The Canaries in the Data Mine

It seems every few days someone asks me whether AI is a bubble. I think that is the wrong question, or at least one with no useful answer. A bubble is only a bubble in hindsight. While you're in one, "too big" is a matter of opinion. The better question is simpler.

In the case of AI, that's not in the chip providers' guidance, which arrives a quarter or two after the orders changed. Not in the hyperscalers' capex plans, which no one wants to be first to cut. Rather, it's somewhere upstream, in a price that clears every day against buyers nobody is financing.

I think there are two such places. The price of memory and the terms of the marginal financing dollar. Both can be observed. And as semiconductors are not a new industry, we can see what they did in prior turns.

Miners carried canaries underground because the bird reacted to bad air before the miner did. Data centers have their own canaries. The suppliers of their two scarcest inputs, capital and memory. And they will know long before the rest of us if the air has changed.

Why these canaries are visible

Let's start with memory, because it is the only transparent price in the AI supply chain.

The AI chips are sold on allocation, with opaque pricing, multi-quarter backlogs, and essentially one provider (NVIDIA) deciding who gets what. We learn about a turn in chip demand when the chip provider decides to tell us. But memory is different. It's a commodity. A given DRAM die is the same part from any maker, it trades daily in a spot market and monthly or quarterly under contract, and the prices are published. In a system where almost everything is presented with black-leather-jacket-clad hype, memory is the unbiased thermometer readout.

Beyond the transparency, three things make memory lead the cycle rather than just mirror it. First, supply can't adjust. Fabs take years to build and can't be idled, so a change in demand shows up almost entirely in price. Second, it is in every part of the build. High-bandwidth memory on the AI chip, DDR5 modules beside the CPU, and NAND for storage. So memory demand tells you how many servers are actually being built, not just how many AI chips were sold. And third, commodity DRAM is the pressure release valve of the system. High-bandwidth memory uses roughly three times the wafer area per bit of conventional DRAM. So when AI orders are strong, makers shift capacity to it and starve the commodity market, whose price spikes. That's what we've seen in 2025 and 2026, and it is what caused memory makers' revenue to triple in a year. When AI orders soften, wafers flow back and the commodity price falls first, while high-bandwidth contracts, negotiated about a year ahead, still look fine until contracts roll. The headline HBM price is a lagging number while the commodity spot price is the leading one.

Late in a cycle, buyers build inventory against a shortage they expect. The first sign of a turn is those buyers quietly stopping their spot purchases and working down what they hold. From there the sequence has run the same way each time. That is, spot rolls over, contract prices reset one to two quarters later, the memory makers' guidance follows, then the chip providers and equipment names, then the broad semiconductor index.

2017-18spot peak Oct 20, 2017 · SMH peak Mar 12, 2018

2021-22spot peak Apr 29, 2021 · SMH peak Dec 27, 2021

2025-26 (live)SMH high Jun 22, 2026 · spot still rising

In 2017 to 2018, spot DRAM peaked in October 2017 and the semiconductor ETF (SMH) peaked in March 2018, about five months later. In 2021 to 2022, spot peaked in April 2021 and the ETF in late December, about eight months later.

One detail is worth mentioning. This works at tops, not bottoms. In the last cycle the index bottomed in October 2022 while memory prices kept falling for another year. Markets tend to anticipate recoveries and deny peaks. So everything through the memory lens is about the top of the cycle and tells you nothing about when to buy.

The 2023 transition

A modern break in this relationship is 2023, although the break actually strengthens the premise. Memory prices collapsed through 2022 and 2023, and the semiconductor ETF rose more than 70% anyway. If memory leads the chip cycle, how did the chip stocks have one of their best years while memory had one of its worst?

