Jensen Huang closed his GTC keynote this year with a choir of robots singing a campfire song, a Disney snowman walking across the stage under its own power, and a digital version of himself waving goodbye. The crowd at the SAP Center gave him a standing ovation. Around 30,000 people attended in person. That number, by itself, tells you something about how quickly the center of gravity in technology has shifted.
GTC used to be a developer conference for people building on Nvidia GPUs. This year it felt more like the opening of a new industrial era. Jensen spent nearly three hours on stage making a single argument: AI is no longer a product you buy. It is infrastructure you build, like power plants and factories. The chips are the easy part. The hard part is everything around them.
Two transcripts sat on my (virtual) desk after the event. One was a recording of Jensen on the All-In podcast, talking through the Groq acquisition, the inference inflection, and what he actually expects from his engineers in an AI-native company. The other was a Moonshots podcast recap with Peter Diamandis walking through the week’s announcements alongside three sharp people who were paying close attention. Both were more candid than the keynote. Both pointed toward the same set of conclusions.
Here is what I took away, and what I think it means.
The Trillion-Dollar Number Is Not What You Think
When Jensen stood at the podium and said Nvidia could see $1 trillion in orders for its Blackwell and Vera Rubin chip systems through 2027, financial analysts filed it under “aggressive guidance” and moved on. The stock dipped. That reaction misses the point.
The number is not a revenue forecast. It is a demand signal. Jensen was describing a world in which the companies building AI infrastructure, the hyperscalers, the sovereign wealth funds, the enterprise IT buyers, are all lined up trying to purchase something that cannot be built fast enough. The constraint is not demand. It is manufacturing capacity.
On the All-In podcast, Jensen explained it this way: you should not equate the price of the factory with the cost of the tokens it produces. A $50 billion AI factory, built on Vera Rubin, will generate tokens at a lower per-unit cost than a $30 billion factory built on slower chips. The premium is not really a premium. It is an efficiency argument. You are paying more for the machine because the machine prints more cheaply.
“I can prove it that the $50 billion factory will generate for you the lowest cost tokens.” -- Jensen Huang, All-In Podcast, GTC 2026
For anyone trying to understand where capital is flowing in AI right now, this framing matters. The biggest buyers are not making product decisions. They are making infrastructure decisions. The question they are answering is not “which model should we run” but “which factory do we want to own.” That is a different kind of bet, with a much longer time horizon and a very different risk profile.
OpenClaw Changed the Cultural Story
Until a few months ago, AI agents were a concept that lived inside enterprise software demos and research papers. Something that sounded transformative in a conference room but felt abstract when you tried to explain it to your parents, your team, or yourself.
OpenClaw changed that. It is an open-source framework that turns any computer running an Nvidia chip into what Jensen called a personal artificial intelligence computer. It has memory, it can run code, it spawns sub-tasks, it connects to external services. Developers downloaded it by the millions almost immediately after it launched. Jensen called it the fastest-growing open-source project in the history of software, growing faster than Linux did in thirty years.
The Moonshots crew observed something worth sitting with: this is not just a technical milestone. It is a cultural one. Claude Code, Anthropic’s AI coding tool, had already introduced the concept of an AI agent to developers and engineers. But it was enterprise-only. OpenClaw did what ChatGPT did for language models in 2022. It put the interface in front of everyone.
Jensen made an analogy on stage that I have been turning over since: OpenClaw is to agents what Android was to smartphones. A platform that anybody can build on. That does not mean every build will succeed, but it does mean the category is no longer gated by a handful of companies. The ecosystem is open, and ecosystems that open up tend to surprise people with what grows inside them.
Nvidia then announced NemoClaw, an enterprise-ready version of the same framework with governance, security guardrails, and compliance tooling baked in. The move is straightforward: let the open-source version spread, then sell the hardened version to companies that need their agents to be auditable. IBM did something similar with Linux twenty years ago. It worked.
