← Briefing history

The capex story stays intact — but the bear case has changed shape rather than grown louder.

Read-only snapshot of Nvidia capex

Sep 28, 2026 · 6 findings · ran 5m 43s

Nvidia Capex Digest

TL;DR

The capex story stays intact — but the bear case has changed shape rather than grown louder. The new argument isn't about whether hyperscalers will spend (they're guiding above $1T for 2027) but what they'll spend it on: memory is eating the GPU share of the wallet, and the credit market is repricing who can afford to fund it (nvidia-ai-capex-cycle-sustainabilityeurasiabusinessnews.comfinance.yahoo.cominvesting.com). Meanwhile, China optionality reopened through an unexpected side door — a workstation GPU — and Anthropic's ~$2T IPO slipped to November, right after Nvidia's own earnings print.

The bear case moved from capex level to capex composition

The demand debate is over; the wallet-composition debate has begun. Mr. P (P Equity Research, on the Logan Jastremski podcast) argues the industry has left the "GPU shortage era" for a "system bottleneck era":

"Next year, hyperscaler capex is expected to land between $1.1 trillion and $1.2 trillion, and memory alone could account for 50% to 60% of that — roughly $500 billion to $700 billion." — BigGo Finance

The institutional spread on memory's share is wide but uniformly rising — JPMorgan at 49% for 2027 and 60% for 2028, UBS at 63-73% — and it shows up in Nvidia's own numbers: Bloomberg-surveyed analysts see gross margin "sliding under 72% by the fourth quarter," with rising memory costs the central factor (nvidia-ai-capex-cycle-sustainabilityeurasiabusinessnews.comfinance.yahoo.cominvesting.com). This matters because the bull counter — Morgan Stanley's 25-50% ROIC projection on $1.4T of hyperscaler spend — answers the level question, not the composition one (nvidia-ai-capex-cycle-sustainabilityeurasiabusinessnews.comfinance.yahoo.cominvesting.com). The Q4 margin print is now the single most informative datapoint in the cycle.

What to watch: Whether Nvidia's Q4 gross margin print holds above the 70% tripwire as memory costs bite.

The financing chain is now the marginal buyer — and it's repricing

The circular-financing web got harder numbers on both ends this cycle: the end customer's dependence on other people's balance sheets, and the price of the debt in the middle. OpenAI's FT-reported projections show $278B of negative free cash flow through 2030 against an $856B compute tab, with ~$700B sitting "mostly on other companies' balance sheets" — as Rory O'Driscoll put it:

"OpenAI only burns $278B because Oracle, Nvidia and others build the data centers and lease the capacity back, with revenue support and backstops behind them." — SaaStr/20VC

The middle of that chain just showed its stress: Oracle's 2056 bonds crossed an 8% yield for the first time, CDS hit a record, and one downgrade would push ~$120B of bonds out of investment-grade indexes (nvidia-infrastructure-leasing-circular-financingbloomberg.comcnbc.comreuters.comstraitstimes.com). Apollo's Torsten Slok's hyperscaler CDS basket climbed above 100bps — an eight-year high — with CDS widening for Meta and Nvidia itself, though the 150bps distress tripwire hasn't fired (nvidia-infrastructure-leasing-circular-financingbloomberg.comcnbc.comreuters.comstraitstimes.com). The demand isn't flinching; the question is who can still afford to fund it.

What to watch: Oracle's credit — a downgrade to junk would be the first hard break anywhere in the financing chain.

China reopened through a workstation GPU, not a datacenter part

For the first time since zero-China guidance was set, there's a concrete, dated reopening channel — and it's a clever one. The Information reports China's MIIT asked ByteDance and Alibaba to report purchase plans for Nvidia's RTX Pro 5500, with the government signaling intent to approve:

"some industry executives expect the RTX PRO 5500 chip to avoid U.S. export restrictions" — Reuters

The RTX Pro 5500 is a Blackwell-generation workstation GPU — which is exactly why it may clear both capitals, sitting outside the US export-control perimeter for datacenter AI chips (nvidia-china-revenuereuters.comthestandard.com.hk). ByteDance plans a 1M-unit order, and the read on why Beijing is moving is telling: domestic substitution names sold off on the news, with coverage framing it as Huawei falling short (nvidia-china-revenuereuters.comthestandard.com.hk). This is optionality, not yet a leg — no revenue is in guidance, and the truce runs only to Jan 10, 2027.

What to watch: Q3 FY2027 earnings on Nov 18 — whether the RTX Pro 5500 channel shows up as restored China revenue or a guidance change.

The de-rate grinds toward the tripwire while the IPO calendar reshuffles

NVDA's multiple keeps compressing against record fundamentals. On the Bloomberg basis it's under 17x forward — cut in half versus 2025 — while lower-end estimates run ~14.3x, "the lowest in a decade," within a turn of the registered 14x tripwire (nvda-derating-decade-low-multiplebusinessinsider.comstraitstimes.com). The relative-performance tell is stark: Nvidia has gained 22% in 2026 while the SOX surged nearly 76%, with Intel and AMD more than tripling (nvda-derating-decade-low-multiplebusinessinsider.comstraitstimes.com). Huang's counter at Goldman's conference: Nvidia is "the world's first and only growth value stock."

The IPO calendar adds a sequencing wrinkle. Anthropic's ~$2T listing slipped from October to November — confirmed by multiple outlets — and the interpretation split immediately. Harry Stebbings: "A company that expects to be 20x or 30x oversubscribed in the biggest IPO of our lifetimes doesn't normally wait" (anthropic-ipo-public-market-testcomptroller.nyc.govfinancefeeds.comfinance.yahoo.comnikkei225jp.com). The benign read, per O'Driscoll: it's about banker logistics and a clean audited Q3, not market stress (anthropic-ipo-public-market-testcomptroller.nyc.govfinancefeeds.comfinance.yahoo.comnikkei225jp.com). The consequence either way: Anthropic now prices after Nvidia's Nov 18 print, so the market gets Nvidia's demand signal first.

What to watch: A forward P/E print below 14x — the line between skepticism and pricing an earnings break.

What surprised us

  • The workaround is a workstation part. Nobody had the China reopening running through an $6,000+ workstation GPU that sidesteps the export-control perimeter entirely — the most creative piece of regulatory engineering of the cycle (nvidia-china-revenuereuters.comthestandard.com.hk).
  • Memory is the new battleground. The bear case refined itself rather than repeating: UBS thinks memory spend could hit $900B — more than all hyperscaler capex combined this year (nvidia-ai-capex-cycle-sustainabilityeurasiabusinessnews.comfinance.yahoo.cominvesting.com).
  • Nscale funded and priced a date. The neocloud leg closed a $3.36B pre-IPO round led by Third Point and is targeting a mid-November IPO at ~$30-35B — another November test stacking onto the same window (nvidia-infrastructure-leasing-circular-financingbloomberg.comcnbc.comreuters.comstraitstimes.com).
  • xAI is the quiet demand monster. Colossus 2 is heading toward ~770K Blackwell-generation GPUs at a single customer, with two more tranches dated within six weeks (nvidia-ai-capex-cycle-sustainabilityeurasiabusinessnews.comfinance.yahoo.cominvesting.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?