Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
Yes — and the bear case just changed shape rather than gained ground. The Rubin production-cut scare dissolved into a stale April re-report while Rubin ships on schedule; Nebius raised GPU rental rates for the second time in three months; and the first real write-down of the cycle arrived at Nvidia itself — a $0.4B charge for excess H200 inventory that is immaterial to a $303B revenue base but tells you exactly where transition risk lives. The story is intact; the debate has migrated from "demand is fake" to "scarcity pricing is temporary."
The Rubin cut rumor died of old age — the ramp is real
The supply-side scare resolved in Nvidia's favor. The claimed 25% cut to 2026 Rubin production (2M → 1.5M units) on HBM4 validation delays traces to April 8–10 reporting — KeyBanc via MLQ, Chosun, TrendForce — recycled by a late-September Chinese piece, not new news (Finding: samsung-custom-nvhbm-and-nvidia-memory-sourcing). Five months of contrary evidence piled up in the meantime: Vera Rubin "began production shipments in August" and management expects it at "about 20% of Data Center revenue in Q3," while Supermicro has been shipping NVL72 racks since Sept 23 (TIKR, Sept 26).
"We have spent years building the liquid-cooling stack, the manufacturing capacity, and the deployment teams for exactly this moment." — Charles Liang, via HPCwire
UBS raised its 2027 Nvidia AI GPU production forecast to 8.8 million units — the opposite direction of a cut — and Rubin's first MLPerf results showed throughput up to 3.7x the GB300 NVL72 (ad-hoc-news, Sept 24). The binding constraint remains delivery, not demand: CFO Colette Kress expects "supply to remain a bottleneck at least through the end of fiscal year 2028."
What to watch: memory, not GPUs — HBM4 supply is the tightest input in the stack, and Samsung's NVHBM qualification remains the swing factor for 2027.
The demand falsifier still hasn't fired — Nebius is hiking, not cutting
Last cycle's cleanest demand test was simple: the first neocloud price cut. Instead, Nebius raised pay-as-you-go rates effective Oct 1 — H100 to $4.50/GPU-hr (~17%), B300 to $9.50 (~21%), memory-backed instances ~41% — its second hike in three months (shattered.io, Sept 23). The pattern break is the tell: "Instead of newer chip generations pushing older ones down in price as supply normalized, the market saw the opposite: broad-based increases across H100, H200, B200, and B300 simultaneously." 2022-era silicon is getting more expensive to rent, and NBIS jumped 11% pre-market because "investors viewed it as a sign of tight capacity, not weak demand" (Finding: big-tech-capex-reaches-710b-in-2026).
Supply forecasts moved the same way: India's Yotta ordered 80,000 Vera Rubin units in a $12B expansion, and Nvidia's own numbers keep confirming — Q2 FY27 revenue $96.22B (+106%), Q3 guide $108B ±2% (TIKR).
What to watch: whether any neocloud cuts a rental rate this quarter — the tripwire remains unfired, and every hike pushes its resolution date out.
The bear case got smarter: "scarcity pricing is temporary"
The overbuild debate didn't disappear; it repositioned to later-dated ground. Rothschild Redburn initiated CoreWeave and Nebius at Sell on Sept 21 — and notably, "Redburn's thesis does not require AI demand to collapse. It only requires the economics of renting GPUs to get less exceptional as supply broadens and big customers build their own infrastructure" (Yahoo Finance). Both stocks rose around the call (Finding: big-tech-capex-reaches-710b-in-2026).
That's a materially different bet than Burry's write-off thesis from last cycle: it's a wager on supply broadening, not demand collapsing — which means the falsifier is also later-dated and harder to observe. The market paying up for scarcity while analysts argue the scarcity premium is temporary is exactly the tension that resolves in data, not narrative.
What to watch: hyperscaler in-house capacity coming online in 2027 — that's the mechanism Redburn's thesis actually runs through.
The first hard write-down arrived — at the vendor, and it's a churn story
The disclosure the overbuild watch has been waiting for finally landed, and it sits at Nvidia itself: "Nvidia took a $0.4 billion charge in the first half for excess H200 inventory as demand for that chip faded" (TIKR, citing the 10-Q) (Finding: gpu-deployment-gap-overbuild-debate). Scale check: $0.4B against TTM revenue of $302.97B at a 74.7% gross margin — immaterial (NVDA market view, Sept 27).
The structure matters more than the size. Demand for the previous generation faded fast enough to strand the vendor's own finished inventory — while, in the same window, Nebius was hiking H100 (2022 silicon) rents ~17%. Old chips can strand on the vendor's balance sheet even as the installed base appreciates in the rental market. Transition risk concentrates at whoever holds unsold inventory of a dying generation; scarcity keeps old-generation rents firm. No hyperscaler or neocloud has yet disclosed a GPU write-down, deployment percentage, or depreciation-schedule reversal — that watch stays open.
What to watch: whether the H200-style charge recurs or grows at the next 10-Q — one is churn; a pattern is the overbuild signal.
What surprised us
- The Rubin cut was a zombie story. The claim that resolved last cycle's open thread turned out to be five months stale — April ramp-mix reporting (Blackwell filling calendar 2026) recycled as September news. A useful reminder that supply-chain "cuts" deserve date-checking before they deserve worry.
- The first write-down of the AI boom is Nvidia's, and it's tiny. After months of Zitron's $50–100B "warehoused at Microsoft" estimates, the actual hard number is $0.4B of H200 at the vendor — and the same week, 2022-era GPUs got a rent hike. The generational-churn version of overbuild looks nothing like the warehousing version.
- Memory is out-pricing GPUs. Nebius raised memory-backed instance rates ~41%, more than any GPU tier — the scarcest input in the stack right now isn't silicon, it's what sits next to it.
- Sell ratings can't dent the tape. Redburn's Sell initiations on both CoreWeave and Nebius were met with the stocks rising — the market is still paying for scarcity and daring the supply broadening to show up.
Open threads worth a vote
- None this cycle — the Rubin verification thread closed with the stale-re-report verdict, and no new steer-able question was opened. The natural candidate for next cycle: whether the H200 charge is a one-off or the start of a pattern.