The week the AI bill came due — and the receipts didn't reconcile

Three numbers, same week. Microsoft down 34% since October — its worst quarter since 2008, driven by shareholders who stopped accepting "AI capex now, returns later." Rate-cut probability flipping from 90% to a 52% hike probability inside thirty days, the fastest sentiment reversal since 2022. And H100 rental prices, which every cloud and every startup financial model assumed would depreciate on a 4-7 year curve, now trading above their October 2022 launch value.

That last one is the quiet bomb. Standard depreciation schedules underwrite GPU clouds, data center deals, every inference cost projection in your pipeline, and the unit economics of every AI-native company you've funded or work at. Reasoning models and agent workloads turned four-year-old chips into appreciating assets. If your 2026 plan assumes compute gets cheaper, your 2026 plan is wrong.

Meanwhile, at JPMorgan's Tech100, Jack Dorsey told a room of capital allocators that using Block's open-source coding agent Goose for a few hours each morning convinced him he could halve his engineering workforce. Ali Ghodsi at Databricks described the same arc independently. Two CEOs, two $40B+ companies, same conclusion, on the record, in front of investors. That's not a productivity-tool pitch — that's a labor-arbitrage thesis being entered into the public record.

The contradiction is the story

Here's where it gets uncomfortable. The capex-revolt narrative says AI ROI is overhyped and shareholders want their money back. The Dorsey narrative says AI is already so productive at the CEO desk that workforces should compress 50%. Both can't be fully true. The resolution everyone is dancing around: capability is real, the productivity gains are real for some tasks, and the economic distribution of those gains is unsettled. Markets are pricing both "AI doesn't pay back" and "AI eliminates half your headcount" simultaneously, and the dispersion inside the Mag 7 — MSFT -34%, Meta -29%, NVDA -20% — is the market trying to figure out who captures the value and who pays for it.

The research undercuts the simple version of the CEO story. AI tools increase competition entry by 42% with no measurable improvement in individual success rates. LLM-generated code ships vulnerable 30% of the time. Chatbots side with users 49% more than humans do, which means LLM-as-judge evaluation pipelines are systematically biased. AI agents now have persistent shells, browser sessions, and Box plugins, but no IAM model treats them as the privileged service accounts they actually are. We are deploying the productivity story into a measurement vacuum — and the people doing the deploying are CEOs working off their own anecdata.

The escape valves opened the same week

While frontier compute got more expensive, the open layer closed the gap. GLM-5.1 scores 45.3 on coding evals against Claude Opus 4.6 at 47.9 — a 5.4% gap that was 30% twelve months ago. Quantized Qwen 3.5-9B runs on a 16GB MacBook Air with 20K context. RotorQuant's Clifford Algebra trick cuts quantization from 16,384 FMAs to roughly 100 — call it a 160x reduction — with cosine similarity at 0.990 versus TurboQuant's 0.991, shipping today as fused CUDA and Metal kernels. A three-line KV dequant change buys 22.8% decode speed at 32K context for anyone running long-context inference.

Validate the benchmarks on your own evals before you ship anything — community-reported numbers aren't peer-reviewed. But the direction is clear. The barbell — frontier labs getting bigger and more expensive, open weights getting good enough on consumer hardware — squeezes the middle. Mid-tier closed-API vendors without hyperscaler backing are the most exposed category in software right now.

Anthropic is the canary. The leaked Capybara tier hints at the next capex escalation. Anthropic is already throttling existing API customers while licensing Claude to power Yahoo Scout's 250 million users — meaning your production traffic now competes with a consumer platform for the same constrained capacity. OpenAI killed Sora overnight and torched a $1B Disney partnership in the process. If Disney's contracts didn't survive an AI vendor's mood swing, yours won't either.

The security backdrop nobody priced in

Iranian APT Handala compromised FBI Director Kash Patel's personal Gmail this week — TechCrunch verified the leak via DKIM signatures. The same group recently wiped tens of thousands of devices at Stryker. CISA is operationally degraded by the DHS funding lapse. LiteLLM, downloaded 3.4M times a day, was found shipping credential-harvesting malware that Karpathy assessed as itself AI-generated. Compliance certifications are eroding as a trust signal — Delve received SOC2 and ISO27001 with allegedly fabricated audit data. The defensive perimeter is thinning at the exact moment that 30%-vulnerable AI-generated code is replacing the humans who would have caught it.

What to do this week

Three moves, in this order.

First, run your own agent productivity audit before someone above you does it with worse math. Pick one engineering team. Instrument tasks-completed, time-to-merge, and lines-per-session with and without coding agents for two sprints. Bring the data to your next planning review. The goal isn't to defend headcount — it's to make sure the headcount conversation happens on numbers you generated, not on a CEO's morning anecdote.

Second, re-underwrite your compute. Pull every GPU contract, every cloud commitment, every inference-cost line in your model. Replace the declining-cost assumption with flat-to-rising. Benchmark RotorQuant and the KV dequant fix against your current quantization stack on your eval suite this sprint — a 22.8% decode speedup from a three-line change is the highest ROI optimization available, and it's free. Build a multi-provider abstraction layer if you're single-sourced on Claude; you're sharing capacity with Yahoo now.

Third, treat AI agents and AI-generated code as the privileged, untrusted entities they are. Mandatory SAST gate on every PR, no exceptions for AI-authored code. Inventory every persistent agent integration — Codex plugins, Box connectors, browser sessions — and apply service-account-grade IAM. Survey C-suite personal email exposure and enforce FIDO2 hardware keys this week. The FBI director's personal Gmail wasn't hardened. Yours probably isn't either.