Inference Is the New Battleground
For most of the last decade, the AI industry competed on training. Who could assemble the most compute, the biggest dataset, and the most talented researchers to produce a better model. That race is not over, but it is no longer the only race.
Inference is what happens after a model is trained, when you actually use it. Ask it a question, run it as part of a workflow, have it take an action in the world. As AI moves from chatbots to agents, inference gets dramatically more expensive. Jensen described the shift on the All-In podcast: when you run an agent, it is accessing memory, using tools, spawning sub-agents, reasoning across a long context. The compute needed is roughly 100 times what a simple question-and-answer interaction requires. Going from chat to reasoning to agentic, across two years, compute demand grew by a factor of 10,000.
That is why Nvidia acquired Groq. Not because of the Groq name, but because Groq built a chip specifically optimized for inference. The Groq 3 Language Processing Unit, announced at GTC, is designed to sit alongside the Vera Rubin GPU rack in a data center, handling the inference workload while the GPUs do the heavier lifting. Jensen said roughly 25 percent of a next-generation data center should be allocated to these inference-optimized chips. He called it disaggregated inference: running different parts of the AI pipeline on the hardware best suited for each.
For startups building products on top of AI, this matters because the cost of running your product is about to change significantly. Not necessarily upward. Inference optimization is the area where the most dramatic cost reductions are happening right now. Sam Altman noted during the week that the cost to produce the same output from OpenAI’s first reasoning model to its current one has dropped by roughly 1,000 times in about 16 months. That curve has not flattened.
The Token Budget Is the New Salary
The most quietly radical thing Jensen said at GTC had nothing to do with chips.
He described how Nvidia thinks about AI spending for its own engineers. If you have a software engineer or researcher earning $500,000 a year, he said, and at the end of the year that person has only spent $5,000 in AI tokens, something has gone wrong. He expected engineers to be spending something close to half their salary in compute. Not because tokens are cheap. Because the productivity gain from using them well is enormous.
“If that $500,000 engineer did not consume at least $250,000 worth of tokens I am going to be deeply alarmed.” -- Jensen Huang, All-In Podcast, GTC 2026
The analogy he used was a chip designer who refuses to use CAD software and insists on working with paper and pencil. The tools exist. Not using them is not a choice, it is a failure mode.
This is not a theoretical framing for what might happen someday. It is a description of how one of the most successful technology companies in history is structuring its internal expectations right now. The implication for anyone building or managing a team is pretty direct. If your best people are not deeply integrated with AI tooling, they are operating at a fraction of their potential. And in a competitive market, that gap is going to cost you.
What This Means for Startups
The conventional wisdom coming into 2026 was that AI-native startups face an existential problem: the foundation models are getting cheaper and better so fast that any advantage a startup builds on top of them can be erased in a product release. Why build a specialized AI tool if OpenAI or Anthropic will just add that feature in three months?
Jensen addressed this directly, and his answer was not what the skeptics expected. He said the enterprise software industry is not about to be destroyed by AI agents. It is about to be multiplied by them. The tools that exist today, the databases, the design software, the ERP systems, are the conduit through which work gets represented back to humans. Agents will not replace those tools. They will use them. A hundred agents will be banging on SQL, on Blender, on Photoshop, because those tools are how the output gets reviewed, revised, and trusted.
The implication for startups is that the moat is not in the model. It is in the data, the workflow, and the relationship with the customer. Jensen put it plainly: deep specialization wins. If you understand a domain more thoroughly than anyone else, and you connect your agents to customers in that domain early, the flywheel that builds is very hard to disrupt. The agents get better the more they are used. The relationship compounds.
This tracks with what the best early-stage founders are actually building right now. The most interesting YC companies this batch are not trying to compete with the frontier labs. They are taking a narrow, well-understood problem in a specific industry, wrapping an agentic workflow around it, and capturing data that nobody else has. That is the play.
