ONE — A NUMBER THAT SUMMARIZES THE DAY

220,000 — the count of NVIDIA H100, H200, and GB200 GPUs that Elon Musk leased to Anthropic this morning, the company he called misanthropic in February. 300 megawatts of new compute, doubled Claude Code rate limits the same afternoon, and an exploratory partnership on multi-gigawatt orbital data centers that ties directly to Musk's $7.5 trillion comp-package vest. "No one set off my evil detector," Musk wrote. Sam Altman was once Musk's protege. Sam Altman is now his courtroom adversary. Number Two has a new face.

THREE — ACTIONS TO TAKE TODAY

Build the abstraction layer before the harness deepens further. Today Anthropic shipped Dreams (agents that review their own past sessions and restructure memory between runs), Routines (Claude Code can be scheduled to run on a cadence prompted by another instance of Claude — "the default is now I will have Claude prompt Claude Code"), multi-agent orchestration in general availability, an outcomes loop, webhooks, and Microsoft 365 Excel/PowerPoint/Word add-ins. The cost of switching the lab today is double what it was on Monday. Every Dream, every Routine, every webhook integration adds switching cost in a direction the platforms — Microsoft Agent 365, Google Agent Mode, Apple iOS 27 multi-model selector, Meta's Hatch, OpenAI's ChatGPT Apps directory, ServiceNow + Accenture's same-day FDE program — are about to challenge. Architect for the day the agent has to speak through someone else's directory service.

Price the missing hardware story before you sign the next SPV. The trillion-dollar Anthropic aftermarket is built on harness ownership. The platforms are coming for the harness. The lab's only structural defense is shipping its own surface. OpenAI bought Jony Ive's design firm in 2024 for a reported $6.5 billion — device expected 2026-2027. Anthropic has nothing public on hardware. xAI has nothing public on hardware. That gap is in no analyst model. Whoever ships the surface owns the moment. Apple owns iOS. OpenAI owns Jony Ive. Anthropic owns silence. Until the silence breaks, the trillion-dollar bet has a structural blind spot you should be sizing.

Take the only AI position that doesn't require you to predict who wins the harness. NVIDIA reported $46 billion in its most recent quarter, $41 billion of it data-center, presold through 2027. Today's news cycle is the receipt: Dario admitted on stage to 80-fold Q1 growth against a 10-fold plan; Pichai said Google Cloud would have been higher with more compute; Microsoft AI run-rate $37 billion up 123% year over year; OpenAI's MRC open networking protocol announced today was built with AMD, Broadcom, Intel, Microsoft, and NVIDIA cooperating around the bottom of the stack. The smart grid runs on Jensen's transformers. The smart router runs on Jensen's transformers. The picks-and-shovels trade just keeps printing. The labs and the platforms can fight over the harness for the next decade. NVIDIA's cap table does not care which side wins.

Today's actions touched the same questions we've been working with clients on this quarter — sizing harness-deepening contracts before they become impossible to walk away from, pressure-testing SPV pitches against missing hardware roadmaps, and helping operators build the abstraction layer the labs would prefer they don't.

FIVE — STORIES TO KEEP YOU INFORMED

Wednesday, May 6

1. Anthropic shipped the harness offensive of the year. (Full analysis below.) Code with Claude in San Francisco today: Dreams, Routines, multi-agent orchestration in general availability, outcomes loop, webhooks for managed agents, Microsoft 365 Excel/PowerPoint/Word add-ins, eight new data partnerships including a Moody's MCP app on 600 million companies. Every feature deepens the harness moat. Every feature is also a more visible target for platform absorption. Both can be true.

2. Google leaked Agent Mode and ServiceNow cloned the Anthropic JV — same Wednesday. Agent Mode appeared in leaked Gemini-app screenshots today: dedicated tab for tasks beyond single chat turns, scheduled actions, skills, inbox triage, ghostwriting. Likely Google I/O reveal on May 19-20. Same morning, ServiceNow and Accenture announced a forward-deployed engineering program at Knowledge 2026 in Las Vegas — 300 pre-built agents on ServiceNow's "AI Control Tower," directly competitive with Anthropic-Goldman-Blackstone-Hellman & Friedman. Ten days after we wrote the JV trade was a new asset class. Goldman's design has a competitor before the JV's first portco engagement closed.

3. Meta's Hatch is training on Anthropic — and intends to swap Anthropic out at launch. The Information broke it; FT and Korea Economic Daily confirmed today. Meta is building a consumer agent that operates DoorDash, Yelp, Reddit, Outlook, and Instagram, currently on Opus 4.6 and Sonnet 4.6, planned to swap to Meta's Muse Spark at general availability next month. The platform is using the lab as a temporary training rail it has publicly committed to commoditizing. Every platform with its own model investment will replicate this play.

