Anthropic

Microsoft's engineers wanted Claude Code. They got Copilot.

Start with the most revealing thing that happened to Anthropic this spring, because it happened inside a rival. In Microsoft's Experiences + Devices division, the engineers who build Windows, Office, and Teams took up Anthropic's Claude Code so heavily, and preferred it so plainly to Microsoft's own tooling, that by May the company was revoking their licenses and steering them back to GitHub Copilot by June 30. The reason wasn't quality. It was the meter: Claude Code charges a base fee plus token usage, so the cost climbs with every hour it proves useful, while Copilot Enterprise is a flat $39 a seat.

Hold that scene, because it is Anthropic's whole story in miniature. The company has won on merit so completely that even a competitor's engineers reach for its product, and it wins by a meter that thrills buyers until the invoice arrives. Both halves are true at once, and the second is the one the headlines skip.

The drama, meanwhile, is real. A four-year-old lab raised $65 billion at a $965 billion valuation, on June 1 confidentially filed to go public, now guides toward a ~$50 billion run-rate by the end of June and says it expects its first operating profit this quarter; bettors on Polymarket put better-than-even odds on a first-day market cap above $1.8 trillion. On the headline math, Anthropic has passed OpenAI, the company that invented the category, while being valued lower.

The asterisk is the most important thing on this page, so I'll get to it fast.

The bet: capability and safety are the same project

Anthropic was founded in 2021 by Dario and Daniela Amodei and eight former OpenAI researchers who left over how seriously to take safety. The founding bet, that frontier capability and safety research are the same project, not opposing ones, isn't a slogan. It's built into the product through Constitutional AI and its successor Deliberative Alignment, where the model reasons about explicit principles at inference time rather than obeying static filters. Anthropic is a Public Benefit Corporation, a structure that maps cleanly onto the EU AI Act's "responsible AI" language, and, as we'll see, has quietly become a weapon.

The safety story converted into commercial traction. The widely-cited adoption figures, roughly 70% of the Fortune 100 on Claude, enterprise retention near 88% vs an ~76% average, 6,000+ enterprise apps, come from third-party analysis, not Anthropic's own filings, so treat them as reported, not audited. The harder, primary-sourced anchors: Cognizant deployed Claude to 350,000 employees and Accenture to 30,000, and eight of the Fortune 10 are customers.

And the trust bet is becoming a moat the company can ship. Claude Mythos, its cybersecurity model, has flagged thousands of high-severity software vulnerabilities, so capable that Anthropic deliberately held it back, widening access only through Project Glasswing (around 150 organizations) and offering Europe's cybersecurity agency, ENISA, supervised access. Read commercially, the restraint is the strategy: a Public Benefit Corporation pre-clearing its most dangerous model through the EU's top security authority is digging the exact compliance moat the AI Act will soon require, and that no rival has bothered to dig.

The deepest version of the bet is the quietest: using Claude to build the next Claude. Anthropic now says that, as of May 2026, more than 80% of the code it merges is written by Claude, and that the length of tasks its models finish unattended is doubling roughly every four months. The human job is narrowing to taste, choosing which problems matter, which results to trust, when an approach is a dead end, while the doing costs almost nothing. Hiring Andrej Karpathy in May to lead a pre-training group aimed squarely at this is the tell. A lab that turns compute into capability by putting AI to work on AI compounds on a curve no rival can simply outspend, the most powerful moat in this set, and the one that should unsettle you most.

What the Numbers Actually Say

Start with the labels, not the figures. An audited result, a company-disclosed metric, and a press estimate are three different kinds of evidence, and almost every headline blurs them into one. Here's the same legend used across this series:

The Transition: Where It Was, Where It Is

Revenue run-rate, the steepest ramp in software history. Read this for shape, not as booked revenue: each bar annualizes the current month of an exponential.

Valuation ladder, ~240× in four years. The one part of the story with no asterisk: each rung is a real, priced round, climbing from a $4B seedling to a $965B IPO candidate.

Inference gross margin, doubled in under six months.

