95% of NVIDIA’s Grace Blackwell GPUs Yet to Be Deployed: Report

More than 95% of the world’s announced NVIDIA Grace Blackwell GPU capacity has yet to be deployed, according to Air Street Capital’s latest State of AI Report Compute Index, which tracks major AI compute deployments worldwide.

The index estimates that 100,128 Grace Blackwell GPUs (GB200/GB300) had been deployed as of July 1, compared with an announced pipeline of around 1.66 million GPUs, meaning less than 5% of the announced capacity is operational. Another 308,640 GPUs are in the installation phase

The accompanying deployment data shows that the largest announced projects remain largely unbuilt. Saudi Arabia’s HUMAIN project accounts for roughly 600,000 announced Grace Blackwell GPUs, while Stargate Abilene, the OpenAI, Oracle, and Crusoe-backed project, accounts for about 450,000.

South Korea’s national AI programme has announced around 260,000 GPUs, while Nscale’s Microsoft deployment totals roughly 200,000 announced accelerators.

In contrast, deployed Grace Blackwell systems remain comparatively small, with the first GB300 clusters only recently entering production.

The report also tracks previous NVIDIA GPU generations. Hopper-based clusters remain the largest deployed AI systems globally, led by xAI’s Colossus 1 with around 200,000 H100/H200/GH200 GPUs, followed by Tesla Cortex with roughly 66,000 GPUs and Meta's GenAI clusters with about 49,000 GPUs. 

Earlier A100 deployments remain concentrated in research compute, with Meta Research SuperCluster leading at around 16,000 A100 GPUs.

The State of AI Report Compute Index is an ongoing dataset maintained by Air Street Capital that tracks publicly disclosed AI compute infrastructure across GPU generations, including A100, Hopper, Blackwell and Grace Blackwell systems. 

According to the project’s methodology, the index compiles information from company announcements, public procurement documents, regulatory filings, hyperscaler disclosures, government announcements, satellite imagery, vendor data and industry sources. 

Each installation is classified as announced, installed, or deployed based on publicly verifiable evidence, with the database updated continuously as projects progress.

The deployment figures come as NVIDIA prepares its next AI platform transition. 

At GTC 2026 in March, the company revealed that the Vera Rubin platform has entered production, scheduled to succeed Grace Blackwell in the second half of 2026, with Rubin Ultra planned for 2027 as part of its annual data centre roadmap.

At the event and subsequent earnings discussions, NVIDIA said it has visibility into at least $1 trillion in revenue from its Blackwell and Rubin platforms through 2027.