy, CodeBuddy, and OpenCode, and is accessible through DeepSeek Harness for custom agent workflows.

• Asymmetric architecture leads to higher cost efficiency in agentic tasks: DeepSeek V4.1 Flash adopts an asymmetric CED (Causal Encoder–Decoder) design, activating 16B parameters during decoding but only 8B during prefilling, out of 552B total parameters. This asymmetry improves cost efficiency for read-heavy, write-light agentic workflows.

• Decouples memory from computation using Engram conditional memory: DeepSeek V4.1 Flash’s 196B Engram conditional memory is queried only when needed and stays out of per-token dense computation, so the model can recall facts and long-context information cheaply without slowing generation.

• Extreme KV Cache Compression (DSA2): DeepSeek V4.1 Flash requires only 892 bytes of KV cache per token (75% lower vs DeepSeek V4 Flash’s 3,514 bytes), cutting HBM demand to 1/4 and SSD demand to 1/8 of the previous generation.

• Attention algorithm innovations lead to lower attention cost: DeepSeek V4.1 Flash adopts Compressed Sparse Attention 2 (CSA2) and Decoder Sliding-Window Attention (Decoder-SWA), where cross-layer KV reuse and Top-K indexing reduce attention cost, while the decoder replays only the last 128 tokens to avoid redundant compute.

• Cheaper and competitive pricing: DeepSeek V4.1 Flash is priced at RMB 0.04/2/8 per million cache/input/output tokens. At roughly USD 0.2 /Mn tokens blended price, it is among the cheapest Chinese models.

Mimo V2.6 Pro: Scaling Up Reinforcement Learning for Self-Improvement

• Xiaomi released its latest flagship model, Mimo V2.6 Pro, on Sep 21, retaining the same 1.02 trillion total parameters and 42B active parameters as its predecessor, Mimo V2.5 Pro. Mimo V2.6 Pro ranks 1st among frontier Chinese models and 5th globally on Artificial Analysis' Intelligence Index. Its performance gains show that post-training alone can still push model capabilities further without increasing parameter count, and mark a key step in Xiaomi’s exploration of recursive self-improvement (RSI).

• Mimo V2.6 Pro adds native multimodal capabilities at unchanged scale. It adds native image, video and audio-to-text modalities, making it a multimodal frontier model, and it further adds computer-use/agent-control capabilities. It also supports 3D spatial reasoning. It can be viewed as an LLM-as-world-model approach to world models in physical AI.

• A key reinforcement learning milestone on the path to RSI. Xiaomi expanded its trial-and-feedback RL loop across coding, office work, design, and cybersecurity, grading candidate solutions against one another for reliability. By comparing attempts relatively within a group, the approach delivers more accurate and diverse reward signals for long-horizon tasks and forms a closed loop for model self-improvement.

• Mimo V2.6 Pro adopts a unified multi-task reinforcement learning approach in post-training. Xiaomi’s AI team trained coding, general-agent, visual, and cybersecurity tasks together across different agent frameworks, enabling the model to transfer skills across tasks and tools.

• Mimo V2.6 Pro’s pricing remains unchanged at $0.2 per million tokens. We believe this combination of higher performance and lower pricing is pushing the Pareto frontier higher and is likely to put competitive pressure on its Chinese peers.

Step 5 Preview: Starting to Catch Up in Model Capabilities and Bringing Competitive Pressure

• StepFun, one of China’s star AI startup labs, released its latest flagship model Step 5 Preview on Sep 20. Notably, the model surpasses Moonshot’s Kimi K3 and ranks as the 4th best Chinese model. Step 5 Preview has 600B total parameters, with 27B active per token, and supports a 1M-token context window and vision input. Its improved performance and lower cost per task imply intensifying competition among Chinese model vendors.

• Step 5 Preview is a proprietary sparse mixture-of-experts (MoE) model with 600B total parameters and 27B active parameters per token—significantly larger than Step 3.7 Flash (198B total and 11B active). Despite this increase in scale, it activates only 4.5% of its parameters for each generated token, 1% below Step 3.7 Flash’s ~5.5% activation rate. This lower activation ratio concentrates computation on the most relevant experts while preserving the broader model’s capacity.

• Step 5 Preview is built for long-context workflows. The model is optimized for complex agentic tasks, including autonomous task execution, software engineering, and large-scale professional research, with a particular emphasis on financial use cases. StepFun has also demonstrated that Step 5 Preview can sustain long-running workflows and use execution feedback to identify issues, revise intermediate results, and improve its final output.

• Competitive pressure increasingly comes from delivering comparable outcomes at a lower cost. Step 5 Preview delivers frontier-model capabilities while offering a lower cost per task than its Chinese peers. It achieves an AA intelligence score of 44 at a cost of $0.72 per task, which is 64% lower than the cost per task of Kimi K3 ($2.0) and GLM 5.3 ($2.01). We believe Step 5 Preview is likely to put competitive pressure on its peers' premium pricing.

Meta Re-enters the AI Race with Muse Spark 1.3 and Launches Its First 2C AI Agent

• With the release of its flagship Muse Spark 1.3 model, Meta is back at the table, reentering an increasingly competitive AI race. The Muse Spark 1.3 flagship model has surpassed Grok 4.7 on Artificial Analysis’ Intelligence Index and ranks 3rd globally. Its model intelligence sits just behind Anthropic’s and OpenAI’s flagship models, signaling intensified competition in the US market. Its consumer app Muse has surpassed ChatGPT to rank 1st on the iOS free download leaderboard in September.

• Meta brings a consumer-facing personal AI assistant to everyday tasks. Meta has launched its new consumer-facing AI agent Muse on Sep 8, it is powered by Muse Spark models and can answer questions, complete tasks, browse the web, compare product prices, and help make purchases with the user’s permission. Unlike coding and workplace-focused agents such as Codex and Claude Code, Meta’s Muse targets the broader consumer market. This gives it a larger potential addressable market and more monetization opportunities across high-frequency daily use cases.

• Muse can leverage Meta’s social network ecosystem and has the potential to reach a broad consumer market. Muse reached approximately 2.8 million downloads within its first 12 days of launch. This strong download figure demonstrates that users are willing to adopt AI agents in their daily lives, rather than using them only for work. Muse can integrate with the apps users rely on every day, such as email and Instagram, and act as a personal assistant that helps them complete tasks across those apps. Meta’s position as a social media leader gives Muse a natural distribution advantage and a strong entry point into consumers’ daily digital routines. Given Meta’s large consumer base, Muse could reach an even broader market and unlock monetization opportunities across payments, advertising, entertainment, travel, and e-commerce.

LLM Market is More Crowded as Several New Challengers Outside of the US and China have Emerged

• We noticed that several non-US/CN models appeared on Artificial Analysis’ intelligence ranking in Sep. These new challengers are from Singapore, Korea and the UAE.

• Agnes 3.0 Flash (Sep 2026): Agnes 3.0 Flash is an open-weight text model developed by the Singapore based company Sapiens AI. It is positioned as a next generation model for agentic coding and tool-driven tasks.

• Motif 3 (Aug 2026): A new language model developed by the South Korea based company Motif Technologies. It is positioned as a large-scale, open-weight foundation model built in-house using a proprietary architecture, rather than an improved version of overseas open-weight models.

• K2 Horizon (Sep 2026): K2 Horizon is a new family of six open-weight foundation models developed by the Institute of Foundation Models (IFM), a global AI lab launched by Abu Dhabi’s Mohamed bin Zayed University of Artificial Intelligence (MBZUAI).

• Apodex 1.1 (Aug 2026): A closed-weight reasoning-first model and online workbench developed by Apodex, a Chinese-capital-funded company based in the US. It is positioned as an execution engine for complex, multi-step professional tasks.

横向滑动查看完整图表

Flagship Models’ Artificial Analysis Ranking Change from Nov 2025 to Sep 2026

• As Artificial Analysis updated its intelligence score methodology, the current intelligence score is no longer comparable with our historical intelligence score from Nov 2025. However, ranking changes among frontier labs are still meaningful.

• Comparting to the ranking in Nov 2025, Anthropic still leading the pack. OpenAI and SpaceXAI remain in top 4 globally, while Meta has entered the frontier cohort with Muse Spark 1.3. Google is falling behind with ranking dropped from no.2 to no.10.

• Among top Chinese model labs, BABA, Z.ai, Moonshot are consistently in the frontier cohort. Xiaomi and StepFun entered the cohort with their new models, while MiniMax and DeepSeek fell out of the top 10.

横向滑动查看完整图表

As Expected, Model Performance Gap Between Best US and CN Models Started to Widen in 2H26

• Currently, the leading Chinese frontier model (Mimo V2.6 pro) has reached ~80% of the intelligence level of the top U.S. frontier model (Claude Opus 5.5).

• The newly added AutomationBench-AA benchmark tests whether AI agents can handle real, multi-step work tasks in business software. AA sets a high bar: an agent only scores if it finishes the entire task without breaking any guardrail.

• Under a consistent index, the model intelligence gap between the US and CN widened from 12% in Aug to 20% in Sep.

• As we previously highlighted, U.S. labs’ growing advantages in access to cutting-edge GPUs (NVDA Rubin), and growing anti-distillation precautions could cause the performance gap between Chinese and U.S. frontier models to widen in 2H26.

Chinese Frontier Open-Source Models Continue to Outperform US peers

• Top Chinese AI labs keep releasing new open-source models that outperform the best U.S. open-source model.

• The leading Chinese open-source model, Mimo V2.6 Pro, now has an intelligence score 65% higher than the best U.S. open-source model (Inkling Small).

• Given low price and high performance, as well as flexibility to fine-tune and be deployed locally, Chinese open-source models have become prevalent globally.

Chinese Models are Likely to Continue to Adopt Open-Source Approach

• Overall, closed source models achieve higher intelligence scores than open-source models, reflecting their technological edge. Open-source models offer flexibility, transparency, low costs and ecosystem benefits. They can be fine-tuned and deployed locally without vendor lock-in, while thriving communities and toolchains accelerate innovation and broaden adoption.

• China's open-source approach is very likely to put pricing pressure on US closed-source models, while making Chinese models affordable for other emerging markets.

• We also note that more non-US/China challengers are adopting an open-weight strategy.

Claude Opus 5.5, Grok 4.7, and Qwen 3.8 Max have the Best Agentic Capability

• As Artificial Analysis stopped publishing its Agentic Index in Sep, we estimate frontier models’ agentic capability by blending their AA-Briefcase, GDPval-AA and AutomationBench scores at a 15:10:5 ratio (matching their weights in the AA Intelligence Index).

• US models lead the pack: Claude Opus 5.5 ranks first, followed by Grok 4.7, Muse Spark 1.3 and GPT-6 Astra. Muse Spark 1.3 was the biggest climber in Sep.

• Among CN models, Qwen3.8 Max scores highest in agentic capability and ranks 3rd globally, followed by GLM-5.3, MiMo V2.6 Pro, Kimi K3, Step 5 Preview and DeepSeek V4 Flash Vision. Models from non-US/CN regions lag behind in agentic capability.

横向滑动查看完整图表

US Models Dominate Image Generation, Chinese Models are Strong at Video Generation (1)

• Besides LLMs, multimode models that include image, video, and audio are becoming more important, as they provide more information for models to comprehend the world similarly to humans.

