TL;DR

Three days after I argued that 2026’s layoffs were burning institutional knowledge to fund AI bets, the count crossed 150,000 and the first receipts started landing: a team that swapped 12 senior engineers for 8 juniors with Copilot, then watched the codebase rot six months later. The savings were real and immediate. The damage is real and lagging. That lag is exactly why this keeps happening — and why the cuts that look smartest this quarter will look worst by next year.

The Number Moved While I Was Writing

When I published The AI Headcount Panic on May 27, the running total was around 114,000 tech jobs cut in 2026. By the time the post had been up a few days, layoff trackers were past 150,000 across more than 500 companies — Oracle’s 20,000–30,000, Amazon’s 16,000, Dell’s 11,000, Meta’s 8,000, Block’s 40% gutting. A single day, May 20, accounted for roughly 11,000 on its own.

I’m not relitigating the cover-story argument here. The receipts in the last post stand: Meta’s capex pivot, Intuit’s CEO saying flat out it had nothing to do with AI, Wix’s spend-discipline correction dressed up as transformation. What changed in the last 72 hours is that the consequence I warned about stopped being a prediction.

The Lag Is the Whole Story

A developer’s post made the rounds in March: her employer cut 12 senior engineers and backfilled with 8 juniors armed with Copilot and Claude. Six months later, the codebase was in serious trouble. Her line — that the savings were real but the institutional knowledge was gone — got shared thousands of times, mostly without comment, because there’s nothing to argue with. It’s just the bill arriving.

This is the part the spreadsheet can’t model. Headcount savings show up this quarter. The cost of losing the person who knew why the retry logic had that weird 800ms backoff, or which customer’s contract has the undocumented data-residency clause, shows up two, three, four quarters later — as an outage, a botched migration, a compliance miss. By then the engineer who’d have caught it in a code review is two jobs away.

Here’s the contrarian part for the people running these decisions: the AI didn’t fail. The juniors didn’t fail. The tooling did exactly what it promised — it let 8 people produce roughly what 20 used to. What it can’t produce is the judgment about which 20% of that output is about to detonate. That judgment was never in the repo. It was in the heads you just expensed.

Why “Smart” Cuts Are the Dangerous Ones

In the rooms where this gets decided — and I’ve sat in a few across Tikal engagements — the most defensible-looking cut is the senior IC who “isn’t shipping that much.” Low commit volume, lots of meetings, mentoring, design reviews. On a velocity dashboard they look like overhead. They’re actually the error-correction layer for everyone else.

Cut the obvious dead weight and you lose a little. Cut the quiet senior who reviewed every risky PR and you lose the thing that kept the juniors-plus-Copilot setup from compounding its own mistakes. The org that did the “smart, surgical” reduction often did the most expensive one — it just won’t see the invoice until the next incident review.

What I’d Actually Do

Not “don’t adopt AI” — adopt it harder, at the level of your real expertise, which was the whole point last time. But if you’re cutting:

  • Map the context, not the headcount. Before a layoff list is final, ask who is the only person who knows X. If the answer to “who else could review this subsystem” is one name, that’s not a cost line, that’s a single point of failure you’re about to remove.
  • Make AI the capture tool, not just the replacement. The same models writing the code are very good at interrogating a departing engineer and turning “why is it like this” into written ADRs and runbooks. If you’re going to lose the person, at least extract the context first. Almost nobody is doing this.
  • Watch the lag indicators. Incident MTTR creeping up, review cycles lengthening, the same bug class recurring — those are the knowledge bill arriving, and they show up before the quarter does.

Conclusion

The headline number will keep climbing, and the announcements will keep crediting AI. The honest version is that a lot of organizations are running an uncontrolled experiment on their own institutional memory and won’t get the results back for a year. The companies that look reckless right now — protecting their senior engineers, treating AI as leverage instead of a layoff line — are the ones who’ll still have a defensible product when everyone else is reverse-engineering their own codebase.

The bill always comes due. AI just changed the payment terms to net-180.