Most training run compute greater than 2e27 FLOP by EOY 2026?

1kṀ2363

Dec 31

92%

chance

ALL

see https://epoch.ai/trends#compute. At market creation, the most recent update put grok 3 at 5e26 FLOP.

This market will resolve to epoch's estimate if it is available at time of market close. If unavailable, the most sensible estimate at that time will be used.

Update 2025-12-29 (PST) (AI summary of creator comment): Distributed training runs are not relevant to this market at 2e27 FLOP - only centralized training runs count.

Market context

Technical AI Timelines

OpenAI

AI Benchmarks

Get

1,000

to start trading!

People are also trading

By the end of 2030, will a 2e29 FLOP training run have ever taken place?

76% chance

Will a machine learning training run exceed 10^27 FLOP in China before 2028?

44% chance

Will a machine learning training run exceed 10^26 FLOP in China before 2027?

86% chance

Will a machine learning training run exceed 10^25 FLOP in China before 2027?

82% chance

Will a machine learning training run exceed 10^26 FLOP in China before 2028?

85% chance

Will a machine learning training run exceed 10^27 FLOP in China before 2030?

79% chance

Will a machine learning training run exceed 10^26 FLOP in China before 2029?

82% chance

Will a machine learning training run exceed 10^30 FLOP in China before 2035?

34% chance

Will the largest machine learning training run (in FLOP) as of the end of 2035 be in the United States?

69% chance

Will the largest machine learning training run (in FLOP) as of the end of 2040 be in the United States?

Sort by:

https://x.com/EpochAIResearch/status/2005686960099348539

@Dulaman that is not relevant to this market

@Bayesian I disagree. That curve can do all sorts of interesting things in 2026

@Dulaman this is completely implausible. they are 1000x behind at least and that doesn't even include the interconnect disadvantage. it's completely incomparable

default trend is scenario B, but if we're in scenario A, then distributed training could catch up to "top 1 centralised" before EOY 2026

@Bayesian you're thinking about it wrong

@Dulaman no you are, this is pure slop

@Bayesian Ok then, what do you think this curve will look like over the next 12, 24 months?

@Dulaman not at 2e27 FLOP

I suspect we're going to get a 10^26 FLOP distributed run by 2027. 10^27 seems less likely, but I could see interesting things happening with the open source models in China, if they start doing massive distributed training runs between companies

like if the technology is there to handle a run that large, and they're all releasing open source models anyway, why wouldn't they pool their resources together?

the "interconnect disadvantage" isn't an argument against large FLOP runs, if anything the opposite is true. You want to burn more compute to overcome the inefficiencies caused by diloco etc