• We believe LLMs will excel in use cases like search, coding, office applications, and customer service. Meanwhile, multi-modal models are likely to power applications in autonomous driving, AR/VR, content creation, advertising and gaming industries.

• Besides text models, US AI labs also lead in text-to-image models. As of Sep 2026, OpenAI’s GPT image 2 still significantly outperforms all competitors in text-to-image generation.

US Models Dominate Image Generation, Chinese Models are Strong at Video Generation (2)

• Both ArenaAI and Artificial Analysis adopted Elo ranking which are based on users’ blind selections to test out different model modes’ multi modal capability.

• The key difference is: ArenaAI compares LLMs based on their ability to generate images that match prompt descriptions, while AA's Elo scores are primarily derived from evaluations on specific content such as cartoon and illustration, or UI/UX design etc.

• In the text-to-video model rankings, Google’s Gemini Omni Flash 1.1 and BABA’s Wan 3.0 ranked no.1 and 2.

Open-Source Models Generally have More Hallucination but Could be Minimized After Fine-tuning

• AA-Omniscience Index rewards correct answers, penalizes hallucinations, and doesn’t penalize refusals. Its scores range from -100 to 100, with 0 indicating an equal number of correct and incorrect answers, and negatives meaning more wrong than right (e.g. A score of -30 means 65 wrong/35 right answers out of total 100 questions).

• The index indicates even the frontier models hallucinate. Therefore, their accuracy would not be acceptable in commercial scenarios that are highly sensitive to hallucination (e.g. finance, healthcare). For now, LLMs need to be fine-tuned or RAG-enhanced to improve accuracy.

• Chinese open-source models are more inclined to hallucinate than U.S. peers. But we believe Chinese models still have advantage in commercial adoption because open-source models can be easily fine-tuned or RAG-enhanced by third-party, which leads to a stronger ecosystem (similar to Android), whereas closed sourced models could only be fine-tuned by model vendor (similar to iOS).

OpenRouter’s Token Consumption Data Showed Strong Growth in Demand

• OpenRouter is a model routing platform. It allows enterprises and developers to access models from various vendors via its platform’s API.

• Currently, OpenRouter is the largest model routing platform worldwide. However, we estimate that OpenRouter’s global share of token consumption is only in the low single digits, which means its market share data may not be representative of the whole market.

• Despite this, we believe OpenRouter’s token consumption market shares still provide a perspective on long-term token consumption trends.

• From Sep 2025 to Sep 2026, total weekly token consumption on OpenRouter increased 24.8X. Weekly token consumption for some representative open-weight models, such as DeepSeek, increased 60.8X over the same period. This shows strong demand growth driven by wide adoption of AI agents and longer model reasoning.

• OpenRouter’s data could underestimate consumption of open-source models given it cannot not track tokens consumed in private cloud and on-premise.

API Pricing Gap Between US and CN Frontier Models Widened Again in Sep

• US – CN frontier model pricing gap widened from 60% in Aug to 70% in Sep, mainly because OpenAI’s GPT-6 Astra launched at premium pricing, while China’s average was dragged down slightly by DeepSeek’s 74% price cut on DS V4.1 Flash.

• The average price of Chinese models has doubled since Jan 2026, driven by both rising inference cost (high-end GPU and memory price hikes) and Chinese models’ improving performance.

• We introduced another version of US API pricing by including lighter versions of US frontier models – we call it US all-in API pricing. China’s API pricing gap vs US all-in API pricing expanded from 47% in Aug to 52%.

US AI Labs Cut Prices by 24%-80% for their Lighter Models to Compete with China

• US frontier models retain stronger pricing power than their Chinese peers, given the current US-CN model performance gap.

• In the US, OpenAI launched its new frontier GPT 6 Astra at premium pricing in Sep, raising blended prices by 150%, but at the same time it lowered prices for GPT 6-Sol and GPT 6-Luna by 50%/56% compared to GPT 5.6 Sol/Luna. Meanwhile, Anthropic lowered its Claude Opus 5.5’s pricing by 24%.

• In China, DS cut the price of DS V4.1 Flash by 74% compared to DS V4 Pro in Sep, while StepFun increased the price of Step 5 Preview by 187%.

横向滑动查看完整图表

API Pricing Gap Between US and CN Frontier Models Widened Again in Sep

• Among AI labs, Anthropic and OAI have been able to maintain price premium as their flagship models are able to boost frontier intelligence consistently, even though Anthropic recently priced Opus 5.5 24% lower than Opus 5.0.

• In China, Moonshot was able to significantly raise API price thanks to Kimi K3’s strong performance. But others such as Z.ai and DS could only raise price so much due to a low base. Price ceiling in China is more prominent given a more crowed and competitive landscape.

• We believe pricing pressure could be a smaller problem for DeepSeek than for other Chinese players, since 1) DS’ strategy is to offer value-for-performance models instead of SOTA models, which means smaller parameter size and less training cost, and 2) DS focuses on inference innovations to improve compute and memory efficiency, which could substantially lower their inference cost, allowing them to still achieve a decent margin despite intense price competition.

Light Versions of US Frontier Models have Become Much More Price Competitive

• AA’s cost-per-task metric measures frontier models’ weighted average cost for answering a single question (task) when going through all the benchmark tests. Cost-per-task depends on both API pricing and inference efficiency.

• OpenAI released GPT-6 Sol and GPT-6 Luna on September 22 at headline prices that are 50% and 56% below those of their predecessors, respectively, citing caching and inference efficiencies.

• In particular, GPT-6 Luna is priced at US$0.07 per task, 74% lower than DS V4.1 Flash’s US$0.27. This means OAI now not only aims to provide SOTA models but also competes directly with Chinese vendors in value-for-performance models.

Chinese Models are Still Cheaper than US Models Despite Tendency to Overthink

• We noticed that although some Chinese models have low blended API prices, their actual cost per task is not as low as it appears.

• This is because cost per task is a function of API price and the number of inference tokens. Given that Chinese frontier models have adopted heavy RL to improve their intelligence, they tend to “overthink” (think longer and multiple rounds) in inference stage in order to ensure accuracy. This would increase the model’s cost per task.

• But even so, leading Chinese models are still substantially cheaper compared to frontier US models.

横向滑动查看完整图表

Claude Opus 5.5, GPT 6 Astra/Sol, Mimo V2.6 Pro, and GPT 6 Luna Now Sit at the Pareto Frontier

• Claude Opus 5.5, GPT 6 Astra/Sol, Mimo V2.6 Pro and GPT 6 Luna now sit at the Pareto frontier as they offers the most intelligence at their respective cost levels

• Claude Opus 5.5 sits at the premium end, offering the highest intelligence scores, but also has the highest cost per task ($5.98/Task).

• GPT-6 Sol now directly competes with frontier Chinese models like GLM-5.3, Qwen3.8 Max and Kimi K3, while GPT-6 Luna offers the most value-for-performance and competes with DS V4.1 Flash.

World Model - The Road for AI to Interact with Physical World

A World Model Answers “What Happens If I Do This” — Not Just “What Happens Next”

• A world model is an AI system’s internal simulation of how the world works, which allows it to predict what will happen next based on an action and plan actions by “imaging” outcomes before acting, much like how people mentally rehearse a move before making it.

• Unlike systems that simply classify inputs or generate outputs, world models aim to simulate dynamic processes, including physics, object interactions, and causal relationships.

• Yann LeCun (ex-Chief AI Scientist of Meta) argues that language carries very little information compared with what a child takes in by watching the world. A system trained only on text, therefore, has no grounding in physical reality, no lasting memory and no real ability to plan. Thus, he believes LLMs are showing real progress but in a wrong direction. He even predicted that in five years, nobody would use GPT-style LLMs.

• World models will be a hugely important new development area because they give AI something LLMs lack, which is the ability to predict what will happen in a physical world if an action is taken.

REQUIRED PROPERTY 1 Representation

WM recognizes the object and knows its location and relationship with other objects. For example, it recognizes there is a glass on the edge of the table.

REQUIRED PROPERTY 2 Prediction

It can simulate and predict what will happen to an object. For example, it knows a glass would fall and break rather than float when it got pushed off the table.

REQUIRED PROPERTY 3 Plan and Action

When condition changes, it will plan and take actions accordingly. For example, it will move the glass to center of the table to avoid it from falling. If the glass is replaced with a cat, it will not take action.

World Model is a Crucial Component of Physical AI that Predicts What Happens Next

• Physical AI is artificial intelligence that perceives, reasons and then acts on the physical world. It includes robots, autonomous vehicles, drones, automated factories etc.

• World model is the predictive component of physical AI: the part that works out what will happen if the machine acts, before it commits.

• Learning without breaking things. A robot’s training data does not exist until it moves, and every bad move damages something. The world model lets most learning take place via simulation, with only the last fraction on real hardware.

• A training-time tool, not yet a runtime one. In most deployments, the world model generates scenarios and validates policies offline; what runs on the robot is a compact vision-language-action (VLA) model that makes the decision.

Short History of World Models

• The idea of world models is eighty years old, but it only became plausible in the past three years

• Cognitive science has used the phrase in roughly this sense since Kenneth Craik argued in 1943 that organisms carry “small-scale models” of reality to anticipate events.

• What changed is that since 2018, we can build useful artificial versions, and since 2024, ones in which a person can walk around in.

1943–1990

Control theory and cognitive science

Craik (1943): organisms carry “small-scale models” of reality to try out alternatives before committing.

Kálmán’s 1960 filter — inside every aircraft and phone — maintains an internal state, predicts, and corrects when reality disagrees.

The internal model principle (1976): to control something well you must contain a model of it.

1990–2018

The neural era begins

Sutton’s Dyna (1990) set the template for model-based reinforcement learning.

Schmidhuber proposed recurrent nets that predict future states from observations and actions.

Ha & Schmidhuber’s “World Models” (March 2018) named the field: an agent learned to drive entirely inside its own hallucinated version of a game, then transferred back.

2018–2023

Proof of capability

PlaNet planned in a learned latent space from pixels.

MuZero (Nature, Dec 2020) mastered Go, chess, shogi and Atari without being told the rules of any of them.

DreamerV3 (Jan 2023) became the first system to obtain Minecraft diamonds from scratch — no human data, no hand-built curriculum, one set of hyperparameters across 150+ tasks.

2022–2026

Became plausible with two diverge tech routes

LeCun’s June 2022 JEPA paper: predicting the future by generating pixels is a category error — better to predict an abstract summary and throw away what cannot be known.

Meta built I-JEPA, V-JEPA and V-JEPA 2. LeCun left Meta in November 2025 and raised $1.03bn for AMI Labs in March 2026 on the explicitly anti-generative thesis.

2025 was when World Models Really Started to Capture Market Attention

横向滑动查看完整图表

Why are World Models Important?

• Yann LeCun argues that language carries very little information compared with what a child takes in by watching the world. A system trained only on text, therefore, has no grounding in physical reality, no lasting memory and no real ability to plan. Thus, he believes LLMs is showing real progress, but in the wrong direction. He even claimed that within five years, nobody will use GPT-style LLMs.

• World models matter because they give AI something LLMs lack, which is the ability to predict what will happen if it takes an action. That ability underpins five things.

• It may be a path to more general intelligence. Researchers such as Yann LeCun and Fei-Fei Li argue that intelligence built only on language lacks grounding, memory, and planning. World models add the ability to imagine “what if I do this?”, which humans use constantly.

• AI that acts in the physical world needs WM. A robot anticipates what its action will cause before acting. Without WM, robots can only follow pre-programmed moves and can’t cope with situations they have not seen.

• Machines can practice safely and cheaply in WM. Training a robot in the real world is slow, expensive, and sometimes dangerous. Nobody wants a car learning to handle a child running into the road by actually trying it in real situation. A world model lets machines rehearse thousands of times in simulation, including rare events.

How World Models Differ From Language Models (1)

• A language model is a machine that has read thousands of texts on a glass breaking when it falls from a table, but does not really understand the physics of it.

• A world model is a machine that has watched the glass break, understands the physics of it, and can predict it (e.g. when it sees a cat push a glass off a table).

• Both turn what they are given, words or camera image, into a compact set of numbers, then use the same kind of engine, a transformer, to look at everything so far and predict what comes next. Both learn that simply by practicing on vast amounts of ordinary material, with nobody labeling anything or teaching them the rules.

• World models, VLA models and LLMs all require compute, memory, and electrical power for hardware infrastructure. They also share the same pool of AI research and engineering talents.

How World Models Differ From Language Models (2)

横向滑动查看完整图表

Will World Models Make LLMs Obsolete?

• World models are unlikely to replace LLMs, they are more likely to complement each other.

• Language is the interface: People control world models with words, and an LLM inside turns those words into instructions the world model can follow. World models don’t replace LLMs; they depend on them to understand what users want.

• Most valuable work isn’t physical. Code, contracts, research, knowledge, customer service all work in language and symbols, where a model of physics adds little. Demand for LLMs would not shrink as robots get better.

• The two are merging, not competing. NVDA’s Cosmo 3 combines reasoning, world generation and action prediction in a single model, and Google builds its robotics models on Gemini.

• LLMs already form a partial picture of the world. Harvard researchers showed that a model trained only on moves from the board game Othello has built an internal map of the board. But those internal pictures are patchy and fragile, which is exactly why connecting LLMs to world models helps.

• The economics favor LLMs for most tasks. A second of 720p video takes 4,000x more tokens than a second of text. OpenAI shut Sora down in Mar 2026, partly to free up compute for its core LLM products. For anything that can be done in words, an LLM will remain far cheaper than a world model.

• Large AI players invested in both LLMs and world models.

• What would change our view: 1) A WM-first system beating LLMs at language and reasoning; 2) AMI Labs or similar lab turning JEPA-style models into a real product; 3) LLM revenue stalling while physical-AI revenue scales.

Differences between a World Model and a Vision-Language-Action (VLA) Model

横向滑动查看完整图表

Role of a World Model and VLA Model in Autonomous Driving Scenario

World Models vs. LLMs: Compute and Memory Intensity (1)

• Route 1 (generative video) and Route 3 (interactive neural engines) clearly cost more compute and memory than LLMs. Both routes use diffusion models to generate video pixels frame by frame.

• Route 4 (3D splatting) and route 5 (physics engine hybrids) are heavier only in one generative step each, but that step could also be compute intensive.

• Compute: generative world models require far heavier compute per second of output than LLMs:

• Training is not where the difference lies: building a large world model takes a training budget roughly comparable to a frontier LLM. (Cosmos: 20m hours of video, reportedly ~10,000 H100 GPUs over ~3 months; Meta Movie Gen: 1.25m GPU-hours across 6,144 GPUs). The gap opens up in inference: the everyday cost of running the model each time a user or robot asks it to generate something.

• An LLM writes text one small piece (a token, roughly a word) at a time, and each piece is cheap to produce. A generative world model generates video, and a single frame is equal to thousands of these pieces (~16,000 per second of 720p, against ~4 for text).

• Most world models also use diffusion: they start from noise and refine the whole block of frames over dozens of passes, so each second of output consumes compute required to generate many more seconds of content.

• The cost gap shows up in real workloads. Generating a 5-second clip with the open-source HunyuanVideo model takes about an hour on a single A100 GPU. It is estimated a 10-second Sora clip would require about 40 GPU-minutes, and costs roughly US$1.30.

World Models vs. LLMs: Compute and Memory Intensity (2)

• Memory capacity: the clearest pressure point

• To keep a scene consistent (a room that stays the same room, an object that doesn't vanish when the camera turns), a world model has to keep a record of everything it has already generated. This short-term memory is the KV cache, and it grows steadily the longer the model runs.

• The numbers get big quickly. The WorldKV paper (May 2026) found that for a mid-sized 14-billion-parameter world model, the required memory exceeds 200GB after just one minute of interactive output. That exceeds the HBM on even the most advanced single GPU available today, so the workload has to be split across several chips.

• As the memory fills up, performance drops. In the same study, generation speed fell from about 8.9 to 3.6 frames per second over that one minute. In practice, memory capacity is what limits how long a world model can stay coherent, which is the central technical challenge the industry is trying to solve.

• Implications

• Generative world model routes need more compute and even more memory capacity, which is positive for GPU and memory players.

• HBM demand sits in the datacenters, where generative models train and produce simulated environments. The routes built for robot decisions (JEPA, Dreamer) are light enough to run on the robot’s local chip, which uses cheaper LPDDR memory, not HBM.

• Route selection, therefore, directly impacts compute and memory demand. If non-generative routes win in robotics, less of the physical-AI buildout turns into HBM and compute demand. Generative demand for content, gaming, and synthetic training data would remain either way.

• Suppliers haven't confirmed the link yet: SK hynix's 2026 outlook does not cite world models as an HBM driver.

• Near term, LLMs likely remain the dominant source of HBM demand, given their far larger user base.

Data: Video is Abundant; Video with Actions Attached is Scarce

• Raw video is not scarce. Epoch AI puts the world’s usable public text at ~300 trillion tokens; Cosmos’s single 20m-hour training set implies several times that entire stock. There is no video wall in any meaningful sense.

• Action-labeled data is desperately scarce. AgiBot World holds over a million trajectories from 100+ robots; DROID is 76,000. Against a billion hours of internet video, the ratio is roughly six orders of magnitude.

• The economics is stark: real demonstration clips cost $50–500 each; synthetic equivalents $1–1.50. That 30–50× gap is why every serious player has a synthetic-data strategy. Unresolved: how much of a policy’s training can be synthetic before it degrades.

• Three bridges to watch: egocentric video from smart glasses; fleet data (Waymo ~200m autonomous miles, Huawei ADS in 1.9m vehicles); and state-organized data collection in China, with a Beijing target of 10,000 humanoids in commercial use by year-end.

• Legal risk is materially higher than for video. Video carries identifiable people, recognizable characters, and studio copyright. Any world-model business trained on scraped video carries an unquantified liability that a text model does not.

Unsolved Problems on the Algorithm Side

• Long-horizon consistency is the central blocker. The best public interactive world model, Google's Genie 3, keeps a scene coherent for a few minutes and remembers only about one minute back. A robot task or a game session needs hours. Small errors pile up frame after frame until the world drifts apart. At least five research papers in 2025–26 tackled this, but none has solved it.

• The world stops when you look away. In a real room, a kettle keeps boiling while your back is turned. In generated worlds, hidden events usually freeze. A June 2026 benchmark of 23 models found this across the board. Scaling one model family from 1.3B to 14B parameters made it worse, which suggests the fix is a design change, not more compute.

• Looking real is not the same as being physically correct. Leading image-to-video models score only about 25–40 out of 100 on Physics-IQ, the main benchmark for whether generated video obeys physics. The benchmark's authors say bigger models and better prompts are not enough to close the gap, and scores need to roughly double before generated worlds are safe to train robots in.

• No agreed way to measure progress. At least eight competing benchmarks (Physics-IQ, WorldModelBench, VBench-2.0, RoboArena, and others) launched within about 18 months, each arguing the others are inadequate. Without a commonly accepted benchmark, buyers can't compare vendors, and vendor claims can't be checked.

• Learning actions from passive video is hard. Models learn from vast amounts of ordinary video, but a robot needs to know what caused each change: which movement, which force. Linking actions to outcomes is still weak, which is why action-labeled data remains the bottleneck (thousands of hours of it, against a billion hours of internet video).

• The non-generative route has its own problems. JEPA's training goal is hard to define: which parts of the future should the model keep, and which should it ignore? Its understanding is also invisible, stored in abstract codes with no picture to check against, so progress is hard to verify or demonstrate.

• The sim-to-real gap persists. Even when a robot learns well inside a generated world, some of that skill is lost on real hardware. How much of a robot's training can be synthetic before performance drops is still unpublished.

The Seven Technological Routes to World Models (1)

• Currently, there are seven technological routes to develop world models.

• Each route is trading against the same constraint — how much of the world you keep track of versus how cheaply you can run.

• The most crowded route is generative video, where Chinese labs are most competitive.

• The likely end-state is a combination of multiple routes: NVIDIA’s Cosmos 3 explicitly combines vision reasoning, world generation, and action prediction in a single mixture-of-transformers model; World Labs’ Atlas combines generation with explicit geometry; Waymo’s world model is a generative video system post-trained to emit lidar.

横向滑动查看完整图表

The Seven Technological Routes to World Models (2)

• We classify the seven technical routes using a four-quadrant chart.

• The horizontal axis runs from implicit understanding of the physical world hidden inside the WM, to explicit WM structure you can inspect, such as geometry, equations, or words.

• The vertical axis runs from generating a visible world (generative route), to predicting an action with no visual output (the predictive route).

The Seven Technological Routes Explained: Generative Video (Route 1)

• Route 1 — Generative video. The model watches enormous amounts of video and learns to continue a video clip given what has happened so far, paint the next frames.

• Its strength is that the training material is already out there in effectively unlimited quantity, the engineering is well understood, and the output is photorealistic and legible to anyone.

• Its constraint is that it is painting appearances, rather than tracking a world. Nothing underneath guarantees that the shelf you just passed is still where it was, objects drift or vanish, and every second of output costs serious compute because the picture is rebuilt frame by frame.

The Seven Technological Routes Explained: Latent Prediction, the JEPA School (Route 2)

• Route 2 — Latent prediction, the JEPA school. Instead of predicting pixels, the model predicts a compressed summary of what happens next — closer to "the box will land on the left edge of the pallet" than to a rendered image of it.

• The argument is that most of a video's detail is noise a decision-maker doesn't need, so throwing it away lets the system learn far more from far less data and run much faster on a robot.

• The constraint is commercial rather than technical: the output is a set of numbers, so there is nothing to demonstrate, no product to sell, and no consumer revenue on the way. It also remains largely unproven at scale. Source: Jefferies

The Seven Technological Routes Explained: Interactive Neural Engines (Route 3)

• Route 3 — Interactive neural engines. Same underlying idea as video generation, but the next frame responds to your input in real time — you press forward, and the model paints what you would see if you walked forward. There is no game engine, no pre-built level, no artist; the world is invented as you move through it.

• The strength is a fully explorable environment from a text prompt, which is what made Genie 3 such a striking demonstration. The constraints are memory and cost. These systems hold only around a minute of consistent world before they start forgetting what they showed you, and each user must be served their own continuous generation, so the cost per userhour is brutal and cannot be spread across a batch. Source: Jefferies

The Seven Technological Routes Explained: Explicit 3D Reconstruction and Generation (Route 4)

• Route 4 — Explicit 3D reconstruction and generation. Rather than painting pictures, the system builds an actual threedimensional model — real geometry, real surfaces, a scene that exists as a file.

