AI Infrastructure Spending — Cycle Digest
TL;DR
The capex cycle hit its consensus peak-growth year — ~$790 billion across the five hyperscalers in 2026, up 86% — and the story shifted from how much to how it's funded: debt, off-balance-sheet commitments, and vendor guarantees are carrying the marginal dollar. Micron's record print extended the memory shortage through 2028, Oracle found a third way to buy nuclear power through a regulated utility, and OpenAI's IPO slipped to 2027 while Anthropic prepares a potentially larger one.
The funding layer is where the cycle now shows strain
Capex guidance is still rising everywhere — but the bond market has started charging different buyers very different prices for the same buildout. Bloomberg consensus pegs 2026 combined hyperscaler capex at ~$790.3 billion, up 86%, which analysts call "the peak growth rate of this cycle," decelerating to +35% in 2027 and single digits by 2028–29 hyperscaler-capex-surge-2026-component-inflation.
"capex now equals 98% of operating cash flow, and three of the five companies have flipped to negative free cash flow." — hyperscaler-capex-surge-2026-component-inflation
The differentiation is stark: Oracle's 10-year G-spread sits near 197bp versus Microsoft's ~40bp, and roughly $229 billion of debt has been issued year-to-date — with Amazon and Alphabet accounting for 78% of it while Microsoft issued zero. Risk is also migrating into structured finance: Nvidia's up-to-$105B residual value guarantee for an Ohio energy project, a Meta subsidiary guaranteeing CleanSpark's $2.28B high-yield bond, and ByteDance's $29.6B syndicated loan. Per the Founder Securities tracker, "debt is migrating from tech giants' balance sheets to project companies, private credit, and related parties — and chain transparency is declining."
That matters because the marginal dollar of AI infrastructure is now leveraged into a 5.24% 10-year Treasury — the first cycle where financing structure, not demand, is the live risk.
What to watch: Hyperscaler Q3 earnings in late October–early November; if capex actuals miss, "the growth deceleration could arrive earlier than anticipated."
Memory: the shortage just got a two-year extension
Last cycle we flagged Micron's fiscal Q4 (Sept 30) as the test of whether the shortage narrative extends or cracks. It extended — emphatically. Micron delivered $54.2 billion in revenue, up 379% YoY, with 87.0% gross margin and EPS of $33.42 against a $31.61 consensus, and guided fiscal Q1 2027 roughly 8% above the Street memory-shortage-dram-hbm-prices.
"we do not have line of sight to when supply and demand will return to balance." — memory-shortage-dram-hbm-prices
CEO Sanjay Mehrotra expects conditions to be "much tighter in fiscal 2027 and 2028 than they were in 2026." The contracting architecture is hardening into structure: 26 multi-year take-or-pay customer agreements covering over 35% of revenue through 2030, ~$150 billion in remaining performance obligations, and calendar 2027 HBM supply "vast majority" sold at significant price increases. Memory is also moving from commodity to co-designed silicon — Micron is building the industry's first custom HBM4E ("NVHBM") for Nvidia's next-generation GPUs.
The market's reaction was the tell: MU closed down 2.1% on the print with RSI at 75 — a beat-and-raise that sold off, consistent with a trade where the shortage was already priced.
What to watch: Whether hyperscaler Q3 commentary quantifies memory-driven component inflation in 2027 capex guides.
Nvidia: the buyer of last resort is now the seller of its own stock
Nvidia enters November's FQ3 print with the strongest demand signal of the cycle — and, notably, as its own largest buyer. The board authorized a $150 billion buyback increase on September 28, bringing the total to $235 billion, "the largest single stock buyback authorization in U.S. corporate history" nvidia-data-center-revenue-marvell-optical-surge. Morgan Stanley reinstated the stock as its semiconductor top pick on October 2, betting the "AI infrastructure cycle has room to run."
The fundamentals back the bulls: data center revenue of $89 billion, up 117% YoY and accelerating for a fourth straight quarter, with FQ3 guided to ~$108 billion — notably assuming zero China data-center compute revenue, a market Nvidia describes as "effectively foreclosed." Rubin began shipping in August and is expected to reach 20% of data center revenue in FQ3.
"Demand for Rubin is so massive that Nvidia is guiding for 70% year-over-year revenue growth in fiscal 2028." — nvidia-data-center-revenue-marvell-optical-surge
The bear case didn't go away — Michael Burry revived his critique of Nvidia's GPU depreciation schedules, invoking a 1960s bubble parallel. But the ~$108B ex-China guide is now the number the Street measures durability against, and memory costs "have yet to make a substantial dent" in 74.7% gross margins.
What to watch: The November FQ3 report — specifically whether Rubin's 20% revenue share lands and whether the depreciation/circularity critique gains traction if growth decelerates.
Power: the regulated-utility subscription arrives
The AI power playbook just added a third structure, and it's the most politically defensible one yet. On October 2, We Energies announced a deal letting Oracle "subscribe" to 10–20% of the output of the Point Beach Nuclear Plant for Project Lighthouse, the Port Washington campus Oracle is co-developing with OpenAI oracle-we-energies-point-beach-nuclear-subscription. Unlike Amazon-Talen (front-of-the-meter PPA) or Oracle-Bloom (behind-the-meter fuel cells), a regulated utility intermediates here — and the deal is being sold as a ratepayer benefit.
"could spare the utility's other customers roughly $300 million in fuel costs over the next five years" — oracle-we-energies-point-beach-nuclear-subscription
Because Oracle's subscription absorbs Point Beach's above-market nuclear fuel costs — the main reason We Energies spends about twice as much on fuel as its sibling utility — the hyperscaler becomes the solution to a ratepayer complaint rather than its cause. The same day, We Energies filed to build a new gas plant that "would more than double the site's electricity output" — the nuclear and gas sides of AI power arriving in one filing cycle. If the Wisconsin PSC approves, every regulated utility with merchant nuclear gets a template.
What to watch: The PSC approval decision — a green light would likely spawn copycat subscriptions in other merchant-nuclear states.
What surprised us
- OpenAI's IPO delay was resolved — and the market treated it as a chip event. Altman confirmed the IPO slips to 2027, officially citing "need for safety-related work"; the real math is a ~33% gross margin and $27B+ burn against a reported $1 trillion valuation floor, with SoftBank's ~$64.6B exposure hanging on it openai-ipo-timing-ai-financing-window
. Chip stocks slid on the news — the largest would-be raise in the AI chain is now a market-moving variable even in absence.
- Anthropic may test the public window first. While OpenAI waits, Anthropic is reportedly preparing for a potential $2 trillion IPO, with ARR up from ~$9B at end-2025 to ~$65B openai-ipo-timing-ai-financing-window
- A record beat-and-raise sold off. Micron's 87% gross margin print sent the stock down 2.1% — when a shortage is fully priced, good news stops working.
- Hyperscalers are now pitching themselves as ratepayer heroes. The Oracle-We Energies framing inverts the cost-shift political attack entirely — a shrewd template.