First — I owe you an update. Life got busy. Between the day job, the options book, and the general chaos of Q1 earnings season, I let the notebook go quiet for a month. That ends now. There's been too much happening in the AI capex world to stay silent, and honestly, the last five weeks have been the most interesting stretch since I started writing this. Let's get into it.
The Big Picture: What Happened Since March 4
When I last wrote — the Broadcom Q1 FY2026 note — the narrative was clean: hyperscaler capex was accelerating, custom silicon was gaining share, and the equipment cycle was heating up. A month later, the picture is messier but arguably more interesting. We got blowout earnings across the board, stocks that sold off anyway, a memory compression scare courtesy of Google, and a real shooting war in the Middle East threatening the physical supply chain. The supercycle isn't over — but it's being stress-tested.
NVIDIA: $68B Quarter, Stock Down 20% From Peak
NVIDIA posted Q4 FY2026 revenue of $68.1 billion— up 73% year-over-year — with full-year FY2026 revenue hitting $215.9 billion (up 65%). Guided Q1 FY2027 to $78 billion ±2%. Gross margins guided to ~75%. By any normal standard, these are staggering numbers.
The stock dropped ~5.5% on the report and has continued sliding. As of this writing, NVDA sits around $177, down roughly 20% from its all-time high.
So what gives? Three things:
- ›Gross margin compression fears. The inclusion of stock-based compensation in non-GAAP metrics starting Q1 FY2027 spooked some analysts. The ~75% non-GAAP margin guide is still excellent, but the market had priced in perfection.
- ›Geopolitical overhang. The Iran conflict is acting as a constant drag on risk appetite for anything touching global supply chains. More on this below.
- ›Valuation gravity. After a 140%+ run from mid-2024, the stock needed to digest gains. The market is asking: “at $215B in annual revenue, how much upside is left?” The answer is probably a lot — but it's harder to model when geopolitics and efficiency gains create noise.
My read: NVIDIA's fundamentals are not the problem. This is a multiple compression event on top of macro fear, not a demand inflection. The $78B Q1 guide alone tells you the spending pipeline is intact.
Micron: 196% Revenue Growth, Stock Dumps 20%+ — The “Earnings Paradox”
Micron's Q2 FY2026 might be the single most absurd beat-and-dump I've seen in semicap. Revenue: $23.86 billion (up 196% YoY). EPS: $12.20(vs. $9.30 consensus — a 31% beat). Gross margins: 74.9%. HBM4 and HBM3E contracts sold out through the rest of the calendar year.
The stock dropped ~6% post-earnings and fell 20%+ from highs over the following sessions. Why?
- ›Capex guidance spooked the market. Micron raised FY2026 capex to over $25B (from $20B), with a “meaningful step up” of $10B+ in 2027. Investors saw margin dilution risk.
- ›Google's TurboQuant — the efficiency scare. On March 24, Google Research published TurboQuant, a compression algorithm that reduces LLM KV-cache memory footprint by up to 6x with no accuracy loss. The market immediately asked: “does this kill HBM demand?” Micron lost $25B+ in market cap in 48 hours.
- ›Peak-margin anxiety. At 74.9% gross margins, the market decided there's nowhere to go but down. Classic semiconductor cyclical thinking — even though the structural demand story here is fundamentally different from prior memory cycles.
My take on TurboQuant
The market overreacted. TurboQuant targets inference KV-cache only — it doesn't touch training memory, which is where the HBM demand supercycle lives. Training workloads are capacity-bound, not efficiency-bound. Every major hyperscaler is still building multi-gigawatt clusters that need enormous HBM stacks. TurboQuant is real and useful, but it's an inference optimization, not a demand destroyer. If anything, making inference cheaper increases deployment, which eventually drives more training. Jevons paradox applies here.
Microsoft & Meta: Same Spending, Opposite Reactions
Microsoft posted Q1 FY2026 revenue of $77.7B (+18% YoY), with Azure growing 40%. But the headline was capex: ~$35B in the quarter (+74% YoY), with plans to increase AI capacity by 80% and nearly double the datacenter footprint over two years. The stock dipped — investors questioned whether the ROI timeline justifies the spend.
Meta, meanwhile, reported $59.9B revenue (beating $58.4B consensus) with EPS of $8.88 (vs. $8.24 expected) and the stock surged 10%+. Full-year 2026 capex consensus sits near $110B. The difference? Meta showed clear evidence that AI spending is flowing to the bottom line. Microsoft hasn't made that case as cleanly yet.
For the capex thesis, both outcomes are bullish. The money is being spent either way. But the market is starting to differentiate between hyperscalers that can monetizeAI infrastructure and those still in “build it and they will come” mode.