I believe the answer is that memory was shifting, or about to shift, to a different market. In 2022 and 2023 most memory went into PCs and phones, and both were working off the glut left by the pandemic buying spree. Memory prices fell because that glut was clearing. Meanwhile the AI wave that started with ChatGPT in late 2022 ran almost entirely through one company's GPUs. The number of AI servers being built was tiny next to the PC and phone market, and the high-bandwidth memory those servers use was a small fraction of what the memory makers produced. So the price of memory told you about legacy uses and had nothing to do with why chip stocks were rising. The room our thermometer was in had yet to be occupied by AI, so to speak.

That has clearly changed, and the change makes the canary worth more now than it was then. Memory is now inside the AI build, not adjacent to it. High-bandwidth memory and the DDR5 that sits beside the server CPU are made on the same advanced production lines, so an AI server order and a commodity memory price are now connected at the fab. The hyperscalers have become the buyers who set the price.

Industry estimates put memory near 30% of 2026 semiconductor revenue, and DRAM industry revenue went from about $27 billion a quarter in early 2025 to about $97 billion a year later. In the two prior downturns DRAM revenue fell 35 to 40% while logic fell 10 to 20%. Memory is once again the swing factor in the industry's profit, and this time the swing is being driven by AI demand.

It's also the one demand read that can't be round-tripped. This cycle's financing is partly circular. Chip providers taking equity stakes in customers, take-or-pay compute deals, debt-funded GPU purchases. A chip provider can support its own revenue by financing its own customers. Nobody can do that to a spot memory price. And guiding capex down is a competitive signal no hyperscaler wants to send first. What they can do quietly is stop taking spot memory, lean on contracted volume, and let inventory build. And that is exactly what showed up in the second quarter of 2026, with contract price growth slowing sharply quarter on quarter.

The memory canary: are the customers still buying?

Here is what to watch on the memory side, in the order it has historically moved.

1. Spot pricing stops rising and slips below contract. Double-ordering buyers stop taking incremental supply. Nothing in any earnings release changes. Spot is the only series free to move, because long-term agreements have frozen contract prices for the largest buyers. The test: spot below the prevailing contract price for four weeks, or 10% below its 60-day high. (This is driven by historical data)
2. The industry's next-quarter contract forecast is revised down instead of up, or an actual prints below its range. Through this up-cycle every revision has gone higher. The first one the other way is the market admitting the marginal bid is gone.
3. Contract growth goes to zero, then negative, a quarter or two later. This is the number the headlines carry, and one that spot has already told you. The earlier read is the second derivative: growth still positive but decelerating for two consecutive quarters, inside a genuine up-cycle. That is what fired earlier this year.
4. Memory-maker inventory days rise and guidance comes in light. Micron reports off-cycle, with fiscal quarters ending in late February, May, August and November, so its print lands three to four weeks before the late-October hyperscaler capex guides. They are the earliest-reporting memory maker.
5. Memory makers cut capex, which reaches the equipment names.
6. Backlog and customer-concentration questions reach the chip providers, and the industry re-rates as a whole.

As the priced proxy, the SMH semiconductor ETF has historically broken between steps two and four. The whole point of watching is to see steps one and two before the market prices step three.

The record

How well does this work? The contract deceleration test, evaluated in real time on reconstructed contract history since 2016, has been met five times. Two were clean hits. August 2018, after which the index fell 22% within about four months, and September 2021, after which the index peaked three months later and then fell 27%. One was early, in November 2017, eleven months before the bust. One was partial in November 2024, followed by a 26% drawdown that was mostly macro-driven. The fifth is live, met on July 3, 2026.

Extended back to 1995 on the Bank of Korea's DRAM export price index as a proxy, the same test with an up-cycle co-occurrence has been met six times with four hits: 2004, 2018, 2022 and 2024.

Now, a single signal shouldn't be trusted blindly. We'd like to avoid being the economist who predicted nine of the last five recessions. Reading it as successive additive signals, like rungs on a ladder, is more useful. So, read as nested events on that index, here are the outcomes since 1995.

