The Week Everyone Wanted a Piece of AI

Three days after launch, the models went dark — and stayed dark for nineteen. On June 12, the U.S. Department of Commerce ordered Anthropic to suspend all foreign-national access to Claude Fable 5 and Claude Mythos 5, days after both launched to widespread praise. Because Anthropic had no reliable way to verify user nationality in real time across its global cloud footprint, it pulled both models everywhere, for everyone, simultaneously — from the Claude Platform, AWS Bedrock, Google Cloud, Microsoft Foundry, Snowflake, and Box all at once. On June 30, Commerce Secretary Howard Lutnick lifted the order. Fable 5 returned globally on July 1; Mythos 5 is back only for a vetted set of U.S. organizations under Anthropic's Project Glasswing.

The trigger was a jailbreak, not a breach. Amazon researchers found a technique that bypassed one of Fable 5's cybersecurity safeguards and flagged it to federal authorities. Anthropic contested the government's framing, telling press the demonstration surfaced a handful of previously known, minor vulnerabilities rather than a sophisticated exploit — but the order took effect regardless, and the company had no legal path to a partial, nationality-filtered shutdown. That's the detail enterprise buyers should sit with: the failure mode wasn't a security flaw. It was the absence of a mechanism to comply narrowly.

For finance, healthcare, SaaS, and critical-infrastructure customers, this was a live-fire test of a risk nobody had modeled. Force-majeure clauses written before 2026 didn't contemplate a government-mandated, instantaneous suspension of a vendor's flagship model across every integration at once. Production workflows built on Fable 5 or Mythos 5 fell back to Opus 4.8 with no warning window. Whatever contingency plan your organization has for "the vendor has an outage" almost certainly doesn't cover "the vendor is legally barred from serving you," and those require different mitigations.

The precedent outlasts the resolution. Anthropic's own restoration announcement frames what emerged as a durable, three-tier structure — full public availability, vetted-partner access, and total suspension — that didn't exist before June 12 and now sits available to regulators for the next frontier release, from any lab. Enterprise architects who treated multi-model fallback as a nice-to-have should treat it as a compliance requirement starting this quarter, regardless of which lab you're standardized on.

Sam Altman wants Washington to become a shareholder in every major American AI lab — starting with his own. The Financial Times reported this week that OpenAI has proposed handing the U.S. government a 5% equity stake, worth roughly $42.6 billion at the company's $852 billion March valuation. Altman's pitch, modeled explicitly on the Alaska Permanent Fund, envisions Anthropic, Google, and Meta contributing matching stakes into a government-seeded "Public Wealth Fund" that would distribute AI-driven economic gains directly to citizens. None of those other labs have signaled interest.

The timing is the tell. The offer landed six days after OpenAI delayed the public rollout of GPT-5.6 at the government's request, and in the same week Anthropic was still working through the aftermath of the Fable 5 and Mythos 5 shutdown. Both events point to the same pattern: the U.S. government is now an active participant in decisions about how and when frontier models ship, and the labs are responding by offering to formalize that relationship on their own terms before it gets formalized for them.

Precedent exists, and it's not comforting for anyone hoping this stays theoretical. Washington already holds a 9.9% stake in Intel, converted from CHIPS Act grants, and negotiated 15% revenue-sharing arrangements with Nvidia and AMD on China chip sales. Senator Bernie Sanders has filed a competing bill demanding 50% public ownership of major AI firms' equity — which makes Altman's 5% look less like generosity and more like an opening bid designed to make a bigger ask look unreasonable by comparison.

For enterprise buyers, the relevant question isn't whether the government takes the stake — it's what strings come attached if it does. Equity typically comes with governance rights: board influence, disclosure requirements, veto power over certain decisions. A frontier lab with a government shareholder is a different kind of vendor than one without, with different incentives around what gets built, what gets restricted, and whose interests get weighed when those two things conflict. That's a vendor-risk conversation that didn't exist on anyone's procurement checklist eighteen months ago.

AWS just made its clearest statement yet on where enterprise AI actually breaks. The company is committing $1 billion to a new Forward Deployed Engineering organization, embedding pods of five to six engineers directly inside customer environments for roughly 45-day engagements to build and ship production agentic systems using the customer's own data and governance controls. Teams are already live with the Allen Institute, Cox Automotive, the NBA, the NFL, Ricoh, and Southwest Airlines. Microsoft answered within days, committing 6,000 engineers to its own "Frontier Company" deployment push.

The bet underneath both moves is specific: model quality stopped being the constraint months ago. Independent research from MIT, McKinsey, RAND, and Gartner converges on roughly the same finding by different methods — somewhere between 73% and 95% of enterprise AI pilots never produce measurable ROI. Not because the models can't do the work, but because nobody on staff can wire agentic systems into real business processes fast enough, under real governance constraints, before the pilot budget runs out. AWS's model is designed to leave customers self-sufficient — knowledge graphs, runbooks, trained internal staff — rather than create ongoing dependency, which distinguishes it from a traditional consulting engagement.

