From the Leaderboard to the Negotiating Table

The race stopped being about who has the best model. It's now about who writes the rules and who prices the risk.

For three years the AI story was a leaderboard — whose model topped which benchmark this month. This week made it obvious the story has moved. The most consequential things happening in AI right now are happening at a negotiating table, a regulator's desk, and a stock exchange. The model race is still on; it's just no longer where the leverage is.

Three rivals, one table

The CEOs of OpenAI, Google DeepMind, and Anthropic are all heading to the G7 summit in France this week — the first time Sam Altman, Demis Hassabis, and Dario Amodei share the same summit. OpenAI says it expects a package of voluntary commitments, with youth safety and frontier cyber/bio risk near the top of the agenda.

Why it matters: when rivals this fierce sit at one table with governments, the terms they agree to become the floor everyone else builds on. Voluntary commitments have a way of hardening into procurement requirements and, eventually, law. If you deploy AI in a regulated space, read the language coming out of this summit closely — it's a preview of the compliance conversation you'll be having in twelve months.

Washington put a clock on it

The backdrop is a U.S. executive order, "Promoting Advanced Artificial Intelligence Innovation and Security," now moving from announcement to implementation. It directs agencies on aggressive timelines — first deliverables due July 2 — and stands up a voluntary framework where developers can give the government early access to "covered frontier models" before wider release. Notably, it stops short of mandatory licensing.

The takeaway for operators: the pre-release-review norm is forming in real time, and "voluntary" frameworks set expectations fast. The governance discipline I keep harping on — knowing what your systems do, where they escalate, and how you'd prove it — is becoming a market expectation, not just good practice.

Anthropic asked for the brakes

In an unusual move, Anthropic publicly argued for a globally coordinated slowdown on frontier development as AI begins to accelerate its own progress — noting that by May 2026, most of the code landing in its own repository was written by Claude, not humans. Coming from a lab also racing to an IPO, it's a striking tension — and it reframes the safety conversation right as the G7 convenes.

You don't have to agree with the proposal to take the signal: the frontier labs themselves are telling you capability is outrunning control. For anyone building on these models, that's an argument for designing your own guardrails now rather than assuming the platform will hold the line for you.

The IPO race is the new scoreboard

The clearest sign the game changed: valuations, not benchmarks, are becoming the scoreboard. SpaceX began trading as SPCX with xAI folded inside — effectively the first AI-infrastructure mega-cap to face public price discovery — while OpenAI and Anthropic are both in the pipeline after confidential filings. Meanwhile the megacaps wobbled: the Magnificent Seven shed roughly $2 trillion in June on fears the IPO wave pulls capital toward pure-play AI.

This matters to builders more than it looks. Public AI companies have to show margins, which means the era of deeply subsidized tokens has a clock on it. I wrote about the circular nature of AI capital and the cost reality starting to bite in Field Note #02 — public markets will only accelerate it. Watch your model bills, and keep a current view of what each model actually costs.

The model flood and the four-way coding race

Underneath the politics, the product race only accelerated. Google's Gemini 3.5 Pro is rolling toward general availability with a 2M-token context window; xAI's Grok 5 is moving toward public beta; and Claude Opus 4.8 already leads the headline intelligence rankings. The agentic-coding market went from two serious players to four: at Build, Microsoft pushed GitHub Copilot toward an autonomous developer, and Google's Antigravity 2.0 now orchestrates multiple agents in parallel — even as Sundar Pichai conceded Google is "a bit behind" and undercut on price.

For teams, the upshot is leverage: more capable models, more competition, falling prices — against a moving target for evaluation. If you're choosing a stack, keep options open and benchmark across models rather than marry one.

The Founder's Take

The center of gravity in AI just moved from the lab to the ledger and the negotiating room. With three CEOs at the G7, an executive order setting pre-release norms, and even Anthropic floating a slowdown, the next month is about who writes the rules — not who tops the index. None of this means the model race stopped; it means the model race is no longer where the leverage is. My advice to operators is what it's been: don't wait for the rules to finalize before you build the discipline. The teams that come through this strongest will be the ones treating governance and cost control as features, not afterthoughts — the same orchestration argument behind the AI Forward Framework. Watch what comes out of France, and watch the July 2 deadline. That's where this month gets decided.

— Megan