AI capex sustained but with China decoupling tail risk
What changed
The headline signal — how confidence moved from the previous snapshot, and why.
Held at 0.87 ± 0.05 (from 2.2.3). The two legs moved in opposite directions and roughly cancel: the demand side added its largest buyer-side confirmation yet (Oracle's $638B RPO, +363% YoY, mostly large AI contracts) plus demand breadth (
Alphabet's 3M+ TPU order at Intel), while the positioning/financing tail kept firing (
Oracle -10% after-hours on funding concerns,
NVDA to 200.42 and AVGO to 372.10 through the war tape). The week's chip weakness was war-and-rates flow, not an AI demand signal -
Monday's rebound showed the bid returns the moment the macro lifts. The 0.87 ± 0.05 set in 2.2.3 already encodes exactly this two-sidedness; no new China-tail development landed to move either bound. Beta(40, 6) ~46 effective observations held. The horizon is quarterly: the thesis is re-tested every earnings cycle, and the export-control invalidation could fire at any policy moment.
The thesis
The claim and where confidence stands now.
Hyperscaler AI capex continues at elevated pace into 2026. AMD Q1 2026 (May 5 AH) materially confirmed the demand side - revenue $10.3B vs $9.88B consensus, Data Center +57% YoY at $5.8B, Q2 guide $11.2B vs $10.3B consensus, Lisa Su quoted "tens of billions" by 2027. NVDA Q1 FY2027 (May 20 5 PM ET) is the next binary tell - consensus $78.8B revenue (+77% YoY) / $1.77 EPS; Q2 guide near $86B is what the after-hours reaction rides on; options price an 8-10% implied move; prediction-market beat probability is ~97% priced in - the asymmetry is unfavorable on the upside, and NVDA has closed lower on 4 of its last 5 earnings reports despite beating revenue 3-4% for six straight quarters. Huang estimates the unrealized China market at ~$50B "effectively gone with no clear return timeline". The May 14 H200 export approval to ~10 Chinese firms (Alibaba, Tencent, ByteDance, JD.com plus Lenovo and Foxconn) holds through the weekend into Monday and through Tuesday morning - the Trump-Xi summit concluded with no AI-policy reversal or extended-export-control reinstatement. Friday May 15 delivered broad chip profit-taking on the oil shock / rate-up backdrop: NVDA 225.32 (-4.42%), AMD 424.10 (-5.69%), INTC -8% to ~$108 (after a YTD +214% run), MU -6.6%. CSCO printed a divergent +13.4% to 118.21 confirming the Wed-AH Q3 print (FY26 AI orders raised to $9B from $5B, HSBC upgrade to Buy with $137 PT). Monday close NVDA $222.32 (-1.33%), AMD $420.99 (-0.73%), QQQ $705.88; Tuesday close NVDA 220.61 (-0.77%),
AMD 414.05 (-1.65%) on pre-earnings positioning, with Wednesday pre-market NQ futures +0.8% recovering ahead of the after-close print. The PHLX Semiconductor Index sits ~32% above its 50-day moving average after a +143% trailing year - a consolidation setup. The Rubin architecture H2 2026 ramp is the positive overhang investors are watching for. Update May 20: NVDA reported Q1 FY2027 after the close and the
initial after-hours reaction was muted on lower-than-expected guidance - the Q2 guide, the operative variable, did not clear the priced-for-perfection bar; the demand side nonetheless stayed firm intraday with
AMD 447.58 (+8.10%) on memory-supply tightness and a broad AI rally. Update May 21: the full figures correct the 2.1.6 "soft guide" read -
NVDA reported record revenue of $81.6B (+85% YoY) and adjusted EPS $1.87 (+140% YoY) with current-quarter guidance of $91B, above the ~$86B Street whisper, plus a
25x dividend increase and a new $80B buyback; the stock still
fell ~1% after-hours on a 45x P/E and a $5.3T cap. The named tell printed bullish for a capex-sustained thesis; the soft tape is a valuation/positioning reaction, not a demand signal. Update May 22: no new capex signal landed -
NVDA closed Thursday at 219.51 (-1.77%), the beat-and-raise shrugged off into the cash close, while
AMD held at 449.59 (+0.45%) on the memory-supply tailwind; the China export-control tail is unchanged. Update May 25: the AI tape split into the Friday close with no new demand signal -
NVDA faded to 215.33 (-1.90%) continuing the post-earnings drift, while
AMD ripped to 467.51 (+3.98%) on the memory-supply tailwind - positioning and valuation, not demand. The China export-control tail is unchanged over the quiet holiday weekend. Update May 26: no new AI demand or capex signal landed over the holiday or into the Tuesday pre-market; the beat-and-raise and the China export-control tail are unchanged. Update May 27: the first cash close brought no new demand or capex signal -
AMD ripped +7.78% to 503.89 on the memory-supply tailwind while
NVDA closed flat at 214.86 (-0.22%) - positioning, not demand - and the China export-control tail is unchanged. Update May 28: the Wednesday cash close brought no new AI demand or capex signal - chips faded with
NVDA 212.60 (-1.05%) and
AMD 495.54 (-1.66%) giving back part of Tuesday's rip - positioning rather than demand, with the China export-control tail unchanged. Update May 29: chips led the Thursday record day -
NVDA 214.25 (+0.78%) and
AMD 518.09 (+4.55%) - positioning into the broad melt-up rather than a new demand signal, the beat-and-raise and the China export-control tail unchanged. Update June 1: BOTH legs of the thesis strengthened. Demand:
Dell reported AI-optimized server revenue of $16.1B (+757% YoY) with a $51.3B AI-server backlog and guided to $60B in FY2027 AI-server sales - total revenue +88% YoY, the fastest in eight years and the strongest third-party hyperscaler-capex confirmation of the cycle, and
Jensen Huang's Computex 2026 / GTC Taipei keynote unveiled the Arm-based RTX Spark PC superchip, the Vera CPU, and a Microsoft Windows-AI-PC co-engineering partnership, lifting NVDA +4.13% to ~$219.87 Monday intraday. China tail:
the Commerce Department closed the loophole that had allowed NVDA / AMD advanced AI chips to reach Chinese firms operating outside China - license requirements now attach to China-headquartered entities regardless of location, exposing the >20% of NVDA FY2026 compute revenue that flowed through intermediaries, a concrete tightening that moves the export-controls-extended invalidation leg closer. Friday itself was profit-taking (
NVDA 211.14 (-1.45%),
AMD 516.10 (-0.38%)) ahead of the Monday Computex pop;
the SOXX semiconductor index gained ~24% in May. Broadcom's June 3 print is the next capex tell. Update June 2: the demand side delivered four independent confirmations in 24 hours.
