The coming corporate war over AI token budgets
- https://www.youtube.com/watch?v=m-bT5v5Tm7w
- Original title: Unfortunately, I Was Right
Prime takes a victory lap: eight days after he predicted companies would stop letting engineers burn unlimited AI tokens, Sam Altman admitted AI cost suddenly became "a huge issue" at the start of 2026. Off that, Prime fires off five more predictions about how corporations will ration tokens — most of which he thinks land within 6–12 months, none requiring AGI.
The core thesis: every company already scrutinizes a $50 expense in a $1,000 meeting, so the era of every employee freely spending $10k/month on tokens cannot last. The five predictions:
- Open-source token donations — you'll be able to gift tokens from your budget to OSS projects to run CI and tooling (cheap models like Kimi, not GPT-5.5). The "nice" one, akin to donating compute to protein folding.
- Token budgets as a benefits category — alongside salary, equity and 401k, you'll get a yearly token stipend; come in under budget and you get a fatter bonus, incentivizing efficiency over slop.
- "Token poker" replaces planning poker — agile estimation becomes guessing how many millions of tokens a feature will cost; a whole new consulting class reintroduces agile as "token agile."
- Org/team-wide token budgets — one engineer can drain the whole team's allotment, breeding resentment, "pair prompting" to save tokens, and a management class whose job is negotiating token quotas. Long-running agent prompts get team-reviewed on GitHub.
- Budget allocated by output — companies review git logs and hand the biggest "slop cannons" effectively unlimited budget (cf. ClickUp's "100x org" framing), which Prime thinks will code some company straight into bankruptcy.
Throughout he ties this to George Hotz's line that adopting AI agents into software dev "will be one of the most costly mistakes in the field's history" — not because agents are bad, but because the spend will delete companies that don't control it.