Two Trades, One Selloff: What the AI Repricing Data Actually Supports
The move, and the shape of it
Since 1 June 2025, the AI complex has delivered one of the widest return dispersions in recent market history. Memory and storage compounded at rates normally reserved for biotech binaries. The hyperscalers actually deploying the technology went sideways to sharply lower. The gap between those two outcomes is the whole story.
| Company | 1 Jun 25 | 28 Jul 26 | 3M | 12M | Since 1 Jun 25 |
|---|---|---|---|---|---|
| Memory & storage | |||||
| SanDisk SNDK | $37.65 | $1,095.85 | +0.0% | +2,458% | +2,811% |
| Western Digital WDC | $51.27 | $464.01 | +6.9% | +492% | +805% |
| Micron MU | $94.21 | $820.68 | +58.7% | +653% | +771% |
| Seagate STX | $116.05 | $747.03 | +10.9% | +380% | +544% |
| Compute silicon | |||||
| Intel INTC | $19.55 | $86.30 | −8.7% | +336% | +341% |
| AMD AMD | $110.77 | $454.85 | +28.4% | +158% | +311% |
| Broadcom AVGO | $240.45 | $380.76 | −8.8% | +30% | +58% |
| Nvidia NVDA | $135.16 | $196.95 | −1.2% | +11% | +46% |
| Foundry & equipment | |||||
| Applied Materials AMAT | $155.39 | $476.45 | +20.9% | +167% | +207% |
| ASML ASML | $730.57 | $1,581.81 | +9.9% | +129% | +117% |
| TSMC TSM | $193.01 | $392.30 | −0.9% | +63% | +103% |
| Networking & hardware | |||||
| Dell DELL | $109.62 | $392.06 | +87.5% | +199% | +258% |
| Vertiv VRT | $107.91 | $269.50 | −17.9% | +85% | +150% |
| Arista Networks ANET | $86.64 | $169.65 | −1.9% | +38% | +96% |
| Hyperscalers & software | |||||
| Alphabet GOOGL | $171.23 | $333.73 | −13.3% | +74% | +95% |
| Palantir PLTR | $131.82 | $123.56 | −11.2% | −22% | −6% |
| Microsoft MSFT | $456.30 | $393.47 | −3.2% | −26% | −14% |
| Oracle ORCL | $164.09 | $119.99 | −25.7% | −52% | −27% |
| Power & electrical | |||||
| GE Vernova GEV | $471.85 | $943.38 | −12.9% | +43% | +100% |
| Constellation Energy CEG | $304.01 | $259.85 | −16.8% | −25% | −14% |
Total returns with dividends reinvested, anchored to calendar month-ends, through the 28 July 2026 close. The 1 June 2025 figure is the total-return-implied entry price, which for dividend payers sits roughly 1% below the actual 30 May 2025 close.
Where the bears have the data: memory pricing
This is the number that matters, and it is dated, published and quantified. TrendForce's conventional DRAM and NAND flash contract price forecasts, quarter on quarter, describe a price series whose second derivative has just turned hard negative.
Micron, SanDisk, Seagate and Western Digital are cyclical businesses earning cycle-peak margins on that series. TrendForce's own framing is that memory revenue is surging because of price rather than volume, with the bits going to PCs and smartphones actually shrinking. That is the textbook setup for consensus earnings estimates being wrong by a wide margin within two quarters, and it is a sufficient explanation for the drawdown in this group without invoking anything about artificial intelligence at all.
Two supporting distortions deserve attention. DDR4 spot pricing has reached $2.10 per gigabit, above HBM3e at $1.70, a legacy price inversion that tells you the market is not clearing normally. And SK Hynix's first-quarter ASP growth was held back by declining HBM contract prices even as commodity DRAM soared. That, not fading AI demand, is the cleanest explanation for the HBM4 delay that triggered the selloff: HBM lost its scarcity premium to conventional memory, so Hynix reallocated capacity toward where the margin had moved. The tape read it the other way.
Where they do not: end demand
Every indicator that would break first if the buildout were cracking is instead at a record.
NVIDIA DATA CENTRE
CONTRACTED BACKLOG
DC VACANCY, N. VIRGINIA
Nvidia reported April-quarter revenue of $81.6 billion, up 85%, with free cash flow of $48.6 billion and a July-quarter guide of $91 billion that assumes no China data centre compute whatsoever. Hyperscaler capital expenditure guidance has moved up through the year, not down: Alphabet raised its 2026 range to $195–205 billion from $180–190 billion mid-quarter, and the top five now aggregate to roughly $750 billion, a 67% annual increase and the third consecutive year of 60%-plus growth.
Where the fear is genuinely earned: the financing
The bear case has a real foundation, but it sits on the balance sheet rather than the order book. The buildout is now consuming free cash flow at companies that have never had to think about it, and the credit market has repriced accordingly.
