The AI premium got marked down
The premium operators built into their AI plans got marked down this week. In four arenas, equity markets, model pricing, US regulation, and energy, an assumption that held through 2025 reversed or stalled. None of the four is a story about a better model. Each is a story about a price: investors, buyers, regulators, and grid operators all started valuing AI on its arithmetic rather than its narrative.
The market stopped paying for AI layoffs
Through 2025 the defensible read was that announcing an AI-driven workforce cut was something investors rewarded. It signalled that management had found real efficiency, that margins would expand, that the company was ahead on automation. Klarna, Salesforce, and a string of others made the cut-and-signal move and were treated as disciplined. A chief executive planning 2026 could reasonably assume that pairing layoffs with an AI rationale would support the share price, not damage it.
On May 17, CNBC reported that of 23 S&P 500 companies which had cut staff with an explicit AI rationale, 13 of them, or 56%, were trading below their announcement-day price, with an average decline near 25%. Salesforce is down about 32% (per CNBC) since it cut 4,000 support roles last September and credited its AI service agents. Fiverr, which went "AI-first" with a 30% cut, is down 54%. Nike, which trimmed about 800 roles for automation, is off 35%. Two days later Standard Chartered said it would remove about 7,800 support roles by 2030, more than 15% of that workforce, and chief executive Bill Winters called it "replacing lower-value human capital" (Fox Business). The backlash forced a written walk-back inside a day, from a bank posting record profitability.
Any executive whose 2026 plan treats an AI-layoff announcement as a positive catalyst should reconsider. The market is now asking for the second number, the productivity or revenue gain the cut was supposed to fund, and discounting the announcement until that number shows up. Finance and investor-relations teams should stop leading with headcount math and lead with output-per-employee math, because investors have started pricing the gap between the two. This is the equity-market version of a point we made earlier, that AI displacement got specific: the vague automation story no longer clears.
Google broke the premium for frontier models
For two years the defensible pricing assumption was that frontier-class AI was scarce, therefore expensive, and that the labs held the pricing power. OpenAI's finance chief described demand as a "vertical wall." The top prosumer tiers settled around $200 a month, ChatGPT Pro and Claude Max among them, and per-token prices for the best models stayed high enough that resellers competed on everything except cost. The working assumption was that the best model commanded a premium, and that the premium would hold.
At its developer conference on May 19, Google cut the top of its AI Ultra plan from $250 to $200 a month and added a new Ultra tier at $100, while keeping its $20 plan. It shipped Gemini 3.5 Flash as the free default for coding and agent work, priced at $1.50 and $9 per million input and output tokens, and positioned it as beating its own Gemini 3.1 Pro on coding and agent benchmarks at roughly a quarter of frontier cost. Google is not claiming the outright capability lead; analysts still place its strongest model behind the restricted frontier. It is claiming the price.
The operators who should be uncomfortable are the ones whose model is "resell frontier tokens at a margin" or "charge a premium because we run the best model." When a near-frontier model is good enough for coding and agents at a quarter of the cost, the premium for "best" shrinks to whatever the specific task genuinely requires. Resellers on thin spreads, and labs whose revenue plans assume pricing power rather than cost leadership, should model a price war rather than a vertical wall. It also sharpens last week's read that the leader trade stopped working: the leader no longer sets the price.
Washington pulled back from regulating frontier models
Two weeks ago the defensible regulatory read was that the United States was edging toward pre-release oversight of frontier models. The White House was openly weighing an executive order modelled on drug approval, under which a model would pass a government security review before release. It was framed as a response to Anthropic's restricted Mythos system surfacing, by the company's own account, tens of thousands of software vulnerabilities. Safety-positioned vendors, and anyone treating compliance readiness as a moat, could reasonably plan around a tightening federal regime.
On May 20, hours before a scheduled Oval Office signing, Trump postponed the order (per CNN). The framework had already been softened to a voluntary 90-day agency review, and even that was pulled after late lobbying, with AI adviser David Sacks reportedly opposed and the president saying he "didn't like certain aspects" and did not want to "dull America's edge" while the country is "leading China." OpenAI and Anthropic had both been working with the White House on the text. As of now there is no federal pre-release review for AI models, voluntary or otherwise, and no date set for one.
Vendors who built a compliance-first story expecting federal rules to validate it should recalibrate. Our own argument that regulatory and compliance barriers are among the most durable moats still holds where the barrier is a real accreditation, a banking licence or a medical-data clearance that takes years to earn. It does not hold where the barrier was an anticipated federal rule that an afternoon of lobbying can postpone. If your safety posture was a bet on regulation arriving, the bet just got longer. If it was a bet on customers valuing safety directly, nothing changed.
Europe got priced out of the buildout
The comfortable assumption for European founders and enterprises was that AI arrives through the cloud, and the cloud abstracts geography: you rent frontier capability by the token from anywhere, so location stops mattering. On the demand side that is still roughly true. On the supply side, where the compute physically gets built, it stopped being true this year, and the gap now looks structural rather than temporary.
Electricity for European data centres runs far above the US. Reporting this month put it near $111 per megawatt-hour in the UK and $89 in Germany (per CNBC) against about $28 in the US, and the International Energy Agency notes European industrial power runs at roughly double US levels. Time is the harder constraint: in the major European hubs a new facility waits 7 to 10 years for a grid connection, up to 13 in the most congested, against up to 5 in the US. The resulting buildout gap is estimated at around 100 to 1. The US has its own friction, with roughly 7 of 12 planned gigawatts for 2026 slipping, but the relative position is not close.
European AI companies that assumed equal compute economics should price in a structural disadvantage on anything training- or inference-heavy, and weight their roadmaps toward work that runs on rented frontier capability rather than owned hardware. The push for "sovereign AI" runs into the same arithmetic: you cannot subsidise your way around a 7-to-13-year grid queue on the timeline AI is moving. The cheapest place to run the model increasingly decides where the capability gets built, and at the moment that place is not Europe.
Read the four together
The four breaks share one mechanism: a soft variable operators had been pricing on narrative hardened into a number this week. Investors stopped paying for the story of AI efficiency and began demanding the margin it was meant to produce. Google stopped charging a premium for the idea of frontier capability and started competing on cost per token. Washington declined to turn AI risk into a federal rule, so the compliance-as-moat story lost the backstop it was counting on. And the cost of building AI showed itself to be an energy-and-grid number that geography decides. The 2025 habit was to extrapolate every AI trend in a straight line. The discipline 2026 rewards is narrower: for each line in your plan, ask what the arithmetic underneath it actually says, because this week it was the arithmetic, not the story, that moved.