Q1 2026 Earnings Scorecard
| Company | Revenue | YoY Growth | Key Signal | Stock Reaction |
|---|---|---|---|---|
| NVIDIA (Q4 FY26) | $68.1B | +73% | Q1 FY27 guide $78B | -5.5%, -20% from peak |
| Micron (Q2 FY26) | $23.86B | +196% | HBM sold out, capex raised to $25B+ | -6%, then -20% over 6 sessions |
| Microsoft (Q1 FY26) | $77.7B | +18% | ~$35B capex, Azure +40% | Dipped |
| Meta (Q1 2026) | $59.9B | Beat | AI monetization visible | +10% |
| Broadcom (Q1 FY26) | $19.3B | +29% | AI rev $8.4B, 6 XPU customers | +2–5% |
| AMAT (Q1 FY26) | $7.01B | -2% | WFE guide $124–134B, >20% growth | Flat |
The Iran War: From Geopolitical Noise to Supply Chain Reality
This is the variable I wasn't tracking a month ago, and it's quickly becoming the most important one. The US-Iran conflict has escalated from a market overhang into a direct threat to semiconductor manufacturing inputs.
Here's what happened: Iranian missile strikes on Qatar's Ras Laffan Industrial City in late February knocked offline roughly 30% of global semiconductor-grade helium supply in a matter of days. Spot helium prices surged 40–100%. Force majeure declarations followed.
Why does helium matter? It's irreplaceable in semiconductor lithography and deposition processes. There is no substitute. The Semiconductor Industry Association has warned that a sustained helium disruption would send shockwaves through the entire global chip manufacturing industry.
The supply chain pressure points:
Helium: ~30% of global semi-grade supply disrupted. TSMC, Samsung, and Intel all depend on steady helium for EUV lithography and CVD processes.
Bromine: Two-thirds of global production comes from Israel and Jordan — both affected by regional instability. Critical for semiconductor etching chemistry.
Energy routes: Taiwan imports nearly all its energy. A prolonged Strait of Hormuz disruption threatens TSMC's ability to maintain output — and TSMC fabricates ~90% of the world's most advanced chips.
Direct threats to tech: Iran's Revolutionary Guard has explicitly called US tech companies “legitimate targets” — Apple, Google, Meta, Microsoft, NVIDIA were all named.
The near-term impact is manageable — Samsung and SK Hynix have helium reserves and HBM contracts locked through the year. But the binding constraint on AI scaling in 2026 is increasingly material and energy infrastructure, not the willingness to spend on GPU clusters. If the conflict escalates or persists, the ~$650 billion in planned global AI infrastructure investment could face real execution delays.
So — Is the Supercycle Over?
No. And I'll tell you why with one number: semiconductor industry 2026 sales are projected at $975B–$1T, with hyperscalers allocating 75% of $602B in capex to AI infrastructure. TSMC's record $56B capex budget for 2026 is unprecedented. Micron's HBM is sold out. NVIDIA is guiding $78B for a single quarter.
What we're seeing is not the end of the cycle — it's the first real stress test. The market ran ahead of fundamentals in late 2025, priced perfection into every name, and is now reconciling with reality: wars happen, algorithms get more efficient, and capex budgets have to be digested. None of that changes the structural demand picture.
My framework from Capex Signal #0 still holds:
- 1.Constraint: Power, helium, and advanced packaging capacity — not demand.
- 2.Intensity: Still accelerating. HBM stacks are getting taller, EUV layers are increasing, and custom silicon is adding complexity.
- 3.Durability: Multi-year. TSMC expects AI chip revenue to grow at a 60% CAGR through 2029. This is not a one-quarter phenomenon.
What to Watch Next
- ›TSMC Q1 2026 earnings (April 16): Revenue guided $34.6–$35.8B (+38% YoY). The most important datapoint for the entire AI supply chain. Watch helium supply commentary closely.
- ›AMAT Q2 earnings (May 14): Will the H2 acceleration materialize? Watch WFE guidance updates and China exposure commentary post the $253M BIS settlement.
- ›Iran conflict trajectory: Helium supply, Strait of Hormuz shipping, and whether threats against US tech companies escalate beyond rhetoric.
- ›TurboQuant adoption: Watch for ICLR 2026 presentation and whether hyperscalers actually deploy it at scale. If inference gets cheaper, model deployment accelerates — which could be net bullish for training infrastructure.
- ›Memory pricing: DRAM and HBM pricing power has been Micron's best friend. Any cracks in ASPs will be the first signal that the market is right to be cautious.
Bottom Line
The AI capex supercycle is alive. The numbers prove it — $68B quarters from NVIDIA, 196% revenue growth from Micron, $35B single-quarter capex from Microsoft, and $56B annual capex from TSMC. What's changed is the risk surface: geopolitical disruption to critical materials, efficiency breakthroughs that spook memory valuations, and a market that's pricing in the possibility (however small) that the spending might not compound forever. Separate the signal from the noise. The capex is real. The demand is structural. The volatility is the price of admission.
I'm back. More signals coming. — J