Microsoft’s AI Chip Shortfall: A Gap Between Plans and Reality

Internal documents obtained by the Guardian show Microsoft has installed roughly 2.2 million AI-focused GPUs across its global data-centre portfolio. That is less than half of the 1.8 million-chip target the company set for the end of 2024 and far below analyst expectations. The shortfall suggests many newly built data-centre shells remain unequipped.

The Guardian

Since 2022, Microsoft has spent about $280 billion on land, buildings and compute infrastructure, including more than $41 billion in the most recent quarter. CEO Satya Nadella pledged to double the global data-centre footprint by mid-2027. AI chips—primarily Nvidia’s H100 and Blackwell models—are central to this expansion, but Nvidia does not disclose sales volumes.

The Guardian

• Microsoft’s reports claim an addition of 5 GW of AI-focused capacity over the past two years, implying a potential total of 10 GW.

The Guardian

• Professor Shaolei Ren (UC Riverside) estimates a 10 GW AI footprint would need roughly 12 million H100-class chips (700 W per chip).

The Guardian

• Using a conservative 80 % of power for AI chips, the effective compute power is 8 GW, or about 6.4 million GPUs.

The Guardian

• Nvidia said its top four customers ordered 3.6 million Blackwell chips; Microsoft’s share is inferred to be near 1 million, yet internal data show fewer than 0.5 million installed.

The Guardian

• In the *All Things AI* podcast, Nadella said the main obstacle is a lack of “warm shells”—completed facilities with sufficient power—leading to inventory of uninstalled chips.

The Guardian

• He announced the “Fairwater” project in Wisconsin was “going live,” but satellite imagery and a local newspaper later indicated the facility remained largely offline.

The Guardian

• A Microsoft spokesperson stressed the firm does not publish specific chip counts and called the Guardian’s estimates “inaccurate” and based on “incorrect assumptions.”

The Guardian

• Internal documents show 2.2 million chips installed, while analysts calculate a need for 4–12 million based on reported gigawatt capacity.

The Guardian

• Sustainability reports suggest AI capacity of roughly 1.2 GW in 2024 (about 4 million chips), contradicting the 5 GW figure in earnings releases.

The Guardian

• Nvidia declined comment on the distribution of Blackwell orders, leaving Microsoft’s exact share unverified.

The Guardian

If a substantial portion of Microsoft’s AI-focused infrastructure remains unequipped, the company could face delays in delivering AI services, potentially ceding ground to rivals such as Amazon, Google and Meta that may have more fully operational chip deployments. The discrepancy also highlights the opacity of the AI-chip supply chain, complicating investors’ and regulators’ ability to gauge AI commercialization.

The Guardian

Microsoft says expanding power-rich “warm shells” is a priority but has offered no concrete timeline for completing chip installation. The next quarterly earnings release should provide updated figures on AI-related power usage and data-centre readiness, which may clarify the shortfall.

The Guardian