Three weeks ago, Anthropic had the most capable AI models in the world publicly available. Then Washington pulled the plug.

On June 12, 2026, the U.S. Department of Commerce sent Anthropic CEO Dario Amodei a letter invoking national security authorities and ordering the company to disable public access to Claude Fable 5 and Claude Mythos 5. The models had launched three days earlier. Commerce Secretary Howard Lutnick signed the order. Anthropic complied the same evening. For 18 days, two frontier AI systems sat offline — not because they failed, but because the government decided who could use them and when.

By June 30, the controls were lifted — partially. Access was restored to a curated list of approved U.S. entities under a licensing framework administered by Commerce, according to Semafor. The government did not name all approved entities publicly, though the arrangement permits Anthropic’s foreign national employees to use the models. A similar dynamic had unfolded at OpenAI: the Trump administration asked the company to stagger the rollout of its own next-generation model to “Trump-approved customers” only, according to reporting by Fortune and The Information on June 26.

What just happened is not a one-off. It is a policy.

What We Know

The sequence of events is this: Anthropic launched Fable 5 and Mythos 5 on June 9. Commerce issued an export control directive on June 12. Anthropic disabled the models that evening. On June 26, the administration selectively restored access to certain U.S. companies and government agencies, per CNBC. Full access was restored on June 30, with a licensing structure in place rather than open availability.

OpenAI’s situation differs in form but not in substance. Rather than an outright order, the administration requested that OpenAI voluntarily limit rollout of its next major model to vetted enterprise and government customers first — effectively a soft control. OpenAI agreed, according to The Information and Reuters (June 25, 2026). The Guardian reported on June 26 that OpenAI framed the arrangement as cooperation with “national security” interests, not coercion.

Separately, on July 2, Microsoft announced a new entity called Frontier Company, backed by $2.5 billion. Its stated purpose: embed Microsoft AI engineers directly inside enterprise customers to help them choose the right AI model — from Microsoft, OpenAI, Anthropic, Google, or the open-source ecosystem. Microsoft CEO Satya Nadella characterized the move as a defense against a world where “a few models eat everything they see.” At the same time, Anthropic released Claude Sonnet 5, priced at $2 per million input tokens through August 31, then rising to $3 per million, per Anthropic’s release notes.

On the infrastructure side, the four largest U.S. tech companies — Amazon, Alphabet, Meta, and Microsoft — are on track to spend a combined $650 billion on AI infrastructure in 2026, with analysts at Goldman Sachs projecting the figure could exceed $1.1 trillion by 2027, per Fortune (June 29, 2026). That capital is flowing into GPU clusters, data center construction, and networking equipment.

What’s Driving It

The government’s sudden interest in controlling frontier AI deployment is not spontaneous. It reflects two converging pressures: fear of capability leakage to China, and a desire to maintain negotiating leverage over the AI companies themselves.

The export control logic is straightforward in theory. Models capable of accelerating scientific research, biological design, or autonomous cyber operations represent a dual-use risk. If such a model is accessible via API to any user globally, foreign nationals — whether independent or state-affiliated — can use it without restriction. Commerce’s invocation of export control law applies the same framework used for semiconductors and encryption to model weights or API access. This is new legal territory, and the administration appears to be testing how far it extends.

The leverage angle is harder to confirm but worth naming as a hypothesis. The Trump administration has been negotiating with major AI labs over a range of issues, including the Stargate infrastructure joint venture, export rules for chips, and access for U.S. companies to foreign markets. Controlling when a model launches — and who gets access first — gives the government a tool it didn’t have before.

Microsoft’s Frontier Company announcement fits a different pattern: a bet that the enterprise AI market is fracturing into a multi-model world. Enterprise buyers who once hoped to standardize on one provider are finding that no single model leads across all tasks. Microsoft’s $2.5 billion play is an attempt to own the integration and advisory layer rather than the model itself. Nadella’s public skepticism toward model concentration reads as both genuine strategy and a signal to regulators.

Meanwhile, Deloitte and NVIDIA’s mid-2026 enterprise AI reports, cited by MarketScale, show that agentic AI — systems that take autonomous multi-step actions — has moved from experiment to production in a measurable share of large enterprises. ROI evidence is accumulating. RBC Capital Markets analyst Rishi Jaluria wrote in June that enterprise AI spending momentum was building toward the second half of 2026, with adoption transitioning “from pilot to production,” per Business Insider. The gap, both reports note, is in oversight and talent, not in model capability.

Implications

For enterprise buyers, the export control episode introduces a risk that didn’t exist six months ago: a model you depend on can go offline on short notice, not because of a product failure, but because of a government decision. The Anthropic situation lasted 18 days. A longer disruption — or one affecting a model embedded in critical workflows — would be materially damaging.

Procurement teams that haven’t yet formalized multi-model strategies now have a concrete reason to do so. That is not speculation; it is the lesson the Anthropic episode demonstrates. It also partially explains why Microsoft is spending $2.5 billion to position itself as the model-agnostic integration layer. If government actions can ground a single vendor’s flagship product, then lock-in to any one model carries new political risk on top of the usual technical risk.

For the AI labs themselves, the picture is more complicated. Cooperation with the government opens doors — preferential access to government contracts, potential inclusion in Annex A lists that bypass future restrictions. But it also sets a precedent. The labs are now, in effect, licensed participants in a managed market rather than open commercial players. The terms of that management are still being written.

For national competitiveness, the picture is grimmer than official framing suggests. IBTimes UK reported this week, citing academics and policymakers, that the U.S., EU, and China have built “structurally incompatible” regulatory regimes. The U.S. has moved toward federal preemption of state rules while using executive action for export controls. China is building its own parallel system. An AI governance architecture that functions across borders does not exist. A piece in Digit.in, drawing on policy analysis, argued that U.S. export restrictions — intended to preserve the American lead — may be inadvertently accelerating China’s push toward AI self-sufficiency by cutting off access to U.S. cloud infrastructure and model APIs.

What to Watch

The most immediate question is whether the licensing framework the Commerce Department established for Anthropic’s models becomes a template. If so, future frontier model releases may require government pre-clearance before public availability. That would represent a structural change in how AI is commercialized in the U.S. — closer to the arms-export licensing model than to the app store model.

Watch for whether OpenAI’s next major model launches with a similar staged rollout. The Information reported the administration made the request; it did not report a formal order. If OpenAI complies informally and the administration then regularizes that request into policy, the precedent expands.

Microsoft’s Frontier Company deserves scrutiny as a business model. The theory is sound — enterprises need help navigating a fragmented model market. The $2.5 billion commitment is large enough to be serious. The question is whether enterprise customers will pay advisory and integration fees on top of model API costs, or whether they’ll absorb those costs in-house as their own AI teams mature.

On infrastructure: $650 billion in annual capex across four companies is historically unprecedented. The returns that justify it depend on sustained enterprise demand growth. RBC’s Jaluria and other analysts are projecting that growth for late 2026. If it materializes more slowly — due to governance gaps, talent shortages, or model disruptions like June’s — the capex cycle will face questions it doesn’t face today.

Finally, watch China’s July 15 deadline. China’s new Interim Measures for the Administration of AI-Based Anthropomorphic Interactive Services take effect that day, per Hunton Andrews Kurth (July 2026). It is a narrow regulation governing AI companion products, but it signals continued regulatory divergence between Beijing and Washington at exactly the moment when a common framework would be most useful.

References