Commentator @teortaxesTex disputes claims that OpenAI's 2,100 MW Stargate cluster will enable near-term 20-trillion-parameter models

A lot of people are going crazy right now that Anthropic has a massive lead with Claude Mythos and they are right if you only think about the presence

Right now Mythos is a few months ahead of GPT-5.5.

But in a few months OpenAI will leapfrog Anthropic with their 15-20T model

Until RSI all defeats are psychological

"OpenAI will leapfrog Anthropic with their 15-20T model" 🤣🤣🤣🤣 OpenAI's pretraining team is a complete shitshow All the good pretraining people are at Anthropic

> But in a few months OpenAI will leapfrog Anthropic with their 15-20T model There is no evidence nor rumor that OpenAI is capable or willing to train a 15-20T model any time soon Lisan believes the path to 100T is paved with Blackwell/Vera Rubin chips, this ain't so simple

@zephyr_z9 so GPT-5.5 sucks for its size and training a 3-5x larger model is impossible for them?

@scaling01 5.5 has really good post training OpenAI sucks at pre training Scars from GPT 4.5 haven't healed

@teortaxesTex i mean if they don't then they deserve to lose lmao

because they have more than enough capacity to train such model

they have like 200k GB200 at their Stargate Abilene + 520k GB200 at Microsoft's linked Fairwater datacenters

@zephyr_z9 My strategy

⬇️

@scaling01 If this were a game where you press X to spend 100M GPU-hours and end up with a strictly stronger model, then everyone involved would deserve to lose because it's too simple a game. I've been hearing of a 100T GPT-[X+1] since X was 3, probably. They might want to avoid GPT-4.5.

@scaling01 I'm not overindexing on anything, I'm telling you that we most certainly won't see a 20T or even 15T OpenAI model until Q4 at best, and you @scaling01 should think more carefully of why they would even waste time on consumer products like 5.2-5.5 instead of a Death Star already.

if compute was everything I would bet all my horses on Google and Meta, but they didn't see the trend and didn't value coding

I'm still long-term bullish on both companies, but as we get closer to RSI I get less and less bullish, because at some point they will not be able to catch up anymore even with a 5x compute advantage

@teortaxesTex everyone is overindexing on GPT-4.5 a shitty 2024 model it's insane

index stronger on GPT-5.2-GPT-5.5 that have been trained with blackwells

@teortaxesTex if they don't have a monster model by end-of-year then I will change my outlook on them

but right now I am very bullish

@scaling01 I discard that and assume comparable sparsity for all models in both lineups, and your 15-20T to practically correspond to "3-4x+ larger than 5.5, 1.5-2x larger than Fable in ways that matter". And I say they won't be there this year

@zephyr_z9 I knew you would mention GPT-4.5; because why else would you make that statement?

if GPT-4.5 is your only data point then you are overindexing

I see GPT-5.3, GPT-5.4 and GPT-5.5 all as experiments for training on GB200s

the next step, another scaled up model, is natural

@scaling01 Yes, they are training a big model for gpt-6 but it ain't 15T-20T

It’s pretty clear that OAI went all in on the ‘consumer’ strategy, which is best suited to a smaller model with different levels of test time compute. However, Anthropic have made it clear that the winning strategy is to go for enterprise + max intelligence.

It’s now clear that all the excess profits come from frontier intelligence, and that the consumer approach is a dead-end economically. If OAI still don’t realise this then yes they do deserve to lose.

@mimrocker well not for them maybe lmao

@zephyr_z9 My Internal Plan is as follows📈

⬇️Details as follows