H200 is the previous generation expansion card (PCIe). The current B200 is ~2x in many ways. Note there is no mention of GH200 (integrated GPU-CPU) and all those fancy
NVIDIA servers.
OTOH, DeepSeek squeezed well their impaired bandwidth H800 cluster. I hope they get a lot of H200s.
Maybe, but a clear Republican bailout of AI might wipe them out for several election cycles / foreseeable future. Big tech isn’t popular, AI isn’t popular and bail outs aren’t popular
What evidence do you have that that's going to be the case? I ask because my entire life, I've seen terrible things done by the Republican Party. And regular people get really hurt. For example, the great financial crisis. Yet, a little bit of time passes, and that 30-some-odd percent goes right back to voting for them.
The public voted for Republicans in their highest-ever numbers in 2020 when the party did everything possible to denigrate public health efforts and scientific research at a time when hundreds of thousands were dying of Covid, with no Mexican wall or Obamacare repeal promises met.
There is no scenario where the American vote for a party will fall below 48%, and elections will continue to be decided by how 3-4 states vote.
Okay, let’s see you guys get passed the inference costs disclosure. According to WSJ it is enough to kill the frontier shop business model. It’s one of the biggest things blocking OpenAI
Yes to IPO you have to submit an S-1 form which requires the last 3 years of your full financials and much more. You can’t just IPO without disclosing how your business works and whether it makes or loses money and how much.
Inference costs aren't a problem, selling inference is almost certainly profitable. The problem is that its (probably) not profitable enough to cover the training and other R&D costs.
You did not parse that article properly. It regurgitates only what everyone else keeps saying: when you conflate R&D costs with operating costs, then you can say these companies are 'unprofitable'. I'd propose with a proper GAAP accounting they are profitable right now; by proper I mean that you amortize out the costs of R&D against the useful life of the models as best you can.
I am not aware of any frontier inference disclosures that put margins at less than 60%. Inference is profitable across the industry, full stop.
Historically R&D has been profitable for the frontier labs -- this is obscured because the emphasis on scaling the last five years has meant they just keep 10xing their R&D compute budget. But for each cycle of R&D, the results have returned more in inference margin than they cost in training compute. This is one major reason we keep seeing more spend on R&D - so far it has paid, in the form of helping a number of companies hit > $1bn in annual revenue faster than almost any companies in history have done so.
All that said, be cautious shorting these stocks when they go public.
reply