The Monday cash session converted the Computex pop - NVDA +6.26% to 224.36 (Massive-verified), the largest single-day NVDA gain of the sequence;
HPE reported a record Q2 on AI-infrastructure enterprise demand, raised its FY outlook, pulled its long-term financial targets forward two years, and is up ~23% Tuesday pre-market;
Jensen Huang called Marvell "the next trillion-dollar company" Tuesday morning (the stock exploded ~20%) and said NVDA has supply to support "very, very robust growth"; and
SoftBank's Masayoshi Son called AI "50X bigger" than the dot-com boom, framing corrections as buying opportunities. No new China-tail development landed in the window - the Commerce loophole closure stands as-is.
AMD faded -1.16% to 510.13 on profit-taking after its May run. Broadcom reports tomorrow (June 3) after the close: guidance points to $22B revenue with AI semiconductor revenue expected near $10.7B (+140% YoY) - the next binary tell, with custom-silicon read-through for the hyperscaler capex picture. Update June 3: the demand side got its most direct confirmation yet - from the buyer side, not the supplier side.
Alphabet priced an upsized $84.75B equity raise ($30B underwritten offerings, a $40B at-the-market program, and a $10B Berkshire Hathaway private placement at ~$350/share) - its first stock offering in 20 years - explicitly to fund over $180B of 2026 AI-compute capex (double 2025), with more planned for 2027. Every prior confirmation was supplier revenue; this is the hyperscaler capex itself, pre-announced and equity-funded. The supplier tape extended:
Marvell closed +32.5% at 290.79 (record Q1 revenue $2.42B, >$10B-by-FY2029 custom-chip outlook),
AMD +2.24% to 521.54,
NVDA touched an all-time intraday high of 232.28 before fading to 222.82 (-0.69%), while
HPE pared its +23% open to a ~+9.5% close at 56.15 - a reflexivity caution inside the euphoria. No new China-tail development landed.
Broadcom closed at 481.57 ahead of its Wednesday after-close print - the next binary tell. Update June 4: the binary tell fired - and split. Fundamentally
Broadcom posted a Q2 blowout: revenue $22.2B (+48% YoY), AI semiconductor revenue $10.8B (+143% YoY, above its own guide), EPS $2.44 beat, Q3 guided to $29.4B with AI semis to grow over 200% YoY, FY2026 AI reaffirmed at ~$56B and FY2027 reiterated above $100B - the strongest demand-side confirmation yet. But
the stock rose ~2.6% after-hours then reversed to roughly -15% pre-market because CEO Hock Tan did not RAISE the FY2027 number, sparking a sector-wide chip rout (CrowdStrike -11%); the supplier tape was mixed Wednesday (
AMD +4.0%,
NVDA -3.6%,
Marvell +3.7% but fading from +11% intraday) and rolled over into
Thursday's pre-market rout. The demand leg strengthened while positioning grew fragile. Update June 5: the fragility became a $1.3T de-rating -
the Philadelphia Semiconductor Index fell ~10.3%, its worst day since March 2020,
NVDA -6.2% to 205.10,
AMD -6.6% to 466.38,
Broadcom to 385.73 - on intact fundamentals, the China-and-positioning tail firing while the demand leg held. Update June 11: the demand-confirmation machine kept printing into a still-hostage tape.
Oracle's Q4 FY2026 print showed remaining performance obligations of $638B (+363% YoY, +$85B sequentially) with OCI revenue +93%, most of the backlog from large AI contracts - the largest buyer-side capex datapoint of the cycle - and
the stock fell ~10% after-hours on the capital raising needed to fund it, the third straight fund-it-first reflex (Broadcom's guide-not-raised selloff, Alphabet's $84.75B raise, now Oracle).
Alphabet ordered over 3 million TPUs from Intel for 2028 (INTC +10%) - demand breadth plus supplier diversification beyond NVDA silicon. The tape itself stayed macro-hostage:
Monday's chip rebound (MU +10%, Nasdaq +0.86%) reversed into
a -3.3% Nasdaq Tuesday with tech -5% on the US strike preparations, and
Wednesday closed NVDA 200.42, AMD 452.40, AVGO 372.10. Demand keeps strengthening; positioning keeps de-rating on war-and-rates flows; no new China-tail development landed (the Commerce loophole closure stands as-is).
Drivers
The underlying macro forces this thesis expresses - the loading mean is how much each force drives the thesis, the stddev our confidence in the mapping.
Supporting evidence
Typed, citation-backed observations across time, grouped by strength. Hover a point for the claim.
What would invalidate this
The machine-evaluable conditions that would falsify the thesis.