A 13-person startup just got a $600M valuation because your AI bills are out of control

Uber burned through its entire 2026 AI budget by April.

Meta employees burned 73.7 trillion tokens in a single month.

Both companies had to slam on the brakes.

Engram just exited stealth with $98M and a $600M valuation — for 13 people.

The reason: when a model reads a 70,000-word contract, its internal memory balloons past 100 gigabytes.

250,000x the size of the original file.

And it rebuilds that from scratch on every single query.

This is not a pricing problem.

This is an architectural one baked into every transformer-based model you're deploying.

Engram's bet: pre-train a compressed memory on your organization's corpus once, then load it on every query instead of relearning everything.

Microsoft is testing it inside Microsoft 365.

Notion is piloting it in custom AI agents.

Harvey is exploring it for legal document environments.

The $600M valuation for a company with no public product and unverified claims tells you where the market thinks the real problem is.

Not model quality.

Token governance.

Your CFO is about to make AI costs a board-level conversation.

Audit your token consumption by function right now.

If your agents are re-reading the same documents on every query, you're burning money on architecture that was never designed for enterprise scale.

---