Anthropic, OpenAi And The Name Of The Game

$1 trillion — Anthropic's reported aftermarket valuation this week, with bid sheets at major private-wealth desks pricing the lab roughly 2.6× its $380B mark from earlier this year. The math underneath is the indictment: a venture-grade 10× return on a $1T entry requires Anthropic to become a $10 trillion company. The most valuable company that has ever existed is roughly $5T. Buying Anthropic at the trillion is a bet that the lab ends the decade twice the size of anything in human history. Mark Hanna would have ordered another martini.

Re-budget your AI deployment from tokens to operators today. Three independent operators published the same number from different angles this morning. Reflex showed a vision agent costs 45× more in tokens than the same model running through a structured API — same Claude Sonnet, the gap is set by the integration choice. Lemkin at SaaStr put two AI VPs on the books at $254/month combined — and immediately walked the headline back with a 0.5 FTE-per-active-agent maintenance rule. Tunguz's structural fact: a three-engineer team running 20 agents loses 33% of its institutional memory the day one engineer quits. Tokens are cheap. The engineer who tunes the agent is the line item. Whoever is selling you a multi-year managed-services contract is selling you a staffing agency dressed as a model license. Budget for that.

Do the $10T arithmetic out loud before you sign any AI SPV pitch this week. Goldman, Morgan Stanley, and JPMorgan private wealth are calling $5M+ balance clients with marked-up secondary access to Anthropic and OpenAI on top of management fees on top of carry. Run the math on the call. Anthropic at a $1T entry needs to clear $10T to deliver a venture-grade return. The most valuable company that has ever existed is roughly $5T. If your fiduciary cannot say out loud that the bet only works at a market cap higher than any company has ever reached, the bet has not been priced. Decide. Tell your spouse the answer. Tell your fiduciary the answer.

Read your AI vendor's product names against what they actually do — and make your team do it too. Anthropic shipped ten financial-services agents Tuesday with names like Pitch Builder, Earnings Reviewer, Statement Auditor, Month-End Closer, Know Your Client. Those aren't software product names. They're job titles. The procurement category Anthropic just opened is AI staffing, not AI software — and the SaaS vendors missed it for two years. The mirror exercise is the one that lands: the same architecture, deployed inside an insurance company, would honestly be named Rejected: Small Print, Rejected: Not Covered, Rejected: No Pre-Clearance, Rejected: High Deductible. Same agent. Honest naming. The agent's name is the business model. Read every vendor's roadmap with that lens before the next contract review.

Today's actions touched the agent-deployment math and the SPV pitch our HNW readers are getting in the inbox. Both are work we have done with clients this quarter — sizing the operator overhead under a managed-services contract and pressure-testing the math on lab-secondary vehicles. If your team is staring at "what does this actually cost us" and isn't sure where to start, that's the conversation we're built for.

Outsider Labs Advisory →

Tuesday, May 5

1. Active Directory, all over again — and the lab just got commoditized one layer further. Microsoft Agent 365 went GA on May 1 at $15 per user per month standalone, or bundled into the new Microsoft 365 E7 tier at $99 per user per month — the same 15% bundling discount that has anchored Microsoft's enterprise lock for two decades. Days later at Cloud Next, Google countered with Agent Identity (a cryptographic certificate per agent) and Agent Registry (a central library of every internal agent and skill). Whoever owns agent identity owns agent policy, owns agent data access, owns the enterprise. Ten thousand seats on Agent 365 is $1.8M a year before you have run a single agent. The model brand just got commoditized one layer further. The lab is no longer the choke point. The directory service is.

2. The diary that wasn't supposed to be a court exhibit. Greg Brockman's testimony this week put his OpenAI stake at roughly $30 billion in the public record, and his own 2017 personal journal — entered as evidence — described the for-profit conversion as "pretty morally bankrupt." Same week, the prosecution surfaced the framing that Elon Musk left OpenAI in 2018 because he gave the company a 0% chance at AGI and wanted the work moved into Tesla as a secret project hidden from shareholders. The two co-founders of OpenAI both have private writings from 2017 that contradict their current public positions — and that record entered the file while the company is being marked at $852 billion on the secondary tape. Every CIO on a 2026 vendor map should ask one question this week: what is in your vendor's working journal, and where is it stored. (Full analysis below.)

3. Coinbase joins the AI-cited layoff list. Brian Armstrong cut roughly 700 jobs Tuesday — 14% of Coinbase headcount — and explicitly tied the move to AI-driven productivity. Crypto-natives were not on the bingo card for this round. They are now. Same week: Microsoft's 2026 Work Trend Index found that 65% of workers fear falling behind on AI while only 13% say their company rewards experimentation. That gap — between the fear and the incentive — is the actual labor story of 2026, not the headline cuts. The senior people who survive these rounds are about to discover their leverage has compounded. The hires Coinbase keeps are the hardest to replace in the market.

4. Old AI keeps beating the doctor — and Goldman is already in the trade. Mayo's AI reportedly identified cancer roughly three years before the human pathologists did, per a study surfaced this week. Same window, AIdoc raised $150M Series E led by Goldman Sachs with Nvidia Ventures participating — AI medical imaging that helps physicians read scans faster. Same Goldman that led the Anthropic JV. Same Goldman that arranges the SPVs. The health-AI implementation play is one trade with the consulting play, and the lead arranger is the same firm. If you are running a health system in 2026, your procurement question is not "is the model good enough." It has been good enough for two years. Your procurement question is whose forward-deployed engineers walk in with it, and what the SPV around them looks like.

