Google's Colossus 1 training run consumed 1e26 FLOP, per Epoch AI analysis. At $4.6B estimated cost, it dwarfs prior records by an order of magnitude.
Key facts
- Colossus 1 used 1e26 FLOP training compute.
- Estimated cost: $4.6B.
- 10x compute of GPT-5's ~1e25 FLOP.
- Model likely exceeds 10 trillion parameters.
- Most expensive single AI training run ever.
Key Takeaways
- Google's Colossus 1 used 1e26 FLOP at $4.6B, per Epoch AI.
- It is the largest known training run, signaling a new capital scale.
The 1e26 FLOP Milestone
Epoch AI estimates Google's Colossus 1 training run consumed 1e26 FLOP of compute, according to their analysis. That is roughly 10x the compute of OpenAI's GPT-5, which used ~1e25 FLOP. The cost, at $4.6B, signals a new tier of capital concentration in AI — comparable to building a small nuclear reactor.
Google did not disclose the exact parameter count or architecture. But 1e26 FLOP at contemporary efficiency implies a model well beyond 10 trillion parameters, possibly exceeding 100 trillion. The run likely used Google's TPU v7 pods, which the company announced in 2025 as capable of 1e26 FLOP sustained.
Capital Escalation
The $4.6B figure is not just a compute cost — it represents the total budget including hardware depreciation, energy, cooling, and engineering labor. Epoch AI derived this from standard industry models: $0.046 per 1e15 FLOP for TPU v7, scaled linearly. [According to Epoch AI], this makes Colossus 1 the most expensive single AI training run ever, surpassing even the rumored $2B cost of GPT-5.
This concentration mirrors broader trends. Google's $1.65 trillion off-balance-sheet AI infrastructure leases, reported July 21, 2026, show the company is betting on ever-larger models. Colossus 1 is likely the first of a new class of "gigamodel" runs that will become routine by 2028.
Benchmark Implications
Epoch AI does not provide benchmark results for Colossus 1 — the model is not yet public. But at 1e26 FLOP, it should outperform GPT-5 on all standard benchmarks by a wide margin. The key question is whether diminishing returns have set in: GPT-5 achieved 89% on MMLU; Colossus 1 might reach 95%, but the incremental gain per dollar is shrinking.
Google's competitors are watching. [According to recent reporting], OpenAI and Anthropic are both planning 1e26 FLOP runs for late 2027. If Colossus 1 delivers only marginal improvements, the industry may pivot to inference-optimized architectures rather than brute-force scaling.
What to watch
Watch for Google's Q3 2026 earnings call, likely in October, where executives may disclose Colossus 1 benchmark results or announce a release date. Also watch for OpenAI and Anthropic to announce their own 1e26 FLOP runs, which would confirm the scaling trend.
Source: news.google.com