• Because the structure is genuinely there, the world has high consistency: walk away and come back, and the room is unchanged, and the result can be exported into game engines, mapping tools, or design software. That is a real business today in games, film, and architecture. The limitation is that these worlds mostly sit still. You get a convincing space but very little of what happens in it — no falling, no contact, no consequence — which is precisely the part a robot needs to learn. Source: Jefferies

The Seven Technological Routes Explained: Physics-engine Hybrids (Route 5)

• Route 5 — Physics-engine hybrids. Here, engineers handwrite the rules — gravity, friction, how a joint moves — inside a traditional simulator, then use AI to make the images look realistic enough that a robot trained inside it works outside it.

• The strengths are correctness and free ground truth: because the world was authored, every position, force and collision is known exactly, which makes the training data perfect. This is what trains humanoids and autonomous vehicles in 2026. The constraint is coverage. Someone has to write each rule, so the approach handles the scenarios anticipated and breaks on the open-ended messiness of the real world, which is why it is a workhorse rather than a path to general capability. Source: Jefferies

The Seven Technological Routes Explained: Model-based Reinforcement Learning (Route 6)

• Route 6 — Model-based reinforcement learning, the Dreamer school. The agent builds a small private model of its own world and then practices inside it — imagining thousands of attempts before making a real one, much as a driver rehearses a maneuver mentally.

• Its strength is sample efficiency. Because most of the practice happens in imagination, the agent needs far fewer real-world trials, and the approach has a long peer-reviewed track record on games and control problems. Its constraint is that those results came from relatively simple settings. Scaling the same trick to the full visual complexity of a warehouse or a street has not been convincingly demonstrated, and there is no product line built on it. Source: Jefferies

The Seven Technological Routes Explained: LLM as World Model (Route 7)

• Route 7 — LLM as WM, so that it predicts how a situation will change by reasoning it through in words (or code) rather than generating video or 3D. This works best where the “world” is made of symbols and rules, such as code, games, or business processes, and a model of physics isn’t needed.

• The appeal is that it costs almost nothing to try, since the infra already exists and the models are already deployed — and for a surprising range of everyday reasoning, it works. The constraint is that describing physics is not the same as modeling it. A model can write a fluent sentence about a glass tipping off a table without any reliable grip on the trajectory, and the evidence that these systems have genuine physical grounding remains thin. It is a useful component for planning and instruction, not a substitute for the routes above. Source: Jefferies

World Model Use Cases, Three Tiers: Paying, Piloting, and Demonstrating (1)

• Already paying today: AV simulation & validation: This has the clearest ROI because it generates rare, dangerous corner cases that real-road testing can't capture. McKinsey-derived figures put physical testing at roughly $4.50–5.50 per km of robotaxi operation against about $0.60 for a private car. Every mile moved from asphalt to simulation is a direct cost saving with no safety downside. Waymo's world model, built on Genie 3, turns ordinary dashcam footage into full camera and lidar simulations. Applied Intuition sells the same capability to 18 of the top 20 automakers and was valued at $15bn.

• Two other areas earn real money but aren’t really new world-model business: 1) Video / creative content: MiniMax (Hailuo), Seedance, Kling, Runway, and Veo are all monetizing. But the unit economics are unproven, and copyright exposure is heavy. OpenAI shut down the Sora app in April 2026, and Runway is pivoting toward robotics. 2) Industrial digital twins: This is the largest existing revenue pool (Dassault, Ansys), but it predates the world-model label. AI makes these tools better, rather than replacing them.

• Likely scaling next: 1) Architecture, real estate & design: This means generating explorable, geometrically accurate 3D spaces from a handful of photos, which fits straight into existing CAD workflows. Examples are World Labs' Marble ($20–95/month), backed by a $200m strategic investment from Autodesk. 2) Defense & synthetic environments: The budgets are large. Anduril booked $2bn of 2025 revenue, and Palantir and Shield AI sell into autonomy programs. In-Q-Tel's investment in Odyssey signals direct government interest, but it's hard to say how much of the revenue comes from world models.

• Commercial but unprofitable: 1) Robotaxis: Waymo runs about 500k paid rides a week across 14 cities and reports 95% fewer serious crashes. Baidu Apollo Go, Pony.ai, and WeRide operate at scale in China. Still, each vehicle costs over $120k, cost parity isn't expected before the mid-2030s, and Cruise burned >$10bn before shutting down. 2) Warehouse & factory robots: Deployments are real but small. Figure has about 740 robots working, including at BMW, Agility has a robots-as-a-service deal with GXO, and Amazon runs more than 1m warehouse robots. The financials are stark: Agility booked $1.8m of revenue against a $140m operating loss. China ships >90% of humanoids, yet Reuters found they "can do kung fu but not factory work."

• Still demos: Interactive gaming & general-purpose humanoids: Project Genie offers only 60-second sessions inside a ~$250/month subscription, with no standalone business model. Humanoid "autonomy" is still largely teleoperated.

World Model Use Cases, Three Tiers: Paying, Piloting, and Demonstrating (2)

横向滑动查看完整图表

US vs China: The Two are Not Competing on the Same Axis of Physical AI

• The US leads the digital half of the loop — chips, training runs, and the software that turns a model into a product.

• China leads the physical half — robots, components, factories, and the deployments that generate interaction data that no compute can substitute for. Its top five humanoid makers took 86% of global H1 2026 shipments (Counterpoint), and ~140 Chinese manufacturers account for over 90% of global humanoid sales.

• On policy, the contrast is the sharpest finding in the section. Embodied intelligence is a named priority in China’s 15th FiveYear Plan, with roughly CNY 1 trillion allocated across national and regional funds targeting robotics, AI, and other emerging technologies. The principal US federal actions in the same window are a $5bn AI-for-science program and an import ban on foreign humanoids.

• The most under-appreciated fact: the American AI buildout depends materially on Chinese physical supply — ~30% of certain US transformers and switchgear categories, over 40% of data center battery imports, and around two-thirds of global optical transceiver units.

Both Sides have a Genuine Bottleneck: US In Manufacturing Scale, China In EUV

• The US is very likely to build the best world models. China is very likely to own the machines that generate the data those models need, and to build them at a price no one else can match.

• Unless one side solves the other’s problem — American manufacturing scale, or Chinese lithography — the equilibrium is mutual dependence enforced by mutual export controls, which is precisely the environment in which supply-chain names outperform platform names.

横向滑动查看完整图表

World Model Still at Early Development Stage, But These Indicators are Worth Following

横向滑动查看完整图表

A Summary of Major Players in World Models

World Labs Builds Consistent and Explorable 3D Worlds

• In 2024, World Labs was founded by Fei-Fei Li, the creator of ImageNet. On Sep 28, AMD acquired World Labs in an all-stock deal which valued World Labs at ~US$8.2bn (~US$5bn in early 2026). AMD announced that Fei-Fei Li will join as executive VP and chief scientist, reporting to Lisa Lu. This deal is expected to close by end-2026.

• Key strategy: 1) Early commercialization: subscriptions plus an API from day one, unlike research-first rivals such as AMI labs. 2) Expand from creative work into physical AI. 3) Strategic investors as distribution: Autodesk is a joint-research partner, starting with film and media before expanding to design, construction, and manufacturing. Sea brings a SEA gaming and e-commerce angle.

• Tech route: World Labs mainly took the route 4 (explicit 3D), which is built on Li’s thesis of “spatial intelligence”. Instead of simultaneously “painting” the world, World Labs builds a persistent 3D environment using explicit 3D/splatting. The environment is explorable and editable. Moreover, World Labs also combines other routes. For example, RTFM tests the interactive, frameby-frame approach (route 3), and Atlas combines video generation (route 1) with 3D.

• Products: 1) Marble: Launched in Nov 2025, Marble is an app that creates lasting, editable 3D worlds from text, images, video or a rough 3D sketch; exports 3D files and video. Worlds are static. 2) Atlas: Launched in Sep 2026, Atlas is the foundation model behind future Marble versions. Generates up to 1min of 1440p video with matching 3D geometry. 3) RTFM: Real-time model that generates each frame as you move through a world, running on a single H100 GPU. 4) Spark: Free, open-source software for displaying its 3D worlds in a web browser.

• Use cases: 1) Gaming and VR: level design and fast environment prototyping. 2) Filming: virtual sets and previsualization. 3) Architecture, interiors, and real estate: explorable spaces built from photos. 4) Robotics: robotics simulation and testing.

AMI Labs Ignores Details and Predicts How a Summary of the World Changes

• Advanced Machine Intelligence (AMI Labs) was founded in Paris in Dec 2025 after Turing Award winner Yann LeCun left Meta. In Mar 2026, AMI raised a US$1.03bn seed round at a post-money valuation of ~US$4.5 billion. Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions led the round, joined by NVDA, Samsung, Toyota Ventures, Sea, and Temasek. LeCun’s published work includes LeWorldModel (2026), which learns a simplified picture of a scene so it can predict what will happen next and plan accordingly, and LeVJEPA (2026), which learns from video and serves as a building block for future world models.

• Key strategy: 1) Research first, on a long horizon: there are no near-term revenue plans, and the aim is to get the science right before building products. 2) Partner with data-rich industries rather than consumers: healthcare first through Nabla, then manufacturing, biomedicine, and robotics.

• Tech route: AMI follows route 2, latent prediction (JEPA), which LeCun proposed in 2022. JEPA doesn’t generate video. It compresses what it sees into a compact summary and predicts how the summary will change, deliberately ignoring details it can’t predict, such as how leaves flicker. It then uses those predictions to plan actions. It learns from video, audio, and sensor data rather than text. AMI is explicitly the anti-generative camp, the opposite to Sora, Genie, and World Labs. AMI’s approach requires less compute and memory because it predicts in a compact latent space rather than generating full pixels/frames, making it more compute-efficient.

• Products: None yet. This is the biggest difference from World Labs. AMI is research-first; CEO expects first product-usable components in ~1 year, with commercial products years away. As of mid-2026, it had published research papers on its approach, but no products or commercial model releases. Its only disclosed commercial link is an exclusive partnership with Nabla, which gets first access to AMI’s technology for healthcare.

• Use cases: 1) Healthcare: clinical AI agents that understand workflows and read audio, imaging, and physiological data, through Nabla. 2) Industrial processes: predicting the behavior of systems with thousands of sensors, such as jet engines, steel mills, or chemical plants. Its compute-efficient architecture could give it an advantage in industrial settings, where systems need fast, reliable decisions without extreme compute demands. 3) Robotics and autonomous driving: from home robots up to fully self-driving cars. 4) Smart glasses: devices that watch what you’re doing and predict your next step to help you.

Google DeepMind is the Most Complete World Model Player

• Google (DeepMind) is the most complete world model player. It has the leading interactive world model, a top video model, Gemini to control them in plain language, agents and robot “brains” trained on top, its own TPU chips, and in Waymo, a real-world user operating at scale.

• Key strategy: 1) Keep everything in-house: research, models, chips, and products, with Waymo as the internal customer. 2) World models as a training ground for AI.

• Tech routes: Google adopted multiple routes. Route 3 (interactive neural engines): Genie 3. Route 1 (generative video): Veo and Omni. Route 6 (model-based RL): DeepMind created MuZero and DreamerV3. Route 7 (language reasoning): Gemini, which works as the control layer on top.