A few other things worth naming for founders specifically. First, the cost of building is dropping faster than most people account for. The teams that will win are the ones that ship at the pace these new tools enable, not at the pace of the previous generation. Second, if you are not thinking about how your product fits into an agentic world, you are building for a version of the market that is already fading. Third, the infrastructure layer is increasingly commoditized. The advantage lives at the edges, in the proprietary data and the customer trust that only comes from doing the work.
What This Means for Venture Capital
The investment thesis that made sense two or three years ago, find the best foundation model company and concentrate capital there, has become more complicated. The frontier model race is expensive, consolidating, and increasingly dependent on relationships with cloud providers and government bodies. It is not the right place for most early-stage funds to be playing.
What GTC clarified is that there are at least three other places in the stack that matter enormously and are far less picked over. The first is the physical AI layer, where robots, autonomous vehicles, and industrial systems are just now reaching a point of genuine commercial viability. Jensen said the physical AI business is already approaching $10 billion annually and growing exponentially. The companies supplying software, simulation tooling, and specialized models into that layer are still early.
The second is the inference tooling layer. As agents become the dominant deployment pattern, the companies that help enterprises run them efficiently, secure them, govern them, and integrate them with existing systems will be foundational. NemoClaw is Nvidia’s bid for this layer. There are many adjacent problems that Nvidia will not solve, and those problems are investable.
The third is vertical software being rebuilt natively for agents. Not retrofitted with a chatbot. Actually redesigned around the assumption that the primary user is an AI system acting on behalf of a human. The playbook Jensen described, deep domain expertise plus early customer connection plus proprietary data, maps directly onto a category of startups that are just beginning to emerge.
One thing worth watching carefully: Anthropic’s enterprise market share grew from roughly 40 percent to 73 percent of first-time enterprise AI customers over three months this year, according to data shared during the Moonshots GTC recap. OpenAI moved in the opposite direction. That is not an obituary for OpenAI. Both companies will be very large. But it does suggest that the enterprise market is rewarding focus, reliability, and trust over breadth and ambition. For investors, it is a reminder that the company with the most impressive capabilities is not always the one that wins the customer.
The Bigger Picture
The All-In podcast hosts asked Jensen whether he was worried about AI doomerism, the growing segment of public opinion that views AI development as dangerous and worth slowing down. His answer was thoughtful and worth unpacking.
He said something like: warning is good, but scaring is less good. Anthropic has done genuinely important work on AI safety and security, and Nvidia is a large customer and genuine admirer of that work. But he argued that technology leaders now have enough influence that their words shape policy, and that catastrophic predictions without evidence can do real damage. The historical comparison he reached for was nuclear energy. The United States has spent decades shutting down nuclear plants based on fear, while China has been building them at scale. He did not want AI to follow the same pattern.
Reasonable people can disagree about where exactly the line is between appropriate caution and counterproductive fear. That debate is worth having. What struck me is that Jensen is making it at all. Three years ago this conversation happened mostly in academic papers and alignment research blogs. Now it is happening at the biggest hardware conference in the world, in front of 30,000 people, and being broadcast to millions more.
The scale of what is being built is becoming visible to people who do not think about AI professionally. That creates both an opportunity and a responsibility. The opportunity is that more people are going to want to engage with this technology and invest in it and build with it. The responsibility is that the framing we use to describe what is happening, and what the risks actually are, will shape what gets built and what gets stopped.
GTC 2026 was, among other things, a very large, very expensive argument for a particular framing. The era of AI as a curiosity is over. The era of AI as infrastructure has begun. Whether that is cause for optimism or anxiety probably depends on where you are standing and what you are building.
Most of the people in that arena in San Jose were builders. They left with a lot of work to do.
Brian Bell is the Founder and Managing Partner of Team Ignite Ventures, an early-stage venture capital firm focused on AI, enterprise software, and founder-led companies. He hosts the Ignite Podcast and writes Ignite Insights on Substack.