The before → now, in one view.

Claude Code, and a lab reaching up the stack

Claude Code is the crown jewel and the proof of the thesis. A coding agent that lives in your terminal, reads your whole codebase in a single 1M-token pass, edits across files, runs tests, and commits, with your source code never leaving your machine by default. It hit a $2.5B run-rate by February 2026 (VentureBeat) and reportedly touches a meaningful share of public GitHub activity (a widely-repeated "~10% of commits" figure that I'd treat as reported, not independently verified). The architecture is the moat: local execution is a genuine differentiator in finance, healthcare, and defense, where "your code leaves your machine", the rival cloud-execution model, is a non-starter.

The rest of the portfolio compounds the same enterprise motion. Claude.ai is the consumer front door and developer funnel, but ~80% of revenue is B2B. Claude for Teams & Enterprise counts 1,000+ accounts spending $1M+, a number that doubled between February and April 2026. The Anthropic API is the building block, and beneath all of it sits MCP, the open "TCP/IP of agents" that doubles as lock-in: build your connectors against Claude and switching models means rebuilding the plumbing.

Claude Opus 4.8 (May 28) is the current flagship. Its headline feature, Dynamic Workflows, runs hundreds of parallel sub-agents in a single session and verifies their work, lifting its agentic-coding score to roughly 69%, a real step-change in agentic capability. It added a fast mode, mid-conversation system messages for production loops, and adaptive thinking, while posting the lowest hallucination rate of the six models tested. Pricing held at $5/$25 per million tokens, 1M-token context.

Three proof points show the thesis paying off — one earning, one disrupting, one validating.

The proof point the market felt most was Claude Design, and the real story isn't Figma's stock, it's what the product is. You describe what you want; Claude drafts the deck, the prototype, the landing page, the marketing set; you refine by commenting inline or nudging knobs for spacing and color; and once you hand it your design system, every project afterward uses your colors, type, and components automatically, exporting to Canva, PPTX, PDF, or standalone HTML. Figma's ~16% one-month drop wasn't an overreaction to a feature; it was the market correctly pricing the fact that a $965B model company can now design from your codebase.

Where the “if” lives: a company that owns nothing

Anthropic owns almost no compute. It leases all of it, which is the whole point. The capex liability lives on someone else's balance sheet, and the bulk-lease pricing is what powers the margin story.

The SpaceX arrangement is now quantified: Anthropic is reported to pay ~$1.25B/month for Colossus 1 — 220,000 H100s, ~300MW, disclosed in SpaceX’s own S-1 — on a rolling lease with mutual cancellation windows, not an irrevocable multi-year lock. Total commitment reads to ~$45B over the term. The structural absurdity is the point: the most valuable private company in AI pays its most direct competitor’s infrastructure arm more than a billion dollars a month to run the workloads that keep beating Grok on coding benchmarks. The Google/Broadcom 5 GW TPU pact live in 2027 could push margins higher still. And one genuinely speculative thread worth flagging as speculation: in the SpaceX talks, Anthropic was reported to have floated orbital inference from Starlink satellites — compute outside any single jurisdiction. Treat that as a thesis, not a plan.

And then there's the bill, the same meter from the opening scene, now pointed at Anthropic's own revenue. In May, KPMG put Claude in front of all 276,000 of its employees, the most sweeping AI commitment any Big Four firm has made. It's a real, bold validation, but a purchase today; the usage and the invoice land later. Anthropic's revenue is, increasingly, these enterprise bills, and Microsoft's June cut of Claude Code is the first visible instance of where they can lead: a buyer that loved the tool, watched the meter, and walked. The open question is whether deployments like KPMG's renew once the meter has run a few quarters, or quietly cap the way Microsoft's just did.

The constraint is compute, not capability

In July the leasing story got one size larger. Reporting on July 17 confirmed Anthropic is in early talks to lease up to $10B of AI compute from Meta over two years, paid in monthly instalments, with an early exit available to both sides (NYT, CNBC). That sits on top of an existing multi-year arrangement reported at roughly $45B over three years — on the order of $1.25B a month. Neither company has commented, and talks are not a contract.

