The week almost nothing was a model
For two years, the question that organized the artificial intelligence industry was which model reasoned best. The first week of June 2026 buried that question without ceremony. In seven days, the biggest names in technology made their loudest moves and almost none was a model. Nvidia put a one-petaflop AI chip inside the personal computer, Anthropic filed to go public, and Microsoft, Google, Nvidia and ServiceNow spent their flagship conferences competing over the same thing: the runtime an autonomous agent runs inside.
The shift is sharp and it has consequences. If 2024 and 2025 were years of choosing a model, 2026 is the year of choosing a platform: a chip, a runtime, a control plane and, increasingly, a security posture. Those platform decisions are far stickier and multi-year than the choice of a model, which today is almost an interchangeable commodity. Whoever picks a model can swap it next month; whoever picks a platform is locked in for years.
For a country that makes no chips, runs no hyperscalers and generates none of the gigawatts this consumes by itself, the turn matters more than it looks. When the contest was in the model, an API key was enough to take part. When the contest is in the wall sockets — the silicon, the electricity, the cloud capacity — taking part requires infrastructure most economies do not own. The front moved from the brain to the body of AI, and the body is physical, expensive and concentrated in few hands.
Nvidia comes down from the data center to the desk
The move that best symbolizes the change happened in Taipei. At Computex 2026, on 1 June, Nvidia unveiled the RTX Spark, a chip its chief executive, Jensen Huang, described as “a new era of the PC,” holding up a small computer able to run an agent around the clock without relying on the cloud. Huang’s phrase — “no meter anxiety,” about not paying for cloud usage — sums up the ambition: bring powerful AI to the edge, to the device, out of the data center.
The market read the play instantly. Nvidia’s announced entry into the PC chip market sent shares of AMD, Intel and Qualcomm lower, as Wall Street recognized the threat. Analysts saw Nvidia moving beyond the data center toward the so-called edge, as smaller devices become capable of running AI workloads without tapping the cloud. Until now, Nvidia’s entire recent run-up was tied to the data center; with the RTX Spark it is trying to own the desktop layer too.
The financial context amplifies the gesture. Nvidia became the second-most valuable company in the world, overtaking Apple. A company that sells picks and shovels for the AI gold rush is now worth more than almost any maker of the final products. It is the clearest signal of where the power in the chain sits: not with whoever makes the model nor whoever makes the phone, but with whoever makes the silicon both need.
Anthropic rings the bell before anyone else
If the RTX Spark showed where the technical power sits, the week’s other big story showed where the financial power sits. Anthropic, the company behind the Claude models, confidentially filed a draft registration statement (Form S-1) with the U.S. Securities and Exchange Commission on 1 June, clearing the path for an initial public offering. The move puts it ahead of its rivals in the race to Wall Street.
The surrounding figures are dizzying and should be treated as what they are: private-market valuations, not audited accounting facts. The filing came less than a week after Anthropic closed a 65-billion-dollar funding round that lifted its valuation to 965 billion, nearly tripling the 380 billion of just three months earlier. The company is one of the three most anticipated AI IPOs of the year, alongside OpenAI and SpaceX, the latter aiming at a valuation of around two trillion dollars and seeking to raise more than 75 billion.
The figure that gives the real measure of the phenomenon is not any single valuation but a banking warning. Bank of America cautioned that the IPOs of SpaceX, OpenAI and Anthropic could push up the technology sector’s weight, and analysts described 2026 as a cycle that will be “the most consequential since the dot-com era or the most expensive lesson on narrative versus fundamentals that public markets have ever taught.” In other words: no one yet knows whether the numbers reflect real demand or inflated enthusiasm, and the public market will be the exam.
The real battlefield: the runtime
Beneath the chip and IPO headlines, the week’s real war was fought on less visible ground: the runtime of agents. Microsoft, Google, Nvidia and ServiceNow spent their flagship conferences competing over the runtime an autonomous agent runs inside. It is not a technical footnote: it is where it is decided which company gets paid every time an AI agent does a task.
The concrete moves illustrate it. Microsoft unveiled at its Build conference the Project Solara platform, meant for other companies to build agent-powered devices on top of it; it showed as references a desktop display and a security badge usable as assistants, and said firms like Target, Best Buy and Levi Strauss are already exploring how to use it. Microsoft is also pushing its own coding model inside the editor, part of its effort to reduce reliance on OpenAI’s technology.
The change in billing rules is the most concrete clue of where the money is heading. GitHub announced that, as of 1 June 2026, all Copilot plans bill on usage-based AI credits, code review consumes Actions minutes, and user-level budgets and an upgrade path to “Copilot Max” are added. The shift from a fixed per-seat price to per-agent-usage billing means teams automating with agents will see their spend tied to usage patterns — tokens and run minutes — that can change the monthly cost abruptly. The economics of AI stop being a subscription and become a meter.
