What Would You Say... You Do Here?

3.5x — what 95th-percentile firms now consume in AI per worker, per OpenAI's first enterprise B2B Signals report released yesterday. That ratio was 2x a year ago. Twelve months from now it's 5x. Frontier firms send 16x as many Codex messages per worker. They are not buying more seats. They are using each seat to do more. The companies running with the right answer to "what would you say you do here" are pulling 3.5x ahead of the companies still measuring AI by license utilization. Bob and Bob are walking the floor again. They have data this time.

Stop measuring AI adoption in seats. Pull the depth quintile by Friday. OpenAI's B2B Signals report says message volume explains only 36% of the productivity gap between top firms and the rest. The other 64% is depth — long prompts, multi-step delegated workflows, agent supervision. Cisco's case in the report: 1,500 engineering hours saved per month, defect-resolution throughput 10-15x, build times cut 20%. Run the same audit internally today. Sort engineers by token-spend-to-output ratio, not by license count. Your CFO can read the bottom-quintile slide before the weekend.

Audit your wage-premium headcount against your AI-depth chart. The MIT Acemoglu paper that hit this morning says automation 1980-2016 explains 52% of US income inequality, 60-90% of productivity gains were burned offsetting wage premium, and the higher the wage the more attractive replacement becomes. DeepL fired 25% of its workforce today citing AI. Cloudflare cut 1,100. AI is now 26% of April's job cuts. The intersection of high-wage plus low-AI-depth is where your CFO is going to start cutting next quarter — unless you've moved that worker to the high-AI-depth column first.

Write the depth-gap slide before your investors ask for it. Q3 earnings calls are six weeks away. Anthropic admitted 80x Q1 growth on stage Wednesday — that supply curve is what's funding the productivity bifurcation everyone else is buying. Public-company analysts are about to start asking some version of "what percentage of your engineering org is operating in the top depth quintile, and how fast is that count growing?" The CEO who can answer with a real number wins the multiple. The CEO who answers with seat-count utilization gets the downgrade. Build the chart that works whether the analyst asks or not.

Today's actions touched the same questions we've been working with clients on this quarter — auditing engineering orgs by depth percentile, designing the role re-org that gets headcount math caught up to the technology, and writing the metric that survives the Q3 earnings call. If you're staring at "where do I start" and don't have the answer, that's the conversation we're built for.

Outsider Labs Advisory →

Thursday, May 7

1. OpenAI just published the depth-gap receipts and almost nobody named the number. (Full analysis below.) The first B2B Signals report says 95th-percentile enterprises now use 3.5x as much AI per worker as typical firms — up from 2x a year ago — and 64% of the gap is depth, not message volume. Frontier firms send 16x more Codex messages per worker. Cisco saved 1,500 engineering hours per month and treats Codex "as part of the team." The seat-count metric every IT department reports up the chain just got reclassified as a 36% explainer. (Full analysis below.)

2. DeepL fired 25% of its own workforce — citing AI — the same day OpenAI's chart landed. (Full analysis below.) 250 jobs gone at the company built on wrapping Claude and GPT for translation, with the press release framed as operations efficiency. The unwritten part: Anthropic and OpenAI now ship translation as a frontier-capability side-effect. Every "operations" press release in the next twelve months is going to have a product-line obituary buried underneath it. Read both lines. Cloudflare cut 1,100 the same week. Challenger Gray's April report names AI as 26% of the month's job cuts.

3. The MIT Acemoglu paper says this is exactly how 1980-2016 played out. (Full analysis below.) Daron Acemoglu and Pascual Restrepo published Automation and Rent Dissipation in the May Quarterly Journal of Economics: automation explains 52% of US income inequality growth since 1980, with 60-90% of productivity gains burned on wage-premium replacement rather than output. "The higher the wage of the worker, the more attractive automation becomes to firms." a16z published a bullish chart this morning saying engineer demand is rising. The fine print says wage growth is rising in AI-augmented systems-design roles and the substituted roles are the ones declining. Same paper, two narratives. Yours depends on which side of the bifurcation your headcount sits.

4. Zyphra shipped a frontier-competitive model trained entirely on AMD. ZAYA1-8B (760M active parameters, 8.4B total, MoE) matches DeepSeek-R1 on AIME 2026 and HMMT, competitive with Claude Sonnet 4.5 on reasoning, closes in on Gemini 2.5 Pro on coding. Trained end-to-end on a 1,024-node AMD MI300X cluster built with IBM — the first frontier-tier model end-to-end on AMD silicon. Pairs with Lisa Su's Wednesday-earnings call: agentic workloads shifting CPU:GPU ratio from 1:8 toward 1:1. The trillion-dollar Anthropic mark is built on NVIDIA-only economics. Two architectural alternatives shipped in ten days (SubQ on Tuesday, ZAYA1 today). The David Silver thesis from Heat keeps printing.

