The main competitions US labs are facing, are from one another at this point, not from open weight models. If anything, the competition from new entrants in the west (Meta, xAI, even Google if you think about it) has pushed things even faster, but regardless, I don’t see how OpenAI and Anthropic would have been any slower if there were no open weight models.
That is something they simply cannot do with Chinese entities. Allegedly smuggling GPUs may be feasible, but that'd probably ruffle to many feathers in both the US and Mainland China, especially as we know they will pop a bubble to their own (temporary) detriment if it gets out of control as seen with Evergrande a few years ago.
Beyond all of that, would that juice be worth the squeeze? Bypass massive sanctions, risk your companies future and standing on the global stage just to provide foreign labs with your GPUs illegally and create a comparatively negligible added demand for your accelerators when compared with the (predominantly) closed-model-labs? That seems very high risk, very low reward in my eyes.
Nvidia would rather sell those big interconnected gpu racks than individual GPUs.
> I don't think it's good for Nvidia if everyone has to go through a few labs APIs to get access to AI inference.
Maybe but isn't a business customer better for Nvidia than regular customers? They buy in bulk and commit to multi year purchases. Bonus points for circular deals
Would that not perhaps hurt their product ecosystem? What if people decide they don't want so much of one product, and would actually like to buy some of those cancelled products at some point in the future? I guess they'd have to go to a different company then!
Their P/E ratio is lower than in 2020 (before AI got rolling, and way below their 2023 peak.
If anything - based on current sales alone - NVDA is undervalued.
If (it’s a big if), NVDA get leapfrogged in hardware capability it will be trivial for the next generation of deployments to stop using them entirely (Google and Microsoft are already developing custom inference chips, as are others)
The same story goes back for years and years in consumer hardware too. AMD GPUs often had more raw compute capability, but fell short in actual game performance (or other applications).
Nvidia's moat is, at least partially, the software to make the hardware useful.
The hyperscalers are one story but research/consumer/“Everyman” inference is likely very different.
i would like to, but won't/can't buy their overpriced crap -> they're losing a customer -> they're doing something wrong -> their valuation is just about 0 to me, ballpark estimate