GTC 2026 Day 1 – Requiring Inference

A keynote that needs human inference to fully understand.

• The press and the market are somewhat nonplussed following Jensen’s keynote, as they did not get the big announcements that tell them where to focus their attention, but under the hood, there is a claim that, if true, will cement Nvidia’s dominance for years to come.
• Jensen kicked off GTC 2026 with a keynote where the main story was its excellence at inference and the claim that no one can fix the business model of compute like Nvidia can, even at 75% gross margins.
• The main takeaways were:
  • First $1tn: where Jensen stated that he could see a runway to the company realising $1tn in cumulative revenues for 2026 and 2027.
  • This is the kind of explicit commentary that the market feeds on, and so the shares jumped 4%, but then they fell back as Wall Street ran the numbers and realised that this was not such a big upgrade after all.
  • Based on where consensus estimates for 2026 and 2027 are, this represents roughly a 10% upgrade to estimates for 2027, which is why they quickly fell back once this was realised.
  • Jensen also said that demand could be higher than this but it is quite possible that he would be unable to fulfil it even if it did come to pass.
  • Hence, what Jensen is signalling is a continuation of the supply constraint, meaning that, like 2026, 2027’s revenues will be dictated by how much capacity Nvidia can reserve at TSMC.
  • All the signs at the moment are that spending on infrastructure is far from over, but a lot of this spending is being financed with debt, which increases risks should there be an exogenous shock that delays demand.
  • Hence, while demand looks like it will continue to grow, the risks to that demand are rising, making each successive year more risky than the one that preceded it.
  • Second, the business model of compute: which I think is the real story of GTC.
  • As AI rolls out, capacity demand is moving rapidly towards inference, and I would not be surprised to see it end up at 90% of the total market for AI infrastructure.
  • Nvidia’s lock-in is not nearly as strong in inference, which is why it bought Groq, why it launched Dynamo and why it is making the case at GTC 2026 that it has the best and most economical offering for inference.
  • I have long argued that the business case of compute is broken, as a 1GW $40bn data centre earning $10bn a year is unable to generate a decent return within its economic life.
  • This means that anyone who can fix this problem is likely to be on the receiving end of a lot of demand.
  • This is precisely what Nvidia is claiming it can do with Vera Rubin and some Groq bells and whistles.
  • The idea is that not all tokens are created equal and that in some use cases, buyers will be willing to pay a significant premium for tokens that are delivered quickly.
  • This is where Groq comes in, as Nvidia has integrated it into the Vera Rubin system, where it will be deployed specifically to deliver these tokens.
  • The combination of higher value tokens, more tokens per GW from Vera Rubin and much higher speed from Groq is what underpins Nvidia’s claim that it can increase the revenue per GW by 5x.
  • This would amount to $50bn per GW, and if it proves to be accurate, it will completely change the economics of compute (stay tuned).
  • Given that Nvidia is clearly leading this sector, it will also cement its dominance in inference as well as training.
  • However, Blackwell was supposed to deliver similar economic benefits, but intense competition meant that the price of tokens fell, resulting in revenue remaining static at $10bn per GW.
  • If past is prologue, then none of this will come to pass, meaning that the business model of compute will not improve with revenues remaining stuck at $10bn per GW.
  •  This in turn, means that the likelihood of financial pressure causing a reset in demand will continue to increase.
  • Unfortunately, Nvidia can do very little about this other than cut its prices, which it is clearly not going to do.
  • Hence, all it can do is sit on the sidelines and hope that Vera Rubin is the generation that delivers the much-needed improvement to the business model of compute.
• The net result is that Nvidia remains in a very strong position as Vera Rubin is at least one generation ahead of everyone else.
• This means that it can improve the economics of computing for AI data centres more than anyone else, even though it is still earning 75% gross margins.
• The real winners here are the memory companies, where demand for their products now looks set to overwhelm supply until the end of 2027 or longer.
• Samsung fared even better, being named as the maker of the Groq LP30 3^rd generation chip that Nvidia will be using, which is a big endorsement for the fledgling foundry business.
• The fact that the market didn’t understand what Nvidia is trying to say means that it will continue to worry about longer-term growth, and the shares may continue to drift despite rising earnings.
• This could represent an opportunity to invest in the best AI company at an increasingly attractive valuation.
• In the meantime, I hold Samsung and Qualcomm for AI, Ouster for robotics and nuclear power for the relentless demand for electricity that the AI boom is producing.