5. China's National AI Industry Investment Fund is putting state money into DeepSeek at $50 billion. Per TLDR-AI today, DeepSeek is in talks to raise from China's $8.8 billion government-backed fund at a $50B mark. Moonshot AI (Kimi) closed a Meituan-led round at $20 billion the same week, quadrupling its valuation in months. Foreign Policy's Wednesday cover argued cutting-edge US frontier models are too expensive for much of the world. The China hedge has receipts now. While every US newsletter today wrote about the SpaceX-Anthropic harness deal, the open-weights, AMD-trained, state-backed alternative stack got two valuation prints and a global-distribution pitch.

Bob And Bob Are Walking The Floor

Yesterday OpenAI released its first B2B Signals report. Firms in the 95th percentile of AI use now consume 3.5x as much intelligence per worker as the typical firm. A year ago that gap was 2x. Message volume explains only 36% of it. The other 64% is depth — longer prompts, richer context, multi-step delegated workflows. Frontier firms send 16x as many Codex messages per worker. They are not buying more seats. They are using each seat to do more.

This is the structural bifurcation we've been writing toward for six weeks. Heat (Apr 30) said the smart money is hedged. Porsche In The Driveway (May 3) said you bought the Porsche and you're driving it like a bicycle. I Drink Your Milkshake (May 4) said the labs are walking into your operations through your PE sponsor's checkbook. Name Of The Game (May 5) said the trillion-dollar aftermarket is three layers of buyers running a brokerage script. No One Sets Off My Evil Detector (yesterday) said the harness offensive at Code with Claude was the lab's answer to Dario's 80x Q1 admission. All of it points to the same place. The AI-native company is not a company that uses AI. It is a company that has reorganized around AI — different role definitions, different headcount math, different review process — and the productivity gap that re-org produces is now measurable at 3.5x and widening.

What's new today is the bookend. Same news cycle: DeepL fires 25% of its workforce citing AI. Cloudflare cuts 1,100. AI is now 26% of April's job cuts per Challenger Gray. Daniela Amodei stands on the Code with Claude stage and declares Claude's transition from "chatbot" to "colleague" complete. MIT just published the Acemoglu paper in the May print issue of the Quarterly Journal of Economics, arguing automation 1980-2016 was used to control wage premium, not boost productivity, and 60-90% of productivity gains were burned offsetting that targeting. The bifurcation is not a contradiction. It is the pattern. The productive companies are getting 3.5x. The non-productive companies are firing 25%. The wage-premium worker is exactly who AI replaces first. Acemoglu's paper is not a forecast. It is the chapter that comes before the one we are now living in.

a16z dropped its bull case yesterday — "the AI job apocalypse is a complete fiction" — twenty-six charts arguing engineer demand is rising and AI is more job-maker than job-taker. They are right at the level they're writing. They are also gracefully glossing over the fingerprint: AI-exposure is driving above-trend wage growth in systems design, hiring is stronger in AI-augmented roles than in AI-substitute roles, and the roles in decline are customer-service representatives and medical transcriptionists. The aggregate count rises because the BLS classifies the agent supervisor and the 1x coder under the same job title. That same line item means a different thing than it did three years ago. Bookkeepers did not disappear when Excel arrived. They rebranded as FP&A analysts at higher comp. What disappeared was the typist. The same compression is rolling through software in 2026. Aggregate engineer count rises. Aggregate engineer wage rises. Output per engineer rises 3.5x. The headcount you are about to lose is the headcount in the middle of your salary band that did not yet learn to drive the agent fleet.

The supply side of the gap is real. Anthropic's Dario Amodei admitted Q1 grew 80x against a 10x plan. The Information reported Anthropic now accounts for more than 40% of Google Cloud's $462 billion revenue backlog. One private company is 40% of the largest cloud-services backlog ever assembled. Three quarters from now, the earnings-call analyst question is going to be a standing item. It will sound like a softball. It will not be one. "What percentage of your engineering org is operating in the top depth quintile, and how fast is that count growing?" The CFO who answers with a real number wins the multiple. The CFO who answers with seat-count utilization gets the downgrade. The 3.5x gap is the new gross margin.

The 3.5x is invisible on a seat-count report. It is the most visible thing on a competitor's payroll line. Bob and Bob are walking the floor. The question is what would you say... you do here.

At COAI today: Full briefing — What Would You Say... You Do Here? — at getcoai.com.

— Harry

Sources: