Nvidia Capex Digest — Is the AI Capex Story Still Intact?
TL;DR
Yes — and the spend got bigger, not smaller: the top four hyperscalers are now expected to drop $1.8T through 2028, roughly $600B more than the cash their businesses generate over the same stretch. The bear case has stopped arguing about whether they'll spend and started auditing whether the spend converts — Ed Zitron's new piece estimates ~$234B of GPUs sitting in warehouses. Meanwhile Apple opened a second, quieter front: local inference pitched as "no cost per token."
The spend keeps repricing upward — the debate is now conversion, not commitment
The capex story didn't just hold this cycle; it got re-rated up. Arete Research, at the AI Infra Summit, put the top four hyperscalers at $1.8 trillion of capital investment through 2028 — about $600 billion more than their underlying businesses generate in cash over that period, meaning capex will consume nearly all of their operating cash flow in 2026 (Broadband Breakfast). The gap gets financed, not foregone.
The scarcity signals all point the same direction. Global X counts data center equipment lead times around 42 weeks, "83% higher than 2019 levels," with capex up ~75% in 2026 and possibly another 30% in 2027 (Global X). And the single most load-bearing datapoint for Nvidia's order book: Nvidia has supply for only about 70% of stated demand, a constraint expected to persist through fiscal 2028 (24/7 Wall St). Downstream, Nebius raised GPU cloud prices 17–21% citing demand, and Micron's CEO says data center demand runs at "approximately 150% of Micron's actual committed supply" (HostingJournalist). The standing tripwire — a first capex cut or order cancellation — remains unfired, and the market agrees: NVDA closed Sept 23 at $228.87, 2.7% below its 52-week high, on TTM revenue of $302.97B (+105.9% YoY) (NVDA market view).
What to watch: whether the $600B financing gap shows up as hyperscaler debt issuance — that's where the first crack would surface, not in guidance.
The bear case got its accountant: Zitron counts what's actually plugged in
The deployment-gap argument finally has a detailed accounting behind it. Ed Zitron's Sept 22 investigation, "Where're All The AI Chips?", argues that "NVIDIA's revenue growth is almost entirely the result of speculative purchases by hyperscalers and neoclouds that take years to install its GPUs" — Where's Your Ed At.
The numbers are the point: "construction in progress" across hyperscalers and neoclouds has grown from $102.9B in 2023 to over $374B today, and Zitron estimates $50–100B of GPUs warehoused or unpowered at Microsoft alone — echoing Nadella's November 2025 admission of "a bunch of chips sitting in inventory that [he couldn't plug in]" (Where's Your Ed At). His demand-concentration math is the sharper edge: OpenAI and Anthropic make up more than 70% of hyperscalers' AI revenue, and backlogs are "more a function of Dario Amodei and Sam Altman's DocuSign accounts" (Where's Your Ed At).
Why it matters: this is the intellectual groundwork for converting a narrative debate into an earnings event. If even a fraction of the warehousing estimate is right, the first disclosed write-down or deployment-disclosure change is the fundamental trigger — and Zitron explicitly invokes the SEC's 2022 Nvidia crypto-mining settlement as precedent for forcing disclosure of operational capacity vs. sales. Caveats apply: the piece is adversarial, the $234B figure is his own estimate ("might be a little generous"), and it doesn't dispute Nvidia's reported revenue — only what that revenue represents (Where's Your Ed At).
What to watch: the newly opened tripwire — any GPU inventory write-down, uninstalled-GPU disclosure, or depreciation catch-up by a hyperscaler or neocloud. That's the cleanest falsifier either way.
Apple opens a substitution front at the desk: "no cost per token"
The demand-quality question picked up a consumer-scale protagonist. With the new Mac line shipping Sept 22, Apple is pitching corporate buyers that Mac Studios — which "can cost nearly $20,000 and handle AI tasks locally" — are "cheaper than renting data centers." Hardware chief Johny Srouji, to Reuters: "Once you have the machine on your desk, you've paid for it... There's no cost per token. You're just using the machine again and again" — Reuters. The technical claim is no longer exotic: Apple demoed four Mac Studios running a trillion-parameter model off one wall outlet.
Context cuts both ways. Apple holds only about 4.6% of the enterprise desktop market versus 91.3% for Windows, and its own most advanced cloud model reportedly runs on Nvidia GPUs through Google Cloud (Reuters; Benzinga/Yahoo). Today this is rounding error against ~$800B of 2026 hyperscaler capex. But it's the second credible mechanism — after cheap-model routing — by which AI usage grows without frontier data-center GPU pull, which is the quiet version of the bear case.
What to watch: enterprise purchasing patterns and any hyperscaler commentary on inference-demand mix — the signal that local inference is eating token revenue rather than adding edge hardware demand.
What surprised us
- The bear case conceded the spending question entirely. Nobody serious argues hyperscalers will cut capex anymore — the argument is now purely about deployment and conversion. That's a real shift from last cycle, where the debate was still about whether the build-out slows (Broadband Breakfast).
- The supply constraint is the bull case's best defense. Nvidia covering only ~70% of stated demand through fiscal 2028 is a harder fact than any backlog number — you can fake a backlog, not a sold-out factory (24/7 Wall St).
- Zitron's strongest number is Microsoft's. ~$265B of capex since 2022 against ~$50B of GPUs in service is the kind of gap that survives adversarial scrutiny better than his complex-wide estimates (Where's Your Ed At).
- Apple's pitch is aimed at Nvidia's economics, not its chips. Huang downplayed competition — but a "no cost per token" sales line is a direct attack on the metered-inference revenue model the whole capex thesis rests on (Reuters).
Open threads worth a vote
- First GPU Inventory Write-Down or Uninstalled-GPU Disclosure by a Hyperscaler/Neocloud — the new tripwire from Zitron's audit; a fire here converts the narrative debate into an earnings event overnight.