Capacity has stages
Secured power, construction, installed systems, and usable compute are different milestones. Announced gigawatts should not be read as live capacity.
THE AI LANDSCAPE 04 / INFRASTRUCTURE & COMPUTE
Explore the clouds, dedicated clusters, and infrastructure programs behind AI. See who operates them and which capacity is available, planned, or still being brought online.
THE BIG PICTURE
Selected coverage. Check each entry for context and evidence.
Try a different name or a broader category.
A manually reviewed snapshot as of September 23, 2026. Colors organize the evidence; they do not score quality. On smaller screens, swipe across the comparison.
CONTEXT MAKES THE DIFFERENCE
Our interpretation of the comparison.
Secured power, construction, installed systems, and usable compute are different milestones. Announced gigawatts should not be read as live capacity.
Networking, storage, orchestration, software, and service access shape the infrastructure a customer actually uses.
An exclusive capacity agreement can give a model developer compute without transferring ownership of the data center.
TRANSPARENT BY DESIGN
Official announcements, product documentation, and model repositories. All linked sources were checked on September 23, 2026. Publication dates and review dates are shown separately.
Selected infrastructure operators and programs, grouped by their route to use. Cloud services are offered to customers; dedicated compute primarily supports an operator or contracted users; buildout programs aggregate facilities and partnerships over time. Status describes the cited milestone, not a verified site-by-site capacity audit. No market-share or total-capacity ranking is implied.
Dates on continuously updated documentation are labeled “Current documentation” unless a specific publication date is available. This page is maintained manually; the review date is not a live-feed timestamp. Company announcements describe the company’s claims and plans, not an independent audit.
Amazon / Cloud services
AWS offers Trn3 UltraServers built around Trainium3, with Neuron software support. Availability of a product family does not mean capacity in every AWS region.
Microsoft / Cloud services
Microsoft’s Maia 200 announcement describes deployment inside Azure for inference. This is evidence of an operating custom-chip program, not general customer access to every Maia system.
Google / Cloud services
Cloud TPU lists Ironwood as generally available and TPU 8t and 8i as coming soon. The next-generation announcement must not be read as broad customer availability.
CoreWeave / Cloud services
CoreWeave reports bringing up and validating Vera Rubin NVL72 and publishing measured results. That establishes a working system, not universal customer availability across its fleet.
Mistral AI / Cloud services
Mistral’s August update combines regional inference services with a longer-term European compute program. Its target of up to 1 GW by 2030 is a plan, not installed capacity today.
Meta / Dedicated compute
Meta describes a network of data centers and a combination of partner accelerators and custom silicon. This row maps the infrastructure role rather than claiming independence from suppliers.
SpaceX / Dedicated compute
Anthropic announced an agreement for all of Colossus 1’s compute capacity, with more than 300 MW expected within a month. The announcement is evidence of the agreement and schedule, not an independent confirmation of delivered capacity.
OpenAI + partners / Buildout programs
OpenAI says Stargate surpassed its target of securing 10 GW. Securing infrastructure is different from bringing every facility online; the program spans partners and sites at different stages.