The industry hit rate climbs with each rung, while the equity hit rate peaks in the middle, because by the time the price index is 15% off its high equities have usually already fallen. Certainty and lead time are bought with the same currency here. In both completed modern cycles, the deceleration plus one confirming read from spot was visible on public data about two months before the 2018 break and about three months before the 2021 to 2022 peak. That's the useful window. But it isn't a large one.

This cycle's complication

U.S. hyperscalers signed multi-year agreements in 2026 that restrict price increases for those clients. When demand turns, the adjustment will show first in volume and inventory, not in the contract price, because the contracts don't let the price move. That is why I weight spot, inventory days, and maker capex guidance ahead of the headline contract number in the current environment.

And one new ratio is worth considering: the premium of the server module over the PC module has roughly doubled since late 2025. A turn in that ratio would be the first sign the AI channel's pricing power was fading. It has no track record yet, though.

The second canary: is anyone still lending?

The second canary is the financing part, and "broken" doesn't need to mean a default. The break is when the price of the marginal dollar crosses the project's return and lenders start rationing by quantity. At a high enough all-in yield the capex cycle slows with no change in end demand required. So the sequence to track runs like this: first the price and terms of the marginal dollar, then the quantity that clears, then the collateral rules, then the counterparties. Capex guides are the lagging confirmation.

Think of the financing as a stack of lenders, each getting hurt in turn as the marginal dollar gets expensive. Here is the list, with the most fragile first.

At the bottom are the GPU-cloud operators and the people lending against their chips. These are companies like CoreWeave that borrow to buy GPUs and rent them out. Their first big facilities ran near 11% floating, secured by the GPUs themselves and the customer contracts attached to them. Since then the market has found progressively cheaper ways to fund them: high-yield notes near 9%, a low-coupon convertible, and the first GPU-backed facility to earn an investment-grade rating. Then in August 2026 the tone changed. A $2.6 billion term loan only cleared once lenders got maintenance covenants, a debt-service test and full amortization, and it priced 100 to 125 basis points wider than the earlier deals. Covenants like that had been close to extinct in leveraged lending since 2010.

One step up is the lease and developer layer. The hyperscalers increasingly don't own their data centers. They sign long leases, and a developer builds and finances the building. By mid-2026 those off-balance-sheet lease commitments across the six largest names were estimated near $1.2 trillion, roughly $660 billion of it signed but not yet commenced. The balance-sheet risk has quietly moved from the hyperscalers to the developers and whoever is lending to them.

Then there is Oracle, the one hyperscaler doing this with borrowed money and the loudest financing canary. It is the largest non-financial borrower in the U.S. high-grade index, its long bonds yield near 8%, and it needs roughly $40 billion of new funding this fiscal year. Whether that program clears at full size and full tenor is the live test as of September 2026.

Above that sits the true investment-grade paper of the other hyperscalers, which most people assume is safe, and mostly is. Direct debt across the six is near $460 billion, and the rating agencies have started warning that capex, $785 billion in 2026 and near $1 trillion projected for 2027, is eroding free cash flow at all of them. But that is free cash flow, so this layer doesn't break so much as slowly get more expensive.

And at the top, equity and convertibles are the cheapest money of all as long as the stock price cooperates. Global convertible issuance hit a 24-year high in 2025. It is the last layer to close, but when the equity stops cooperating it closes quickly, and the funding demand rolls down onto the layers below.

What the lender-side data actually leads

Most of what happens in this layer has no composite data series behind it. These are events you read about, not numbers you can chart. So unfortunately there is tedious work involved. Depending on your exposure, this might be worth it. But there is one exception with real history, if we want to build a proxy.

Business development companies are listed funds that make private loans, the same kind of loans that finance data centers and GPU purchases. Because they trade on an exchange, the market prices them every day, and that price can be compared with the stated value of the loans they hold. When the shares trade below that value, the market is saying it trusts the loans less than the lender does. That discount is the closest thing there is to a daily quote on private-credit stress. Placed alongside the ordinary high-yield bond market, the rough signal becomes private-credit vehicles trading at a deep discount while the high-yield market is still calm. That pairing says the stress is specific to private lending rather than a general credit selloff, which is exactly where the AI financing lives.