• Products: 1) Genie 3 (Aug 2025): real-time interactive world model: type a prompt and explore a world at 720p and 24 frames a sec, with ~1min of memory; 2) Project Genie (Jan 2026): consumer app combining Genie 3 with Gemini and Nano Banana Pro image model to create and explore worlds. 3) Veo 3.1/ Gemini Omni: Veo 3.1 is video generation model, Gemini Omni creates and edits video from any mix of text, image, audio, and video. 4) SIMA 2 (Nov 2025): learns from human gameplay in video games, then self-improves through play, including inside Genie 3-generated worlds. 5) Gemini Robotics 2 (Jul 2026): a VLA for wholebody control and dexterity, an embodied-reasoning model (ER 2) for planning and multi-robot coordination, and an on-device VLA. 6) Waymo WM (Feb 2026): Genie 3 adapted to produce camera and lidar data, used to simulate rare road events.

• Genie 3 learns from video to simulate interactive 3D worlds that respond to actions. It generates a world frame by frame, it learns a compact sense of what a scene looks like, then keeps drawing what comes next.

• Gemini Robotics ER2 and VLA models handle physical reasoning and robot control. Gemini robotics ER2 acts as higher-level brain, reasoning about what could happen next in the real world and plans tasks, it then hands off motor execution to a lower-level VLA model.

• Use cases: 1) training AI agents in generated worlds. 2) self-driving simulation: Waymo tests rare, dangerous scenarios it could never collect safely on real roads. 3) Robotics: Gemini Robotics 2 robots can tidy a cluttered room or tie knots. 4) Gaming: building and exploring playable worlds in Project Genie.

Meta Spread Across Several World Model Routes Rather than All-in on One

• Meta’s world model work is spread across three groups: its FAIR research lab, Reality Labs (VR, glasses, and robotics), and Meta Superintelligence Labs (frontier AI unit). Meta’s own research labs created the JEPA approach that AMI is now built on. That team was hit hard when Yann LeCun left in Nov 2025, and several of his senior colleagues followed him to AMI.

• Key strategy: 1) Feed world models into Meta’s own hardware. 2) License robot software to others rather than compete on hardware.

• Tech route and products: Meta is spread across several routes rather than all-in on one. V-JEPA 2 is route 2 (latent prediction). V-JEPA 2 (Jun 2025) has 1.2bn parameters, and was trained on 1m+ hours of video plus 62 hours of robot data. These models learn from video and predict how a scene will change in a compressed latent space, focusing on what changes rather than every pixel, then links those predictions to possible actions so a robot can plan its next move. The key distinction between Meta and AMI’s approach is that V-JEPA is video-first, while AMI aims to extend the approach to broader sensor inputs and real-world control.

• Code World Model (Sep 2025) is route 7 (LLM as world model). It is a 32bn parameter LLM that predicts what a program will do, line by line. WorldGen (Nov 2025) generates walkable 3D game worlds from a text prompt, as standalone 3D meshes that work in Unity and Unreal. Hyperscape Capture (Sep 2025) is route 4 (explicit 3D). It scans a room with a Quest 3 headset and turns it into a photorealistic 3D VR scene. Mango is route 1 (generative video). It is an image and video generation mode.

• Use cases: Meta’s V-JEPA focuses on physical AI. Meta has shown V-JEPA enabling robots to reach, grasp, pick, and place, and navigate in unfamiliar environments, supported by capabilities including action anticipation and depth understanding. Compared with AMI Labs, Meta has an advantage in scale, with larger training data and stronger compute and product infrastructure. Other than physical AI, Meta’s world model also entered other use cases like smart glasses, VR, metaverse, coding, and content.

NVIDIA Gives Away Cosmos that Focuses on Physical AI to Sell More Hardware

• NVDA doesn't need to back a winning approach, because it sells what every world model vendor needs: chips to train models and run models inside robots and cars. NVDA’s world models are free to download. They exist mainly to make its hardware more valuable.

• Key strategy: 1) Cover the three key parts of physical AI: NVDA systems to train the model, Omniverse and Cosmos to simulate the world, and Jetson Thor or DRIVE chips to run it on the machine. 2) Give models away to sell the hardware.

• Tech routes: NVDA covers almost every route at once. Cosmos’ video generation is route 1. Isaac Sim, Omniverse and Newton combined with Cosmos are route 5 (physics-engine hybrids). Cosmos Reason, which explains scenes in words, is route 7. Cosmos 3 then merges reasoning, generation and action into a single model.

• Nvidia’s Cosmos platform integrates physical reasoning, simulation, and action prediction within a unified multimodal framework. It is designed to support robotics, autonomous vehicles, and industrial simulation, with an emphasis on synthetic data generation.

• Nvidia’s world model technology combines learned world simulation with physical reasoning. Cosmos 3 pairs two specialized parts. One makes sense of what the machine sees and what it is told to do, and the other imagines what the scene will look like and what action to take next. Cosmos is released as an open-weight model, developers can tailor the models with their own robot, camera and sensor data.

• Use cases: Nvidia’s Cosmos world models focus on robot training and testing, autonomous driving and industrial vision. Its key strength lies in its vertically integrated stack, spanning compute hardware, software frameworks, simulation environment and deployment within its own autonomous driving ecosystem (Nvidia DRIVE).

Tesla: Vertically Integrated Player that Builds WM for its own Autonomous Driving and Humanoid

• Tesla doesn’t sell a world model. It builds one for internal use, to train and test the software that drives its cars and, increasingly, its Optimus robots. The world model sits at the center of a closed, vertically integrated system. Tesla owns the cars that collect the data, the chips, the training computers, the driving software and the robots.

• Key strategy: 1) data flywheel: millions of customer cars send back video, and “smart triggers” capture rare events. The simulator trains on that footage and then tests and trains new FSD versions, and the cycle repeats. Tesla also has the largest real-world driving dataset of ~14bn FSD miles. No other companies collect driving video at this scale. 2) vertical integration: its own inference chips, training compute cluster on NVDA GPUs, software, and factories give it control over cost and speed. 3) Cameras only, with minimal hand-coding: no lidar and no HD maps. Tesla is taking the approach that more data and compute will win, which keeps hardware cheap and easy to scale. 4) One AI stack for cars and robots: the same world simulator and training approach carry over from FSD to Optimus.

• Tech route: Tesla uses route 1 (generative video), built to behave like route 3 (interactive neural engine). Its simulator learns from massive fleet videos and generates realistic camera footage that responds to the car’s actions. Tesla also builds generative video models to create interactive driving simulations for training and evaluation. The same video-generation system is now being used for Optimus. The driving software itself works like a robot “brain” (VLA model), going straight from cameras to controls with little hand-written code.

• Products: 1) Neural world simulator: a learned model that generates synthetic video from all 8 of a Tesla’s cameras, showing how the road responds to the car’s actions. 2) FSD: one end-to-end neural network that turns camera images directly into steering and braking. 3) Robotaxi: driverless ride-hailing using Model Ys, 4) Optimus V3: Humanoid robots.

• Tesla is applying simulation to driving and humanoid robotics. The main applications of Tesla’s world models are evaluating its own FSD and Robotaxi driving behavior, checking whether software updates introduce failures, and generating rare or challenging situations. Tesla has also demonstrated extending its video-generation approach to Optimus, including indoor environments and manipulation scenarios, making robotics another application direction.

Tencent Adopts Same Route as World Labs and Focuses on Game Development

• Tencent is China’s most active open-source builder of 3D world models. Its world model is anchored in gaming, Tencent’s core business, and sits under the Hunyuan AI brand. Robotics is handled separately by its Robotics X lab and the Tairos platform.

• Key strategy: 1) Open source the world models to become the platform for developers, and make money through Tencent Cloud and its own products. 2) Empower gaming first: design world models to fit existing game production workflows, serving Tencent’s own studios and wider industry.

• Tech route: Tencent’s world simulation roadmap has evolved from interactive video generation to explicit 3D/splatting reconstruction. Route 4 (explicit 3D): HY-World 2.0, its flagship model is the closest to World Labs’ Marble, but free. It creates a full 360-degree view of a scene, decides where a camera would travel, fills in the extra angles, and then turns everything into a proper 3D model through explicit 3D/splatting. Route 3 (interactive neural engines): HY-World 1.5, Tencent’s equivalent to Genie 3. It keeps drawing the virtual world as the user walks around it with a keyboard and mouse, and it remembers what earlier spots looked like, so places and objects stay the same when the user comes back to them. Route 1 (generative video): HunyuanVideo.

• Products: 1) HY-World 2.0 (Apr 2026): Flagship world model that generates, rebuilds and simulates 3D worlds from text, images or video. Unlike versions 1.0 and 1.5, it outputs editable 3D files that load straight into Unity and Unreal. 2) HY-World 1.5/ WorldPlay (Dec 2025): Real-time interactive world model that paints the scene at 24 frames per sec. 3) HunyuanVideo/ Hunyuan3D: Video generation and 3D object generation models. 4) Tairos + Hy-Embodied models (Jul 2026): Robotics platform offered as “Embodied-AI-as-a-Service” through Tencent Cloud.

• Its application direction emphasizes games, immersive media, and simulation environments. Hunyuan World targets game development, virtual reality, interactive content creation, and physical simulation. Tencent’s advantage lies in its deep expertise in gaming and 3D content, making explicit, persistent 3D world a natural technology pathway. Its gaming ecosystem also provides clear downstream use cases for world generation and interactive experiences.

Alibaba: Aims for Physical AI and Builds on Qwen’s Capabilities

• Alibaba (BABA) is China’s full-stack AI player. It combines open models, a cloud business to monetize them, and robot-brain ambitions built on Qwen. It is strongest in video generation and has an unusual spatial-data asset in Amap.

• Key strategies: 1) Build everything on Qwen: robot models reuse Qwen’s language and vision abilities. 2) Open source to win developers, but keep flagships closed: the Rynn models, Abot and earlier Wan versions are open, while its best consumer video model (Wan 3.0) is proprietary. 3) Invest in hardware partners rather than build robots at scale.

• Tech route: BABA’s approach is world models as the data and brain layer for robots, built on Qwen. Route 1 (generative video): Wan is positioned as frontier video generation models worldwide. 2) LLM-based world models (route 7): Qwen-RobotWorld connects Qwen’s understanding of images and language with prediction of what happens next. 3) Physics-grounded simulation (route 5): Abot-World uses a physics engine to generate training data.

• Products: 1) Wan: video generation model. Wan 2.1 and 2.2 (2025) were released open-source and became widely used; Wan 3.0 (Aug 2026) makes 30s, multi-shot clips at up to 1080p and can turn a script document straight into video. 2) HappyOyster turns a written description or a picture into a real-time interactive world that users can explore.; 3) Qwen-Robot series (Jun 2026): Three robot models: RobotNav for navigation and driving, RobotManip for manipulation, and RobotWorld a world model that predicts physically consistent outcomes. 4) DAMO Rynn series: RynnVLA-002 (Nov 2025) combines a robot brain and a world model in one open model; RynnBrain (Feb 2026) is an open robot brain built on Qwen. 5) Amap Abot (Apr 2026): A full robotics stack. Abot-world is a world model that generates training data video, depth, 3D point clouds, movement paths), alongside navigation and manipulation models, and a four-legged robot.

