Hyperscalers are buying all the chips to then rent them to us later

This may sound alarming, but it appears as simple math to me

1. Chip price increase applies to everyone, including hyperscalers 2. I assume most of consumers refuse to pay for such ridiculous prices. 3. Hyperscalers pay the price because they expect return on investment 2. We, consumers, are actually the end-users of the chips they bought

It appears then that Hyperscalers are outbidding us at buying chips, then will make us pay back the ridiculous price they outbid us with through renting them back to us.

The way out seems simple: don't pay for AI and wait till one can afford chips for local (or federated) AI.

But there's the whole b2b market that I don't know what to think about

What are your thoughts?

6 points

・

adelks

・

a day ago

3 comments

stogot ・ a day ago

Since when do you buy a direct A15? Or as Tranium? Or Maia? Or cobalt? Or graviton? Or a TPU? The hyperscalers are selling to enterprises not consumers. The apple chips are sold to consumers but they sell the whole hardware not just the chip

  • stop50 ・ a day ago

    in the Price for an nvidia card is not only the price for the card. Its also the limitations of producing it. The Hyperscalers have much deeper pockets than the consumers, so they compete via the price to buy as much as possible from the limited batches nvidia/amd/... sell. As a result, manufacturers are changing the number of different chip types, so that instead of 4 graphics cards and 1 CUDA card, there is now 1 graphic card and 3 CUDA cards.

    • adelks ・ a day ago

      My current understanding is that regular consumer GPUs are very much capable if they had enough VRAM, and buying like 2 isn't out of this world.

      A 64GB consumer GPU not existing has no technical nor financial (buyer side) reason to it, other than market segregation. If a 24GB one costs 1000 bucks, people are ready to pay quadruple that for a 32GB GPU. And VRAM is the main bottleneck for running larger LLMs, not performance. And some resort to offloading to RAM etc...

      Segregation between pro and consumer hardware are in a big part artificial I think, for fatter margins.

      Otherwise, yeah Hyperscalers have deep pockets, but they can only have it by getting back money from their users/customers, and we're gonna give them that by using their products and paying for it (indirectly or directly).

      I don't think it's a good thing that only select companies gatekeep AI and it feels to me like it's going that way with chip prices.