That change carries a strategic implication beyond the bill. A fixed per-seat cost is predictable and budgeted once a year; a usage cost turns every automation decision into a live financial variable. The more useful an agent is, the more it is used, and the more it is used, the more it costs. The incentive inverts relative to traditional software: it is no longer about squeezing a purchased license, but about watching an open tap. For a small organization, that unpredictability can be the difference between adopting agents or giving up on them, and the decision moves out of its hands into those of the provider that sets the per-token rate.
Energy as the physical limit
The least glamorous component of the new front is also the most decisive: electricity. Data center projects in the United States are running into the hard limits of power grids and into local opposition. AI is not ethereal: every agent that runs, every model that trains, consumes real energy that comes from real grids with finite capacity.
The size of the energy and compute commitments being signed this season is telling. In a regulatory filing, SpaceX reported that Google will pay 920 million dollars a month to rent 110,000 Nvidia GPUs, CPUs and memory, from October 2026 through June 2029. A single compute-rental contract between two giants moves figures that exceed the annual technology budget of entire countries. Compute capacity has become an asset leased the way land or oil is leased.
Europe read the dependence and reacted. France is courting more than 110 billion euros in AI investment, and the European Union is pushing for tech sovereignty. The European message is explicit: whoever does not control part of the physical infrastructure will be at the mercy of whoever does. It is exactly the lesson the shift from the model to the wall sockets leaves for the rest of the world.
The models continue, but as backdrop
It is best not to overstate the burial: models did not vanish from the week, they just stopped being the main story. There were moves, and reading them together confirms the thesis. Google brought its Gemini 3.5 Flash model to general availability; OpenAI confirmed it will retire GPT-4.5 from ChatGPT on 27 June; and Microsoft used its Build conference to embed its own coding model in the editor. They are model announcements, yes, but all subordinate to a platform: the Flash that runs in Google’s cloud, Microsoft’s editor, the ChatGPT that retires one version to push another. The model no longer sells alone; it sells as the front door to an environment.
Meta’s case is the most illustrative of the shift in priorities. Meta repeatedly delayed the launch of the API for its Muse Spark model — unveiled in April as the first model from its Superintelligence Labs and meant to close the gap with rivals — with no new date announced as of 4 June. And on the same day as that delay it did what it did treat as a priority: Meta unveiled the Meta Business Agent and a companion platform, expanding agent capabilities across WhatsApp, Messenger and Instagram so businesses can answer questions, recommend products, book appointments and, in some cases, close sales; it said it would offer it globally with paid subscription tiers in coming months.
The reading is plain: Meta can afford to delay its frontier model, but not its commercial agent layer. The first is a prestige trophy; the second is where the business is. The company that for years defined its value by the muscle of its models now races to place agents in the three messaging channels half of humanity already passes through. The model waits; the agent that monetizes does not.
The return of the silicon makers
While Nvidia grabbed headlines, the rest of the chip industry repositioned on the same physical ground. Intel announced at Computex the availability of its Xeon 6+ processors, aimed at agentic, cloud-native and network-intensive AI workloads, with greater performance density and power efficiency. Intel and Foxconn also announced a collaboration on AI infrastructure, and Intel’s list of vertical alliances added Siemens, Hitachi and others for sector-integrated solutions.
The pattern emerging from these announcements fits the general shift: chip companies no longer compete over who has the fastest processor in the abstract, but over who delivers complete “chip-to-rack” solutions, ready for the data centers that run agents. The week’s narrative, extended from data centers and racks down to chip level, was precisely about integrating silicon into vertical customer solutions. Value moved from the loose component to the integrated system.
That race for integration has a side effect rarely named: it raises the entry barrier even higher. If competing is no longer selling a chip but delivering a full rack with its software, its cooling and its certified energy efficiency, the number of actors able to play shrinks further still. The concentration of power the open model seemed to avoid reappears, intact, one layer down, in the hardware. And hardware is not downloaded: it is built, cooled and plugged in.
The week’s balance
The conclusion the first week of June 2026 leaves is that the model-spectacle phase is over, and the infrastructure phase has begun. Models will keep improving and competing, but they are no longer where power is decided. Whoever spends 2026 mistaking the model race for the main event will spend 2027 wondering why their pilots never reached production.
For countries that make no chips and run no clouds, the reading is not resignation but focus. The fight over silicon and runtime is lost in advance for whoever arrives late and without capital, but the trust layer — the agent that understands local regulation, sensitive data and the processes of a specific institution — stays open and does not commoditize. The region’s challenge is not to compete with Nvidia over the chip, but to avoid being reduced to a perpetual tenant of platforms that bill by the meter. The week that moved the AI war to the wall sockets made clear, above all, where it does not pay to arrive late.