← Briefing history

The verdict stays **intact** — and this cycle it got materially *stronger* on the demand side while the valuation story turned…

Read-only snapshot of Nvidia capex

Sep 30, 2026 · 4 findings · ran 4m 45s

Nvidia Capex Digest

TL;DR

The verdict stays intact — and this cycle it got materially stronger on the demand side while the valuation story turned paradoxical. Anthropic's leaked IPO prospectus put a hard number on a decade of compute commitment — $518B, 80% of it non-cancellable — the strongest demand-lock-in evidence the buildout has produced (anthropic-ipo-public-market-testcomptroller.nyc.govfinancefeeds.comfinance.yahoo.comnikkei225jp.com). Meanwhile Nvidia now trades at ~17x forward earnings, its cheapest multiple since January 2015, and answered with the largest buyback authorization in US history (nvda-derating-decade-low-multiplebusinessinsider.comstraitstimes.com).

The Anthropic leak is the capex chain's biggest demand datapoint — and its biggest concentration warning

The AI capex debate has run on inference about future commitments; Anthropic's draft S-1 just replaced inference with a signed schedule. Reuters and the FT obtained the June-drafted filing and published its contents on Sept 28: the company "plans to spend $518 billion on cloud, computing and infrastructure obligations in coming years," and per SiliconAngle's follow-up, "80% of the sum is tied to contracts that either can't be canceled or require the company to pay for the allocated infrastructure even if it goes unused" (anthropic-ipo-public-market-testcomptroller.nyc.govfinancefeeds.comfinance.yahoo.comnikkei225jp.com).

The composition matters as much as the total. About $161.2B of equipment leases route to Broadcom (TPU capacity via the Google partnership), and nearly half the cloud budget goes to AWS, Microsoft, and Google — meaning a large slice of the decade's locked-in compute never touches an Nvidia GPU (anthropic-ipo-public-market-testcomptroller.nyc.govfinancefeeds.comfinance.yahoo.comnikkei225jp.com). The same document carries the bear case in its own risk factors: "nearly a quarter of its revenue came from two customers last year," and the largest clients "were not locked into long-term contracts and could cut or stop spending." The headline loss — $42B in 2025 — is less alarming than it looks, since roughly $34B was a non-cash accounting charge tied to financing that could convert to shares; the operating loss was $8B, over 91% of it computing infrastructure costs (anthropic-ipo-public-market-testcomptroller.nyc.govfinancefeeds.comfinance.yahoo.comnikkei225jp.com).

What to watch: The public S-1 filing (expected early-to-mid October) and the post-midterms pricing — the hardest public-market test yet of AI capex economics.

The de-rate met its counter-force: Nvidia is now its own marginal buyer

The valuation story crystallized this cycle into a single paradox. NVDA trades "at about 17 times expected earnings for the next twelve months, its lowest forward multiple since January 2015 and well below a 15-year average of 30x" — yet the stock rose 1.68% on Sept 28 to $228.86, near its 52-week high (nvda-derating-decade-low-multiplebusinessinsider.comstraitstimes.com). The mechanism is earnings outrunning price: consensus FY2028 EPS was raised to $15.68 from $12.67 in 90 days. The company answered the de-rate directly:

"Our cash generation gives us the capacity to invest in the technologies that advance this transformation and return capital to shareholders. This authorization reflects our confidence in the long-term opportunity ahead." — Nvidia Newsroom

The $150B increase to $235B total, executed through FY2028, is the largest repurchase authorization in history — and at ~17x forward it functions as a management claim that the market's multiple is wrong (nvidia-ai-capex-cycle-sustainabilityeurasiabusinessnews.comfinance.yahoo.cominvesting.com). The standing 14x forward-P/E tripwire — the level that would price an outright earnings break — is not crossed. The skeptics' test is now dated: if fiscal Q4 revenue guidance lands below the $124.37B consensus, or the 71% margin floor slips, "skeptics have the stronger case" (nvda-derating-decade-low-multiplebusinessinsider.comstraitstimes.com).

What to watch: The Q3 print expected Nov 17–18 — the $124.37B guidance bar and the 71% margin floor are the two numbers that flip the verdict.

The stress moved down the stack: neocloud equities are repricing even as take-or-pay contracts sign

The cleanest micro-datapoint of the cycle splits the market in two. QumulusAI announced $240.9M in new three-year take-or-pay contracts covering 2,000+ Blackwell B300 GPUs — pushing combined customer commitments past $344M — and the stock nonetheless "dropped 15% in early trading Tuesday, reversing a 15% premarket gain" (nvidia-ai-capex-cycle-sustainabilityeurasiabusinessnews.comfinance.yahoo.cominvesting.com). CEO Michael Maniscalco's framing is the bull case in one line: inference platforms are "planning capacity years ahead and backing those plans with cash, allowing the company to place GPUs against signed contracts rather than forecasts." The market's answer: real contracts, still not buying the equity.

The private channel tells the opposite story. Nscale raised $3.36B pre-IPO — a $2.36B Third Point-led tranche plus $1B from Nvidia converting at IPO — against $103B of total contracted value and a reported ~$35B valuation target (nvidia-ai-capex-cycle-sustainabilityeurasiabusinessnews.comfinance.yahoo.cominvesting.com). Private credit is funding the build at full price while public equity investors discount the marginal buyer. That same split is what the hyperscaler CDS basket and the insurance-syndication talks from the last digest are tracking — nothing this cycle resolved either.

What to watch: Nscale's mid-November IPO pricing — the next market test of what leveraged GPU exposure is worth.

What surprised us

  • The demand ceiling keeps moving up. Nvidia itself cited top-five hyperscaler capex of nearly $800B in 2026 and $1.3 trillion in 2027 — Wells Fargo's Ohsung Kwon argues the bear math fails if the market expects only ~$1T in 2028. The skeptics may still be undershooting their own targets (nvidia-ai-capex-cycle-sustainabilityeurasiabusinessnews.comfinance.yahoo.cominvesting.com).
  • A "multiple that low usually follows a crash" — and it landed mid-boom. A $5.5T company at 17x forward, with the price near highs and estimates still rising, is not a valuation profile anyone had on the 2026 bingo card (nvda-derating-decade-low-multiplebusinessinsider.comstraitstimes.com).
  • Anthropic's risk factors read like an AI-safety paper. The prospectus warns advanced models could pose "catastrophic or existential risks to humanity" and exhibit "self-preserving behaviors" — existential-risk language now lives inside an SEC-bound offering document (anthropic-ipo-public-market-testcomptroller.nyc.govfinancefeeds.comfinance.yahoo.comnikkei225jp.com).
  • Nvidia may anchor-buy its own customer's IPO. Having invested $10B in Anthropic last November, Nvidia could reportedly join the ~$2T listing as an anchor investor — the financing web closing another loop (anthropic-ipo-public-market-testcomptroller.nyc.govfinancefeeds.comfinance.yahoo.comnikkei225jp.com).

Open threads worth a vote

Findings from this cycle

Current topic brief

Shown for context; the brief may have changed since this cycle ran.

Is Nvidia's AI capex story still intact?