5. Jack Clark put a 60% probability on automated AI R&D by end of 2028, on the record. Anthropic's policy chief — the person whose job is publicly managing the safety message — published an essay yesterday arguing the public benchmark trends imply a roughly 60% chance that a frontier model autonomously trains its successor by the end of 2028, with a 30% probability by end of 2027. Same week, Continual Learning Bench 1.0 shipped as the first benchmark designed to measure self-improvement directly. The recursive-R&D claim is now falsifiable. If you are signing a 36-month contract with a lab whose Director of Policy has publicly placed a 60% bet that the model in your contract will be designed by an earlier version of itself with no human in the loop, you have a different kind of vendor risk than you had last week. Plan for the bridge.

Mark Hanna Would Have Raised His Glass

Anthropic is trading at a trillion dollars in the aftermarket. To make money buying it there at the venture-grade returns the firms writing those checks have to clear, the lab has to become a $10 trillion company. The most valuable company that has ever existed is roughly $5T. Buying Anthropic at the trillion is a bet that the lab ends the decade twice the size of anything in human history. Mark Hanna's three-act script in the diner scene of The Wolf of Wall Street is now playing inside every private-wealth office in America. The labs are running the pitch. The labs are also running the table.

The casino has three layers of buyers. The labs are pot committed — none of them can lower a valuation without triggering a death spiral on the prior round, none can stop announcing features without losing the next round. The institutional investors are averaging up to defend prior marks: writing the next check at $1T is the only way to keep their existing $50B mark from going to zero. The retail money is reaching — Goldman, Morgan Stanley, JPMorgan private-wealth desks calling $5M-balance clients with marked-up SPV access on top of carry on top of management fees, plus the public IPO behind it. Pull any chair out and the entire structure collapses, which is why nobody pulls any chair. The prisoner's dilemma is not at the lab level. It's the architecture of the trade.

Hanna's three movements run in order. First, admire the mission — the labs sell AGI, alignment, safety, "too dangerous to release," every frontier release a "new idea, special situation, special idea" delivered on a capital-raising cadence rather than a product-development one. Then, the reveal. Greg Brockman's testimony this week put his OpenAI stake at roughly $30 billion in public record alongside his own 2017 journal calling the for-profit conversion "pretty morally bankrupt." Every other senior partner in OpenAI's history has walked. The one who stayed is the one Sam Altman cut into a personal vehicle in 2017. Meanwhile the Anthropic-Goldman-Blackstone-Hellman & Friedman joint venture announced Monday is — literally, not metaphorically — a special-purpose vehicle. Goldman takes an arranging fee. Blackstone takes an implementation fee. The JV bills the operating company. Anthropic books the inference. Same engineers powering all three magic acts. Three valuation bumps, one cause.

Then, the meat. The agent has a job title. Pitch Builder. Statement Auditor. Month-End Closer. Know Your Client. The procurement category Anthropic just opened is AI staffing, not AI software. The honest version of that pitch — written by anyone who has spent forty minutes on hold appealing an insurance denial — would name the agents differently. Rejected: Small Print. Rejected: Not Covered. Rejected: No Pre-Clearance. Same architecture. Honest naming. The token, by the way, is not where the cost lives. Reflex showed a vision agent costs 45× more than an API agent on the same Claude Sonnet — the gap is set by the integration. Lemkin runs 20+ agents with 3 humans against his own 0.5-FTE-per-agent rule that says he should have ten. Tunguz showed a 33% institutional-memory loss when a three-engineer team running 20 agents loses one engineer. Tokens are cheap. The operator is expensive. The labs figured this out twelve months ago. That is why they are launching consulting firms instead of cutting prices on inference.

There is one cleaner version of this trade running on a different field. SpaceX confidentially filed for IPO on April 1 at a $1.75-2T target with $50-75 billion of new capital — the largest IPO in history. The board approved Musk's pay package in January and it is now in the IPO filing: 200 million super-voting restricted shares contingent on $7.5 trillion market cap and one million people on Mars, with a second tranche tied to 100 terawatts of orbital compute. Musk just literally wrote "$7.5 trillion" into a board comp package. SpaceX needs 4.3× to clear that mark. Anthropic needs 10× to clear $10T. SpaceX has Starlink ($20B revenue, $14B EBITDA), defense lock-in, and xAI inside it. Anthropic has software margins and counterparty risk to AWS and Microsoft. Musk delivers — Falcon reusability, Tesla's $50B-to-$1T arc, Starlink. The man who said "we will go to Mars" has the more credible bet on $7.5T than the man who said "we will achieve AGI" has on $10T. And the more credible bet is itself a bet that the company being IPO'd ends the decade fifty percent above the most valuable company in history.

The pitch every consumer-facing financial product has run for fifteen years is the same: we are making this easier. DraftKings made it easier to bet on sports. Polymarket made it easier to bet on every public outcome. Robinhood made it easier to trade — and the customer who paid $9.99 to E*Trade now pays nothing on the line item and roughly the same in slippage to Citadel and Jane Street. In every case, "easier" was the seller's pitch, not the buyer's outcome. AI is now running the same play at the operator layer. Cursor and Codex and Claude Code make it easier to ship code. Anthropic ships ten financial-services agents on a Tuesday morning to make it easier to deploy AI inside your finance department. Soon, the SPV in the inbox makes it easier to invest in AI. The seller books the friction as revenue. The buyer didn't see the fee.

The casino is open twenty-four hours a day, three hundred and sixty-five days a year. The lab is on one side of the trade. You are on the other. The name of the game has not changed since 1993. Mark Hanna would have raised his glass and ordered another martini.

At COAI today: Full briefing — Anthropic, OpenAI And The Name Of The Game — at getcoai.com.

— Harry

Sources: