The AI industry spent the past three years telling enterprise buyers that the hard part was getting started. This week’s news suggests the hard part is actually right now.

Three distinct storylines converged in mid-June 2026. The White House signed an executive order on June 2 giving the intelligence community a formal role in AI model vetting—then days later ordered Anthropic to restrict foreign national access to two of its most advanced systems. Separately, Microsoft launched seven proprietary AI models under its MAI brand, breaking years of near-total reliance on OpenAI’s technology. And a stream of enterprise case studies confirmed what many CIOs had quietly been saying: the token bills are too high, and the returns are hard to measure.

None of these stories is surprising in isolation. Together, they mark a real shift in the industry’s operating environment.

What We Know

On regulation: Executive Order 14409, signed June 2, creates a voluntary framework requiring AI companies to share frontier models with the U.S. government up to 30 days before public release. The order gives intelligence agencies enhanced authority to test those models for national security risks. White House chief AI adviser David Sacks pushed hard to keep the process voluntary, reportedly to prevent what he calls “regulatory capture” by large labs, according to Axios reporting on June 12. The voluntary framing is consequential: companies that participate get earlier federal engagement; those that don’t face no formal penalty under the order.

That same week, the Trump administration separately ordered Anthropic to restrict foreign national access to two models: Claude Mythos Preview and Claude Fable. Nextgov/FCW reported June 14 that Anthropic responded by disabling the models for all customers while it works out an access control mechanism. Anthropic CEO Dario Amodei had published an essay June 10—covered by Politico—calling for mandatory testing that would allow governments to block deployment if an independent auditor deems a model too risky. His proposals go substantially further than the voluntary EO framework.

On models: Microsoft launched seven MAI models on June 10, led by MAI-Thinking-1, a reasoning-focused system the company says is competitive with top third-party offerings. The launch also included a formal announcement that Microsoft is building what it calls a “superintelligence lab”—a research division and development philosophy intended to define the next phase of the company’s AI work. This represents Microsoft’s clearest public break from its OpenAI dependency, though the two companies remain deeply commercially entangled through Azure.

Google DeepMind released Gemma 4, an open-model family of four systems (E2B, E4B, a 26B mixture-of-experts variant, and a 31B dense model), all licensed under Apache 2.0. DeepMind also announced a $10 million research fund focused on multi-agent safety, citing concerns about emergent risks when millions of autonomous agents interact at scale, per MIT Technology Review.

On enterprise costs: Gartner revised its 2026 AI model spending estimate upward to approximately $32.6 billion, roughly double its 2025 figure. But the headline number obscures a growing tension. Uber exhausted its entire 2026 AI budget by April, according to Forbes, driven primarily by token consumption on Anthropic’s Claude Code coding assistant. Uber’s COO subsequently capped AI tool usage, a notable reversal for one of the industry’s more aggressive technology adopters. CIO magazine reported that for every dollar enterprises spend on AI licenses, they’re spending three to five dollars on implementation and integration work—costs that rarely appear in vendor ROI projections.

On infrastructure: KKR launched Helix Digital Infrastructure on June 11 with more than $10 billion committed to hyperscale data center development, led by former AWS CEO Adam Selipsky. A Guardian investigation published June 8 found that a majority of planned U.S. AI data centers are sited on drought-stressed land, with one proposed project in Box Elder County, Utah—the Stratos Project—spanning approximately 40,000 acres and targeting up to 9 gigawatts of power. VanEck launched a data center supply chain ETF (ticker: RACK) on June 2, explicitly benchmarked to the AI infrastructure build-out.

What’s Driving It

The regulatory moves reflect competing pressures. The White House wants to preserve U.S. AI dominance while preventing adversaries from accessing frontier capabilities—goals that sometimes conflict. Export controls on Anthropic’s models respond to the second concern. The voluntary testing framework tries to address both without triggering the kind of mandatory licensing regime that Sacks has publicly opposed.

Amodei’s call for mandatory blocking authority is an unusual move for a sitting AI CEO. It’s worth noting that Anthropic benefits commercially from tighter safety standards applied uniformly to all labs—a barrier structure that advantages companies already investing heavily in safety research. That doesn’t make the safety argument wrong, but it’s a real incentive to acknowledge.

Microsoft’s MAI launch is partly competitive positioning and partly supply-chain risk management. Full dependence on OpenAI’s API creates pricing and availability exposure. Building in-house models, even ones that aren’t yet frontier-class, gives Microsoft negotiating leverage and fallback capacity for Azure customers.

The enterprise cost crisis has a structural cause. Token-based pricing creates costs that scale with usage rather than with value delivered. Developers who use AI coding assistants extensively—which is the entire point—generate enormous token volumes. Uber’s situation isn’t unique; it’s a leading indicator. Becky Trevino, chief product officer at FinOps vendor Flexera, told CIO that value measurement is now the central challenge, not adoption.

Implications

For U.S. businesses, the Uber story is the most immediately actionable signal. Any enterprise running AI coding assistants, customer service agents, or document processing workflows at scale needs a FinOps function specifically for AI token costs. Without usage governance, the math does not close.

The Anthropic export control situation creates a near-term planning problem for multinational companies that use Claude Mythos or Fable in global deployments. Access restrictions tied to user nationality are difficult to enforce cleanly in enterprise environments where contractors and employees span dozens of jurisdictions. Expect Anthropic to release a tiered access architecture; expect it to be clunky initially.

The Microsoft MAI announcement matters most for large Azure customers. If Microsoft’s in-house models reach competitive capability levels, procurement teams will have real price pressure leverage for the first time in years. That day isn’t here yet—MAI-Thinking-1 is new and benchmarks against leading models haven’t been independently validated—but the trajectory is clear.

The infrastructure investment surge, despite environmental headwinds, signals that major investors still expect AI compute demand to grow substantially. KKR’s $10 billion Helix vehicle, the VanEck ETF launch, and the scale of hyperscaler India investments all point to a bet that current capacity will be insufficient within two to three years. That bet could be wrong. Power and water constraints are real, and the Stratos Project in Utah faces the kind of environmental review that has delayed or killed comparable facilities.

What to Watch

Anthropic’s access control rollout. The company needs to restore Claude Mythos and Fable for legitimate enterprise customers without running afoul of the export control order. How it segments access—by contract type, user verification, geography—will set a precedent for how other labs comply with similar future orders.

Whether OpenAI signs onto the voluntary EO framework. The order is voluntary. OpenAI, whose relationship with the current administration is complicated by ongoing corporate restructuring, has not publicly committed. If the major frontier labs diverge on compliance, the framework’s credibility erodes quickly.

Enterprise AI budget freezes through Q3. Uber’s reset is visible because the company is public and the story leaked. Similar decisions are likely happening across the Fortune 500 without public disclosure. Watch for enterprise software vendors reporting lower-than-expected AI consumption in Q2 earnings calls.

Multi-agent safety research outcomes. Google DeepMind’s $10 million research fund is small relative to the scale of agent deployment, but it’s a signal that the lab’s safety team sees the multi-agent interaction problem as urgent. The research timeline is 18 to 24 months. Companies deploying multi-agent workflows now are operating ahead of the science.

Water and power policy for data centers. The Guardian’s reporting on drought-sited facilities will generate regulatory and activist attention. Permitting delays or water-use restrictions in Western states could materially shift where new capacity gets built—and push costs higher.

References