Pierre Ferragu

I'd love this to be true, but the math doesn't work. However large Nvidia is, they may have enough capital to pre-train a frontier-scale model, but they won't have what the labs do: the resources to conduct frontier research. Pre-training is only a fraction of the total cost of being at the frontier. The bulk goes into the research that enables you to train the next generation—the path-finding. This is also why replicating 95% of a frontier model's performance ex-post is cheap: the path has already been found. I can see Nvidia outperforming the best open-weight models. But if they don't keep their models proprietary and sell tokens at a premium, they can't sustainably finance frontier research. It may appear feasible today because they capture profits upfront on capital, but over the long run they won't earn enough to compete with frontier labs. Frontier labs make 70 cents per dollar of revenue after accounting for inference costs; Nvidia will make only 15 cents in gross margin. Unless the frontier fragments, Nvidia lacks the scale to compete there—and neither does any open model. Today Nvidia makes this 15 cent upfront so they look like taking thlion's share of the profit, but it is an illusion. In the long run they earn less than the frontier players can. Nvidia can nevertheless dominate the open-weight space. The dynamics flip completely: they have far more to invest in pre-training open, best-in-class lagging-edge models than any of their clients can. That's where they win. Thoughts?

$NVDA - too large to be true? Remember the similar debate on $AAPL in the early 2010’s. What would it take to get Nvidia to $ 10tn market cap? https://t.co/fT2PfJMFTz

$NVDA - too large to be true? Remember the similar debate on $AAPL in the early 2010’s. What would it take to get Nvidia to $ 10tn market cap? https://t.co/fT2PfJMFTz

Cool segment last week about Nvidia @ $ 1tn run-rate end of next year, etc. https://t.co/gZnqe3OPgJ

Who is going to the Tesla even in Austin on Thursday? When and where do we get together? what for?

You’d measure demand elasticity with volume increase / price decrease. You are happy with elasticity above 1, and I do not remember elasticity of more than 2. Here we are talking a coefficient of 14/2 = 7 … unheard of. The token is the new transistor. https://t.co/3v5In4c4cu

Are we in the singularity, or in a profit bubble? That is the question. https://t.co/3xSIYHMm3I

Tesla could have written 5,6,7,8,9,10. If I may - what is the point you are trying to make here?

Love it. Waymo getting defensive. Nothing makes me more bullish about Tesla. Arguments they bring forward are poor. Check👇 for details. My read: Waymo built a driving gas plant that AI at scale makes irrelevant, and now falls into incumbent rhetoric. Innovator dilemma 101. https://t.co/fSD5cR3vcR

Just saw Tesla’s Cybercab testing in Manhattan. Operator at the wheel, gold two-seater slicing through Midtown. Felt like I time-skipped.