/TECH

Frontier AI Pretraining Costs Hit 20 Million Dollars, Limiting Access

WB

Combined views

17.9K

2 posts, first seen 3h ago

303 likes26 comments

Combined views

17.9K

2 posts, first seen 3h ago

303 likes comments

64 saves

13 reposts

Sentiment

Positive69.2%30.8%Negative

Positive Read

Positive users see $20M frontier pretraining costs as accessible enough for more countries and competitors while negative users worry the barriers are still too high and will drive further consolidation.

Based on 13 visible X reactions from 25 accounts; directional sample.

Digg Deeper

Ask a question below.

Published answers will appear here.

Commentary on X

KP Kislay Parashar@KislayParashar1

2:36 AM · Jul 17, 2026

@willccbb At these numbers a failed training run isn't a bug fix anymore, its a company ending event for anyone outside a few dozen labs.

HO Hank Oslo@xctlot

4:48 AM · Jul 17, 2026

@willccbb "only", that's dirt cheap. That's 100 relatively inexpensive ftes for a year or a 20 person acqui hire

F0 F00M@0xVonNeumann

2:49 AM · Jul 17, 2026

@willccbb So you are saying that we could have thousands of competitors for the frontier? Nice!

AL Allan@AllanatrixQ

2:44 AM · Jul 17, 2026

@willccbb Yeah that makes sense it will consolidate further the barrier to entry is absurdly high

View all

FO Felipe Oviedo@ands_zen

4:42 AM · Jul 17, 2026

@willccbb your posting has been 🔥 recently

SU Satyaki Upadhyay@satyaki_u

2:26 AM · Jul 17, 2026

@willccbb Few thousand is _very_ accessible

WB will brown@willccbb

1:54 AM · Jul 17, 2026

573TECH

it's worth noting that due to rising costs of pretraining, the frontier isn't *that* accessible. training a 5T-A100B model on 50T tokens on GB300s costs like $10 million for the hero run. double it for RL. there's only a few thousand companies in the world that could justify it

WB will brown@willccbb

2:00 AM · Jul 17, 2026

573TECH

6 months ago, the biggest model successfully trained for under $20M all-in was arcee's 400B-A13B trinity large. the theoretical limit is now 10x higher, which is great, but a new lab would need to raise upwards of $100M to do this with low risk, only a few dozen of these so far

View all

FO Felipe Oviedo@ands_zen

4:42 AM · Jul 17, 2026

@willccbb your posting has been 🔥 recently

SU Satyaki Upadhyay@satyaki_u

2:26 AM · Jul 17, 2026

@willccbb Few thousand is _very_ accessible

WB will brown@willccbb

1:54 AM · Jul 17, 2026

573TECH

it's worth noting that due to rising costs of pretraining, the frontier isn't *that* accessible. training a 5T-A100B model on 50T tokens on GB300s costs like $10 million for the hero run. double it for RL. there's only a few thousand companies in the world that could justify it

WB will brown@willccbb

2:00 AM · Jul 17, 2026

573TECH

6 months ago, the biggest model successfully trained for under $20M all-in was arcee's 400B-A13B trinity large. the theoretical limit is now 10x higher, which is great, but a new lab would need to raise upwards of $100M to do this with low risk, only a few dozen of these so far

The Conversation on X

2 posts

26

64 saves

13 reposts

Sentiment

Positive69.2%30.8%Negative

Positive Read

Positive users see $20M frontier pretraining costs as accessible enough for more countries and competitors while negative users worry the barriers are still too high and will drive further consolidation.

Based on 13 visible X reactions from 25 accounts; directional sample.

Digg Deeper

Ask a question below.

Published answers will appear here.