Alphabet reported its first negative free cash flow quarter since its 2004 listing, at minus $5.9 billion. It raised roughly $70 billion in combined equity and debt, suspended its buyback, and doubled long-term debt to $98.2 billion. Second-quarter capital expenditure of $44.9 billion was up 100% year on year.
| Company | 2026E capex | Capex / revenue | Funding position |
|---|---|---|---|
| Oracle | $50B | 86% | BBB− since 13 July, one notch above junk; trailing FCF negative $23.7bn |
| Meta | $125–145B | 54% | $30bn bond October 2025 plus a ~$27bn off-balance-sheet vehicle; no external cloud revenue to defray |
| Microsoft | ~$190B | 47% | Most resilient FCF of the group; GPU contracts written for the full useful life of the asset |
| Alphabet | $195–205B | 46% | First negative FCF quarter since IPO; buyback suspended; long-term debt doubled |
| Amazon | ~$200B | 25% | Lowest intensity of the group; $24.9bn issued July 2026 after $15bn in November |
Source: CreditSights, company guidance and filings, July 2026. Capex-to-revenue ratios are 2026 estimates.
Barclays is explicit that Oracle's CDS is now traded as a liquid hedge on AI capital expenditure and data centre spending generally, rather than on Oracle's own fundamentals. That makes it the single most informative price in this complex, and it is at a record.
Two further items sit under the reported numbers. Nikkei has identified roughly $1.65 trillion of off-balance-sheet obligations across the major hyperscalers, largely lease commitments that prevailing accounting keeps off the balance sheet. And the depreciation wave has not arrived yet.
The revenue gap, and why the data is weak
Frontier model revenue is growing at rates without precedent, and the reported figures are close to unusable as evidence. Published estimates put OpenAI's annualised run rate anywhere between $18 billion and $25 billion, and Anthropic's between $6.5 billion and $30 billion, depending on the source and the month. The dispersion is not sloppiness so much as definitional: the gap between contracted ARR and recognised revenue can reach 70% at companies growing this fast.
Even at the generous end of those ranges, combined frontier-lab revenue is a single-digit percentage of 2026 hyperscaler capital expenditure. The comparison is unfair, since it excludes hyperscaler cloud AI revenue and internal deployment, but it is the anchor the bear case uses. The weakness of the underlying data is itself part of why the market discounts the growth.
GPU economics cut both ways
| Indicator | Reads as demand | Reads as deflation |
|---|---|---|
| Rental pricing | H100 one-year contract pricing rose roughly 40%, from $1.70/hr in October 2025 to $2.35/hr by March 2026, with on-demand capacity sold out across all GPU types | H100 on-demand rentals fell from $8–10/hr in 2024 to $1.80–3.50/hr by 2Q26 |
| Hardware residuals | New H100 street pricing has held, with HBM and packaging costs acting as a floor under manufacturing cost | Used H100 cards fell from about $40,000 in late 2023 to as little as $6,000 by mid-2026 |
| Lead times | 36 to 52 weeks against a backlog of roughly 3.6 million units; the bottleneck moved from fabrication to memory, it did not disappear | Longer lead times also delay revenue recognition and extend the window in which demand can change |
The spot decline is generational obsolescence rather than slack. But it does mean depreciation life assumptions carry more weight in these valuations than any terminal growth rate, and that is precisely the assumption nobody is auditing.
Asia: where this lands
The transmission into Asia is faster and more direct than in the private credit episode we examined in Issue 02, because in this cycle Asia is not the recipient of the shock. It is the origin of it.
The epicentre on both legs. The HBM4 reallocation originated here and the index carries the concentration: the KOSPI fell 5.7% on 24 July with SK Hynix down about 11% and Samsung close to 8%, triggering circuit breakers. Samsung's record quarterly operating profit was met with a 7% share price decline because revenue missed by a rounding error. When a market sells record earnings, the earnings are not the problem. Won sensitivity to dollar risk-off compounds the equity move for unhedged holders.
Direct, first orderEquipment and test names carry a lagged exposure: their order books reflect memory capacity decisions taken two to four quarters ago, so the deceleration reaches reported revenue well after it reaches the share price. The Nikkei fell 2.7% on 24 July on that read. Separately, MUFG, SMBC and Mizuho hold the syndicate exposure to US AI and neocloud financing mapped in Issue 02, which now sits at the intersection of both the credit and the capex questions.
Lagged, two channelsFoundry utilisation is structurally the last thing to break, and TSMC's 12-month total return of 63% on a low-twenties multiple reflects the market's understanding of that. The near-term risk is not demand but mix and customer concentration. The structural risk is the 28 July catalyst: visible progress in advanced Chinese chipmaking compresses the long-run pricing power that the entire island's earnings base assumes.