This lands in the same news cycle as a record 87,714 AI-attributed job cuts through May 2026, the highest total Challenger, Gray & Christmas has tracked for any calendar year on record. Thousands of new, high-paying deployment-engineering roles are opening at the exact moment AI is credited with eliminating other roles — not a contradiction, but the defining employment paradox of this phase: the same technology displacing workers in one function is creating acute scarcity in the function that makes it actually work.

For CIOs, the practical decision is no longer "build or buy the model" — it's "who deploys it, and do they leave you capable of running it yourself." A forward-deployed engineering engagement that ends with your team owning the runbooks is a fundamentally different vendor relationship than one that ends with a support contract renewal. Ask any vendor pitching an FDE-style engagement this quarter exactly what you own when the 45 days are up.

Alex Karp didn't hold back, and the market rewarded him for it. In a combative CNBC interview timed to a new Palantir-Nvidia partnership, the Palantir CEO accused leading AI labs of overcharging customers, harvesting their business data, and putting U.S. national security at risk by running per-token pricing on models that see enterprises' most sensitive information. “Something has gone completely wrong,” Karp said. “I'm going to chillax and waste my time with tokens. I'm going to get no value, and they're going to get my IP.” Palantir shares rose more than 9% that morning.

Strip away the theater and the argument is a real one that CIOs are already having internally. Karp's core claim is that value in AI accrues to whoever owns the layer where the model actually gets put to work against your data — not to whoever built the underlying model. He's selling Palantir's Ontology layer as the alternative: a governance boundary that sits between your data and any model, controlling exactly what the model can see and do. Self-interested, obviously — but the underlying worry, that per-token pricing obscures where your proprietary data and workflows are actually flowing, is one enterprise security teams have been quietly raising for at least a year.

The timing sharpens the point. Karp's rant landed the same week Anthropic was restoring Fable 5 and Mythos 5 after a government-mandated shutdown, and OpenAI was pitching Washington on a government equity stake. Karp explicitly invoked the tension: “Are we really going to outsource the battlefield of this country to the consensus view in Silicon Valley?” Whatever you think of Palantir's own government entanglements, the underlying question — who actually controls the layer where sensitive enterprise and government data meets a frontier model — is now squarely on the table, and it isn't going back off it.

For enterprise technology leaders, this is a governance-architecture question disguised as a personality story. Whether or not you buy Palantir's Ontology pitch specifically, every organization running frontier models against proprietary data should be able to answer, concretely: what data reaches the model, what does the vendor retain, and what would need to change if that vendor's incentives shifted. If the honest answer is “we're not entirely sure,” that's the finding — not Karp's delivery.

Vlad Tenev isn't being subtle about the ambition. “The idea behind agentic trading,” the Robinhood CEO told CNBC this week, “[is] every capability a human can do will be available to an AI agent.” Robinhood introduced tools in May allowing AI agents to execute stock trades on users' behalf and expanded that capability to crypto markets in early July, enabling 24/7 autonomous execution. Tenev, who ran programmatic trading at an institutional level before founding Robinhood, is explicit that the end state is putting hedge-fund-grade automated trading infrastructure into a consumer app with a few hundred dollars in it.

The context makes the bravado worth noting rather than dismissing. Robinhood cut 10% of its workforce earlier this month and missed Q1 profit forecasts amid crypto-market volatility — yet the stock jumped 8% the day before Tenev's comments and is sitting at a $98 billion market cap. The company is leaning into agentic autonomy precisely while under pressure to show growth, which is exactly the condition under which risk controls get under-resourced relative to the feature rollout.

The risk isn't hypothetical, and it isn't really about trading at all. An AI agent operating with real money, around the clock, in a market that can move sharply in minutes, is a live-fire test of exactly the autonomy question every enterprise deploying agentic AI is wrestling with internally: what guardrails hold when the agent encounters a scenario nobody explicitly anticipated. Robinhood's own framing — user-defined limits, not full autonomy — is the right instinct. Whether the limits are well-designed only gets tested in a flash-crash scenario, which is not when you want to discover the gap.

For any enterprise granting an AI agent standing authority to act with money, contracts, or customer commitments, the Robinhood rollout is a useful public test case to watch rather than a retail curiosity to ignore. The scoping question — what, specifically, is this agent allowed to do without a human in the loop, and what happens at the edge of that scope — is identical whether the agent is trading a retail portfolio or approving a supplier invoice. Robinhood is running that experiment at scale, with real money, in public. Take the data point.