4. Anthropic's Q1 grew 80x against a 10x plan — Dario admitted it on stage. "That is the reason we have had difficulties with compute. We're working as quickly as possible to provide more." The SemiAnalysis $44B annualized run rate from Tuesday's TAI brief now has CEO confirmation. Anthropic in talks at $900B per CNBC — the aftermarket has already moved through to $1T. 80x is also a number that cannot be sustained. Q2 won't match. The ratio that fueled the casino in Q1 will produce the headline that ends the casino in Q3.

5. Anton Korinek's NBER paper says automating AI research produces a singularity in ~6 years. Paper #w35155 published this week. Korinek argues that automating software R&D plus just 5% automation elsewhere is enough to overcome diminishing returns and produce explosive growth under empirically grounded calibrations. Pairs with Jack Clark's 60% probability of automated AI R&D by 2028 from Monday. The academic backstop and the lab-policy-chief estimate now agree on the recursive-loop regime being plausible by end of decade. Yann LeCun on the same day called the doom narratives "ridiculously stupid." The argument has not been settled. The fact that it is being argued at this credentialing tier is the news.

Don't get attached to anything you can't walk out on in thirty seconds flat.

Musk just walked out on a moral position in twelve weeks because the cap table needed him to!

SEVEN — SIGNAL / NOISE

Number Two Has A New Face

Three months ago Musk said Anthropic hates Western Civilization. This morning he leased them 220,000 GPUs, called the senior team impressive, and announced an exploratory partnership on multi-gigawatt orbital data centers — the literal moonbase that triggers his $7.5 trillion comp-package vest. The principle didn't change. The cap table did. SpaceX is targeting June 28 at $1.75 to $2 trillion, the largest IPO in history; bookrunners need a marquee AI customer in the prospectus. Anthropic was the only buyer big enough to take 220,000 GPUs in one transaction without procurement falling out. SpaceX moved its xAI training to Colossus 2 once Colossus 2 came online; Colossus 1 became the asset that needed monetizing. Both sides needed the deal. Both sides closed it.

While Musk was pivoting on Anthropic's values, Anthropic was running its first developer conference and shipping the most aggressive harness offensive of the year. Dreams: agents review their own past sessions, extract patterns, restructure stored memory. Each user's deployment compounds because the agent learns the user. Routines: Claude Code can be scheduled, prompted by another instance of Claude. "The default is now I will have Claude prompt Claude Code," Boris Cherny said on stage. The user is no longer in the prompt loop. Multi-agent orchestration: harness now manages multiple specialist agents under a single supervisor. Microsoft 365 add-ins: Claude reads and writes Excel, PowerPoint, Word. Eight new data partnerships: Dun & Bradstreet, Fiscal AI, Guidepoint, IBISWorld, SS&C, Third Bridge, Verisk, plus a Moody's MCP app on 600 million companies. Read as a list, a normal product day. Read as a category, the smart grid grew legs. Each feature deepens the harness. Each feature increases the cost of swapping the lab.

Yeah, but — every other major platform spent the same Wednesday building its own router for that grid. Google leaked Agent Mode for Gemini (likely I/O reveal May 19-20). Meta's Hatch — training today on Anthropic models, will swap to Meta's Muse Spark at launch next month. ServiceNow and Accenture announced an FDE program at Knowledge 2026 that competes directly with the Anthropic JV — same forward-deployed engineering pitch, same 300 pre-built agents, ten days late. Microsoft Agent 365 went GA last week at $15 per user per month. OpenAI's ChatGPT Apps directory now hosts Spotify, Zillow, Canva, Stripe — and Replit — making it an app store inside ChatGPT. Apple's iOS 27 multi-model selector lets users pick Claude or Gemini per Apple Intelligence feature, the same week Apple paid $250 million to settle the Apple Intelligence misrepresentation suit. The lab is sprinting up the stack. Six platforms are sprinting down toward the same harness layer. The collision is the news.

Apple has the bigger stick — unless the labs ship their own edge devices. OpenAI bought Jony Ive's design firm in 2024 for a reported $6.5 billion. Device expected 2026-2027. Whoever ships the surface owns the moment. Apple owns iOS. OpenAI owns Jony Ive. Anthropic owns silence. That gap is the most under-priced strategic question on the trillion-dollar bet — and the day the silence breaks, the harness fight changes hands.

The trade unambiguously winning is the one selling shovels to both sides. NVIDIA presold through 2027. Dario admits 80x Q1 growth and points at the loading dock. Pichai says Google Cloud would have been higher with more compute. AWS Bedrock customer spend is up 170% quarter over quarter. The hyperscalers are committing $725 billion of 2026 capex and the line item is transformers. The smart grid runs on Jensen's transformers. The smart router runs on Jensen's transformers. The picks-and-shovels trade just keeps printing.

The casino is open. Dr. Evil and his new Mini-Me are on the moonbase looking down at OpenAI's courtroom. Number Two has a new face. The smart grid grew legs and the smart router is loading. NVIDIA prints. Apple watches.

At COAI today: Full briefing — No One Set Off My Evil Detector — at getcoai.com.

— Harry

Sources:

Reply

Avatar

or to participate

Keep Reading