Gigawatts, custom chips and multi-cloud: how OpenAI, Anthropic, Google and xAI are fighting over the platform

⚡ The gist in 5 seconds - The battle among AI companies is no longer over individual model releases, but over compute capacity, custom chips, presence in clouds and the developer ecosystem. - OpenAI has secured over 10 GW under contracts, exceeding its Stargate target for 2029; Anthropic is assembling capacity via AWS, Google, AMD and Azure; xAI holds about 1 GW on NVIDIA. - Cloud exclusivity is gone: OpenAI, Claude and Grok models are now available across several major cloud platforms at once. ### 🔍 What was found In May 2026, Anthropic announced it was getting Colossus 1 — a cluster with more than 300 MW of capacity and over 220,000 NVIDIA GPUs. The cluster was built for Grok, and it's owned by SpaceX, which acquired xAI in February. One model developer rented a supercomputer from another, and the market took it in stride — for 2026 this is a normal picture: model releases come every couple of months, while the real battle is fought in gigawatt contracts, custom chips, cloud catalogs and per-million-token pricing. OpenAI discloses its capacity in more detail than anyone: 0.2 GW in 2023, 0.6 GW in 2024, and about 1.9 GW in 2025. In April 2026, the company announced contracts totaling more than 10 GW: NVIDIA for 5 GW of Vera Rubin systems, AMD for 6 GW, Broadcom for 10 GW of its own accelerators, and AWS for roughly 2 GW of Trainium (don't add the numbers up — much of it is chips for the same data centers). According to Epoch AI's estimate, in mid-April only one of the seven Stargate sites was operational, in Abilene, at about 0.3 GW; expansion beyond 1.2 GW was halted in March due to grid connection delays. Anthropic is promised up to 5 GW from AWS (about 1 GW on Trainium by the end of 2026), up to a million TPUs from Google, 1 GW of Ironwood in 2026 and 5 GW of TPU 8i in 2027, up to 1 GW in Azure, and up to 2 GW on AMD. xAI built its first Colossus in 122 days; Colossus and Colossus II together deliver about 1 GW. 2026 capex: Alphabet — $195–205 billion, Amazon — about $220 billion, Microsoft — about $175 billion, Meta — $130–145 billion. On the chip front: NVIDIA posted revenue of $96.2 billion for the quarter through July 2026, of which $89 billion came from data centers (+117% year over year). Google unveiled its eighth-generation TPUs in April — the 8t for training (12.6 PFLOPS FP4, 216 GB HBM) and the 8i for inference (10.1 PFLOPS, 288 GB HBM), claiming up to 2.7x better performance per dollar in training than Ironwood; there are no independent benchmarks. Google will also start supplying TPUs to individual customers for their own data centers. Amazon released third-generation Trainium (144 GB HBM3e) in December 2025, and Microsoft launched Maia 200 in January, which also serves GPT-5.2. OpenAI ordered the Jalapeño inference accelerator from Broadcom: by its measurements, 1.5–1.9x more work per watt than GB200 and GB300; shipments have begun, with a 1.3 GW deployment planned for 2027. xAI still counts everything on NVIDIA. ### 💡 Why it matters Cloud exclusives have all but disappeared: in April, OpenAI and Microsoft rewrote their agreement — OpenAI can now sell its products through l

🔗 Read on habr.com