Is AI a bubble yet? Our five gauges say no

Since our last update, AI revenues have continued to rise, reaching $126 billion over the last twelve months as of July. We also experienced a jumpy market, which led to a severe correction in semiconductor stocks, somewhat cooling public valuations. On our side, we have improved the methodology for counting AI capex (we show both the published and restated series below).

That demand has smacked headlong into a tight supply of compute capacity, which is being met by increasing investment in infrastructure. And with that comes more risk. While the hyperscalers are still using a large share of their cash reserves, they are increasingly scouring the globe for capital, both straight-up debt and increasingly intricate financing vehicles. As Michael Parekh argues, this “gaming of the system” is not only rational; it is necessary, as long as revenue is compounding. But these structures can become brittle if it slows.

Funding quality has deteriorated since Sep 2025. In our base case, we expect it and economic strain to turn red during 2027.

The full analysis shows where the tension is building.

Economic strain: 2027 is a key year

Economic strain measures how heavily a given investment boom weighs on the economy as a whole.

Global AI capex is climbing rapidly, and some analysts forecast it might reach $1 trillion in 2027.

Looking at historical boom-and-bust cycles, investment in a single infrastructure sector exceeding 2% of GDP is often associated with the later stages of a cycle and can warn of an impending reversal. In the railroad busts of the 1870s and 1880s, infrastructure investment exceeded 3.5% of America’s GDP.

For the AI wave, we look at the US economy and apportion its share of global AI capex. We’ve updated out methodology to isolate just AI capex. Under our old methodology, economic strain has reached 1.32%, roughly double its level last year. However, we’ve improved our methodology to exclude warehouses, delivery networks, and non-AI cloud infrastructure. On our new, more precise measure, it sits at 0.86% (the purple line below.)

There is no sign of the companies slowing down. The big four raised capex guidance, again: Alphabet to $195–205 billion, Meta to $130–145 billion, Amazon to some $220 billion, and Microsoft boosted its investment plans with real spend up 69% year on year.¹ On our estimates, some two-thirds is now AI capex. As Andy Jassy, Amazon’s boss, pointed out in the Q4 2025 earnings call: “We have…a fair bit of experience over the years in AWS of forecasting demand signals and doing it in such a way that we don’t have a lot of wasted capacity and that we also have enough capacity to serve the demand that’s there.” In other words, Amazon views its investment as disciplined capacity planning.

These headline numbers don’t necessarily reflect actual spend. A growing share of commitments is stuck in the capex logjam. More thae $315 billion across the disclosing hyperscalers, up from $281 billion a quarter earlier, is waiting in the construction queue.

If capex commitments continue to rise as many expect, this indicator will move into the red at the end of next year.

Industry strain: is the revenue keeping up?

Industry strain compares current investment in AI infrastructure with revenues being generated in the sector. It is a rough measure for how far commercial returns lag the build-out. It is not a full calculation of whether the assets will ultimately pay back.

Across historical cycles, investment sustained at more than six times revenue has warned of a possible bust. For comparison, the ratio peaked at roughly ten times during the 1990s dotcom bubble.

Our proprietary revenue modeling shows second-quarter 2026 deduplicated AI revenue was $43 billion. For the full year to the end of July 2026, revenues were $126 billion, 3.1 times higher than a year ago.² I expect we’ll end the year somewhere between $190 billion and $200 billion.

Under our new AI-capex method, capex runs at 3.6 times revenue, well into the amber and below the telecom bubble’s peak. Under our original all-capex method, it is 5.6 times. (Institutional clients can request access to our full depreciation modeling here).

The Industry Strain indicator is healthier than we expected because revenue growth beat our expectations. This gauge should enter green during 2027.

Two caveats:

• Part of the improvement comes from our methodology change (AI capex against AI revenue) rather than from drastic changes on the ground.
• We estimate that in Q2 2026, 75% of this spend comes from traditional companies and customers buying AI services. The remainder is from the labs themselves buying compute to train models and from venture-backed startups. If you discounted those numbers, industry strain would sit somewhat higher at 4.8x but still in amber.

Revenue momentum: strong, but...

Demand is not slowing down. We measure revenue momentum as the time it takes for revenues to double. To our surprise, revenue has grown faster in the past few quarters than at the latter stages of 2025. Doubling time is now 0.58 years compared to a slower 0.66 years in March 2026.

For the full year of 2025, AI revenues globally ex-China grew 210% to $63 billion. In the 12 months to July 2026, they increased to $126 billion, up 212% from a year earlier.

Our estimates for July 2026 put OpenAI at roughly $40 billion and Anthropic at $65 billion in annualized revenue. That’s three and thirteen times higher, respectively, than when we published our first Boom or Bubble update last September.

It is worth noting that OpenAI’s revenue growth has lagged the sector’s overall growth. The firm added 18% in sales in Q2 2026. We would caution against reading too much into a single quarter that is wildly above or below the trend.

Can this indicator remain strong in the long term, or could it weaken abruptly? In the past six months, token prices have fallen 67%, while revenues have grown 84%. In other words, it’s increasing volume that is doing all the work.

Open-weight challengers like K3 and Qwen3.8 could put pressure on the frontier lab revenues. But hyperscalers still earn good money serving open models. We’ll turn to the impact of potential shifts towards open-weight models in a future update.

Valuation heat: how reasonable is the market?

We measure valuation heat as the price-to-earnings ratio of the NASDAQ-100. The NASDAQ-100’s trailing P/E fell from 33.4 to 29.4 since our last update. We explain why we use this ratio rather than the CAPE Shiller PE ratio here.

Valuation Heat has returned to green. But lower valuations don’t mean lower expectations. Meta’s stock price fell 10% and Alphabet’s 5%, even as the social network company grew revenue 28% year on year and Alphabet’s cloud business grew 82% with a $514 billion backlog.

