AI Infrastructure Spending — Cycle Digest
TL;DR
The slowdown debate finally got its market test — and it produced a structural trade, not a capex cut: chipmakers sold off on Anthropic's safety call while the hyperscalers who buy the chips actually rose. Meanwhile the memory shortage tightened into the cycle's binding constraint, and the power side of the buildout went vertical, with Amazon handing Generac an equity warrant to lock in backup generators. The throughline: demand guidance keeps rising, but everything now runs through supply, power, and financing costs.
The slowdown trade: chips sold, buyers bought
The market's first live rehearsal of "what if the frontier race slows" produced a pair trade, not a capex cut. After Dario Amodei's safety-slowdown call the weekend of September 13–14, Monday's tape split cleanly: Nvidia fell more than 3% while Intel, AMD, and Marvell dropped 5–6%, yet Alphabet rose almost 2% and Microsoft added 1.6% — the buyers of AI infrastructure outperforming the sellers.
"They'll just all stop building data centers and just digest what they have." — ai-capex-slowdown-debate
D.A. Davidson's Gil Luria was skeptical the slowdown is real — "Until they say that, they're not actually slowing down" — and Seeking Alpha's post-mortem confirmed the market was trading sentiment, not guidance: "Monday's market reaction was sentiment-driven; no hyperscaler cut capex guidance, and demand for compute-intensive inference workloads remains robust." By Friday, NVDA had closed at $222.27, just 5.5% below its 52-week high ai-capex-slowdown-debate. The macro backdrop made the reversal more notable: the Fed hiked 25bp to 3.75%–4.00% on September 16 — its first in over three years — with the 10-year touching 5%, a direct financing headwind for a buildout funded by debt, ATM equity, and private credit.
What to watch: The FOMC's October 28 decision — another hike would raise financing costs across the leveraged AI buildout, from compute-financing platforms to data center REITs.
Nvidia's demand signal: double the chips
Guidance keeps extending, not rolling over. Nvidia's fiscal Q2 2027 delivered $96.22B in revenue (net income $59.69B, a 63.7% margin), its fourth straight beat, with data center revenue of $89B up 117% YoY and Q3 guided to roughly $108B nvidia-data-center-revenue-marvell-optical-surge.
"I expect Nvidia to sell twice as many chips as this next year as we do this year." — nvidia-data-center-revenue-marvell-optical-surge
Jensen Huang made that doubling call from a September 17 summit in Scotland, and the company separately flagged 70% growth for the fiscal year ending January 2028 — about $673 billion. On the safety debate, Huang drew the line at product, not pace: "When a product is not safe, we should hold it back and keep engineering it." The constraint on the doubling ambition is no longer demand — it's HBM supply and power for AI factories, which is where the rest of this cycle's action sits.
What to watch: Whether the Q3 guide of ~$108B data center revenue lands, and whether Huang's doubling language gets formalized in guidance.
Memory: the binding constraint, and Intel's Ohio shell becomes an asset
The memory shortage is now the tightest link in the hardware chain, and it's inflating hyperscaler capex directly. Per TrendForce, a DRAM chip that cost $1 in January costs roughly $3.50 today, and HBM spot prices have hit roughly four to five times long-term contract levels memory-shortage-dram-hbm-prices. Amazon's July $20B capex raise to ~$220B was "largely due to higher memory prices" per Andy Jassy. Micron's fiscal Q3 showed what that means for suppliers: revenue of $41.5B, up from $9.3B a year earlier, with gross margin at roughly 85% memory-shortage-dram-hbm-prices
.
The shortage is redrawing supply chains — and rescuing a stalled asset. Reuters reported September 16 that SK Hynix is in talks with Intel to manufacture memory in the U.S. for the first time, either leasing part of Intel's delayed Ohio fab or forming a venture with Intel and cloud firms keen to lock in supply.
"Intel shares surged 5.5% at the open on Wednesday, while SK Hynix's stock finished 4.1% higher in Seoul." — intel-foundry-ohio-memory-talks
Washington's tariff pressure (Commerce Secretary Lutnick has threatened up to 100% tariffs on South Korean and Taiwanese firms unless they produce on U.S. soil) is colliding with Seoul's technology-protection review regime — so the deal is far from done. But the strategic reversal is striking: Intel's Ohio complex, delayed to 2030 and 2031, may become a foundry-as-lessor monetization path for the AI buildout rather than a stranded bet.
What to watch: Micron's September 30 fiscal Q4 report (guided ~$50B revenue, ~86% gross margin) for whether management extends the shortage narrative into 2027–2028 or signals a top — plus any confirmation or collapse of the SK Hynix–Intel Ohio talks.
Power: warrants, flexibility, and the colocation layer
The power constraint is now valuable enough that Amazon is paying for supply security with equity. Generac disclosed a long-term backup-generator deal for Amazon's data centers — about $2.4 billion in initial 2027–2028 deliveries, up to $8 billion in aggregate — with a warrant for Amazon to acquire up to 1.69 million GNRC shares at $200.93, vesting as purchases land generac-amazon-backup-generators-data-centers. Generac shares surged more than 40% in extended trading. Rich Miller's framing lands the point: "Just 18 months after entering the data center market, Generac has landed a huge deal with Amazon Web Services."
Meanwhile, the industry's coordinated answer to grid scarcity and the ratepayer backlash arrived September 16: the AI Energy Management Alliance, launched by Google, Nvidia, and Emerald AI with 20 members including Anthropic, National Grid, and Constellation, wants data centers rewarded for cutting grid demand when supplies are tight. Google has committed 1 gigawatt of curtailable demand ai-energy-management-alliance-grid-flexible-data-centers. The political urgency is real: 84% of Americans are concerned about data centers' impact on local electricity prices, and 79% of Democrats and 76% of Republicans support making developers pay for grid upgrades.
The scarcity is also reaching the colocation layer. Equinix now plans $5B–$7B of annual data center development investment — roughly double its prior expectation — and raised its long-term AFFO-per-share growth outlook to 9%–12% through 2029, with enterprise deals moving "from 250 kVA deals once being considered large to megawatt-scale deployments becoming more common" data-center-infrastructure-capacity-pricing-power. AI labs including Anthropic and OpenAI are reportedly seeking smaller 20–30 MW deals in the UK and Nordics — the labs themselves are becoming colocation customers.
What to watch: Whether regulators actually create the "fast lane" for power-flexible data centers — the coalition's whole thesis depends on it.
What surprised us
- The buyers outperformed the sellers on a slowdown scare. The instinctive trade — sell everything AI — was wrong within hours; the market priced hyperscalers as beneficiaries of a pause, echoing Luria's Amazon-warehouse-pullback analogy ai-capex-slowdown-debate
.
- The customer-takes-equity playbook reached the generator aisle. Amazon has taken warrants from Plug Power and ATSG before; extending the mechanism to backup generators shows how far down the supply chain scarcity pricing now runs generac-amazon-backup-generators-data-centers
- Intel's delayed Ohio fab flipped from liability to asset. The same memory shortage squeezing everyone may hand Intel its first foundry tenant — with the U.S. government as part-owner cheering it on intel-foundry-ohio-memory-talks
.
- Micron's numbers look like a typo. Gross margin of ~85% versus 39% a year ago, on sales that more than quadrupled — and the classic top signals (cracking spot prices, excess inventory) still aren't present memory-shortage-dram-hbm-prices
.