# Half of Nvidia's sales may be in warehouses
_Ed Zitron tells The Tech Report Microsoft has only 2 GW of AI chips online, leaving tens of billions in GPUs warehoused._
**Published:** 2026-09-18
**Source:** https://www.startuphub.ai/hardware/half-of-nvidia-s-sales-may-be-in-warehouses
---
Ed Zitron told [The Tech Report](https://www.youtube.com/watch?v=8C-J2sRBMkQ) that half of Nvidia's revenue could be sitting in warehouses, uninstalled and undepreciated. If he is right, he said, this will be an absolute scandal that humiliates the largest companies in the world.

The interview was built around a Bloomberg report that only about 2 gigawatts of [Microsoft](/startups/microsoft-stock)'s current 12-gigawatt capacity is centered on AI-specific chips. Zitron, who writes the Where's Your Ed At newsletter and hosts the Better Offline podcast, said his own math put Microsoft at 1.993 GW, worth about $50 billion. He estimated Microsoft has spent over $265 billion in capex and, assuming half is GPUs, has $82.5 billion of GPUs not installed at all.

If Microsoft cannot install them, who can.

Zitron framed the timeline as the tell. Microsoft built the first large-scale AI supercomputer in 2020 with Nvidia A100s networked with Mellanox, and today claims it has added a gigawatt of capacity every quarter for three quarters. He argued the gigawatt headlines conflate general compute with AI silicon, and that the actual AI slice is a fraction of what investors assume. The same gap, he said, extends to Oracle's claimed several hundred megawatts and Stargate Abilene, where he doubts a fourth of eight buildings is online, and to [CoreWeave](/startups/coreweave-6-billion-lancaster-data-center), [OpenAI](/startups/openai-gpt-5-6-powers-kiro-for-devs) partners and other hyperscalers now racing to buy Blackwell while Vera Rubin and whatever follows are already on the roadmap. Nvidia, meanwhile, is guiding to double sales next year even as installation lags by years.

What has to happen next is where the risk compounds. Chips bought today for data centers that will not be ready for two GPU generations will sit, age, and face power and cooling uncertainty. Zitron said the accounting clock may not start until gear is placed in service, which lets depreciation stretch, but any shock to a payer like [OpenAI](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/openai-gpt-5-6-powers-kiro-for-devs) would force write-offs in the tens of billions, spread and softened but still real. Warehoused silicon also carries insurance, degradation and obsolescence costs that shareholders are not shown, because hyperscalers do not disclose how many GPUs they bought or how many are online.

The point is not that Nvidia broke the law selling to willing buyers. The point is valuation. If demand is really pre-orders driven by fear that [Jensen Huang](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/jensen-huang-agi-practically-already-g20) will cut off future supply, then the boom is a Kickstarter for capacity that may never come online in time to matter, and every gigawatt update becomes noise until companies report GPU counts and online counts separately. Without that, Zitron argued, the hangover is either a glut that hits at once or a slow roll of markdowns dressed up as efficiency.
---
Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory.