Block's 24% surge and the Pentagon's Friday deadline are the same story

Two numbers from this week, sitting next to each other:

Block cut roughly 4,000 people — about 40% of its workforce — explicitly citing its internal agent Goose. The stock jumped 24% in after-hours trading. Same week, Defense Secretary Pete Hegseth gave Anthropic until 5:01 PM Friday to agree to "all lawful use" of Claude by the Pentagon, or face Defense Production Act invocation and a supply-chain-risk designation previously reserved for Huawei.

The market and the state are saying the same thing from opposite ends. AI capability has crossed a threshold where it's now worth coercing — by capital allocators who want headcount removed, and by governments who want guardrails removed. The middle position — "we'll deploy AI thoughtfully, with our values intact" — is the one being squeezed from both sides.

The Block math doesn't add up, and that's the point

Block claims Goose saves 8–10 hours per worker per week and eliminates 20–25% of manual work. They cut 40%. No A/B test, no methodology, no throughput numbers. The gap between the productivity claim and the headcount cut is enormous, and the market didn't care. Block reported $6.25B in Q4 revenue with gross profit up 24% YoY before the cuts — this wasn't distress. It was margin engineering wearing an AI costume, and Wall Street wrote the template in real time.

Jack Dorsey's prediction that "the majority of companies will reach the same conclusion within the next year" is the part to take seriously. Not because he's right about the productivity gains — he probably isn't, at the magnitudes claimed — but because every public-company board now has a 24% comp to point at. The capability-reliability gap that researchers keep documenting (agents climbing benchmarks, production reliability flat, ChatGPT Health failing >50% of serious medical cases, frontier models nuking in 95% of war games) doesn't matter to a CFO running the Block playbook. The narrative is sufficient. The numbers don't have to clear.

Which means if you can't answer "what's the FTE-equivalent value of our ML investments" with specifics by March, you're handing finance the pen.

Anthropic is the canary, and the cage is shaking

The Pentagon standoff is the more important story, and it's getting the less coverage it deserves outside policy circles. "All lawful use" sounds reasonable until you remember there are essentially no federal statutes governing military AI, autonomous weapons, or AI-assisted surveillance. In a legal vacuum, "all lawful use" means "all use." Google, OpenAI, and xAI have already agreed. Anthropic — holder of the only classified Pentagon AI contract — said no.

And then, on the same day, Anthropic quietly dropped its core Responsible Scaling Policy commitment. Jared Kaplan called maintaining it "unilateral disarmament." The company that built its brand on safety credentials abandoned its signature safety commitment while publicly refusing the government — a dual-track move that should make any enterprise customer who paid the safety premium look at their procurement contracts again.

The precedent is what matters. If the DPA — a Korean War munitions law — successfully gets used to compel a software company to modify its product for state use, then encryption backdoors, cloud access mandates, and algorithm modifications are all now legally plausible. Every AI vendor's acceptable-use policy stops being a control and becomes a suggestion. Multi-vendor isn't a best practice anymore; it's a hedge against your provider being conscripted.

The infrastructure beneath both stories is rotting

While this is happening, the trust model underneath your stack is failing in ways that compound the AI risk. CVE-2026-20127, a CVSS 10 zero-day in Cisco Catalyst SD-WAN, was exploited for three years before the Australian Signals Directorate — not Cisco, not any customer — found it. Chinese APT UNC2814 ran C2 through Google Sheets across 53 organizations in 42 countries for up to nine years, indistinguishable from normal business traffic. Truffle Security found 2,863 live Google API keys silently escalated into Gemini credentials because enabling the API on a project retroactively granted access to every existing key — including keys Google's own docs explicitly told you were safe to embed in client-side JavaScript.

Meanwhile data-theft-only extortion just passed ransomware at 57% of cyber insurance claims. Your backups are defending against the previous war.

What to do this week

One thing, this sprint, that builders should not push: build the FTE-efficiency dashboard for your own team before someone in finance builds it for you using Block's numbers. Specifics matter — tasks automated, hours saved by role, error rates against human baseline, quality-adjusted output. Survey ICs and managers separately, because the productivity asymmetry between them is where the busywork-vs-value gap hides. If you only get one number, get the one that translates AI usage into FTE-equivalents on workflows you can name.

That number is the only thing that competes with the Block narrative when it lands in your CEO's inbox. Vague "AI is making us more productive" is what gets headcount cut. Quantified contribution by team, instrumented monthly, is what gets it kept.

And while you're at it: audit your GCP keys, patch your edge devices, and stop assuming your AI vendor will be allowed to keep saying no.