Frontier AI model compute costs remain very high, limiting profitability for AI companies despite high GPU vendor margins.

Elon Musk's proposed space-based data centers would require approximately $100 per B200 GPU-hour to make financial sense, compared with $2-3 spot pricing on Earth.

“Elon is focused on moving the chips to moving the data centers to space and I think on the numbers that he has I think they are looking at like a $100 an hour per GPU hour for a B200 for it to make sense.”

Frontier labs spend 10 to 20 times more on compute than on data.

“My sense is that the split is something like 20 to 1 or 10 to 1.”

Decagon runs 90% of its workflow on open-source models, not frontier models.

“Today 90% of our workflow is on open source.”

Using a general-purpose scaffold with OpenAI's o3/o4 model ('5.5') to disprove the Erdős unit distance conjecture would cost between $1,000 and $100,000.

“It would probably cost I just ballpark like thousand to 10 to $100,000.”

There will be 3 to 6 companies building frontier models, spending somewhere between $200 billion and $2 trillion per year on model development.

“You're going to have, pick a number, three to six companies making a Frontier model, no one knows, no one honest knows, like something between $200 billion and $2 trillion a year on building these models.”

A common emerging workflow is to use a high-quality, expensive model such as Claude Opus or GPT to generate a taste file for a project, then use cheaper models for all subsequent development work on that project.

“A lot of people what they're doing is they're building one project with a really high quality LLM like Opus or GPT They're building a taste file and then you know, super cheap models to continuously build on that more, you know, project with that taste file.”

Within the next year, teams will be able to dial in compute per task, cutting token costs by over 90%.

“I think you'll get to a place probably over the next year where you can like really dial in hey how much compute do I want to spend on this because I have certain like cost considerations and certain latency considerations and get to like the exact optimal amount of cost. and so if you do that like your token costs go down 90% plus.”

Compute can be converted into proprietary training data, meaning a company without partner data is not necessarily blocked from building specialized models.

“There are ways to turn compute into data and get more and we're we're doing those, right?”

Grok 4.5 uses just one-third the tokens of GPT-5.5 or Fable while achieving a similar intelligence score.

“Grok 4.5 uses just one-third the amount of tokens as GPT-5.5 or Fable while achieving a similar score.”

Training a model close to frontier capability is not technically hard today, and many actors are achieving this not just via distillation.

“It's not hard to train a model that is close to frontier capability.”

Replicating Nvidia's Neatron Nano pre-training run costs approximately $100,000, making it affordable for nonprofits to experiment with pre-training interventions.

“We're looking at scaling up pre-training filtering and going to be doing not full but pretty close to replic full replicas of something like Nvidia's Neatron Nano and it only costs maybe like $100,000 per run.”

OpenAI can afford model training by generating modest margins on trillions of dollars of inference revenue, rather than requiring high margins on training itself.

“We will have so much usage of our models that we do not need to be a gigantically high margin business to be able to afford model training. Like so much of our future compute plans will be used to sell inference to customers >> that even if we can enjoy a modest margin on trillions of dollars of revenue, we can go afford to train some.”