Things Are Cheaper (If You Know Where to Look)

1. Certain Things Are Cheaper (For Now), If You Know Where to Look

The recent CPI print has put some of the teeth-gnashing about raising rates to bed (even if it really shouldn’t matter, one way or another).

Random Walk continues to maintain that if monetary policy is not the problem, then monetary policy is not the solution. In this case, if the AI buildout is driving up the costs of technological components (and it is), and the impulse is driven by the spending decisions of some of the largest, most profitable companies the world has ever known (which it is), then the obvious cure to high prices is high prices.

For what it’s worth, Goldman Sachs agrees with me:

The combined price effects of tariffs, energy prices and tech hardware demand are expected to fade over the next year.

Putting that aside, I have this half-baked theory about why self-correcting inflationary pressure is still raising the hackles of at least some members of the Fed (other than the fact that letting everyone knows you’ve got hackles is the counter-Walsh tantrum de jour).

It occurs to me that perhaps part of the disconnect may arise from a feature of how we measure inflation.

You see, ordinarily, the reason that non-monetary (and non-fiscal) inflation is self-correcting is that higher prices suppress demand—either for the thing itself, or for some other ‘nice to have’ that gets squeezed out by the more expensive ‘need to have,’ (aka, ‘the cure to high prices is high prices.’) In the latter case, one component of the CPI may rise, but then another should fall. One inflation is offset by another deflation.

There’s deflation, it’s just not measured by the CPI

In this case, however, while that’s indeed what is happening—one price rises, while another one falls—it turns out that the deflation is not a measured component of the CPI (or PPI).

That’s because the thing that’s deflating are asset-prices:

Tech multiples, especially for those companies driving up the cost of memory etc. with their considerable outlays, have compressed to the middle of their historical ranges.

In other words, all that capex is driving the price of memory up, and the price of hyperscaler equity down. Put yet another way, the high price of “need to have” AI infrastructure has soaked up all the free cash flow, such that it’s squeezed out the “nice to have” of returning money to shareholders (mostly via buybacks).

One price goes up, and another one goes down. Self-correction working just fine.

But, here’s the thing: the price of equity is generally not a component of measured inflation—the component that comes closest is asset management fees (as a % of AUM), but they’re a relatively small piece of the puzzle. Now, there are plenty of good reasons why we don’t think of asset prices as part of the CPI, but in this case, it remains the case that the feedback loop of demand elasticity is functioning, just not within the usual framework for how we measure it.

And I think that it may be obscuring an important part of the inflation picture. Over time, the rising cost of capital should itself soften the appetite for spending (so the theory goes), but probably not on any timeline that would make 2% benchmark-watchers satisfied.

Idk. It’s a theory, and I’m not sure why it’s wrong per se (even if it might be), so much as “that’s just not how we measure inflation.”

Discount on aisle hyperscaler

Anyways, there’s another reason to bring this all up: it’s once again worth noticing that the largest, most-profitable companies the world has ever known are kinda cheap:

Other than Apple (which has largely opted out of the capex game), forward earnings multiples for ‘big tech’ are at or near the bottoms of their 10-year ranges.

On a monthly basis, the picture is largely the same, although Google is actually above its 10-year median and average:

Google re-rated off the Deepseek sell-off towards the end of ‘25, and while it’s come down from its peak, it’s still somewhat elevated relative to the historical norm. But Meta, Microsoft, and Amazon share prices are definitely feeling the cashflow burn.

The reason, of course, that these companies are so discounted is the point above:

Capex is expected to exceed 100% of cash flow this year and the next. That’s a lot of cash flow that’s (a) not being returned to shareholders; and (b) flowing into the AI infra coffers . . . and so investors are putting their dollars elsewhere.

It’s either short-termism, opportunity cost, or fundamental skepticism about capex payoff, but either way, hyperscaler equity is trading for a relative song. (Another possibility is that investors are concerned that AI cloud is just a less-profitable business than non-AI cloud and therefore a re-rating is the new-normal.)

That’s part of why this is perhaps one of the most consequential charts in equity pricing right now:

Hyperscaler free cash flow is expected to tank in 2027, and then come roaring back with a fury in ‘28-’29.

If the estimates are right (and that’s a big if), then the current price of the hyperscalers is very wrong. For Random Walk’s part, I’m still long hyperscaler (for all the heads-I-win-tails-you-lose reasons laid out before), but you do you.

2. Open Weight Open For Business (cont): When Cheapest Isn’t Always Cheapest

When the Kimi K3 Open Weight model kerfuffle exploded a few weeks back, Random Walk was the first (and perhaps only) person to state the obvious: if Chinese labs are going to freeload on lab capex to release distilled “same but cheaper” knockoffs, then it’s inevitable that frontier model progress would slow.

