Aerial photos and other analysis of SPACEXAI colossus 2 indicate that there are 440,000 B200 and B300 chips active as of 2 weeks ago.
Brian Wang is a Futurist Thought Leader and a popular Science blogger with 1 million readers per month. His blog Nextbigfuture.com is ranked #1 Science News Blog. It covers many disruptive technology and trends including Space, Robotics, Artificial Intelligence, Medicine, Anti-aging Biotechnology, and Nanotechnology.
Known for identifying cutting edge technologies, he is currently a Co-Founder of a startup and fundraiser for high potential early-stage companies. He is the Head of Research for Allocations for deep technology investments and an Angel Investor at Space Angels.
A frequent speaker at corporations, he has been a TEDx speaker, a Singularity University speaker and guest at numerous interviews for radio and podcasts. He is open to public speaking and advising engagements.
Can you point to where Google, Microsoft/OpenAI, Meta, or Amazon/AWS are making huge deals to rent their excess capacity back to large Frontier AI models?
SpaceXAI seems to be the only one.
Which is Brian’s point.
Not that the others aren’t building massive amounts of AI compute. We know they are spending more and building more than SpaceXAI.
It’s just that they’re so slow. By the time their Data Centers are complete, Nvidia (and soon others) will have moved on to the next GPU platform, giving these huge, slow-moving Data Center buildouts a shorter useful lifespan. Meanwhile, SpaceXAI keeps accelerating their rates of implementation, better able to incorporate the latest/greatest GPU platforms.
At least that’s how I read it.
Brian, please correct me if I’m wrong.
Amazon and Google are building to rent to Anthropic. Mix of investment and cash. Microsoft did the investment deal and get loads of AI rents from Openai.
Anthropic uses a multi-vendor strategy (AWS Trainium primary for training, Google TPUs, Nvidia GPUs via clouds/neoclouds) and is shifting toward more direct control of facilities.
anthropic.com
AWS (Primary cloud & training partner): Project Rainier cluster (nearly 500k Trainium2 chips) live since ~Oct 2025. April 2026 deal: up to 5 GW total capacity (Trainium2/3/4 + Graviton). Nearly 1 GW online by end-2026. Commitment: >$100B spend over 10 years. Already leasing heavily; phased expansions through 2026+.
Google Cloud (TPUs): Oct 2025 expansion — up to 1M TPUs / >1 GW (tens of billions value), well over 1 GW online in 2026. April 2026 deal (with Broadcom): additional multiple GW (~3.5 GW next-gen TPUs per Broadcom filing), starting 2027. Google investing up to $40B in Anthropic (cash + compute support). Anthropic also pursuing direct US data center leases (>1 GW combined preliminary agreements across >12 sites); Google providing financial backstops/guarantees for ~5 sites, unlocking ~$35B financing. Part already leasing/using; major new capacity 2026–2027+ (with Google backing for direct leases).
techcrunch.com
SpaceX / xAI Colossus (Memphis, TN): May 2026 deal for full access to Colossus 1 (~300 MW / hundreds of thousands of GPUs per reports). $1.25B per month through May 2029 (~$45B total). Active leasing (ramping 2026).
wired.com
CoreWeave: Multi-year GPU capacity deal (phased online later in 2026+).
I will put up an article on this in a few minutes
“Compute Estimate: Because the Blackwell architecture (B200/B300) is substantially faster than Hopper, this translates to a staggering \(\sim 1.11 \text{ million}\) H100-equivalent chips in pure compute capacity.
• Google: Widely estimated as the overall market leader in cumulative AI compute. Industry analysis, such as those from Epoch AI, places Google’s overall capacity—combining both NVIDIA equivalents and their custom, in-house TPUs (like TPU v5)—at the equivalent of 5 million H100 chips (approx. 1.24M H100-equivalents in pure NVIDIA/third-party data). [1, 2, 3]
• Microsoft / OpenAI: Backed by massive Azure investments and specialized chips, estimates indicate a combined capacity of around 3.4 million H100-equivalents, making them the primary rival in cloud AI infrastructure. [1, 2, 3]
• Meta: Operating vast, internally developed clusters (like their Llama clusters), Meta commands an estimated \(\sim 2.3 \text{ million}\) H100-equivalents. [1]
• Amazon / AWS: Utilizing a mix of NVIDIA GPUs and in-house Trainium accelerators, AWS holds an estimated 2.5 million H100-equivalents of computing power. [1]
While xAI’s Colossus 2 cluster is astonishing for a single, centralized facility, the major hyperscalers maintain broader, multi-billion-dollar global fleets of both GPUs and proprietary ASICs. ”
So they have big centralized cluster, but in terms of total GPU they lag a bit behind.