Week of 2026-06-01 to 2026-06-07 · AI’s subsidies got called in at every level at once — and that’s the healthiest thing that’s happened to the industry in years

For three years the AI buildout has run on a comfortable fiction: somebody else is paying. Investors subsidized the labs, the labs subsidized your coding agent, utilities socialized the power bills, and the labor market hadn’t yet presented its tab. This week, with eerie synchronization, every one of those subsidies got called in — on Wall Street, on your GitHub invoice, on the grid, and in the jobs report. Pull on any one of this week’s stories and the others come with it, because they are not separate stories. They are one story: AI is being repriced from a faith-based capex narrative into an economy, and economies have line items.

Start at the top of the capital stack. On Monday, Anthropic confidentially filed an S-1 — the first frontier lab to start the IPO process, days after closing a $65B round at a ~$965B valuation, on a revenue run-rate reportedly past $47B, up from ~$9B at the end of 2025. By Thursday, the merged SpaceX–xAI entity was running the largest IPO roadshow in history, targeting up to $75B at a valuation near $2T. The private-capital era of frontier AI — the era in which a handful of funds absorbed all the risk while the rest of us watched demos — formally ended this week.

The public markets wasted no time demonstrating what that means. On Tuesday, Broadcom reported a record quarter — revenue up 48%, AI silicon up 143%, Q3 AI guidance of +200% — and the stock fell roughly 15%, because management declined to raise the full-year target. Then Friday’s jobs report came in at double expectations, rate-hike fears ignited, and the Nasdaq dropped 4% in its worst session since April 2025, wiping more than $1T off chip stocks. Read those together and the message is precise: the market is no longer pricing AI growth, it is pricing AI acceleration, and anything flat gets sold. With hyperscaler 2026 capex north of $700B effectively load-bearing for the whole index, the AI trade now moves on Fed policy and payroll prints, not product launches. The labs wanted Wall Street’s money; they now get Wall Street’s nerves. Every AI vendor you depend on is about to become more margin-disciplined, more pricing-honest, and less generous — which brings us one layer down, to the part of this story that hit your own invoice.

On June 1, GitHub Copilot’s 4.7 million paid seats moved to metered, token-based billing, and the screenshots were brutal: projections jumping from $29 a month to $750, power users estimating 10–50x increases for agentic workflows, the Claude Opus multiplier quietly raised from 7.5x to 27x, the free fallback model removed, and code review now burning both AI credits and Actions minutes. The same week, Cursor split its Teams pricing into $32 and $96 seats, X force-migrated its remaining API subscribers to pay-per-use, and Anthropic’s June 15 change — moving Claude Code’s headless, CI, and Agent SDK usage onto separate credit pools metered at API list prices — had teams re-running cost spreadsheets for every pipeline that shells out to claude -p. If you run Claude Code in CI, that last one is your homework this week, not next: the subscription you budgeted in January does not cover what your automation does in July.

It’s tempting to read all this as gouging. It’s more honest to read it as the Broadcom selloff operating one layer down. Agentic inference is genuinely expensive; the flat rates were venture-subsidized customer acquisition; and companies marching into the public markets cannot keep selling dollars for quarters. AI coding tools are being repriced from software — flat, cheap, forgettable — to compute: metered, budgeted, FinOps-reviewed. The tell is that the supply side responded in the same news cycle. Microsoft used Build to ship MAI-Code-1-Flash, a 5B-parameter coding model at 51% on SWE-Bench Pro wired straight into VS Code — explicitly framed as cheaper and less OpenAI-dependent. Nvidia used Computex to announce RTX Spark, a Windows PC superchip co-built with MediaTek that puts serious local inference on every 2027 corporate laptop. When the metered price of a thing becomes visible, the market immediately starts building the cheap substitute. Metering isn’t the end of this story; it’s the starting gun.