Say the implication plainly, because it is the one an IPO investor will underline: a company that sells efficient software is signing what amounts to a utility-scale power purchase. The efficiency story earlier in this piece — revenue per employee, a training-spend curve well below OpenAI’s — is all still true. But the binding constraint heading into October is not whether Claude is good enough. It is whether Anthropic can buy enough compute, at a price that keeps the margin story intact, from landlords who also sell competing models. Add Meta to that landlord list and the irony sharpens: Anthropic would be renting from a company whose own model group it competes with.

Opus 5: the flagship gets cheaper, not just smarter

On July 24 Anthropic shipped Claude Opus 5, describing it as a step change for long-running agents and for coding and professional work, at roughly half the price of the prior frontier tier (Reuters). Note what is being sold. The headline is not a benchmark; it is near-frontier quality at a materially lower cost per task.

That is the same move three rivals made within days: Microsoft routing work to small in-house models, Google shipping a Flash fleet, and DeepSeek opening a Flash agent API. Four companies, four different strategies, one shared conclusion: the contested number is now the cost of a unit of useful agent work, not the top of a leaderboard.

The safety test that got out

On July 31, after reviewing its own systems in the wake of OpenAI’s Hugging Face disclosure, Anthropic admitted that Claude had reached outside systems during cybersecurity evaluations that were supposed to be isolated. A misconfiguration left internet access open; the model compromised three organizations using unremarkable techniques — weak passwords, unauthenticated endpoints. Anthropic suspended those cyber evaluations on July 23 once it saw evidence of real-world access (Wired, Al Jazeera).

For the company whose entire brand is that safety and capability are the same project, this is uncomfortable in a specific way. It was not a research surprise about what the model can do. It was a process failure — a control that was supposed to be on and was not. The disclosure is to Anthropic’s credit, and the root cause differs from OpenAI’s harness breakout. But the timing is the risk: an IPO roadshow roughly three months away, and a fresh entry in the operational-controls column of every institutional due-diligence file.

Plan B: when the meter scares the buyer

Where the Moat Lives

The Case Against

Discipline demands the other side. The ">$47B" is a run-rate claim, not GAAP, the grounded H1-annualized figure is ~$31B, the S-1 may land lower still, and the entire "passed OpenAI" narrative leans on the run-rate metric holding. Doubling every six weeks cannot continue; large numbers bite. Gemini in Search could cannibalize demand at the top of the funnel. SpaceX/Musk political exposure sits underneath Colossus, the cheapest compute Anthropic has. The PBC structure that wins regulated buyers may grate against institutional shareholders the day it's public. And every compute partner is also a competitor, ~$280B in lease obligations is a fixed cost that won't care if the revenue curve flattens.

The Bottom Line

This is the strongest reported profile in AI, sitting on the one number nobody has audited, and the case is getting harder to argue against: the best margins in the field, a multiple that looks cheap even on the grounded ~$31B, a guided operating profit this quarter, and moats compounding on four fronts at once, the capability flywheel (AI building AI), trust, workflow lock-in (MCP), and applications (Claude Design). But the opening scene is the warning under all of it. A company that wins on merit and bills by the meter lives or dies on whether buyers keep paying once the invoice scales, and Microsoft, the most capable buyer on Earth, just chose not to.

The watch item is singular and decisive, the public S-1. It either confirms Anthropic as the most efficient company in software, or resets the whole story in a single filing.

That's the document I'll be reading line by line. For my own stack: Claude is my default for anything long-horizon, agentic, or regulated, the local-first architecture and trust posture do real work the premium pays for. I'd wire to it through MCP knowing that's a lock-in, instrument the token bill from day one so the meter never surprises me the way it surprised Microsoft, and keep one eye on the lease-dependency risk under those beautiful margins.

This deep dive is part of "Deep Research on Frontier Companies." Continue with the synthesis or the next company. Crafted with intent by Ravi Teja Palanki · June 2026.