June 1, 2026
The Token Tax
A week when AI valuations spiked and AI tooling shifted to metered per-token billing: Copilot's June 1 usage-based switch, Cognition's $1B raise at a $26B valuation, the $1.5B Bartz v. Anthropic settlement heading to payout, the EU's still-provisional AI Act Omnibus, and DeepMind's lab-confirmed Co-Scientist liver-fibrosis results.
Two things peaked in the same week. Money poured into AI companies at valuations that would have read like typos a year ago. And the tools those companies sell quietly started charging by the token. Call it the token tax: the shift from flat monthly fees to metered compute, where every prompt has a price and your bill scales with how much you build. Here is what moved in the week ending June 1, and what to do about each item.
TECH: AI in production
GitHub Copilot switched every plan to usage-based billing on June 1.
Copilot used to count "premium requests." Now it meters raw tokens: input, output, and cached, priced at each model's API rate. Every plan ships with a pool of GitHub AI Credits. Copilot Pro+ is $39 a month with $39 in credits. Business is $19 per user. Run past your pool on a paid plan and you buy more.
The lesson for builders is blunt. Heavy model use now lands on the invoice. Sending a routine task to an expensive model is money on fire.
The money play: Pull your token usage this week. Route small edits to a cheaper model and save the frontier model for the hard 10%. Same output, smaller bill.
Sources: GitHub Blog | GitHub Docs
BIZ: build, fund, scale
Cognition raised more than $1 billion at a $26 billion valuation on May 27.
The maker of Devin, the coding agent, was worth $10.2 billion eight months ago. The round was led by Lux Capital, General Catalyst, and 8VC. Cognition says it hit a $492 million annualized revenue run-rate, and that 89% of the code its own engineers commit is written by Devin.
Read past the number. A coding-agent company is using its own product to write nearly all its own code. That is the proof, not the valuation.
The money play: If you are building an AI product, pick one workflow your own tool can run end to end, then measure the share it handles unassisted. That figure is your real pitch.
Sources: TechCrunch | Cognition
LAW: rights and fights
The $1.5 billion Bartz v. Anthropic settlement is heading to payout.
The case covered roughly 500,000 books pulled from pirate libraries to train AI. The settlement pays about $3,000 per work. The administrator is scheduled to calculate each author's share on June 11, and payments follow. The deal covers past training data, not model outputs.
The rule courts keep drawing: training on books can be fair use, but pirating the copies to do it is not. How you source data is now the whole ballgame.
The money play: If you train or fine-tune on outside data, keep receipts. Licensed or public-domain sources, logged.
Sources: Authors Guild | Norton Rose Fulbright
POL: policy, tax, borders
The EU AI Act timeline may shift, but the change is not law yet.
The Digital Omnibus is a provisional agreement to simplify the AI Act. It is not in force. Gibson Dunn notes the changes "only take legal effect upon formal adoption and publication of the Omnibus in the Official Journal, expected before 2 August 2026." If adopted, the heaviest high-risk system rules (Annex III) would move to December 2, 2027. Existing systems would get until December 2, 2026 before the watermarking duty for AI-generated content applies.
This reaches any builder with EU users, not just EU companies.
The money play: The direction holds even if the dates move. If you ship AI output to EU users, build clear "AI-generated" labeling into your pipeline now. Cheaper to bake in early than to retrofit under a deadline.
Sources: European Commission | Gibson Dunn | Gunderson Dettmer
SCI: breakthroughs to build on
Google DeepMind reported lab-confirmed biology results from its AI Co-Scientist, detailed on May 19.
The system runs a loop of agents that generate hypotheses, debate them, and evolve the best ones. In one liver fibrosis test, a repurposed drug it flagged blocked 91% of a damage response that can drive liver scarring, and two of its picks helped liver cells regenerate. These are early lab results, not clinical proof.
The design is the takeaway. Generate, critique, refine, on repeat, with compute spent at test time instead of in one shot.
The money play: Copy the pattern, not the biology. On your next hard problem, run one model to propose, a second to attack it, a third to merge the survivors. Cheap to try, and it beats a single prompt.
Sources: Google DeepMind | Google DeepMind: liver fibrosis
Before you go
The pattern this week: capability keeps climbing while the meter starts running. Build like both are true. If you want help wiring agent loops into your own stack, that is what we do at nomadtechnologist.com. Kobe, the dog, remains unimpressed by all of it and would like his walk.
Not legal, financial, or tax advice.
Get this every weekday.
Free. 5 minutes. Unsubscribe whenever.