Who Trains the Biggest AI ? OpenAI vs Google vs xAI (2020–2026)

How much computing power does it take to train the biggest AI models? This bar chart race tracks the training compute of 20 AI models quarter by quarter, from Q2 2020 to Q3 2026, starting at GPT-3 (314 ZFLOP) and ending at GPT-6 Astra (estimated 1,000,000 ZFLOP), an increase of more than 3,000x. What you’ll see: The #1 spot changes hands 8 times: GPT-3, Gopher, Chinchilla, PaLM, GPT-4, Gemini 1.0 Ultra, GPT-4.5, Grok 4, and finally GPT-6 Astra. The company race: OpenAI, Google DeepMind, xAI, Meta, Anthropic, and more. Unit: zettaFLOP (1 ZFLOP = 10²¹ floating-point operations). Data notes: All values are estimates. Epoch AI flags many models as “Speculative,” which can be off by up to 30x. GPT-6 Astra’s training compute has not been officially published. OpenAI has only said it was trained on 100,000+ GPUs, so the figure shown is an estimate. Only models that ranked in the top 5 by compute at release are included, so 2020–2022 data is incomplete. Each model appears in its release quarter and holds its value afterward. Sources: Epoch AI, Frontier AI Models dataset (CC BY 4.0); “Emergent Abilities of Large Language Models” (arXiv 2206.07682) for GPT-3, Gopher, Chinchilla, and PaLM. Want a bar chart race on another tech topic? Tell us in the comments. Music: ZAY YEZ by ZiMPL / zimplmusic License: Creative Commons — Attribution 3.0 Unported — CC BY 3.0 Free Download / Stream: https://links.al/rMB Music promoted by Audio Library: https://links.al/youtube