Jensen Huang's GTC keynotes have become the closest thing our industry has to a state-of-the-union. Not because NVIDIA is the only company that matters — it isn't — but because Jensen is one of the very few operators at that scale who will say, on stage, what is actually true about compute economics and where the puck is going in the next 24 months. Reuters and The Information had decent coverage of the headlines, but the headlines are the part you can get from a press release. I care about the subtext.
I watched this one twice. Once live. Once with the mute on, pausing every 20 seconds to read the slides. The second viewing is where the interesting stuff showed up. Here are the three signals founders should be pricing into their 2026-2027 plans — and the one thing almost everyone missed.
1. Compute per unit of intelligence is falling faster than your plan assumes
The headline numbers Jensen flashed on cost-per-token and cost-per-inference at the new rack scale are roughly 30x better than 18 months ago. If you built a financial model for your AI product in 2024 or early 2025, your gross margin assumptions are already obsolete — in your favor, but in a way that creates a different problem.
Here's the problem. Your competitors' gross margins are also getting 30x better. Any product whose moat was "we can run this at better unit economics than anyone else" no longer has a moat. The moat has to come from somewhere else — data, distribution, workflow integration, relationships. Something that doesn't live on the GPU.
Founders who spent 2024 bragging about their fine-tuning cost structure should be having a completely different conversation in the 2026 boardroom. If your deck still opens with "we're 4x more efficient than the frontier labs," throw it out. That slide just became a liability.
2. The robotics convergence is no longer speculative
The segment that got the least press coverage is the one that matters most for people building companies. The general-purpose robotics stack. Jensen showed a credible path from "humanoid demo" to "humanoid in a warehouse doing useful work" on a two-to-four year timeline, not a ten-year timeline. I've watched the same kind of pivot before. I was around when the web went from academic novelty to commercial infrastructure in two years flat. The shape of this curve rhymes.
Almost every founder I know is building in software. Not wrong — software is still the best risk-adjusted bet in the near term. But the founders who should be paying the most attention to the robotics side of the keynote are the ones in logistics, manufacturing, healthcare delivery, physical-world services. If your model assumes human labor at today's cost structure for the next five years, you need a sensitivity analysis you probably haven't run. I'm running it on the marketplace's warehouse assumptions this week.
The question isn't whether humanoid robotics will change physical labor economics. It's whether it happens in 2028 or 2031. Your strategy should work in both.
3. The model-scale plateau nobody wants to discuss
Here's the footnote that got no coverage. Buried in the middle of the keynote, on a slide that flashed up for maybe 12 seconds, was a chart showing frontier model training compute growth over the last 18 months — and it is visibly flattening compared to the 2022-2024 exponential.
Translation. The biggest labs are getting more efficient, not just bigger. That shift is enormous for founders. The assumption "the biggest labs will keep pulling further ahead because they have more compute" is no longer automatically true. Smaller labs, specialized models, and domain-tuned systems have a credible window in 2026-2027 they did not have in 2023.
If you're building an AI product that competes with the frontier labs on general intelligence, that's still a terrible bet. If you're building on a specific domain with a specific dataset and specific workflow integration — the math is quietly getting better for you, not worse. This is the single most bullish signal in the keynote for the founders who have been told for two years that "you can't compete with OpenAI or Anthropic." You can. Just not on their turf.
The thing everyone missed
The single most important thing in Jensen's keynote was not on a slide. It was in the language he used about "agentic" compute. He kept talking about workloads that run persistently — hours, days, weeks — instead of single-shot inference. Every serious cloud AI pricing model is built around short inference calls. Nobody has figured out how to price persistent agents yet.
If you're building an agent product right now, your biggest near-term risk is not model quality. It's that the pricing model for agent compute is about to get restructured, possibly multiple times, in ways that will change your unit economics by 5x or 10x in either direction. You should be stress-testing your financials at both extremes. If your financials only work in the current pricing regime, you don't have a business. You have a bet.
What I'm doing with this information
Three concrete decisions I'm making across my companies in the next 30 days.
- Stop writing cost-of-compute into any 2027 marketing page. The number will be wrong by the time the page is indexed. Any moat you build on "cheap compute" is evaporating in real time.
- Start pressure-testing the agent-first versions of our roadmap. Persistent, long-running agents are moving from "prototype" to "product" faster than I would have bet six months ago. The pricing risk is real, but the opportunity is larger.
- Revisit the robotics exposure of the marketplace's warehouse and fulfillment assumptions. Not because robots arrive tomorrow. Because the cost curves I was planning against are the ones that need a second look, and second looks take weeks.
GTC keynotes are marketing events. They're also the clearest window we have into where the infrastructure layer is actually going. Founders who only read the headlines will be surprised twice — once when the cost structure moves, and once when the competition that was supposed to be impossible shows up anyway.
So here's the challenge. Go watch the keynote with the mute on. Pause every 20 seconds. Screenshot every slide that contains a number. Then ask yourself which of your current assumptions is now wrong. If the answer is "none," you didn't watch it carefully enough. Watch it again. The next 24 months are going to separate the founders who read the footnotes from the ones who only read the tweets.
Related: my tour of how I actually use AI inside the operator's cockpit.