Training Data, Not Compute: Ex-OpenAI Researcher Bets $100B

Scaling laws had a good run, but the next big money inflow in AI might not be going to NVIDIA. Ex-OpenAI researcher Andrew Ho is putting his chips on training data — to the tune of $100 billion in targeted collection spending across AI

Story tracker · related coverage

[OpenAI Slashes GPT-5.6 Luna Price 80%: A Pricing Deep Dive 26m ago](/en/news/4489/)

[GPT-5.6 Revenue Surge: How July Beat Q2 10h ago](/en/news/4448/)

OpenAI and the US Government Just Rediscovered the "Blank Map" 13h ago

[Lilian Weng's Return to OpenAI 13h ago](/en/news/4424/)

[LLM API Price Drops: How to Cut Costs by 50% 15h ago](/en/news/4418/)

[Sam Altman's White House Talks: A Call to Decelerate AI? 22h ago](/en/news/4379/)

[Next OpenAI Slashes GPT-5.6 Luna Price 80%: A Pricing Deep Dive →](/en/news/4489/)

All Replies （3） #

M

Underrated angle: data quality caps what any model can learn, no amount of compute scales past that.

0

N

Spent weeks cleaning our dataset, improved results more than any compute bump ever did.

0

Q

Had a similar experience—deduplicating our corpus gave a bigger jump than adding nodes.

0