Comment by chatmasta
20 hours ago
Training cost is actually not that high — it’s fixed and amortizable across the lifetime of the model. Inference is expensive, and open weights don’t solve that problem — in fact, they might even encourage it, since a high cost of entry means consumers will pay for inference directly from the labs anyway.
Unfortunately it seems likely the winner will be the cloud providers. If anyone can run inference on open models, then profit will flow to the vendors who can afford the capital to run them. That’s the CSPs.
(It’s basically the same business model as pharmaceutical R&D, but the major difference is that nobody has even talked about patenting the models like a pharmaceutical company patents each new drug. I’m surprised about that, tbh — why give all the leverage to the cloud platforms? They aren’t training frontier models…)
Winners are hardware companies, GPUs,XPUs, HBM, memory, connectivity. Even CSPs are just compute renters, they charge a margin to make sure their hardware purchases can be made back. But given there are more and more AI CSPs, traditional, and neocloud, pricing competition is inevitable, and given the huge expense of hardware, CSPs are squeeze by the hardware companies and users seeking to lower their own costs.
Inevitably the CSPs will make their own hardware, especially as we start to see specialized chips for specific models or generic inference. This is already happening with Google and TPUs.
It’s easier for the CSPs to move into hardware than it is for Nvidia to move into cloud hosting.
Although as a middle ground I’ve been quite happy with Nvidia Brev for on-demand GPU instances from a select marketplace of CSP offerings. It’s a well kept secret IMO — great product (from an acquisition iirc).
CSPs making own hardware still needs hardware companies, they reduce the Nvidia tax but still need the likes of TSMC, Broadcom, micron/sk hynix, Marvell, the truth semi-companies. CSPs will not have the patents, IPs and talent to replace any of them.
Also, not sure how well CSPs inference stack is compared with vllm + nvidia. A lot of open weight models uses MoE, making the inference stack more complex.
1 reply →