Funding quality: worsening, by design

Our final indicator asks whether the financing behind the build-out is of sufficient quality to absorb a revenue shock. As we move through an investment cycle, capital risk shifts from incumbents’ equity to debt providers and, finally, to structured vehicles that aggregate less specialist capital.

It is a composite score based on four sub-indicators, each scored 0 to 3:

• the funding mix (cash versus debt),
• intermediation discipline (who lends, under what rules),
• tenor match (whether the money’s timeline matches the assets’)
• systemic loss absorption (who eats it if things break).

More details of the methodology are here.

Poor funding quality is often the dry tinder that ignites a crisis, as it was in the housing, telecoms and railroad busts. On our scoring, for example, the telecoms bust of 2000 had a funding quality of 2.8, well above our red threshold of 2. Because poor funding quality is such a tell in an overheated market, we’ll dig into it in more detail.

Funding mix. Funding quality is highest when a company spends its own cash. The AI buildout was largely funded from companies’ own balance sheets last year. Since then, the external debt has increasingly picked up the slack. The cash-funded share of hyperscaler capex has fallen from 85% in 2024 to roughly two-thirds this year.

Big Tech has chewed through its spare cash flow, from $62 billion in Q3 2025 to $21.1 billion in Q1 2026. Even Alphabet, the strongest balance sheet in tech, has so far raised $49.6 billion in equity, its first equity raise in over two decades.

Given the strength of demand for compute, we view these external sources of capital as necessary. Without it, customer demand will go unfulfilled. The trouble is, every new financing structure brings to the table investors who are less able to understand what they are buying.

Intermediation discipline. Labs and their suppliers have found themselves looking like banks: guaranteeing leases, backstopping chips, financing customers. This is because demand has materialized faster than conventional financing was ready to support. Nvidia, for instance, has recently partnered with Wall Street firms to mobilize more than $500 billion in third-party capital for AI infrastructure.³ But part of the agreement is that if the GPUs in the deal are ever worth less than assumed, Nvidia may cover up to 25% of the shortfall (decided on a project by project basis). Risky if the revenues don’t show up, but so far they have.

Tenor match. The debate over whether GPUs only last a couple of years has now (finally) been put to bed. Google is running eight-year-old TPUs at full utilization. CoreWeave, a neocloud, has customers who have bought capacity for six-year-old A100s through 2029. For now, capacity is so constrained that even short leases can support the underlying investment. Google’s SpaceX deal pays $920 million a month for access to roughly 110,000 GPUs through June 2029.⁴ It is worth noting that, unusually, both Google and Anthropic (also contracted to buy from SpaceX) have 90-day cancellation terms on their contracts.

Neoclouds face a slightly different tenor problem. CoreWeave has $10.6 billion of principal due over the next two years, while it is still loss-making ($600 million in the last quarter). As a result, its five-year CDS (insurance against default) has risen from roughly 360 basis points in September 2025 to around 855 basis points at the end of July 2026. Today, its GPU utilisation and revenues remain high.

Systemic loss absorption. AI-linked credit in general is getting more expensive. Goldman’s basket of AI-leader credit has doubled to roughly 150 basis points over Treasuries. It’s noticeable that these exposures are being priced deal by deal — xAI’s chip debt has actually risen in value. This is a sign of a maturing market engaging in more discriminatory underwriting. We’ve also yet to see evidence that market interconnections are dense enough to cause a systemic failure.⁵

Overall, this means that Funding quality is worsening. Cash funds less of the buildout, and the expansion of corporate backstops and debt vehicles is behind that. In September 2025, this gauge sat at 1.1. Today it is at 1.6.

As the chart below shows, we project funding quality to continue to worsen, reaching our red zone in mid-2027.

What this reading means

The AI investment wave remains a revenue-driven boom built on a chronic infrastructure scarcity. Strong end-customer demand is driving high utilization, which supports companies selling that scarce compute capacity and the firms that supply them with chips, memory, networking, and power.

But there is fragility in the financing layer. Companies are doing nearly whatever they can to secure financing to buy the capacity customers are demanding. This has become riskier than the cash-funded capex of the past few years. Companies are now guaranteeing other firms’ obligations and taking on more debt to finance it.

In our model, two indicators turning red signals that the boom has entered a later stage. It does not predict an imminent bust. Economic Strain and Funding Quality share the same capex driver, so we treat them as reinforcing warnings rather than independent confirmation.

In the case of the dotcom bubble, two gauges sat in the red for more than a year before the market peaked. In the case of the 19th-century railroad bubbles, admittedly a much slower era, it was closer to three years.

By the end of 2027, two indicators are likely to be in the red, even as revenues continue to increase.

What would change our call?

Nothing in the current data changes our call.

For that to happen in the short- to medium-term, either revenue growth would have to slow materially, or we would have to uncover deep fraud somewhere in the financing chain. Neither is visible today. Funding quality is deteriorating. But those obligations are still supported by real revenues and scarce, heavily-used capacity.

For now, this remains a boom, not a bubble.

Notes

1. Microsoft’s headline 2026 guidance actually fell from ~$190 billion to ~$175 billion — but entirely because it reclassified future finance leases as operating leases and extended data-center accounting life from 15 to 25 years.
2. We measure value added — revenue net of what AI companies pay each other — to avoid double-counting the same dollar as it passes from lab to cloud. On a gross basis, the figure would be $161 billion.
3. Currently non-binding.
4. The agreement also has an early-termination clause starting in 2027.
5. We are aware of the July 2026 BIS Working Paper, The AI investment race, which is a good theoretical read. However, its key numbers are a function of its input assumptions, which are reported as findings.