Three things happened last week that, taken together, say more about the current state of AI than any product launch. Google lost two of the most consequential AI researchers alive within 48 hours of each other. The U.S. government forced Anthropic to pull its two most capable models from every market on earth. And a survey of 1,550 AI decision-makers found that while 72% of large enterprises are running agentic AI in production, 60% have no formal governance over it.

None of these are isolated events. They are the same story at different altitudes.

What We Know

The Google departures. On June 18, Noam Shazeer — co-author of the 2017 “Attention Is All You Need” paper that defined modern AI architecture, and co-lead of Google’s Gemini models — announced he would join OpenAI ahead of that company’s anticipated IPO. One day later, John Jumper, Nobel laureate in chemistry and VP Engineering Fellow at Google DeepMind, announced he was leaving for Anthropic. Google CEO Sundar Pichai and DeepMind CEO Demis Hassabis both posted public farewells; neither offered any indication of what, if anything, would slow the drain. Losing Shazeer alone would be a significant event. Losing both within 72 hours has no recent precedent in the industry.

The Anthropic export-control order. In mid-June, U.S. Commerce Secretary Howard Lutnick issued an order imposing export controls on Anthropic’s Fable 5 and Mythos 5 models — its two frontier-tier systems — restricting access for any foreign national, including Anthropic’s own non-American employees. Anthropic concluded it had no practical way to enforce nationality checks inside shared cloud infrastructure in real time, so it shut both models down globally. The company has not disclosed a timeline for restoration. The EU called the move a shock; French politicians renewed calls to accelerate support for Mistral, the continent’s only home-grown frontier AI developer. Anthropic is simultaneously navigating a legal and regulatory battle with the U.S. government, details of which have not been fully disclosed publicly.

The agentic AI governance gap. The Agentic AI Institute published survey data this week showing 72% of enterprise firms have agentic AI in production, yet 60% lack formal governance frameworks for it. Separately, Publicis Sapient released its 2026 Global Enterprise AI Report on June 17, drawn from 1,550 AI decision-makers across six markets. Its headline finding: AI adoption is near-universal, but only 10% of respondents said AI is core to their operations. Broad deployment without strategic integration is now the default condition for most large organizations.

Infrastructure math. Hyperscalers are spending more than $600 billion in 2026 on AI infrastructure, according to data from Data Centre Digest. Despite that investment, more than 36 planned projects worth a combined $162 billion have been blocked or significantly delayed due to power availability constraints. The bottleneck is electricity, not capital.

The G7 signal. On June 17, Sam Altman, Dario Amodei, and Demis Hassabis joined Trump and other G7 leaders at a working lunch in Evian-les-Bains, France. The seating arrangement was, as CNBC noted, “a signal of where power sits.” AI executives are now regular participants in sovereign-level geopolitical conversations. That is new.

What’s Driving It

The Shazeer and Jumper departures are not simply about compensation, though compensation almost certainly matters. They reflect a structural shift: OpenAI’s pending IPO creates a liquidity event that Google cannot easily match from inside a company that has been publicly traded for two decades. Anthropic, meanwhile, is aggressively building a research organization capable of competing on the frontier — and Jumper’s AlphaFold-era expertise in AI-for-science is directly relevant to where that competition is heading.

Google’s position is complicated. It has the model capability, the compute, and the distribution. What it lacks, apparently, is the ability to retain people who believe the action is elsewhere. That belief may be rational or may be recency bias; either way, the departures make it harder to dismiss.

The export-control order on Anthropic’s Fable 5 and Mythos 5 reflects a sharpening U.S. policy posture. The Trump administration’s stated position — articulated in a June 22 White House executive order on “Promoting Advanced Artificial Intelligence Innovation and Security” — is that America leads AI because it refuses to stifle innovation with regulation. That framing sits awkwardly alongside an order that forced Anthropic to take its own products offline worldwide. The administration appears to be running two distinct AI policies simultaneously: domestic deregulation paired with aggressive export control. Those are not necessarily contradictory, but the friction they create for U.S. companies with global user bases is real and growing.

The infrastructure bottleneck is, at root, a permitting and grid problem. Capital is available. Power is not. More than 36 projects worth $162 billion are stalled — not because investors withdrew, but because utilities cannot deliver the load on the timelines AI data center buildouts require. This is a structural constraint that will not resolve on a software-cycle timeline.

On enterprise adoption: the Publicis Sapient finding that only 10% of firms consider AI core to operations, despite near-universal deployment, suggests most organizations are still running AI at the periphery — productivity tools, copilots, back-office automation — rather than redesigning workflows around it. The Agentic AI Institute’s governance-gap data reinforces this: agentic systems that take autonomous actions are in production at most large enterprises with minimal oversight frameworks in place. That combination of broad deployment and thin governance is where enterprise liability risk accumulates.

Implications

For U.S. enterprise technology buyers, the Anthropic export-control episode raises a question that will not go away: what happens to a vendor dependency when the U.S. government can require a company to shut its products off globally with no warning? Most enterprise procurement teams do not currently model this as a risk category. They should.

The governance gap is a near-term legal exposure. EU AI Act full enforcement begins August 2, 2026 — six weeks from today. The Act’s extraterritorial provisions apply to any company whose AI outputs are used within the EU, regardless of where the system was built or hosted. U.S. firms that have deployed agentic AI broadly without formal governance documentation are not prepared for this. Berkeley Law’s analysis of the Act, published this week, noted that practitioners “can no longer treat artificial intelligence as a monolithic regulatory object” — the distinctions between generative, agentic, and robotic systems carry concrete compliance weight.

For national competitiveness, the talent drain from Google is a signal worth reading carefully. The company that produced the Transformer architecture, AlphaFold, and Gemini is now losing the people who built those systems to rivals. That does not mean Google loses its lead — it has structural advantages that individuals cannot replicate — but it does mean the frontier is contested in ways that were less visible six months ago.

What to Watch

Anthropic’s model reinstatement timeline. The company has not indicated when or whether Fable 5 and Mythos 5 will return for non-U.S. users. If the restriction holds, it forces enterprise customers to choose between capability and compliance — and accelerates European interest in Mistral as a sovereign alternative.

OpenAI’s IPO. Shazeer is joining ahead of it. If the offering proceeds as anticipated, the equity event will likely accelerate the talent competition across the industry. Watch for further departures from Google, Meta, and potentially Anthropic in the months following.

EU AI Act enforcement actions. August 2 is the full-applicability date. The first enforcement actions — and the companies named in them — will define how seriously the Act is treated in practice. U.S. companies without compliance documentation are the most exposed.

Power procurement. Whether the stalled $162 billion in data center projects clears depends on U.S. federal energy permitting and utility negotiations. Any legislative movement on AI-specific grid access — already discussed but not enacted — would be a material unlock for the infrastructure build.

Google’s response. The company has said little publicly about the departures. Its next major model release, the response to the Gemini co-lead vacancy, and any retention restructuring will be indicators of how seriously leadership is treating the talent situation.

References