What can be seen is that this series leads the levered builders, not the semiconductor index. Twice in the current cycle the discount reached 10% with high yield calm, on October 9, 2025 and February 2, 2026. Over the months that followed, the three most levered AI builders, Oracle, CoreWeave and Nebius, fell by an average of roughly 50% the first time and 25% the second.

There is a caveat regarding how much this is worth. Since CoreWeave listed in March 2025, those three stocks have been so volatile that a 20% average drawdown followed almost nine out of ten trading days within nine months. Not candidates for value investors, to say the least. A signal with a perfect record against that backdrop is not telling you much; it fired inside a fall that was coming anyway. Where it did separate from the background was depth. A 40% drawdown followed fewer than a third of all days, but it followed every deep-discount reading whose window has closed. And, against the backdrop of the sheer financing size of this boom, which as a percent of GDP is eclipsed only by the housing mania of the 2000s and the railroad boom of the 19th century, it qualitatively serves as a meaningful lever.

Truthfully, financing breaks have not typically coincided with memory supply breaks. But in the current environment that precedent may be broken given the size and breadth. And the hubris is almost palpable. Just consider Google co-founder Larry Page's internal quote from late 2024:

But caveat emptor: rate cuts can reverse this layer entirely, since a cutting cycle that lowers all-in yields resets the first three stages of this process. And memory potentially becomes the one-man band. But the Fed MUST cut, right?

Well, go ahead and see FedPulse.io for my thoughts on that.

How the two canaries fit together

The financing signal leads the levered names by weeks and moves on events. The memory signal leads the fund by roughly a quarter and drifts. The full chain goes: financing tightens, orders get cut, spot memory falls, contract follows, guidance follows. The two signals are its two observable beginnings.

A financing break with memory still calm would say the demand data should follow within a quarter or so (historically speaking). Memory turning with financing calm would say the cycle is ending on inventory, as prior ones did, rather than on credit. The dangerous pairing is both moving together. A funding cost that cuts orders and a spot price that confirms the orders were cut would be an unprecedented combination. But in the current setup I am inclined to say it is the most likely one, assuming a break at all.

If not when, then how far?

Again, we don't want to be the economist who predicts nine of the last five recessions. And worse yet would be the perma-bear who is always calling for the end of modern civilization. In both these cases there is never any discussion of a reasonable downturn size. It is usually just an unhelpful "bad" proclamation. And if we are not concerning ourselves with timing per se, then the valuation multiple is the sizing lever.

The price to sales multiple on a top-15 basket of the semiconductor ETF stood at about 15.7 times sales in September 2026, the 96th percentile since 2011, off a peak near 18.6 times in May.

What prior signals did to the multiple is worth splitting into the part that came from the multiple compressing and the part that came from sales falling.

A large multiple leg with a small sales leg, as in 2018, is a valuation reset. Both legs together, as in 2022, is a bust. Today's starting multiple is two to four times the level at the prior signals, and the prior troughs sat at roughly a quarter of today's multiple. Those figures tell us how far. Or, at least, a rational how far.

Now, the clock is more consistent than the depth in the historical data. Nevertheless, if we are to stop subjectively pontificating and try to look for the best historical anchor points for the cycle to oscillate to, 70% is a startling figure. But, of course, demand could show up and usher in a new regime.

And truthfully, the usefulness of AI is proven. So a return to a pre-AI median multiple may not be the correct view. But, even at a rosier 8x, that's still a 50% drop from September 2026 multiple levels.

To be seen. But, I don't really concern myself with that because memory will more than likely strip the need for tea leaves in the exercise and let me know without having to be an expert.