Resilient near termReported progress in advanced domestic chipmaking was one of two proximate triggers for the 28 July leg down, alongside AI debt concerns. Nvidia's April quarter recorded zero data centre compute revenue from China against $4.6 billion a year earlier, so the export restriction has already been absorbed into the numbers. What has not been absorbed is a credible second source at the leading edge, which is a margin question for the entire Western complex rather than a volume question.
Structural, not cyclicalLocal direct exposure is second order but real on three fronts. Sovereign and institutional allocations to AI infrastructure and neocloud credit face mark-to-market. Data centre REIT cap rates are supported by the tightest vacancy on record, though rising financing costs work against that. And the SGX semiconductor complex, UMS, AEM, Frencken and Micro-Mechanics among them, is levered to the same equipment order book as Japan with less liquidity to exit through.
Second orderThe overlooked exposure runs the other way. Memory price inflation of the magnitude seen since 2025 is an input cost shock to regional electronics assembly and handset manufacture, compressing margins in businesses with no AI revenue offset. If contract prices now flatten, that pressure eases. The larger risk remains dollar funding: any broad risk-off that strengthens the dollar hits refinancing for the same borrowers identified in Issue 02.
Cost reliefUSD riskIndicators to monitor
The single most important number in this complex. Deceleration from +95% to +15% is survivable for memory earnings. A negative quarter-on-quarter print breaks consensus estimates outright, and the group is priced on those estimates.
Revenue and guidance will dominate the headlines. The informative lines are inventory, which rose to $25.8 billion, and supply-related commitments of $119.0 billion. Rising inventory against decelerating sequential growth is the classic pre-correction signature in semiconductors.
Every guide this year has moved higher. One downward revision would validate the entire bear thesis in a single sentence and would cascade fastest through Oracle, then the neoclouds, then power and cooling, then memory.
S&P moved Oracle to BBB− in the week of 13 July. One further notch forces mechanical selling by investment-grade-only mandates. Given how the market is using Oracle CDS as an AI capex proxy, that event would transmit well beyond the issuer.
Server schedules have been extended repeatedly across this industry. Any reversal, or any disclosure that shortens useful lives, moves reported earnings across four megacaps simultaneously and is the least-watched risk in the complex.
Positioning implications
Memory and storage carry a quantified, dated risk. These are cyclical businesses at cycle-peak margins on a price series that has demonstrably stopped accelerating, held by investors who have been trained by fourteen months of price action to buy every dip. SanDisk returning 2,811% on a commodity price spike is the definition of a position that unwinds faster than it built. This is the part of the complex where we would be most cautious on drawdown, irrespective of the long-term AI view.
The hyperscaler risk is a balance sheet question, and it is being priced in the right market. Credit spreads are the leading indicator here, not the equity. Alphabet at a record 67 basis points after one negative cash flow quarter is a proportionate response. Microsoft, with the most resilient cash generation and contracts written to the useful life of the asset, is the least exposed to the specific mechanism that is worrying people.
Nvidia's reported numbers currently contradict the tape outright. Ninety-two percent data centre growth with China switched off, 75% gross margins and $48.6 billion of quarterly free cash flow is not the profile of a company at the end of a cycle. The stock has compounded at roughly 34% annualised since June 2025 while Micron did eight times. Whatever the market is repricing, it is not Nvidia's earnings power.
For Asian allocators, the sequencing matters more than the direction. Korea takes the first-order hit and has already taken it. Japan's equipment complex takes a lagged hit that has not yet reached reported revenue. Taiwan's foundry exposure is the most defensible near term and the most exposed structurally. And the regional electronics assembly base gets quiet margin relief from exactly the price deceleration that is hurting the memory names, an offset that almost nobody is underwriting.
Goldman Sachs estimates AI capital expenditure moving from roughly 1.6% of US GDP in 2025 to above 3% in 2026. At that scale this has stopped being a sector question. The equity market is currently treating a memory price cycle, a corporate financing cycle and a monetary policy cycle as one trade. They will not resolve together, and the dispersion in how they resolve is where the opportunity sits.
Price and total return data derived from split- and dividend-adjusted month-end closes through 28 July 2026 · TrendForce memory pricing surveys (January, March, June and July 2026) · NVIDIA Corporation Form 8-K and CFO commentary, Q4 FY2026 and Q1 FY2027 · CreditSights hyperscaler capex estimates (July 2026) · LSEG and Bloomberg credit default swap data (July 2026) · S&P Global Ratings (13 July 2026) · CBRE Global Data Center Trends 2026 (Q1 data) · Nikkei off-balance-sheet obligation analysis (July 2026) · FactSet debt-to-equity data · SemiAnalysis GPU rental price index · Goldman Sachs AI capital expenditure estimates · Federal Reserve Summary of Economic Projections. Frontier model revenue figures are drawn from multiple secondary sources whose estimates diverge materially and are presented as ranges for that reason.