Lots of people were gleefully celebrating what they perceived to be the demise of lab margins (and/or were angrily protesting any lab efforts to protect those margins), without considering the fact that labs without margins could not keep pouring billions of dollars into training a new frontier, if the pirated copy would be released a few weeks later.

While I can’t say for sure that that’s what’s happening, there was some indication at the time that it was, and now there’s this:

OAI announces that its slowing frontier training . . . because of security and monitoring. I mean, sure, that’s possible . . . or it could just be a euphemism for “we’re not going to release a new frontier model until we figure out a better way to prevent distillation.”

Anyways, that’s not really the point, here. The point is really upstream of frontier progress, and it’s the question of whether open models are actually “better.”

The thing is that models are developing so quickly on so many different dimensions that pure apples-to-apples comparisons are difficult to impossible. Among other things, there are different modes to consider (e.g. text, code, video, voice, etc.), different capabilities (e.g. context window size, caching, latency, etc.), and different success rates, depending on both the task and the degree of difficulty.

Random Walk had promised a chartapalooza to make the point, but things have evolved so much in such a brief period of time, that it became kind of silly, so I’ll offer just three anecdata instead.

The first comes from Pitchbook, which ran an experiment to illustrate precisely the point that the cheapest token does not necessarily mean the cheapest model. The reason is pretty straightforward: a cheaper model may cost less per token, but if it burns more tokens on the same task (or really, string of tasks to accomplish a goal), then it’s functionally more expensive:

According to pitchbook’s experiment, Opus 4.8 and Gemini 3.1 Pro were the cheapest options, even though Opus 4.8 cost $25/token and Gemini just $12/token.

The difference comes down to what Pitchbook calls the “reliability premium.”

That is, every time a model reads and re-reads (and writes and re-writes) context to generate an output, it burns tokens (and consumes memory). An agent is basically doing that over and over again—it starts with the token-consumptive “prefill,” (i.e. the core context behind the assigned task), writes new tokens (which get added to the context), evaluates the growing corpus of context, and writes again (and again) until the task is done.

The more mistakes it makes and/or the less efficient the caching is, the more token consumptive (and costly) the overall task becomes. The net result is that two models with very different $/token prices can have roughly equivalent $/tasks.

Doordash ran a similar experiment, in this case routing different levels of code review to different models:

For Doordash, the “best” model varied by the complexity of the problem, where different models had different success rates, leading to different optimal model routing.

3Fourteen Research ran an experiment of their own, once again demonstrating that price/task is not the same as price/token:

Kimi K3 turned out to be ~30% more expensive that Opus 5 on a task basis.

You get the idea: just because open models (or any model) has a cheaper sticker price, does not mean it’s in fact cheaper. There’s a substantial amount of nuance involved, and it depends on the nature of

the task, the hardware, the software running in the background, and I strongly suspect, the skills of the agent-makers.

Don’t get me wrong: the cost advantages of open models are real, but it doesn’t obviously follow that they are as-good or cost-competitive in all cases. Not by a long shot.

In the big scheme of things, though, one thing is clear: computational demand continues to increase, likely driven in large part by “agentic” workflows:

Spot prices for GPU rentals continue to climb across the board, even for the oldest GPUs.

The longer-term unit economics for GPUs are still very much a moving target, but the possibility that compute is “overbuilt” is becoming more and more remote.

3. Private Credit has a Software Healthcare Problem?

Private Credit continues to not-apocalypse, and everyone seems to have mostly lost interest for now.

Things never really got that bad, and most of the badness more or less stabilized. It certainly could get worse, and it’s definitely the case that the great software repricing is making life uncomfortable for some loans, but for the most part, it’s fine.

The worst deals were from the ZIRP vintages, and those shoes have been dropping over the past year and a half:

The vast majority of recent foreclosures came from the peak-zirp vintages.

Again, mistakes were definitely made, but no apocalypse has ensued. To be sure, there is still some white-knuckling left to go, as ‘28 and ‘29 are banner years for software term-loans coming due:

~$100B+ isn’t a huge deal in the big scheme of things, but these were junk-rated to begin with, and likely were issued before software re-rated at the beginning of 2026.

So, yeah, there will be some painful conversations. But melt-down? Doesn’t seem like it’s in the cards.

One interesting aside, however, is that software is arguably not the biggest problem-child in private-credit land:

Random Walk has made this observation before, but healthcare and life-sciences are pacing default rates, with a substantial uptick since the end of ‘25.