Reading the above in order, the contract-price deceleration test (step three of the memory canary) has been met since July 3, 2026, with contract growth going from roughly doubling in the first quarter to roughly half that in the second quarter to a mid-teens forecast for the third. Nothing else on the memory side has moved. Spot is at an all-time high and well above contract, and maker inventories are described as extremely low. Cloud buyers did build inventory in the second quarter, and 2026 high-bandwidth memory contract prices are already declining, with 2027 negotiations landing in October and November.

On the financing side, terms have tightened and an anchor holder has sold, with more than $14 billion of data-center debt repricing wider in the second quarter. Oracle's curve is not inverted. The private-credit discount is well inside the 10% level that preceded the levered-name drawdowns, and narrowing from a 17% low in March. And as of writing, no benchmark deal has been pulled.

Important note: the data above is dynamic and updates regularly; the two paragraphs between it and this note do not, so they may become stale. The email update will contain any narrative updates.

In Summary

What would change the read toward a turn

• Spot 10% below its 60-day high, or below the contract price for four weeks.
• The quarterly contract forecast revised down, or an actual below its range.
• Memory-maker inventory days rising for two consecutive quarters, or "extremely low" disappearing from supplier commentary.
• A memory-maker capex cut, or NAND prices turning before DRAM.
• A benchmark AI financing pulled, downsized or shortened in tenor.
• Private-credit vehicles at a 10% discount to asset value while high-yield spreads stay calm.
• Convertibles across the complex trading to their bond floor.
• GPU advance rates cut, or a GPU-facility refinancing that clears only with the chip provider's support.
• Lease commencements deferred against the signed-but-not-commenced stock.

What would change the read away from a turn

• The contract forecast revised up again, or spot making new highs.
• The funding program clearing at full size and full tenor without concessions beyond the market's.
• The private-credit discount narrowing toward zero.
• A cutting cycle that lowers all-in yields for the levered layer.
• A genuine 2027 server-memory shortage. Early industry estimates have server module bit supply growing well below server CPU shipment growth, and real scarcity sustains prices even as capex growth slows.

Now, the canary doesn't tell the miner what to do. It tells him the air has changed, before he would have noticed on his own.

The suppliers of memory and capital are the two places a turn in this cycle has shown up before it reached the chip stocks. They have proven to be useful canaries in the past. And, given the structure of the current boom, they might be even more useful now.

What to do is another topic. But given the lofty levels, one can almost be assured of some attractive tail risk opportunities being served up.

---

About this analysis

Memory spot and contract prices are published market prices for commodity DRAM parts; the long-history price index is the Bank of Korea's DRAM export price index, used for growth rates only. Quarterly contract-price growth figures are the industry's published quarterly estimates, reconstructed on a publication-dated basis and quoted rounded. Inventory days and net asset values are drawn from SEC filings; bond yields, spreads and Treasury rates from FRED and market data; Korean export values from customs statistics. The semiconductor ETF is the VanEck Semiconductor ETF, with the Philadelphia Semiconductor Index used for history before the ETF existed. The basket multiple is price to trailing-twelve-month sales for the ETF's fifteen largest holdings, built from filings so that each month reflects only what was known at the time. Historical counts are as described in the text and are small; they describe a mechanism, not a rate. Companies referenced: Oracle, CoreWeave, Nebius, Micron, and SK hynix. Figures in the dated block are as of the dates shown and refresh from public data.

Important disclosures

This material is provided for informational and educational purposes only. It does not constitute investment advice, an offer, or a solicitation to buy or sell any security, and should not be relied upon as the basis for any investment decision. Any securities mentioned are for illustration only and are not recommendations. Estimates and analyses reflect judgments as of the date shown and are subject to change without notice; forward-looking statements are inherently uncertain. Past performance and historical patterns are not indicative of future results. New West Capital and its clients may hold positions in the securities referenced. Recipients should consult their own advisers before acting on any information herein.