• Use cases: 1) e-commerce and content: product videos and ads for merchants, short videos, and turning scripts into video. 2) Robotics: household and industrial robots that navigate and handle objects, through Qwen-robot and RynnBrain. 3) Self-driving: RobotNav and RobotWorld both cover driving scenarios. 4) Synthetic training data: Abot-World generates the data robots learn from. 5) Maps and mobility: robots that use Amap’s maps as memory.

ByteDance: Leverages its Strengths in Real-world Video Data and Top Video Models

• ByteDance (BD) named world models as one of four AI priorities in 2026, alongside Seedance, coding, and its Doubao assistant. Seed team’s leader Wu Yonghui set an explicit goal of releasing at least one world model by year-end, and benchmarks it against Google’s Genie 3.

• Key strategy: 1) Make world models a top priority and spend heavily on data to catch up with Google. 2) Build on its video lead: Douyin/TikTok hold enormous libraries of short and real-world video, plus data on how users interact with it. Also, BD could extend Seedance into interactive 3D, launched first on its own Pico headsets. 3) Start with consumer entertainment.

• Tech route: BD is video-first, moving towards interactive worlds. It is strong in route 1 (generative video) through Seedance and is trying to extend it into route 3 (interactive neural engines) with the reported real-time 3D model. It is also building route 4 (explicit 3D) tools and robot brains alongside. This is the same path Google took from Veo to Genie, which is why Genie 3 is its benchmark.

• Products: 1) Seedance: leading video generator. 2.0 (Feb 2026) went viral for its realism; 2.5 (Jul 2026) makes 30s videos in one pass, takes up to 50 reference images, clips or audio files at once. 2) Real-time 3D world model (expected to launch in Oct 2026): Generates explorable 3D spaces on the fly in response to a user’s voice and movement, at 20 frames/sec, built on Seedance, per Bloomberg. 3) Depth Anything 3/ Seed3D 2.0: Tools that rebuild 3D scenes from images, and a 3D-object generator for high-quality geometry and texture. 4) GR-3/ GR-RL: Robot brain models. GR-3 (Jul 2025) handles long, two-handed tasks, including soft items such as clothing. GR-RL (Dec 2025) used real-world practice to raise shoe-lacing success from 45.7% to 83.3%. 5) Pico: VR headsets, the planned hardware entry point for its world model.

• Use cases: 1) Content creation: short dramas, ads, and film pre-production, with Seedance 2.5’s green-screen and mock-up tools aimed at professional studios. 2) Interactive entertainment (planned): live streams, dramas, and games as 3D spaces you can explore on Pico; 3) 3D assets for games and e-commerce (Seed3D 2.0). 4) Robotics: household tasks and fine manipulation (GR-3, GR-RL), plus investments in robot makers.

AI Applications: ByteDance and BABA Lead in AI Assistant; WeChat AI has Great Potential

LLM App a Centralized Consumer Gateway that Unifies Fragmented Apps and Fosters AI Agents

• We believe LLM Apps are important for model vendors mainly because:

• Securing the AI Entrance and Ecosystem: AI vendors’ core ambition is to create an omni-consumer gateway or "AI OS" (Operating System) of the future, which is expected to replace traditional access points, such as browsers or app stores. The Internet giants aim to build a "Super App" or platform gateway to unify previously fragmented applications. Alibaba, for instance, views its "Qwen" app as a "life service entrance" designed to integrate its vast ecosystem, including commerce, payment, and local services, into a single user entrance.

• Platform for Agentic AI: The development of large language models has progressed from merely understanding language to being capable of executing complex tasks (known as Agentic AI or AI Agent). AI vendors could provide AI assistants that not only chat with users, but also genuinely "do things" for the user, such as generating elaborate PPTs or completing complex workflows.

• Driving New Revenue and Growth: LLM apps could provide a new source of revenue generation. Although unlike US vendors that are already monetizing via subscription on different tier accounts, Chinese vendors could still find ways to monetize via VAS, such as voice recording and transcribing, video creation, etc. The AI gateway could enhance an Internet platform’s ecosystem by raising the market share of its various Internet services.

Doubao Still Lead on Mobile APP MAU, Qwen Narrows the Gap with Doubao in Aug

• LLM apps are crucial to model vendors as they serve as key access points for consumers. These apps unify fragmented services into a single platform (e.g., Alibaba’s Qwen as an “omni access point”). LLM apps also enable agentic AI assistants that not only chat but execute complex tasks and workflows.

• In Aug, ByteDance’s Doubao remained the leader on AICPB.com’s list as China’s largest LLM app, with its mobile MAU reaching 314m, but MAU slightly declined in the past month.

• Alibaba’s Qwen’s mobile MAU reached at 272m in Jul, and is the only LLM app that received consistent MAU growth in the past three months.

横向滑动查看完整图表

Chinese LLM Apps have Accumulated Large User Bases, But Monetization Lags US Peers

• ChatGPT is still the world’s largest LLM app, tracking 1,030m MAU globally in Aug, increased by 3% MoM , according to AICPB.com.

• Other major US LLM app including Gemini, Grok, Perplexity, and Claude bagged 189m/68m/24m/43m of MAUs in Apr, and grew by 8.5%/-2.9%/-7.7%/4.3% MoM respectively.

• Major Chinese LLM apps have already accumulated sizable user bases comparing to their US counterparts. However, Chinese LLM apps still trail far behind their US peers in terms of commercialization, given their low API pricing strategy.

Appendix

Glossary

横向滑动查看完整图表

Coverage List (1/2)

横向滑动查看完整图表

Coverage List (2/2)

横向滑动查看完整图表

We would like to thank Jon Zhang, employee of Evalueserve Inc., for providing research support services to our preparation of this report. We would like to thank Connie Yan, employee of Evalueserve Inc., for providing research support services to our preparation of this report.

Company Valuation/Risks

For Important Disclosure information on companies recommended in this report, please visit our website at https://javatar.bluematrix.com/ sellside/Disclosures.action or call 212.284.2300.

Analyst Certification:

I, Edison Lee, CFA, certify that all of the views expressed in this research report accurately reflect my personal views about the subject security(ies) and subject company(ies). I also certify that no part of my compensation was, is, or will be, directly or indirectly, related to the specific recommendations or views expressed in this research report. I, Matt Ma, certify that all of the views expressed in this research report accurately reflect my personal views about the subject security(ies) and subject company(ies). I also certify that no part of my compensation was, is, or will be, directly or indirectly, related to the specific recommendations or views expressed in this research report. I, Annie Ping, CFA, FRM, certify that all of the views expressed in this research report accurately reflect my personal views about the subject security(ies) and subject company(ies). I also certify that no part of my compensation was, is, or will be, directly or indirectly, related to the specific recommendations or views expressed in this research report. I, Nick Cheng, certify that all of the views expressed in this research report accurately reflect my personal views about the subject security(ies) and subject company(ies). I also certify that no part of my compensation was, is, or will be, directly or indirectly, related to the specific recommendations or views expressed in this research report. I, Jacky He, certify that all of the views expressed in this research report accurately reflect my personal views about the subject security(ies) and subject company(ies). I also certify that no part of my compensation was, is, or will be, directly or indirectly, related to the specific recommendations or views expressed in this research report. Registration of non-US analysts: Edison Lee, CFA is employed by Jefferies Hong Kong Limited, a non-US affiliate of Jefferies LLC and is not registered/ qualified as a research analyst with FINRA. This analyst(s) may not be an associated person of Jefferies LLC, a FINRA member firm, and therefore may not be subject to the FINRA Rule 2241 and restrictions on communications with a subject company, public appearances and trading securities held by a research analyst. Registration of non-US analysts: Matt Ma is employed by Jefferies Hong Kong Limited, a non-US affiliate of Jefferies LLC and is not registered/qualified as a research analyst with FINRA. This analyst(s) may not be an associated person of Jefferies LLC, a FINRA member firm, and therefore may not be subject to the FINRA Rule 2241 and restrictions on communications with a subject company, public appearances and trading securities held by a research analyst. Registration of non-US analysts: Annie Ping, CFA, FRM is employed by Jefferies Hong Kong Limited, a non-US affiliate of Jefferies LLC and is not registered/ qualified as a research analyst with FINRA. This analyst(s) may not be an associated person of Jefferies LLC, a FINRA member firm, and therefore may not be subject to the FINRA Rule 2241 and restrictions on communications with a subject company, public appearances and trading securities held by a research analyst. Registration of non-US analysts: Nick Cheng is employed by Jefferies Hong Kong Limited, a non-US affiliate of Jefferies LLC and is not registered/qualified as a research analyst with FINRA. This analyst(s) may not be an associated person of Jefferies LLC, a FINRA member firm, and therefore may not be subject to the FINRA Rule 2241 and restrictions on communications with a subject company, public appearances and trading securities held by a research analyst. Registration of non-US analysts: Jacky He is employed by Jefferies Hong Kong Limited, a non-US affiliate of Jefferies LLC and is not registered/qualified as a research analyst with FINRA. This analyst(s) may not be an associated person of Jefferies LLC, a FINRA member firm, and therefore may not be subject to the FINRA Rule 2241 and restrictions on communications with a subject company, public appearances and trading securities held by a research analyst. As is the case with all Jefferies employees, the analyst(s) responsible for the coverage of the financial instruments discussed in this report receives compensation based in part on the overall performance of the firm, including investment banking income. We seek to update our research as appropriate, but various regulations may prevent us from doing so. Aside from certain industry reports published on a periodic basis, the large majority of reports are published at irregular intervals as appropriate in the analyst's judgement.

Investment Recommendation Record

(Article 3(1)e and Article 7 of MAR) Recommendation Published October 1, 2026 2:37 A.M. Recommendation Distributed October 1, 2026 2:37 A.M.

Company Specific Disclosures

Jefferies is involved in the deal where Volta Energy Solutions S a r l is selling Circuit Foil Luxembourg S a r l to Jiujiang Defu Technology Co Ltd 301511 CH Solus Advanced Materials Co Ltd 336370 KS is the parent company of Volta Energy Solutions S a r l

For Important Disclosure information on companies recommended in this report, please visit our website at https://javatar.bluematrix.com/ sellside/Disclosures.action or call 212.284.2300.

Explanation of Jefferies Ratings

Buy - Describes securities that we expect to provide a total return (price appreciation plus yield) of 15% or more within a 12-month period. Hold - Describes securities that we expect to provide a total return (price appreciation plus yield) of plus 15% or minus 10% within a 12-month period. Underperform - Describes securities that we expect to provide a total return (price appreciation plus yield) of minus 10% or less within a 12-month period. The expected total return (price appreciation plus yield) for Buy rated securities with an average security price consistently below $10 is 20% or more within a 12-month period as these companies are typically more volatile than the overall stock market. For Hold rated securities with an average security price consistently below $10, the expected total return (price appreciation plus yield) is plus or minus 20% within a 12-month period. For Underperform rated securities with an average security price consistently below $10, the expected total return (price appreciation plus yield) is minus 20% or less within a 12month period. NR - The investment rating and price target have been temporarily suspended. Such suspensions are in compliance with applicable regulations and/or Jefferies policies. CS - Coverage Suspended. Jefferies has suspended coverage of this company. NC - Not covered. Jefferies does not cover this company. Restricted - Describes issuers where, in conjunction with Jefferies engagement in certain transactions, company policy or applicable securities regulations prohibit certain types of communications, including investment recommendations. Monitor - Describes securities whose company fundamentals and financials are being monitored, and for which no financial projections or opinions on the investment merits of the company are provided.

