Where're All The AI Chips? (Ed Zitron)
Where’re All The AI Chips? (Ed Zitron)
Ed Zitron’s long-form investigation into the gap between NVIDIA’s GPU revenue figures and actual operational AI data center capacity — arguing the entire AI data center buildout is being systematically misrepresented to investors.
Connections
Ed Zitron builds on prior reporting (his Guardian/Microsoft GPU investigation) to argue that Nvidia’s revenue growth is “almost entirely the result of speculative purchases by hyperscalers and neoclouds that take years to install its GPUs.”
Core findings:
- Microsoft has ~$50B of GPUs operational at ~2GW AI capacity, but has spent ~$265B in total capex since 2022. That implies ~$106.7B in short-lived, uninstalled assets (GPUs and associated gear) sitting in warehouses or unpowered data centers.
- Across tracked hyperscalers and neoclouds (Google, Meta, Oracle, Amazon, SpaceX, Tesla, CoreWeave, IREN, Core Scientific, Applied Digital), there is $374B+ in “construction in progress” assets — and that excludes Microsoft’s contribution entirely.
- Total estimated uninstalled GPUs/TPUs: ~$234B, or roughly 50% of all AI chips sold since 2023.
- Nvidia + Broadcom have sold ~$561.5B in AI chips since early 2023. Zitron estimates majority of Blackwell GPUs are yet to be installed.
The weasel-wording problem: Microsoft claims 12GW of capacity but only 2GW is AI-specific. “Capacity” conflates active AI compute, power secured but not connected, buildings under construction, and non-AI data centers. Zitron argues all capacity reporting is now “functionally useless.”
Anthropic and OpenAI are identified as the distortion agents — their $1.3T in compute commitments and near-infinite capital create the illusion of “insatiable demand” when they account for 70%+ of hyperscaler AI revenue. Remove them and outside demand is ~$22B.
The bottom line Zitron draws: “I believe we are now in an inevitable overbuild situation, one with no neat, tidy Dot-Com Bubble-style exit story.” CoreWeave, Nscale, Lambda — neocloud debt is project-financed against contracts that are contingent on OpenAI/Anthropic raising more capital. A Blackwell GPU flood is incoming and will suppress prices before most capacity is even energized.
Related: Texas Teachers’ CIO Warns AI Capex Boom Echoes Past Busts makes the same overbuild concern from a pension fund CIO’s perspective; Treasury Has an Internal Report Warning About the Dangers of an AI Bubble is the government’s internal version of the same worry.
Key numbers
| Company | CIP / uninstalled | Detail |
|---|---|---|
| Microsoft | ~$50B–$100B GPUs warehoused | 2GW AI capacity; 12GW “total capacity” claimed |
| $122.8B “assets not yet in service” (Q2 2026) | 60% of capex is servers (GPUs/TPUs) | |
| Oracle | $48.5B CIP | Stargate Abilene: only 3 of 8 buildings operational as of June 2026 |
| Amazon | $71.7B CIP (end 2025) | + likely $25B+ added in 2026 |
| Total tracked | $374B+ CIP | Excludes Microsoft, Firmus, Sharon AI, sovereign AI projects |
| Estimated uninstalled chips | ~$234B | GPUs/TPUs sitting in warehouses or unpowered DCs |
The weasel-wording taxonomy
Zitron documents four patterns companies use to obfuscate:
- Conflating total capacity (CPU, cloud storage, AI) with AI-specific capacity
- Reporting power secured (a land/power contract) as operational capacity
- Announcing one or two buildings as a data center campus “open”
- Extending server depreciation schedules to suppress reported costs
Questions Zitron says every journalist/analyst should ask
To NVIDIA: How many Hopper/Blackwell GPUs are operational and generating revenue? What % of FY25/26 sold GPUs are operational?
To hyperscaler CEOs: How much AI-specific data center capacity is operational? How many GPUs by type are in storage vs. installed?