The AI race is becoming a gigawatt race.
Frontier labs are no longer competing only on model architecture. They are competing on power contracts, chip supply, data-center speed, inference limits, and who controls the scarce infrastructure beneath the agents.
Compute scarcity has moved from training-room problem to product-surface problem.
When Anthropic added SpaceX capacity, it immediately doubled Claude Code five-hour limits, removed peak-hour reductions for Pro/Max, and raised Opus API limits. That is the tell: inference capacity is now visible to users as rate limits, latency, product tiers, and degraded peak-hour behavior.
Who has the most compute today?
Approximate H100-equivalent capacity, normalized to Google as the widest bar. These are public/industry estimates, not audited fleet counts.
Headline commitments are not the same as available 2026 capacity.
The hard question is timing: how much capacity is live when users need it, versus committed on paper for later buildout?
Forward compute scale
Why the SpaceX/xAI deal is weird
- Good for Anthropic: immediate scarce capacity from a rival, with product limits improved the same day.
- Bad signal for Anthropic: the cloud deals were not arriving fast enough to prevent user-visible constraint.
- Good for xAI/SpaceX: Colossus can monetize like a neocloud, strengthening the infra story.
- Odd for xAI-the-lab: leasing a whole cluster to a rival suggests Colossus 1 was less urgent for Grok than the cash/strategy value.
Anthropic: relieved, diversified, still behind OpenAI’s scale.
Its own announcement ties new SpaceX capacity to higher Claude Code, Pro/Max, and Opus API limits — visible production inference pressure.
Colossus 1 adds more than 300 MW and 220k+ NVIDIA GPUs within a month: operationally huge, but only 0.3 GW in a gigawatt race.
AWS Trainium, Google TPUs, NVIDIA via Azure/xAI, and Fluidstack sites reduce single-vendor risk, while increasing lease/partner dependence.
Four signals matter more than any one GPU count.
Google controls TPUs end-to-end. OpenAI and Anthropic mostly rent. That changes margins, reliability, and bargaining power.
Capacity promised for 2027–2029 does not fix today’s rate limits. Near-term inference relief is a competitive product feature.
Interconnects, substations, gas turbines, cooling, and data-center construction now constrain how fast chips become usable compute.
Limits, peak-hour throttling, API ceilings, latency, and model availability are infrastructure constraints translated into UX.
AI compute terms in plain English.
Quick definitions for readers who have heard “H100” and “gigawatt” but do not live inside data-center procurement.
What is an H100?
An H100 is NVIDIA’s flagship data-center GPU generation from the current AI boom. It is a specialized accelerator used to train and run large AI models much faster than a normal CPU. When people say a lab has “H100s,” they usually mean it has access to high-end AI compute.
What does H100-equivalent mean?
H100-equivalent is a rough translation layer. Different chips — NVIDIA H100s, GB200s, Google TPUs, and others — have different performance. Analysts convert them into “about how many H100s worth of AI compute” so fleets can be compared directionally.
Why do gigawatts matter for AI?
A gigawatt is a power measure. Massive AI clusters need chips, but chips only become useful when a data center has enough electricity, cooling, networking, land, and grid interconnection. That is why the AI race increasingly sounds like an energy and construction race.
What is the difference between training and inference compute?
Training compute is used to create or improve a model. Inference compute is used every time users run the model. Claude Code limits, ChatGPT availability, API rate limits, and peak-hour slowdowns are mostly inference-capacity symptoms.
Why is Anthropic renting SpaceX/xAI capacity important?
Because Anthropic’s usage limits improved immediately after the deal. That suggests the bottleneck was not abstract PR or future training capacity; it was live production capacity that affected paying users.
Is Anthropic behind OpenAI on compute?
Directionally yes on total forward dedicated scale, based on public commitments. Anthropic has strong supplier diversity and rapid lease velocity, but OpenAI’s Stargate-style program appears larger and more dedicated if it is funded and executed.
Does more compute automatically mean a better model?
No. Data, algorithms, product focus, post-training, inference systems, and distribution all matter. But at the frontier, insufficient compute can cap training runs, product reliability, pricing, limits, and experiment velocity.
Appendix: sources and caveats
Built from Peter’s Obsidian note AI Compute Landscape — Major Players 2026, plus live checks of Anthropic/OpenAI/Epoch pages. Public numbers mix GPUs, H100-equivalents, TPUs, megawatts, capex, and partner commitments; treat the exact figures as directional.
- Anthropic — Higher usage limits and SpaceX compute deal
- Anthropic — Up to 5 GW Amazon compute agreement
- Anthropic — Google/Broadcom TPU partnership
- Anthropic — Microsoft/NVIDIA strategic partnerships
- Anthropic — $50B American AI infrastructure investment
- OpenAI — Stargate infrastructure for the intelligence age
- OpenAI — Five new Stargate sites
- OpenAI — Oracle Stargate 4.5 GW expansion
- Epoch AI — Computing-capacity estimates
- Epoch AI — OpenAI Stargate site breakdown
- xAI — Colossus official page
- Google Cloud — Eighth-generation TPU for the agentic era
- Google Cloud — AI infrastructure at Next 26
- Meta Engineering — Building Meta’s GenAI infrastructure
- Meta AI — Llama 4 announcement
- Data Center Dynamics — Stargate Abilene GB200 report
- CNBC — Meta cloud AI compute report
- CNBC — Anthropic/SpaceX data-center capacity report
- SemiAnalysis — Multi-datacenter training analysis
Share framing: AI agents are not just a software category; they are a load profile. The winners will be the labs and clouds that turn promised gigawatts into reliable inference before users feel the squeeze.