[mNovak]

The wild thing to me, is that they're serving $47B run rate worth of requests on maybe 2-3 GW of compute currently [1], of which only a fraction goes to inference, vs R&D and training. Obviously there have been complaints on token limits and such so they're stretched a bit thin, but nonetheless.

Hard to imagine what a world with 100GW of compute looks like.

[1] https://epochai.substack.com/p/frontier-labs-dont-use-most-a...

^^ This quotes 1.4GW at the end of 2025. Add 0.3GW at Colossus 1, and some initial fraction of 1GW Trainium2 from [2]

[2] https://www.anthropic.com/news/anthropic-amazon-compute

[jillesvangurp]

It gets better; most of their incoming requests don't actually require a frontier model to handle. There's a huge potential for future optimization in this space. Anthropic, OpenAI, Google and a few other companies are going to be well positioned to scale in the few years. A 65$ billion round to finance operations over the next few years isn't that controversial if you look at the growth and profit potential.

I think token counts and GW are a gross over simplification here. Not all tokens are the same in the amount of GPU time they consume or the size of the GPUs they require or the amount of energy they consume. There's a huge optimization potential here once these companies get serious about consolidating the business they have and executing much more efficiently. Given enough time, these companies can heavily optimize their operations. Short term growth and not slamming the brakes on that is their primary concern.

[0xDEAFBEAD]

Where's the moat though? What prevents a race to the bottom with competing AI providers, everyone trying to undercut one another?

[epolanski]

I'm also thinking the same.

I have been trying Claude Code with DeepSeek 4 apis, and the experience is barely different. In fact the margin of error is so small that harness and prompting account for the most impact in output quality.

But, here's the catch: I spend barely more than a handful of dollars per day of regular usage. In fact DS4 via api is cheaper than Claude 100$ subscription.

I really think that very soon many will start realizing that the alternatives are extremely close in performance but dramatically different in pricing.

[port11]

Claude includes or at least promises ZDR in some situations, whereas DeepSeek is explicitly using output to train models. The subsidising might be done with your data.