Here is the part that should actually worry you, though, and it isn’t the money. The same week AI code review became a metered, paid resource, Anthropic published a report claiming Claude now writes more than 80% of the code merged into its own codebase, with engineers merging roughly 8x more code per day than in 2024. Discount the vendor self-report as much as you like — the direction matches what every high-adoption team describes. Generation is no longer the bottleneck; review, verification, and trust are. And into exactly that gap, twice in one week, walked the attackers. Miasma poisoned 32 Red Hat npm packages through a single compromised employee account, using orphan commits that bypassed human review entirely and stolen OIDC tokens that minted valid SLSA provenance — then evaded install-script scanners via a 157-byte binding.gyp. Days later, IronWorm hit 36 packages while specifically harvesting OpenAI and Anthropic API keys — credentials that, in the metered era, map directly to spendable compute. Sit with the symmetry: the industry just put a price on reviewing code at the exact moment review became the scarcest resource in the system, and criminals started stealing the new currency. Rotate your CI secrets this week; argue about the philosophy later.

Zoom out one more level and the same who-pays fight is running on the literal power grid and in the labor market. This week brought dueling studies over whether data centers are raising household electricity bills, just as PJM — whose own market monitor blamed data centers for ~40% of December’s $16.4B capacity auction — heads into its first uncapped auction in years, while Vantage and Liberty Energy simply opted out of the queue with a 1GW bring-your-own-power deal. Challenger’s May data made AI the #1 stated reason for US job cuts — 87,714 AI-linked cuts in five months, more than 2024 and 2025 combined, with profitable companies cutting headcount explicitly to fund AI infrastructure. Note the strangeness: aggregate employment ran hot enough on Friday to spook the Fed while AI displacement accelerates underneath it. Headcount and compute now compete for the same budget line, and this week compute won again. Politics, smelling all of this, moved too: a bipartisan House draft — the Great American Artificial Intelligence Act — offered a three-year freeze on state AI laws in exchange for frontier-lab transparency, Commerce extended chip export controls to Chinese firms’ offshore subsidiaries after admitting the old perimeter leaked, and Beijing answered the IPO wave in its own dialect with DeepSeek’s first-ever external raise — ~$7.4B led by Tencent and CATL with a state fund aboard.

So, the take. This was the healthiest bad week AI has had. Metered pricing, hostile earnings reactions, ratepayer fights, and labor-market accounting are what it looks like when a technology stops being a story and starts being an economy — with unit costs someone has to own, in public. The companies that thrive from here are the ones whose products survive being priced honestly; the discipline the industry kept promising regulators would come from self-governance is arriving instead from the bond market, the grid operator, and the monthly invoice. For you, practically: get a token budget and a model-routing strategy before your CFO assigns you one, treat AI API keys with the paranoia you reserve for cloud credentials, and re-read your CI’s Claude usage before June 15. The metering question is big enough that it gets this week’s deep dive: can AI coding tools survive honest pricing at all — or does the cheap-substitute wave make this whole pricing regime a two-year interregnum?

Also this week

  • Anthropic expanded Mythos, its deliberately unreleased zero-day-hunting model, to ~150 more organizations including critical-infrastructure operators — “capable but withheld” is quietly becoming a product category.
  • OpenAI’s models landed on AWS, ending Azure distribution exclusivity right before its own S-1 — revenue-maximizing behavior for the filing, awkward optics for the partnership.
  • MiniMax launched M3 claiming frontier coding performance as “open-weight” — except the weights hadn’t shipped by week’s end; treat benchmarks as marketing until they’re on Hugging Face.
  • TSMC is deploying Nvidia-accelerated AI inside its fabs — AI chips speeding up the manufacture of the next AI chips is how a hardware cadence becomes a moat.
  • Apple reportedly signed a ~$1B/year deal for a custom Gemini model to power the rebuilt Siri — pre-WWDC reporting; the keynote fell on June 8, so full analysis lands next issue.
  • Cognition retired the Windsurf brand for Devin Desktop and adopted the open Agent Client Protocol — the IDEs churn, but ACP and MCP keep winning as the interop layer.
  • Python 3.15 hit feature freeze with beta 2 — library maintainers, the October target is now real.
  • Trump reportedly floated direct US government equity stakes in OpenAI, Anthropic, and xAI — single-sourced for now, but consistent with the Intel playbook, and worth filing next to the S-1s.

One thing to watch

Does GitHub blink? Prediction, stated falsifiably: within 30 days (by ~July 5), GitHub partially retreats on Copilot pricing — extending promo credits beyond August, restoring a free fallback model, or cutting the Opus multiplier — without reversing metering itself. If they hold the line and churn stays low, flat-rate AI tooling is gone for good and every vendor follows by Q4. Scoring next issue, alongside WWDC fallout and OpenAI’s June 8 S-1.