Three Marches in San Jose, and the line that beat me

Trip 15 to 19 March 2026 · Written up 19 March 2026 · Last expanded 29 September 2026

The first time I came to San Jose for GTC, in March 2024, I spent an evening listening to Christian Szegedy talk about the formalization of mathematical reasoning inside language models. It intersects with research I am contributing to at GT, so I was always going to be there, but I did not expect how far past the topic he would range: whether open software is necessarily synonymous with good AI (coincident, I think, but not atomically linked), and how close reinforcement learning and synthetic data generation really are to the same thing.

I have used objects of his for years, batch and inception, and had recently applied an inception net in a way fairly novel to its own inception (no pun intended). Riding the wake of that wave during my journeyman phase with deep nets and then meeting its crest in person was oddly satisfying.

The rest of the week held that pitch. Llion Jones announced forthcoming results on model fusion learned from the Hugging Face corpus with evolutionary algorithms, Jensen sat down with the authors of Attention Is All You Need, Michael Dell turned up with a surprise hardware announcement, and I got close enough to Jensen that his bodyguards shooed us away.

A singular moment in tech history

Whoever was in San Jose that week probably witnessed a singular moment in tech history. The guest list ran from the original Transformers authors through a good share of the luminaries in AI reasoning, plus a lot of worker bees like me, contributing to that field or vying to, and it felt academic, epochal even.

GTC25 was not that. There were twenty five thousand plus attendees, and Nvidia had annexed a whole block of San Jose, taking the tech museum for impromptu meetups and the park as a food space, which was an improvement on last year, though I still averaged nineteen to twenty thousand steps a day. It is the only tech conference that feels like a college ball game, and Jensen was at the mobile Denny’s before most of us were awake. Early bird gets the worm!

The substance moved too. Bill Dally interviewed Yann LeCun, who challenged the salience of LLMs in his gentle adversarial way, and Jensen argued that scaling laws are intact and accelerating (against Sutskever and others, though not Amodei) and showed one exaflop inside a single liquid cooled rack. Mostly, though, the week was about making deals and getting deals, which is excellent for an MVP and much less so for a researcher, and most of my sessions left no room for questions (Nvidia should hire some Georgia Tech industrial engineers to fix the overflow rooms).

The first Quantum Day

What did feel new was Quantum Day. Jensen inaugurated it by inviting the executives of the new cohort of quantum firms (every major modality and, tellingly, no IBM or Microsoft) to explain why he was wrong. It was tongue in cheek, and it worked as a mea culpa, though nothing fundamentally new came of it.

His real argument was a rebrand: call them quantum instruments instead of quantum computers, since nothing around the QPU resembles a computer in the Von Neumann sense. The DWave and IonQ chiefs pushed back, and nobody said from the stage that while it is a better mental model for hybridization, it would also deflate capital inflows into the firms sitting on it. The second conversation was far better, mostly because nobody on it ran a public company and they stayed on error correction instead of modalities, and that is where John Levy put us at the 1948 moment of classical computing.

What Jensen did not lead with

By GTC26 the tell was in the running order. Jensen did not open with robotics, the Physical AI story that justifies the forward earnings multiple; he opened with tapping structured data through open source LLMs, in partnership with IBM, my old company (glad to see it). That is consolidation, and it matches what I see advising mature enterprises, where the first problem is never the model but the structured data scattered across pools, lakes and silos. Unlock that and AI investment accelerates past the Mag 7.

Then the chip crunch beat me. I wanted another RTX 5090 for my cluster, honestly half the value of showing up in person, but the queue had started at one or two in the morning, and those poor saps sat there until seven for the privilege of paying 2100 USD (one hundred and fifty a day). I did not join it.

I once put all of this on Bruce Tuckman’s stages, forming as everything before GTC24 and storming as the grappling that followed, and consolidation looks a great deal like norming. Jensen said in 2025 that he hoped San Jose would expand along with the economic activity, and after three Marches on that block I suspect he gets it. If you were there for all three, I would be curious which year you think we will end up citing.

Who I met

• Christian Szegedy, AI researcher. Talked at GTC24 on formalizing mathematical reasoning in language models, and on whether open software and good AI are the same thing (coincident, not atomically linked). I have used his batch and inception work for years, so this one mattered to me.
• Llion Jones, Sakana AI. Previewed results on programmatic model fusion over the Hugging Face corpus using evolutionary algorithms. Many have tried it, Microsoft included, and the network and systems ablation is what makes it expensive.
• Jensen Huang, NVIDIA. Three keynotes, three arguments: scaling laws intact and accelerating, quantum computers as instruments, and in 2026 structured data ahead of robotics.
• Yann LeCun, Meta FAIR. Named physical AI, persistent memory, reasoning and planning as the four things to advance, and made Meta’s case that open source is the Web 2.0 platform strategy applied to foundation models.
• Bill Dally, NVIDIA. Interviewed LeCun at GTC25 on the longevity of Transformer based LLMs. The third Dally talk I sat through in a year.
• Matt Kinsella, Infleqtion. Put the threshold inflection at roughly one hundred error corrected logical qubits, against a room that kept saying a million.
• John Levy. Argued we are at the 1948 moment of classical computing with respect to quantum, so the architecture work should start now, and put latency to decoherence at 500 to 800 nanoseconds.
• Rajeeb Hazra. Proposed custom foundation models trained on QC output to learn the nuances of system perturbation, in place of synthetic data generation.
• Will Oliver. Gave the clearest quantum engineering primer I have seen, at GTC26, and drew a bigger house than the two deeper technical talks before him.

Photographs

Links

Llion Jones on evolutionary model fusion over the Hugging Face corpus, the GTC25 keynote on scaling laws and the single rack exaflop, the first Quantum Day panels, and Jensen’s GTC26 keynote on structured data. To be added.

Open entry
This page is not finished, and is not meant to be. I add photographs, people and links as I dig them out, so it keeps growing after publication. Why the whole site works this way is on the colophon.