Healthcare also has some of the widest dispersion in spreads, of any category of BDC loans:

Median spreads are pretty much the same everywhere (but Utilities), but the inter-quartile range for healthcare and social assistance is the largest (and much larger than software’s).

In other words, there’s far more turmoil and uncertainty running ripshit through healthcare than there is in software.

Why that is, I’m not entirely sure, but it could have something to do with the fundamental unsustainability of providing “healthcare services.” With downward pressure on reimbursements, I do wonder if some business models are simply upside down. It’s pure speculation, however.

4. China’s Beggar Thy Neighbor Trade (cont.)

Random Walk periodically riffs on China’s efforts to be the maker of all things to everyone everywhere, while buying nothing from nobody.

• It seems to reflect a strategic approach to identify what’s selling, and then make that, better and/or cheaper than whomever was making it before (at least partially by dint of subsidies).

• Random Walk would posit that the impetus for the approach is demographics: China (and the OECD) is getting older, and perhaps smaller soon, and so if neither domestic markets, nor global markets are getting bigger, then share-capture is the logical next move. In other words, if we’ve decided to weather our democratic winter by selling our debt to subsidize our consumption, then China has decided to sell their stuff, by subsidizing their production.

• the observation matters for understanding the state of play more broadly, but also with respect to trade, and it reflects something of a challenge to trade econ 101, insofar as the theory says e.g. Germany is supposed to enjoy the consumer surplus, while its manufacturing base is brought to ruin, which may indeed prove net beneficial to Germany in the longer-run, but it’s obvious to see why the claim is a harder sell.

Anyways, some charts to add to the theme.

First, yes, China is in fact selling more stuff everywhere, because it’s own domestic market has no juice left to squeeze:

Chinese exports have surged past its imports, all the while Chinese retail sales continue to fall.

And yes, once China decided to swallow Germany’s industry, that’s exactly what it did:

China went from importing from Germany to exporting to Germany in a heartbeat. Perhaps Chinese cars really are better, and perhaps this makes Germany better off. I suppose we shall see.

Bigger picture, none of this should come as any surprise, nor should the tepid growth of durable goods sales everywhere be a surprise:

China is getting older (and even more childless) very quickly.

It used to be a thing where ‘experts’ would periodically remark that ‘Chinese consumers just need to consume more because they’re under-consuming’ or whatever. Maybe so, but plainly ‘under-consuming’ is not the main problem here for aggregate retail demand.

What China does do is make a lot of energy, and a lot of robots:

China has (ironically) ramped up its robot patent game a lot. Robots are a key ingredient when it comes to doing more with less, and it seems like something China prioritized well before the West. We’re playing catch-up now, and while we’ll get there, a lot of the components and technology come from China, so it will take some time.

Another thing that China has done has moved its manufacturing to other countries, like Vietnam, but also, Mexico:

Mexico got very good at manufacturing dishwashers, and more recently, has gotten very good at manufacturing the hardware critical to the AI infrastructure buildout.

Here’s another look at how our neighbors have fared in relative import growth to the US:

Mexico has been doing great on the AI trade, while Canada, not so much.

In this case, however, China is probably not the prime mover. It’s Taiwan (not China) that’s been setting up the AI infra-shop in Mexico.

Why would Taiwan build in Mexico rather than the US of A? Presumably labor costs are a big part of the story, but I do not know that for sure. Claude seems to agree:

Fully loaded assembly labor in Mexico runs roughly $5.70–7.80/hr depending on the region, against about $31.60/hr in Texas. That gap would matter less if this were an automated process, but it isn't — an NVL72-class rack is thousands of hand-routed cables, connectors and manifolds, then burn-in and test. It's one of the most manual stages left in the whole chain, which is exactly the stage where a 5x wage differential dominates.

A 5x gap in labor costs is definitely a worthwhile consideration.

But, back to China for a moment, the great Michael Cembalest just published a big chart pack titled The Year of the Trojan Fire Horse: China’s imbalanced economy and unrelenting mercantilism.

It’s interesting throughout, but the gist of it is that China’s (subsidized) export machine is running roughshod over the rest of the world, including its “trading partners,” of which there are basically none.

“Overcapacity is essentially the hallmark of China’s impact on the global economy”

“~30% of Chinese industrial companies are unprofitable zombies that survive due to a broad-based subsidy approach that dwarfs the rest of the world.”

Anyways, you get the idea. China can produce way more than the world can consume, much to the chagrin of anyone else aspiring to make much of anything. It appears to be a matter of policy, as well, as latent capacity is kept afloat by state largesse.

One interesting aside, is that there is some evidence that the subsidies may be winding down:

Chinese state investment in fixed assets is flirting with negative yoy growth (and private investment has been negative for about 2 years).

If subsidies really are winding down, that certainly could change things quite a bit. It would force prices to rise, and perhaps give non-Chinese manufacturers more of a fighting chance. Now, don’t kid yourself too much, because China may well win with a level playing field, as well, but at least we’ll know.

5. No, “restrictive permitting” did not cause a renovation boom

A few weeks back, the great Arnold Kling posted something a little silly about home renovation spending.

Ah yes, the great zoning barrier is keeping people from spending on new houses, so they’re renovating instead.

Nonsense. New construction spending has regressed because demand has softened, and higher interest rates have depressed purchasing power.

Like you, I have no recollection of YIMBY victories in 2020, nor the NIMBY retrenchment in 2023, that apparently drove construction spending up and then down, respectively:

I do however recall that the non-existent cost of borrowing caused a building surge, and then a dramatic increase in the cost of borrowing, caused an equally dramatic building flatline.

There is no housing shortage and/or affordability crisis, despite what the discourse wants you to believe. Rates (i.e. the demand subsidy) are the story here—and population non-growth—which shouldn’t be some deep contrarian insight, and yet the commitment to the ‘housing shortage’ narrative is indefatigable:

My friend Aziz Sunderji made this neat chart showing the weight of attention that zoning receives as a driver of home values, relative to the actual prime-mover, demand.

The reality is that all public builders are offering historically high concessions to move inventory, operating at GFC-level margins, while sellers continue to outnumber buyers, and inventory builds. You can holler “shortage” all you want, but the data is very much not on your side.

Once again, take a gander at one of Random Walk’s favorite series on home values:

Opendoor shows ~1.5+ of steady home value depreciation. Since they’re in the business of actually selling homes (albeit somewhat regionally biased), it’s data worth paying attention to.

In general, new homes are built and sold at the prices that people are able to pay, such that there are net-new people to pay them. For existing homes, there is a bid-ask spread driven by sticky-prices and un-motivated sellers (who offload their “unrealized losses” to their mortgage lenders).

And yes, when homeowners don’t sell and/or trade-up, they may renovate a bit, but that too is effected by rates, and the data does not support a booming renovation industry:

Both “business activity” and the renovation backlog (as measured by Houzz) have been depressed since rates went up, and sales for renovation goods and materials tell the same story: higher rates are keeping renovation activity (often funded by HELs) on backburner.

Anyways, in terms of when things go “back to normal,” the answer is that with time, and rising incomes, the bid-ask spread will close (and it is, ever so slowly):

Home values will hold steady and/or decline moderately, while some combination of higher incomes and lower interest rates gradually closes the gap.

For now, existing homes are mostly just not selling. And, to repeat the point above, they won’t sell, until lower rates and/or higher incomes move prices up and to the right, which is indeed, the historical pattern:

Existing home sales tend to rise when values go up (or sellers otherwise become motivated, e.g. during a massive recession).

In other words, low-liquidity is hiding the fact that prices would be lower, if sellers were actually motivated to meet the bids where they are. Where there is liquidity, as with the new builds, prices are indeed lower.

It shows up elsewhere in the data, as well. If you look at the historical discount for forced foreclosure sales, the discounts in the Northeast are rather high—some of that is consistent with relative decrepitude of some foreclosures, but the other part of that are market prices held up by low-liquidity. A foreclosure sale, though, is a forced sale that doesn’t get to pick its spots.

The high relative discount to the Northeast is still true, but look at where discounts are way above trend:

Northeastern foreclosure discounts are high, but well within their normal ranges . . . Florida and CA, on the other hand, are historically lower, but are both currently running very hot (at nearly ~97-98th percentile of the normal range).

It makes sense that more liquid markets would have less of a forced-seller discount (because the presence of a forced-seller is less of anomaly). It also makes sense that currently, even those more liquid markets are facing steep discounts . . . because Sun Belt demand has peaked (as all the builders will be the first to tell you).

For the Northeast, it’s closer to business as usual. Homes don’t sell frequently, because people are aging in place, often without a mortgage, and they won’t sell, unless they get a price that meets their fancy. Foreclosure sales, however, take what the market gives, and that’s considerably less than the usual low-volume bid.

Random Walk is an idea company dedicated to the discovery of idea alpha. It’s not investment advice, conflicts of interest abound, it’s totally unreliable, and it reflects the thoughts and ideas of Random Walk, only.