Valuation Methodology

Jefferies' methodology for assigning ratings may include the following: market capitalization, maturity, growth/value, volatility and expected total return over the next 12 months. The price targets are based on several methodologies, which may include, but are not restricted to, analyses of market risk, growth rate, revenue stream, discounted cash flow (DCF), EBITDA, EPS, cash flow (CF), free cash flow (FCF), EV/EBITDA, P/E, PE/growth, P/CF, P/FCF, premium (discount)/average group EV/EBITDA, premium (discount)/average group P/E, sum of the parts, net asset value, dividend returns, and return on equity (ROE) over the next 12 months. Jefferies Franchise Picks Jefferies Franchise Picks include stock selections from among the best stock ideas from our equity analysts over a 12 month period. Stock selection is based on fundamental analysis and may take into account other factors such as analyst conviction, differentiated analysis, a favorable risk/reward ratio and investment themes that Jefferies analysts are recommending. Jefferies Franchise Picks will include only Buy rated stocks and the number can vary depending on analyst recommendations for inclusion. Stocks will be added as new opportunities arise and removed when the reason for inclusion changes, the stock has met its desired return, if it is no longer rated Buy and/or if it triggers a stop loss. Stocks having 120 day volatility in the bottom quartile of S&P stocks will continue to have a 15% stop loss, and the remainder will have a 20% stop. Franchise Picks are not intended to represent a recommended portfolio of stocks and is not sector based, but we may note where we believe a Pick falls within an investment style such as growth or value.

Risks which may impede the achievement of our Price Target

This report was prepared for general circulation and does not provide investment recommendations specific to individual investors. As such, the financial instruments discussed in this report may not be suitable for all investors and investors must make their own investment decisions based upon their specific investment objectives and financial situation utilizing their own financial advisors as they deem necessary. Past performance of the financial instruments recommended in this report should not be taken as an indication or guarantee of future results. The price, value of, and income from, any of the financial instruments mentioned in this report can rise as well as fall and may be affected by changes in economic, financial and political factors. If a financial instrument is denominated in a currency other than the investor's home currency, a change in exchange rates may adversely affect the price of, value of, or income derived from the financial instrument described in this report. To the extent prices are shown in non-US currency, please note that our local currency price targets are based on a currency conversion using an exchange rate as of the prior trading day (unless otherwise noted). Should there be fluctuations in the exchange rate after this date, that will affect the non-US target prices and should no longer be relied upon. In addition, investors in securities such as ADRs, whose values are affected by the currency of the underlying security, effectively assume currency risk.

Other Companies Mentioned in This Report

• AAC Technologies Holdings Inc. (2018 HK: HK$39.44, HOLD)

• Advanced Micro-Fabrication Eqp Inc China (688012 CH: CNY318.55, BUY)

• Alchip Technologies Ltd (3661 TT: TWD3,780.00, BUY)

• Anhui Tongguan Copper Foil Group Co., Ltd. (301217 CH: CNY99.56, BUY)

• Apple Inc. (AAPL: $333.02, UNDERPERFORM)

• AsiaInfo Technologies Ltd (1675 HK: HK$4.04, BUY)

• Avary Holding (Shenzhen) Co., Ltd. (002938 CH: CNY79.55, BUY)

• BYD Electronic (International) Co., Ltd. (285 HK: HK$22.98, HOLD)

• Chanjet Information Technology Company Ltd. (1588 HK: HK$4.22, BUY)

• China Communications Services Corp (552 HK: HK$4.11, HOLD)

• China Mobile Limited (941 HK: HK$79.45, BUY)

• China Telecom Corp Ltd. (728 HK: HK$4.50, BUY)

• China Tower Corporation Limited (788 HK: HK$9.00, BUY)

• China Unicom (Hong Kong) Ltd. (762 HK: HK$5.46, BUY)

• Cowell e Holdings Inc. (1415 HK: HK$22.50, BUY)

• Delton Technology (Guangzhou) Inc. (001389 CH: CNY176.66, BUY)

• Delton Technology (Guangzhou) Inc. (1989 HK: HK$119.60, BUY)

• Faraday Technology Corp (3035 TT: TWD212.00, HOLD)

• Global Unichip Corp (3443 TT: TWD7,990.00, BUY)

• Glodon Software (002410 CH: CNY8.54, UNDERPERFORM)

• Goertek Inc. (002241 CH: CNY22.79, BUY)

• Hangzhou Silan Microelectronics Co., Ltd. (600460 CH: CNY30.75, UNDERPERFORM)

• Hua Hong Semiconductor (1347 HK: HK$105.40, HOLD)

• Hundsun Technologies Inc (600570 CH: CNY19.40, UNDERPERFORM)

• InnoScience (Suzhou) Technology Holding Co., Ltd. (2577 HK: HK$39.38, BUY)

• Insyde Software Corp (6231 TT: TWD235.00, BUY)

• Jiujiang Defu Technology Co., Ltd. (301511 CH: CNY95.25, BUY)

• Kingdee International Software Group Co Ltd. (268 HK: HK$7.35, BUY)

• Lens Technology Co. Ltd. (300433 CH: CNY32.31, BUY)

• Lens Technology Co. Ltd. (6613 HK: HK$21.60, BUY)

• Lingyi iTech (Guangdong) Co. (002600 CH: CNY11.88, HOLD)

• Luxshare Precision Industry Co., Ltd. (002475 CH: CNY49.43, BUY)

• Maxscend Microelectronics Co Ltd (300782 CH: CNY70.17, BUY)

• OmniVision Integrated Circuits Group, Inc. (603501 CH: CNY82.21, UNDERPERFORM)

• QiAnXin (688561 CH: CNY29.90, HOLD)

• SICC Co., Ltd. (688234 CH: CNY104.39, UNDERPERFORM)

• Sangfor Technologies (300454 CH: CNY137.42, HOLD)

• Semiconductor Manufacturing International Corporation (981 HK: HK$60.65, BUY)

• Shanghai Baosight Software Co., Ltd. (600845 CH: CNY17.69, BUY)

• Shengyi Technology Co Ltd (600183 CH: CNY129.96, BUY)

• Shennan Circuits Co Ltd (002916 CH: CNY373.01, BUY)

• Shenzhen Everwin Precision Technology Co., Ltd. (300115 CH: CNY22.48, HOLD)

• Shenzhen Fastprint Circuit Tech Co., Ltd. (002436 CH: CNY40.94, BUY)

• Sinosoft (603927 CH: CNY13.64, BUY)

• StarPower Semiconductor Ltd. (603290 CH: CNY81.21, BUY)

• Sunny Optical Technology (Group) Co., Ltd. (2382 HK: HK$63.35, HOLD)

• Tuya Inc. (TUYA: $1.70, BUY)

• VNET Group (VNET: $5.55, BUY)

• Venustech (002439 CH: CNY14.68, UNDERPERFORM)

• WUS Printed Circuit (002463 CH: CNY115.93, BUY)

• Wingtech Technology Co., Ltd. (600745 CH: CNY16.43, HOLD)

• Xiaomi Corporation (1810 HK: HK$25.24, HOLD)

• Yangtze Optical Fiber and Cable Joint Stock Ltd. Co. (6869 HK: HK$168.10, HOLD)

• Yonyou Network Technology Co., Ltd. (600588 CH: CNY9.22, UNDERPERFORM)

• Yusys Technologies Co Ltd (300674 CH: CNY15.04, BUY)

• Z.AI (2513 HK: HK$641.00, HOLD)

• ZTE Corporation (763 HK: HK$21.86, UNDERPERFORM)

• ZWSoft CO., LTD (688083 CH: CNY50.11, HOLD)

• Zhejiang Crystal-Optech Co., Ltd (002273 CH: CNY24.40, BUY)

• Zhejiang Lante Optics Co.,Ltd. (688127 CH: CNY61.05, BUY)

• Zhejiang Supcon Technology Group Co., LTD (688777 CH: CNY80.05, HOLD)

