March was the month AI stopped being just models

The three frontier model releases of March, GPT-5.4 on the 5th, Gemini 3.1 Ultra on the 20th and Grok 4.20 on the 22nd, were the most visible story of the month. The more consequential one happened around them. March 2026 was the month the AI industry stopped being about models and became, at once, about infrastructure, policy and law, and the three events that defined it did not come from the labs.

The infrastructure story came from NVIDIA. GTC 2026, the company’s annual developer conference, was not primarily a hardware event; it was a platform announcement. NVIDIA unveiled the Agent Toolkit, an open-source runtime for enterprise agents, and a governance layer, alongside partnerships with seventeen major enterprise software companies, including Salesforce, ServiceNow, SAP and Adobe. The Model Context Protocol, the open standard for connecting AI to enterprise software, crossed 97 million installs around the same time. The significance is that a new layer of the industry is forming: not models, but the operating systems that orchestrate agents inside companies, and NVIDIA, whose hardware was already the industry’s floor, has moved to own that layer too.

The policy story came from Washington. On 20 March, the White House published a national AI legislative framework covering six objectives: protecting children and empowering parents, safeguarding communities and energy security, respecting intellectual property, preventing censorship, enabling innovation and building an AI-ready workforce. The document is notable less for any single proposal than for its premise, that the United States should have a coherent federal AI law rather than a patchwork of state rules, and for its timing, arriving just as the frontier releases made the technology’s commercial shape impossible to ignore.

The regulatory story came from Brussels, and it is the one with a date attached. The EU Council agreed on 13 March to consider delaying the high-risk AI obligations of the AI Act from August 2026 to December 2027, after it emerged that only eight of twenty-seven member states had designated the contact points required for enforcement. The delay, if confirmed, would push the EU’s flagship regulation beyond the electoral and legislative calendars that produced it, and it exposes a familiar gap: the law was written to apply uniformly, but the machinery to apply it was never built at the same speed. The United States, meanwhile, passed the AI Accountability Act and NIST released version 1.1 of its AI risk management framework, which means the two largest AI markets are now moving in opposite directions, one legislating at national level, the other postponing what it already legislated.

Put together, the month describes an industry entering adulthood. The model releases established that capability has become a commodity that can be scheduled into a three-week window. The infrastructure announcements established that the money is now in the layer that deploys agents, not the layer that trains them. The policy moves established that governments have noticed, and that their responses are diverging. An industry that was a research field in 2022 and a gold rush in 2024 is, in 2026, a regulated, capital-intensive sector with an operating layer, a legal framework in Washington, a postponed one in Brussels and seventeen enterprise platforms agreeing on shared infrastructure.

The convergence is worth stating plainly: the model race, the platform war and the regulatory scramble are the same contest. Whoever controls the agent runtime controls where models run; whoever defines the rules controls what they are allowed to do; and the labs, having spent the decade making models, now discover that the real struggle is over who owns the rails they run on. March did not resolve any of it, but it made the map visible, and for an industry that likes to think of itself as moving faster than its institutions, that is progress of a kind.