Compute got expensive, capital got expensive, and CEOs started doing headcount math

Three numbers from this week that don't fit the standard model of 2026: H100 rental prices are now higher than at their October 2022 launch. Microsoft is down 34% since October, its worst quarter since 2008. Fed futures flipped from 90% odds of a rate cut to 52% odds of a rate hike — in thirty days.

The assumption underneath every AI budget written in the last two years was cheap money plus falling inference costs. Both halves broke this week. And in the middle of that break, Jack Dorsey stood up at JPMorgan's Tech100 and told the room that using Block's Goose agent a few hours each morning convinced him he could cut engineering headcount by roughly half. Databricks' Ali Ghodsi described the same experience unprompted. Two CEOs of $40B+ companies, same conclusion, on the record, to the people who allocate capital.

That's the story of the day. The rest is context for what you do about it.

The compute floor moved up

Start with the H100 number, because it's the one most people will get wrong. GPUs are supposed to depreciate on a 4–7 year curve. That's the assumption underwriting every data center deal, every GPU cloud unit economic, every "inference costs will fall" slide in every AI startup deck. It's now wrong. Reasoning models and agent workloads want longer contexts, bigger KV caches, more concurrent sessions — and the chip shortage compounded on top. Four-year-old silicon is a more productive asset than it was on ship day.

The efficiency side is real but it's not a rescue. RotorQuant's Clifford Algebra rotors cut quantization FMAs from 16,384 to about 100 — a 160x reduction, 10–19x faster than Google's TurboQuant, with cosine similarity within 0.001. That ships today as fused CUDA and Metal kernels. TurboQuant's own three-line KV-dequant sparsity trick gets you +22.8% decode at 32K context. Open weights closed the coding gap to 5% — GLM-5.1 at 45.3 against Claude Opus 4.6 at 47.9, up from a 26% gap on the prior generation. Qwen 3.5-9B runs on a 16GB MacBook Air with 20K context.

So the barbell: frontier compute is getting more expensive and more capacity-constrained (Anthropic is throttling paying customers while onboarding Yahoo Scout's 250M users), while the "good enough" tier is getting radically cheaper and easier to self-host. The margin squeeze lands on anyone in the middle — mid-tier closed API vendors, and any product whose unit economics assumed the frontier would stay subsidized.

Yes, but — one honest counter-reading: Google funding Anthropic's data center buildout and SoftBank's $40B bridge loan for OpenAI mean the capex arms race isn't slowing, and once new capacity lands in 12–18 months, unit costs at the frontier probably do fall again. Fair. Your Q2 planning still has to survive the interim.

The Dorsey moment is the real news

The headcount claim is easy to dismiss as CEO theater. Don't. What matters isn't whether Dorsey actually cuts Block by 50% — it's that he said it publicly, at Tech100, to investors who now have that number in their heads when they look at every other portfolio company. Within two quarters, every board in tech will ask some version of "what's our Goose plan?"

The supporting evidence is thinner than the confidence. Research this week put LLM-generated code at a 30% vulnerability rate. A separate study found AI tools boost competition entry 49% without measurably improving individual success rates. Sycophancy runs 49% above human baseline. LLM-as-judge diverges from human graders in ways nobody has fully characterized. The CEO-level enthusiasm and the measured quality are moving in opposite directions, and nobody in the boardrooms is looking at the measurements.

That gap is the opening. If you run an engineering org and you wait for the top-down cut, you get cruder math than you deserve. If you instrument your own team's agent-augmented output now — tasks completed, time-to-merge, review coverage, defect rates on AI-authored code — you own the narrative when it reaches your board. If you don't, someone with a spreadsheet and a Dorsey quote does.

The security posture didn't get the memo

One piece of context that reframes the headcount conversation: Iranian APT Handala compromised FBI Director Kash Patel's personal Gmail this week. TechCrunch verified the leaked messages via DKIM signatures. Same group executed destructive wiper attacks on Stryker earlier in the month — tens of thousands of medical devices bricked, not encrypted. CISA is operationally degraded by the DHS funding lapse. LiteLLM, at 3.4M downloads a day, was found shipping credential-harvesting malware that Karpathy assessed was itself AI-written.

The convergence: AI-generated code at 30% vulnerability, vendors planning to fire the humans who'd catch it, foundation model labs quietly commoditizing entire security categories (the rumor of Anthropic's cyber-capable model was enough to tank security equities last Friday), and state-sponsored actors escalating from espionage to destruction. Your attack surface is expanding faster than the workforce protecting it, and the compliance signals you'd normally rely on — Delve reportedly got ISO27001 with fake audit data — are degrading in parallel.

What to do this week

One action, specific, this week: pick your three highest-spend inference workloads and re-run their unit economics under two scenarios in parallel. Scenario one: H100 rates hold or rise 20% through year-end, Anthropic throttling continues, your current API mix stays. Scenario two: you route the non-frontier portion (call it 60–70% of volume for most coding and summarization work) to a self-hosted Qwen 3.5-27B or GLM-5.1 stack with RotorQuant or TurboQuant KV sparsity applied. Put the delta in dollars, per correct completion, on one page.

That page is what you take to your CFO before the next budget review — the one where "AI" stops being a magic word. It's also the page that lets you answer the Dorsey question with data instead of vibes when your board asks. And it's the single artifact that makes the case for keeping the security review headcount you're about to be pressured to cut, because you can show the margin you just recovered from the optimization work funds it.