横向滑动查看完整图表

Other Important Disclosure

Other Important Disclosures Jefferies does business and seeks to do business with companies covered in its research reports, and expects to receive or intends to seek compensation for investment banking services among other activities from such companies. As a result, investors should be aware that Jefferies may have a conflict of interest that could affect the objectivity of this report. Investors should consider this report as only a single factor in making their investment decision. Jefferies Equity Research refers to research reports produced by analysts employed by one of the following Jefferies Financial Group Inc. ("Jefferies") companies: United States: Jefferies LLC which is an SEC registered broker-dealer and a member of FINRA (and distributed by Jefferies Research Services, LLC, an SEC registered Investment Adviser, to clients paying separately for such research). Canada: Jefferies Securities Inc., which is an investment dealer registered in each of the thirteen Canadian jurisdictions and a dealer member of the Canadian Investment Regulatory Organization, including research reports produced jointly by Jefferies Securities Inc. and another Jefferies entity (and distributed by Jefferies Securities Inc.). Where Jefferies Securities Inc. distributes research reports produced by Jefferies LLC, Jefferies International Limited, Jefferies Japan Company Limited, or Jefferies India Private Limited, you are advised that each of Jefferies LLC, Jefferies International Limited, Jefferies Japan Company Limited, and Jefferies India Private Limited operates as a dealer in your jurisdiction under an exemption from the dealer registration requirements contained in National Instrument 31-103 Registration Requirements, Exemptions and Ongoing Registrant Obligations and, as such, each of Jefferies LLC, Jefferies International Limited, Jefferies Japan Company Limited, and Jefferies India Private Limited is not required to be and is not a registered dealer or adviser in your jurisdiction. You are advised that where Jefferies LLC or Jefferies International Limited prepared this research report, it was not prepared in accordance with Canadian disclosure requirements relating to research reports in Canada. United Kingdom: Jefferies International Limited, which is authorized and regulated by the Financial Conduct Authority; registered in England and Wales No. 1978621; registered office: 100 Bishopsgate, London EC2N 4JL; telephone +44 (0)20 7029 8000; facsimile +44 (0)20 7029 8010. Germany: Jefferies GmbH, which is authorized and regulated by the Bundesanstalt fuer Finanzdienstleistungsaufsicht, BaFin-ID: 10150151; registered office: Bockenheimer Landstr. 24, 60323 Frankfurt a.M., Germany; telephone: +49 (0) 69 719 1870 Hong Kong: Jefferies Hong Kong Limited, which is licensed by the Securities and Futures Commission of Hong Kong with CE number ATS546; located at Level 26, Two International Finance Center, 8 Finance Street, Central, Hong Kong; telephone: +852 3743 8000. Singapore: Jefferies Singapore Limited, which is licensed by the Monetary Authority of Singapore; located at 10 Collyer Quay #41-01, Ocean Financial Centre, Singapore 049315, telephone: +65 6551 3950. Japan: Jefferies Japan Company Limited, which is a securities company registered by the Financial Services Agency of Japan and is a member of the Japan Securities Dealers Association; located at Tokyo Midtown Hibiya 30F Hibiya Mitsui Tower, 1-1-2 Yuraku-cho, Chiyoda-ku, Tokyo 100-0006; telephone +813 5251 6100; facsimile +813 5251 6101. India: Jefferies India Private Limited (CIN - U74140MH2007PTC200509), licensed by the Securities and Exchange Board of India for: Stock Broker (NSE & BSE) INZ000243033, Research Analyst INH000000701 and Merchant Banker INM000011443, located at Level 16, Express Towers, Nariman Point, Mumbai 400 021, India; Tel +91 22 4356 6000. Compliance Officer name: Sanjay Pai, Tel No: +91 22 42246150, Email: spai@jefferies.com, Grievance officer name: Sanjay Pai, Tel no. +91 22 42246150, Email: compliance_india@jefferies.com. Registration granted by SEBI and certification from NISM in no way guarantee performance of the intermediary or provide any assurance of returns to investors. Australia: Jefferies (Australia) Pty Limited (ACN 623 059 898), which holds an Australian financial services license (AFSL 504712) and is located at Level 20, 60 Martin Place, Sydney NSW 2000; telephone +61 2 9364 2800. Dubai: Jefferies International Limited, Dubai branch, which is licensed by the Dubai Financial Services Authority (DFSA Reference Number F007325); registered office Unit L31-06, L31-07, Level 31, ICD Brookfield Pace, DIFC, PO Box 121208, Dubai, UAE. This report was prepared by personnel who are associated with Jefferies (Jefferies Securities Inc., Jefferies International Limited, Jefferies GmbH, Jefferies Hong Kong Limited, Jefferies Singapore Limited, Jefferies Japan Company Limited, Jefferies India Private Limited), and Jefferies (Australia) Pty Ltd; or by personnel who are associated with both Jefferies LLC and Jefferies Research Services LLC ("JRS"). Jefferies LLC is a US registered broker-dealer and is affiliated with JRS, which is a US registered investment adviser. JRS does not create tailored or personalized research and all research provided by JRS is impersonal. If you are paying separately for this research, it is being provided to you by JRS. Otherwise, it is being provided by Jefferies LLC. Jefferies LLC, JRS, and their affiliates are collectively referred to below as "Jefferies". Jefferies may seek to do business with companies covered in this research report. As a result, investors should be aware that Jefferies may have a conflict of interest that could affect the objectivity of this report. Investors should consider this report as only one of many factors in making their investment decisions. Specific conflict of interest and other disclosures that are required by FINRA, the Canadian Investment Regulatory Organization and other rules are set forth in this disclosure section. * * * If you are receiving this report from a non-US Jefferies entity, please note the following: Unless prohibited by the provisions of Regulation S of the U.S. Securities Act of 1933, as amended, this material is distributed in the United States by Jefferies LLC, which accepts responsibility for its contents in accordance with the provisions of Rule 15a-6 under the US Securities Exchange Act of 1934, as amended. Transactions by or on behalf of any US person may only be effected through Jefferies LLC. In the United Kingdom and European Economic Area this report is issued and/or approved for distribution by Jefferies International Limited ("JIL”) and/or Jefferies GmbH and is intended for use only by persons who have, or have been assessed as having, suitable professional experience and expertise, or by persons to whom it can be otherwise lawfully distributed. Jefferies LLC, JIL, Jefferies GmbH and their affiliates, may make a market or provide liquidity in the financial instruments referred to in this report; and where they do make a market, such activity is disclosed specifically in this report under “company specific disclosures”. For Canadian investors, this material is intended for use only by professional or institutional investors. None of the investments or investment services mentioned or described herein is available to other persons or to anyone in Canada who is not a "permitted client" as defined by National Instrument 31-103 Registration Requirements, Exemptions and Ongoing Registrant Obligations, as applicable. This research report is a general discussion of the merits and risks of a security or securities only, and is not in any way meant to be tailored to the needs and circumstances of any recipient. The information contained herein is not, and under no circumstances is to be construed as, an offer to sell securities described herein, or solicitation of an offer to buy securities described herein, in Canada or any province or territory thereof. Any offer or sale of the securities described herein in Canada will be made only under an exemption from the requirements to file a prospectus with the relevant Canadian securities regulators, if applicable, and only by a dealer properly registered under applicable securities laws or, alternatively, pursuant to an exemption from the dealer registration requirement in the relevant province or territory of Canada in which such offer or sale is made. The information contained herein is under no circumstances to be construed as investment advice in any province or territory of Canada and is not tailored to the needs of the recipient. To the extent that the information contained herein references securities of an issuer incorporated, formed or created under the laws of Canada or a province or territory of Canada, any trades in such securities must be conducted through a dealer registered in Canada. No securities commission or similar regulatory authority in Canada has reviewed or in any way passed judgment upon this research report, the information contained herein or the merits of the securities described herein, and any representation to the contrary is an offence. In Singapore, Jefferies Singapore Limited (“JSL”) is regulated by the Monetary Authority of Singapore. For investors in the Republic of Singapore, where this material is prepared and issued by a Jefferies affiliate outside of Singapore, it is distributed by JSL pursuant to Regulation 32C of the Financial Advisers Regulations. The material contained in this document is intended solely for accredited, expert or institutional investors, as defined under the Securities and Futures Act 2001 (Singapore). If there are any matters arising from, or in connection with this material, please contact JSL, located at 10 Collyer Quay #41-01, Ocean Financial Centre, Singapore 049315, telephone: +65 6551 3950. In Dubai, this material is issued and distributed by Jefferies International Limited, Dubai branch, and is intended solely for Professional Clients and should not be distributed to, or relied upon by, Retail Clients (as defined by DFSA). A distribution of ratings in percentage terms in each sector covered is available upon request from your sales representative. In Japan, this material is issued and distributed by Jefferies Japan Company Limited to institutional investors only. In Hong Kong, this report is issued and approved by Jefferies Hong Kong Limited and is intended for use only by professional investors as defined in the Hong Kong Securities and Futures Ordinance and its subsidiary legislation. In the Republic of China (Taiwan), this report should not be distributed. The research in relation to this report is conducted outside the People’s Republic of China (“PRC”). This report does not constitute an offer to sell or the solicitation of an offer to buy any securities in the PRC. PRC investors shall have the relevant qualifications to invest in such securities and shall be responsible for obtaining all relevant approvals, licenses, verifications and/or registrations from the relevant governmental authorities themselves. In India, this report is made available by Jefferies India Private Limited. In Australia, this report is issued and/or approved for distribution by, or on behalf of, Jefferies (Australia) Securities Pty Ltd (ACN 610 977 074), which holds an Australian financial services license (AFSL 487263). It is directed solely at wholesale clients within the meaning of the Corporations Act 2001 (Cth) of Australia (the “Corporations Act”), in connection with their consideration of any investment or investment service that is the subject of this report. This report may contain general financial product advice. Where this report refers to a particular financial product, you should obtain a copy of the relevant product disclosure statement or offer document before making any decision in relation to the product. Recipients of this document in any other jurisdictions should inform themselves about and observe any applicable legal requirements in relation to the receipt of this document. This report is not an offer or solicitation of an offer to buy or sell any security or derivative instrument, or to make any investment. Any opinion or estimate constitutes the preparer's best judgment as of the date of preparation, and is subject to change without notice. Jefferies assumes no obligation to maintain or update this report based on subsequent information and events. Jefferies, and their respective officers, directors, and employees, may have long or short positions in, or may buy or sell any of the securities, derivative instruments or other investments mentioned or described herein, either as agent or as principal for their own account. This material is provided solely for informational purposes and is not tailored to any recipient, and is not based on, and does not take into account, the particular investment objectives, portfolio holdings, strategy, financial situation, or needs of any recipient. As such, any advice or recommendation in this report may not be suitable for a particular recipient. Jefferies assumes recipients of this report are capable of evaluating the information contained herein and of exercising independent judgment. A recipient of this report should not make any investment decision without first considering whether any advice or recommendation in this report is suitable for the recipient based on the recipient’s particular circumstances and, if appropriate or otherwise needed, seeking professional advice, including tax advice. Jefferies does not perform any suitability or other analysis to check whether an investment decision made by the recipient based on this report is consistent with a recipient’s investment objectives, portfolio holdings, strategy, financial situation, or needs. By providing this report, neither JRS nor any other Jefferies entity accepts any authority, discretion, or control over the management of the recipient’s assets. Any action taken by the recipient of this report, based on the information in the report, is at the recipient’s sole judgment and risk. The recipient must perform his or her own independent review of any prospective investment. If the recipient uses the services of Jefferies LLC (or other affiliated brokerdealers), in connection with a purchase or sale of a security that is a subject of these materials, such broker-dealer may act as principal for its own accounts or as agent for another person. Only JRS is registered with the SEC as an investment adviser; and therefore neither Jefferies LLC nor any other Jefferies affiliate has any fiduciary duty in connection with distribution of these reports. The price and value of the investments referred to herein and the income from them may fluctuate. Past performance is not a guide to future performance, future returns are not guaranteed, and a loss of original capital may occur. Fluctuations in exchange rates could have adverse effects on the value or price of, or income derived from, certain investments. This report may contain forward looking statements that may be affected by inaccurate assumptions or by known or unknown risks, uncertainties, and other important factors. As a result, the actual results, events, performance or achievements of the financial product may be materially different from those expressed or implied in such statements. This report has been prepared independently of any issuer of securities mentioned herein and not as agent of any issuer of securities. No Equity Research personnel have authority whatsoever to make any representations or warranty on behalf of the issuer(s). Any comments or statements made herein are those of the Jefferies entity producing this report and may differ from the views of other Jefferies entities. This report may contain information obtained from third parties, including ratings from credit ratings agencies such as Standard & Poor’s, and information derived from third-party or proprietary generative artificial intelligence (Gen AI) models. Jefferies does not guarantee the accuracy, completeness, timeliness or availability of this information, and is not responsible for any errors or omissions (negligent or otherwise), regardless of the cause, or for the results obtained from the use of such content. Neither Jefferies nor any third-party content providers, including providers of Gen AI models, give any express or implied warranties, including, but not limited to, any warranties of merchantability or fitness for a particular purpose or use. Neither Jefferies nor any third-party content provider shall be liable for any direct, indirect, incidental, exemplary, compensatory, punitive, special or consequential damages, costs, expenses, legal fees, or losses (including lost income or profits and opportunity costs) in connection with any use of their content, including ratings. Credit ratings are statements of opinions and are not statements of fact or recommendations to purchase, hold or sell securities. They do not address the suitability of securities or the suitability of securities for investment purposes, and should not be relied on as investment advice. Reproduction and distribution of third party content in any form is prohibited except with the prior written permission of the related third party. Jefferies research reports are disseminated and available electronically, and, in some cases, also in printed form. Electronic research is simultaneously made available to all clients. This report or any portion hereof may not be copied, reprinted, sold, or redistributed or disclosed by the recipient or any third party, by content scraping or extraction, automated processing, or any other form or means, without the prior written consent of Jefferies. Any unauthorized use is prohibited. Neither Jefferies nor any of its respective directors, officers or employees, is responsible for guaranteeing the financial success of any investment, or accepts any liability whatsoever for any direct, indirect or consequential damages or losses arising from any use of this report or its contents. Nothing herein shall be construed to waive any liability Jefferies has under applicable U.S. federal or state securities laws. For Important Disclosure information relating to JRS, please see https://adviserinfo.sec.gov/IAPD/Content/Common/crd_iapd_Brochure.aspx? BRCHR_VRSN_ID=483878 and https://adviserinfo.sec.gov/Firm/292142 or visit our website at https://javatar.bluematrix.com/sellside/Disclosures.action, or www.jefferies.com, or call 1.888.JEFFERIES. © 2026 Jefferies