Eight signals from the AI+ Expo 2026

I spent three days last week at the Walter E. Washington Convention Center for the Special Competitive Studies Project's third annual AI+ Expo, which convened more than twenty thousand attendees and close to two hundred exhibitors. The program swung from fusion reactors to humanoid robots to the question of who controls the next decade of compute.

I went because three days of walking a floor like that is the fastest way to recalibrate your mental model of things to come. Reading panel summaries is not the same as standing in the room and absorbing the information with all your senses.

Here is what I came back with.

1. The federal government has fully woken up to AI

The Department of Energy ran the largest single presence on the floor — an entire section anchored by the new Genesis Mission and staffed by people from a dozen national labs. DARPA, NIST, DCSA, the FBI, the State Department, and the National Science Foundation were all there, all briefing and recruiting. Service chiefs spoke from the main stages.

Five years ago, federal AI was a side conversation at industry events, but this one was practically built around it. The center of gravity for serious AI investment, talent, and procurement is moving inside the Beltway, and the consequences of that for cloud capacity, hardware lead times, and the price of GPU compute will be felt across the rest of the industry whether you sell into government or not.

2. China was the room

The Expo literally opened with a full-day forum titled "The China Challenge: The View from Taiwan," co-hosted with a Taiwanese research institute. There were panels on Chinese semiconductor packaging, the PRC's robotics ecosystem, drone supply chains, lithium batteries, and AI-driven influence operations.

The exhibitor floor reinforced it. A Dutch firm called Datenna had an entire booth dedicated to commercial intelligence on Chinese companies, which is the company's whole business: mapping the ownership, financing, and government ties of PRC tech firms for Western customers. Exiger is another company selling supply-chain transparency tools built around the same problem. The phrase "non-red supply chain" stopped sounding like jargon and started sounding like a market category.

3. Energy was in every conversation, whether the panel said so or not

Every serious AI panel eventually became an energy panel. The Fusion Industry Association held a dedicated session on AI and fusion. Zap Energy was on the demo stage. The Washington Post hosted a briefing called "Energy is AI's Next Frontier." The Institute for Critical Infrastructure Technology ran "From Strain to Strength: Securing the AI-Driven Energy Buildout." Even the panels nominally about agentic workflows, geopolitical analysis, or defense applications kept circling back to whether the grid could supply the watts.

The line I kept hearing, in various forms: AI converts watts to intelligence, and we don't have enough watts. Fusion startups treated their work as part of the AI conversation. A gas turbine vendor bragged about importing and refurbishing decomissioned turbines from Turkey. The AI industry has become the insatiable fledgling devouring the grid, and it happened in a single year.

4. Direct liquid cooling is the new baseline

A year or two ago, liquid cooling was something we thought we'd need in the future. Today, we're discovering it's something we needed yesterday. At this Expo, every datacenter-grade piece of hardware I saw was pre-plumbed with liquid loops as standard equipment. Facility design panels, including one called "The Modular Alternative: Rethinking Data Center Deployment", presupposed liquid as the baseline rather than the exception. The thermal density of modern AI hardware has crossed the line where forced air can't keep up, and the industry has quietly adjusted to that fact.

It is not a flashy observation, but it is the kind of infrastructure shift that will quietly determine which buildings are usable for the next generation of computing and which ones aren't.

5. CUDA is no longer the moat it was

AMD had a serious booth this year. So did Arm. Latent AI was there pitching cross-platform deployment of trained models. The reason all of this matters is simple, and a few years overdue: AI code generation has gotten good enough that targeting AMD's platform no longer requires the deep specialist expertise it required three years ago. The compiler tools have matured, and the inference runtimes are interoperable. And increasingly, the model checkpoints themselves are portable across hardware vendors.

NVIDIA's position is still dominant, but it is now dominant on the merits like performance, ecosystem, and momentum rather than on the software lock-in that quietly made the dominance feel permanent. That is a real change.

6. Private and local AI was where the actual momentum was

This was the strongest signal I picked up all week. EdgeRunner AI, webAI, ConfidentialMind, Latent AI, Cambridge Inference — the sheer number of companies selling open-source, private-cloud, edge-based AI solutions was telling. The Edge AI Foundation ran a tactical edge panel. MBZUAI's president gave a keynote titled "The Strategic Importance of Sovereign AI Models," and webAI hosted a "American-Led Open Models" panel.

The pattern they are all selling is essentially the same: open-weight models running on hardware you own, in environments you control, on data that never leaves your network. After three years of renting intelligence from a handful of frontier-model providers, the industry is quietly assembling a serious alternative. The performance gap is closing every quarter. By the end of 2026, a meaningful subset of what businesses currently pay OpenAI or Anthropic for will run on a workstation in a closet for roughly the price of a couple of years of API spend.

This is the shift I expect to look most consequential a year from now, and almost nobody outside the show floor is paying attention to it yet.

7. The robots were a letdown

Boston Dynamics was there. So were 1X Technologies, Engineered Arts, Path Robotics, and RISE Robotics. A Robot Renegades combat arena that ran all day Saturday. The crowds were large, but the capability on display was less impressive than the marketing videos suggest.

If you wanted to see what useful, deployable robotics actually looks like in 2026, the honest answer is that you needed to be in China, and that is essentially what most of the China-focused panels kept saying. The American humanoid robotics scene is making real progress on actuators, hands, and locomotion, but the gap between the demo loop and the deployable system is still measured in years. The press coverage is running well ahead of the reality.

8. The OpenAI booth was strangely quiet

I kept walking past it. There was almost always more foot traffic going somewhere else, like Rhombus Power's geopolitical platform or Exiger's agentic supply-chain demo. OpenAI had set up a clean, modest stand and was running a demonstration of Codex agents. It just was not where the energy in the room was.

Of course, OpenAI is not in trouble. But it is interesting that at the biggest AI policy event in Washington, the company most people associate with frontier AI was a quiet aside rather than the center of gravity. The energy was at the sovereign-AI booths, the applied-AI demos, and the companies showing what AI does for a specific high-stakes workflow. The market has moved past the model land grab and into the harder, slower question of what to actually do with the models once you have them. The models are becoming commodities, and the real value is in the applications.

---

Closing

If I had to compress three days into one observation: the AI industry's conventional wisdom from 2024 is already aging poorly: NVIDIA forever, the frontier-lab oligopoly, humanoid robots imminent, the cloud as the natural home for AI workloads. Every one of those assumptions is under serious pressure.

The interesting question is no longer which large model wins. It is which physical, geopolitical, and architectural constraints are going to shape the next phase of the industry, and whether the people making the decisions are paying attention to the right ones.