Last month I argued that the real AI story was not the frontier-model leaderboard but the services layer wiring those models into delivery. The first days of June 2026 made the point even more bluntly. In a single week, the biggest names in technology made their loudest moves — and almost none of them was a model. Anthropic filed to go public, Nvidia put a one-petaflop AI chip inside the personal computer, and Microsoft, Google, Nvidia and ServiceNow spent their flagship conferences competing over the same thing: the runtime that an autonomous agent runs inside. The contest has visibly moved beyond the model.

That matters for anyone running AI inside a large enterprise, because it changes what you are actually buying. If 2024 and 2025 were about choosing a model, 2026 is about choosing a platform — a chip, a runtime, a control plane, and increasingly a security posture — and those are far stickier, multi-year decisions. Here is how I read the week, and what I am telling my own teams.

One week, three tells — and not one was a model launch

Start with Wall Street. On 1 June, Anthropic confidentially filed a draft S-1 with the SEC, getting out ahead of OpenAI in the race to the public markets. It comes days after the company closed a round at a roughly 965-billion-dollar valuation, overtaking OpenAI, on a reported 47-billion-dollar revenue run-rate. Whatever you think of the number, the signal is unambiguous: the AI boom is now mature enough — and cash-hungry enough — to test ordinary public investors, not just venture balance sheets.

Next, silicon. At Computex, Nvidia pushed off the data-center floor and into the PC with the RTX Spark superchip, and confirmed that its data-center Vera CPU is in full production, with OpenAI, Anthropic and SpaceX among the first customers. And third, the platform vendors. Microsoft Build, Nvidia's GTC Taipei and ServiceNow Knowledge all landed in the same window, and they converged on a single message that I will come back to: the agent runtime is the product now. Three tells in one week, none of them a chatbot.

The model has quietly become the commodity input

The week's actual model news was almost mundane by comparison. Google's Gemini 3.5 Flash reached general availability; OpenAI confirmed it is retiring GPT-4.5 from ChatGPT on 27 June; and Microsoft used Build to ship its own coding model into the editor. The most revealing framing came from the market itself: reporting now describes Microsoft and Google explicitly chasing Anthropic and OpenAI in coding models. Read that carefully — the fight is not over which model is cleverest in the abstract. It is over which vendor owns the surface where developers actually work. The model is the ammunition; the distribution is the war.

For an enterprise buyer, the practical conclusion is liberating: treat the model as a swappable, commoditising input and design for portability. The supplier that is ahead this quarter will not necessarily be ahead next quarter, and increasingly it does not matter, because the value you capture is determined by the layers you wrap around the model — not the model itself.

The agent runtime became the product

This is the development I would not let any enterprise AI leader miss. Across three conferences in one week, the largest vendors stopped selling models and started selling the governed environment an agent lives in. At Build, Microsoft pushed Agent 365 to general availability and, per its keynote coverage, positioned Azure AI Foundry as the orchestration hub for fleets of agents and Windows itself as an agent host. Nvidia, meanwhile, released an open-source Agent Toolkit built around an orchestration layer and a secure runtime, and extended its ServiceNow partnership to deliver governed autonomous agents with auditability wired into every action. Google had already put Managed Agents into the Gemini API, spinning up agents inside isolated sandboxes.

Strip away the brand names and these are the same product: parallelism, isolation, identity, audit, spend limits, policy. In other words, the control plane is now the thing being sold — the point I made last month, except it is no longer a thesis; it is a price list from the four biggest vendors in the industry. The strategic catch is lock-in. Standardising your agents on a runtime is not like calling a model API; it is closer to choosing a cloud. The question on the table for every CIO is no longer "which model," it is "whose runtime do I build my agent estate on, and what does it cost me to leave."

AI moved onto the device — and that rewrites the data conversation

The Nvidia RTX Spark deserves its own paragraph, because it is easy to file under "gaming hardware" and miss the enterprise point. Reporting describes a part that pairs an Arm CPU with a Blackwell GPU and 128GB of unified memory to deliver roughly a petaflop of local AI, aimed at turning Windows into an "agentic" operating system, and shipping this autumn in laptops from Microsoft, Dell, HP, Asus, Lenovo and MSI.

When that much inference can run on the endpoint, the data-residency calculus shifts. A regulated bank, hospital or manufacturer can keep sensitive prompts and documents on the device instead of round-tripping every token to a frontier API. For India and other data-sovereignty-conscious markets, on-device agentic compute is not a gimmick — it is a genuine unlock for workloads that compliance teams have so far refused to send to the cloud. I expect "where does the inference physically happen" to become a board-level question within a year.

Cyber-capable models are now a procurement option, not a research curiosity

The most sobering item of the week: Anthropic expanded access to its cyber-capable "Mythos" model to roughly 150 more organisations across more than 15 countries through Project Glasswing, reaching into power, water, healthcare, communications and hardware — sectors that were under-represented at launch. The same reporting notes the model has identified large numbers of previously unknown software vulnerabilities in testing.

This is dual-use in the truest sense. Defenders genuinely need it, and critical-infrastructure operators getting early access is, on balance, a good thing. But a model that can find and chain zero-days is now something an adversary can rent too, which resets the offence-defence balance in real time. The enterprise takeaway is the one I keep repeating, now with more urgency: an agent acting in your name carries your liability, and your security posture has to assume the other side has the same tools you do. The control plane — identity, sandboxing, audit, kill-switches — is the product, not the compliance afterthought.

My take

I am reading this week as confirmation that the durable value in AI has migrated decisively away from the model and toward everything around it. So I am giving my own organisation four instructions. First, choose the agent runtime deliberately and treat it as a multi-year lock-in — evaluate it on governance, portability and exit cost, not demos. Second, keep the model layer loosely coupled, because it is commoditising and you want the freedom to switch suppliers without re-platforming. Third, plan for on-device and edge inference now, because it is both a cost lever and a data-sovereignty unlock, especially in regulated Indian sectors. Fourth, upgrade the security posture to a world where cyber-capable models are commercially available to everyone, attacker and defender alike.

There is an India-shaped opportunity threaded through all four. The model is global and commoditised; the runtime wars are being fought by hyperscalers; but the governed, domain-wrapped agent layer that a bank or insurer can actually trust still has to be built by people who understand how those institutions run. That is the layer where Indian talent and Indian compliance know-how compound — and cheaper edge silicon only widens the opening. The window I described last month has not closed; this week it widened. The firms that spend 2026 mistaking the model race for the main event will spend 2027 wondering why their pilots never shipped. The ones that invest in the chip, the runtime and the trust layer will own the part of AI that does not commoditise. I know which side I intend my organisation to be on.