SIGNAL / NOISE

He's Thirty Minutes Away. He'll Be There in Ten.

Tomasz Tunguz added up a bill on Monday that nobody had bothered to total: roughly $10 billion committed in twelve months, across five companies, to forward-deployed engineering. Human engineers, embedded inside customers, making AI actually work. OpenAI raised $4 billion for a standalone Deployment Company at a $14 billion post. Anthropic pulled $1.5 billion from Blackstone, Goldman, and Apollo. Microsoft stood up a 6,000-person internal deployment army the same week it cut 4,800 jobs. And the LinkedIn data underneath it all: FDE job postings are up 42x since 2023. The industry selling labor replacement is in a bidding war for labor.

You know this character. Jules and Vincent have a body in the car, and a man in a tuxedo shows up at 8:30 in the morning: "I'm Winston Wolf. I solve problems." He takes over the kitchen, runs the cleanup, and is gone within the hour. That's a forward-deployed engineer. And the detail everybody skips: the Wolf is temporary by design. Nobody keeps him on payroll.

Now watch who's buying and who's renting. Alex Karp went on CNBC last week to torch token pricing — "if it was so valuable, wouldn't I say I'll make you $1 billion and I want 30 percent?" — and the actual news he was there to announce was Palantir plus Nvidia: free, open-source Nemotron models running under Palantir's application layer. Fortune caught the self-serving part (outcome pricing is literally Palantir's rate card), but the structure of the trade is the story. Karp can rent a frontier model off the shelf. The labs cannot rent what Karp spent twenty years building: trusted access to the customer's building. That asymmetry is why OpenAI is handing PE investors a 17.5% guaranteed return floor for a route into their portfolio companies. Read the fine print on who gets paid when: the floor pays out whether your deployment works or not.

Here's the pattern the job boards won't tell you. Prompt engineer, 2023, dead inside 18 months. Workflow specialist. Harness engineer. Now the FDE. Every one is an interface job — a human-shaped shim between an immature technology and an organization. History is brutal to shims. "Computer" used to be a job title. Every electrified factory in 1910 had a chief electrician, until electricity became the walls and the title dissolved into facilities. The roles that survive manage permanent complexity, not temporary friction — and a fleet of agents is permanent complexity.

We've seen this movie, and it's older than the movies. When railroads made operations too complex for owner-operators in the 1850s, the economy invented the professional manager — Alfred Chandler called it the Visible Hand, and it built the modern firm. Agents just recreated that coordination problem at machine speed. Yesterday we showed you the agent's mind is readable. Today the job listings agree on what reading it is called. Stop hiring interpreters. Start training managers.

At COAI today: the full Signal/Noise — with the three deployment models, the shim graveyard, and the Chandler argument — is live at getcoai.com.

Which of your new AI titles are scaffolding and which are load-bearing. That's the org-chart exercise we run at Outsider Labs. If the FDE knocking on your door is forcing the question, that's the conversation we're ready for.

Outsider Labs Advisory →

ONE — A NUMBER THAT SUMMARIZES THE DAY

42x — the growth in forward-deployed engineer job postings since 2023, per LinkedIn data, against 13x for traditional AI engineering. The fastest-growing job title in artificial intelligence is a human being who installs it. The labs backed that title with roughly $10 billion in twelve months, because intelligence stopped being the bottleneck and deployment became it. Titles like this are scaffolding, and scaffolding comes down. The Wolf always leaves.

THREE — ACTIONS TO TAKE TODAY

Sort every AI title on your org chart into shim or load-bearing — today. The test: does the role smooth temporary friction (interface work the next model release erodes) or manage permanent complexity (evals, permissions, agent budgets)? Rent the shims — contractors, an FDE engagement, the Wolf. Hire and promote the load-bearing roles, and give them a management track, because that's what they become.

Give one manager an agent headcount before the end of the day. Not a pilot, a span of control: named human, named agents, on the org chart. The instrumentation shipped this week — Anthropic's new enterprise controls give admins cost-versus-output dashboards and model defaults by role. That's management tooling. Treat it like you'd treat a new manager's first budget review.

Reprice any deployment contract on outcomes before you sign it. OpenAI's Deployment Company carries a 17.5% return floor for its PE backers — they get paid whether your rollout works or not. McKinsey says more than 30% of its fees are already priced on client outcomes. Run Karp's own test on every FDE pitch: accepted outcomes, not embedded bodies. If the vendor won't take the bet, that's your answer.

FIVE — STORIES TO KEEP YOU INFORMED

Wednesday, July 8

  • The $10B handshake: every lab is building a Palantir. OpenAI ($4B, TPG-led), Anthropic ($1.5B Blackstone JV), Amazon ($1B), Microsoft (6,000-person Frontier Co.), Google Cloud ($750M) — all now field deployment armies. Ben Evans's Sunday column called the category: automation requires manual labor. (Full analysis above.)

  • Microsoft is quietly routing Copilot prompts off OpenAI and Anthropic. Bloomberg reports tens of thousands of Excel and Outlook prompts moved to Microsoft's cheaper MAI models — which Redmond admits trail the frontier. Same day, Anthropic shipped spend alerts and budget caps for CFOs. The token vendor is now selling the brakes; believe the tape, not the benchmarks.

  • JadePuffer is the first ransomware campaign run end-to-end by an AI. Sysdig documented an LLM that broke in through Langflow, stole credentials, pivoted, encrypted, and wrote the ransom note — and when one exploit failed, it shipped a corrective payload in 31 seconds. Your attack surface is now your AI stack, and your SOC's response time just became the whole game.

  • Anthropic signed a lease that outlives most mortgages. Twenty years, $19 billion, 401 megawatts in Hawesville, Kentucky — more contracted revenue than landlord TeraWulf's entire $12 billion market cap. A bitcoin miner just became an AI utility. The landlord keeps getting paid; the question is what the tenant's pricing power looks like in 2046.

  • Illinois just gave AI safety subpoena-grade teeth. SB 315, signed Monday: mandatory third-party audits for $500M+ labs, $1M fines rising to $3M, passed unanimously, endorsed by Anthropic and OpenAI. First state law to take auditing out of the labs' own hands. Compliance officers, your 2027 calendar just filled up.

MARK TO MARKET

Where the cycle caught up to us this week.

  • Outcome pricing eats the billable hour. The unit of account flips from effort to accepted outcomes — run Karp's 30% question on every vendor (us, Jul 1, "Show Me the Money," email edition) → consultants must reprice from bodies-and-hours to outcomes; McKinsey already books "more than 30%" of fees on client results (Benedict Evans, No. 650, Jul 5).

  • Intelligence is a component, not a strategy. Own the routing table; assemble, don't pledge allegiance (us, Jun 23, "Central Casting") → "What happens when intelligence becomes a component?" — with a P.S. advising clients to go model-agnostic (Shelly Palmer, Jul 7).

The tape doesn't lie. We just read it early.

— Harry and Anthony

Sources:

Reply

Avatar

or to participate

Keep Reading