Anthropic

Phase A — Understand the business

Lens 1 · Company Overview

Anthropic is a frontier-model AI lab — it trains and sells access to the Claude family (Haiku / Sonnet / Opus, plus a limited "Claude Mythos" tier in 2026) via a usage-based API, first-party apps, and the major clouds. The defining commercial fact: ~80% of revenue is enterprise/business, not consumer — Anthropic "never really had a consumer phase," it built its base on enterprise API contracts and cloud-marketplace distribution. As of April 2026 it had 300,000+ business customers and 1,000+ accounts each spending >$1M/yr (doubled from ~500 in under two months; up from "roughly a dozen two years ago"). The breakout product is Claude Code, a terminal-native agentic coding tool at a reported ~$2.5B annualized run rate by February 2026 — over half of it enterprise (Netflix, Spotify, KPMG, L'Oréal, Salesforce named). Contract structure is the double-edged core of the thesis: usage-based consumption, not seats — it scales vertically with agentic token burn, but it can be "cut, optimized, or shifted to a cheaper model quickly". Whether Claude behaves like recurring software or discretionary compute is the single most important unresolved question about the revenue (carried into Lens 13).

Lens 2 · Supply Chain

Upstream of Anthropic is raw compute, and its supply chain is its most strategically distinctive asset: a deliberate three-lane, multi-vendor compute platform that no other frontier lab matches.

• AWS / Annapurna Trainium — Project Rainier trained Claude on >1M Trainium2 chips; a 2026 expansion commits Anthropic to up to 5 GW of AWS capacity and >$100B of AWS spend over 10 years; Amazon invests $5B now + up to $20–25B more.
• Google Cloud TPU — access to up to 1M TPUs, ~1 GW online in 2026 under the Oct-2025 agreement, +~3.5 GW from 2027 under an April-2026 Google–Broadcom deal; Google invests up to $40B.
• NVIDIA + Microsoft Azure — Nov-2025 deal makes Claude the only frontier model on all three hyperscalers; NVIDIA invests up to $10B, Microsoft up to $5B, against a $30B Azure compute commitment; initial ~1 GW of Grace Blackwell / Vera Rubin systems.

Downstream: cloud marketplaces (Bedrock — 100,000+ customers run Claude there — and Vertex) are the distribution rails, plus the first-party API and apps. Chokepoint: unlike a fabless chip company tied to one foundry, Anthropic has engineered away single-vendor compute risk — but it is now structurally dependent on the aggregate willingness of three hyperscalers (who are also its investors and, via Bedrock/Vertex, its competitors) to keep pouring multi-GW capacity at it. Names or it didn't happen: Amazon/Annapurna, Google/Broadcom, NVIDIA, Microsoft, TSMC (upstream of all the silicon). The diversification is real and rare; the dependence simply moved up a level.

Lens 3 · Competitive Advantages (moats)

Four candidate moats, in descending durability:

1. Enterprise trust + the safety brand. Anthropic's whole identity — the eight founders left OpenAI in 2021 precisely because they thought it was "commercializing too fast" with safety that "couldn't keep up" — converts directly into the thing enterprises buy: a vendor they trust with production workflows. Constitutional AI, the Responsible Scaling Policy (ASL-3), published system cards, and interpretability research are simultaneously a mission and a go-to-market moat in regulated industries. This is the most durable advantage and the hardest for OpenAI/xAI to copy without contradicting their own positioning.
2. Coding dominance + workflow embedding. Claude holds an estimated ~54% of the enterprise coding-model market vs OpenAI's ~21%, up from ~42% six months earlier. Claude Code embedded in developer workflows (and downstream tools) is the closest thing to switching costs Anthropic has — if it deepens into recurring infrastructure rather than swappable API.
3. Multi-cloud distribution (Lens 2) — being the only frontier model on AWS + GCP + Azure widens the funnel and hedges any single hyperscaler turning hostile.
4. Talent / research density — the founding team and a culture that has repeatedly shipped the Arena-leading model. The honest weakness: frontier model quality itself is commoditizing — Gemini 3.1 Pro closed to within ~4 Elo of Claude Opus 4.6 on Arena within days of launch, and GPT-5.5 looms. The model is a depreciating moat; the brand + embedding + distribution are the durable ones. Bargaining power is mixed: strong over customers (who need frontier capability), weak-ish over compute vendors who are also funding it.

Lens 4 · Segments

No audited segment disclosure exists (private). What's sourceable:

• By type: ~80% business/enterprise, ~20% first-party/consumer-ish.
• By product: Claude Code ~$2.5B annualized (Feb 2026) is the only broken-out line — i.e. coding alone is ~8% of a $30B run rate (or ~5% of the $47B bull number), and it's the fastest-growing slice (>10x in three months post-GA). The rest is general-purpose API consumption (agents, RAG, chat) plus app subscriptions.
• By geography / channel: undisclosed; cloud-marketplace mix is material (100,000+ Claude customers on Bedrock alone). Trend: accelerating and broadening — the >$1M-ARR cohort doubling in two months says spend is deepening within accounts, not just adding logos. Cause: the agentic-coding wave converting Claude from "a model you call" into "the engine your engineering org runs on." Mix-shift risk: if growth concentrates in usage-based coding, revenue quality is more cyclical than a seat-based SaaS comp would imply.

Phase B — Measure performance

Lens 5 · Funding & valuation trajectory (+private swap for "earnings")

There is no P&L to print; the scoreboard is the round history, and it is the steepest in the history of private markets [all web, unaudited]:

Lens 6 · Communications / narrative trend (+private swap → founder interviews, not earnings calls)

No earnings calls. The narrative arc through 2026 is "the disciplined, enterprise, safety-first lab that is quietly winning the economics" — Dario Amodei repeatedly framing growth as having outstripped Anthropic's own forecasts ~8x, and management leaning on the revenue-per-dollar-of-compute story ($2.10 by 2028 vs OpenAI's ~$1.60 ) to differentiate from OpenAI's consumer-subsidy burn. What management says more of: enterprise, agents, coding, "race to the top on safety," interpretability. What it's careful about: it filed the S-1 confidentially and a spokesperson stressed it has "not decided when or even if it will go public" — keeping optionality and damping IPO-hype expectations. The tone shift over 2026 is from "research lab with a product" to "the most commercially serious frontier company" — a deliberate pre-IPO repositioning.

Lens 7 · Cap table & secondary marks (+private swap for "comps")

Syndicate quality is the tell, and it is pristine — the inverse of a concentration problem. Crossover/public-market money is already in: Fidelity, BlackRock-affiliated funds, T. Rowe-class names, sovereigns (GIC, Temasek, MGX, Qatar via prior rounds), tier-1 VC (Sequoia, Founders Fund, ICONIQ, General Catalyst, Lightspeed), and crossover specialists (Coatue, Altimeter, Dragoneer, Greenoaks, D.E. Shaw, TPG, Dragoneer). A syndicate this thick with crossover funds is a textbook IPO-proximity signal. Secondary marks (the real-time price discovery):

• Feb-2026 employee tender priced at ~$350B.
• By late May, secondaries implied ~$1 trillion — Forge ~$1T, Hiive listings ~$1,447/share, with some pre-IPO blocks ~$572/share at a discount for timing risk. The market is marking Anthropic above its last primary ($965B) and above OpenAI on secondaries. Translation: on the management-stated $30B run rate, ~$965B is ~32x ARR; on the $44–47B bull number, ~20–22x ARR. For context, that is a software multiple applied to a company still guiding negative-to-thin margins until 2028.

Lens 8 · Price action / catalysts (+private swap → funding & product events)

The "price" is the valuation step-ladder, and the pattern reveals what moves this name: (1) revenue run-rate prints (each doubling re-rated the round — $9B→$30B drove $380B→$965B in three months); (2) compute mega-deals (the Google $40B / Amazon $25B / MSFT+NVIDIA $15B announcements each signaled supply secured and validated demand); (3) the confidential S-1 (June 1) as the IPO-proximity catalyst; (4) model-leadership events (Opus 4.6 holding Arena #1). What the secondary market reacts to is almost purely growth + IPO timing — there is no margin or cash-flow gate disciplining the mark yet, which is itself the risk (Lens 12/13). Forward catalysts: the public S-1 (financials become real and ``→ filed-fact), the IPO itself (as early as Oct 2026), the next model generation vs GPT-5.5/Gemini 3.x, and any disclosure that quantifies the compute-commitment vs revenue gap.

Phase C — Judge people & books

Lens 9 · Management

Founder archetype, mission-grade, and unusually credible. CEO Dario Amodei — Stanford physics, Princeton biophysics PhD, Google Brain → VP of Research at OpenAI where he led GPT-2/GPT-3 — and President Daniela Amodei (ex-VP Safety & Policy, OpenAI) co-founded Anthropic in 2021 with eight defectors from OpenAI (incl. Jared Kaplan, Jack Clark, Chris Olah, Tom Brown, Ben Mann, Sam McCandlish). This is one of the deepest founding benches in tech, and the "we left because safety couldn't keep up" origin is the brand's load-bearing asset (Lens 3).

• Track record: built a frontier lab from zero to ~$30B+ run rate and Arena-leading models in four years — quantitatively the fastest revenue ramp on record.
• Capital allocation: raised ~$132B and is deploying it into $180B+ of multi-vendor compute commitments + $50B of US data centers. Aggressive but deliberately diversified — the multi-cloud strategy is a capital-allocation choice that hedges vendor risk (a contrast to OpenAI's concentrated bets).
• Skin in the game / control: founders have ceded long-term board control to the Long-Term Benefit Trust (see Regulatory/governance) — an inversion of the usual founder-entrenchment dual-class playbook.
• Red flags: the chief governance question for a public investor is exactly that ceded control (mission can legally override shareholder return); plus the related-party density of the vendor-investor ring (Lens 10/13). No promotional-CEO or related-party-self-dealing flags of the classic kind.

Lens 10 · Forensic red flags & exclusivity (+private — no audited statements; reason from disclosed structure)

No financials to forensically dissect — so the red flags are structural, and they are real:

1. Circular / reflexive financing (the headline risk). Anthropic's largest investors are simultaneously its largest suppliers: Amazon (invests up to $25B; sells $100B of AWS), Google (invests up to $40B; sells TPU capacity), NVIDIA ($10B; sells GPUs), Microsoft ($5B; sells $30B Azure). Investment dollars flow in and compute dollars flow back out to the same counterparties. This is the same circular-AI-financing pattern drawing scrutiny across the sector — it inflates both "investment raised" and "revenue/spend" optics and makes arm's-length pricing hard to verify.
2. The margin/burn credibility gap. Gross margin went -94% (2024) → ~40% (2025), ~10pts below prior target, with a 77% target for 2028. Break-even slipped a year to 2028; The Information reports the gross-margin projection was lowered even as revenue soared. A skeptical write-up ("Anthropic's Profitability Swindle" ) argues the "profitability" framing leans on definitions that exclude the true cost of the compute commitments. ~$180B+ committed compute through 2029 (~1.7x the Series H) against a company not yet at break-even is the number to stress (Lens 13).
3. Revenue quality / recognition (unauditable here). Usage-based, hyperscaler-marketplace-billed revenue with related-party suppliers — when the S-1 lands, the first things to read are revenue-recognition policy, related-party transactions, customer concentration, and how compute commitments are carried (off-balance-sheet purchase obligations vs. recognized).

Regulatory findings (required sub-section).

• SEC EDGAR (LR + AAER): 0 findings — Anthropic has no CIK and is not an SEC registrant; no EDGAR enforcement search is possible. (This changes the moment the S-1 is public.)
• Copyright — material. Bartz v. Anthropic: a $1.5B class settlement over 500,000 allegedly pirated books used in training ($3,000/book) — near final approval May 2026; Anthropic already paid $300M with another $300M due five days post-approval. This is the largest publicly reported AI copyright settlement to date and sets the cost-of-training-data precedent for the industry.
• New litigation: a $3B music-publishers suit (Universal, Concord, BMG) over alleged "mass torrenting" of songbooks, amended with piracy claims using Bartz evidence.
• Non-SEC (FTC/DOJ/etc.): no material enforcement action surfaced in web search as of 2026-06-17; standard AI-sector antitrust/safety scrutiny applies but nothing case-specific found.
• Bottom line: No accounting-fraud or securities findings (none possible pre-registration). The material legal exposure is IP/copyright — $1.5B settled + $3B at risk — a quantifiable, sector-wide liability that the S-1 will have to provision for. Verified via SEC EDGAR EFTS (0 findings), web search, and the absence of any 10-K (private) as of 2026-06-17.

Phase D — Project & stress-test

Lens 11 · IPO-readiness & path-to-tradeable (+private swap for "forward projection")

This is the be-early payoff lens, and Anthropic scores at the top of the ladder — IPO-readiness 5/5.

• Stage: Confidential S-1 filed 2026-06-01 (Wilson Sonsini engaged); IPO floated as early as October 2026, though management publicly keeps the "if/when" optionality. The catalyst chain that unlocks a tradeable security is essentially complete except the public filing and roadshow.
• What unlocks the listing: (1) the public S-1 (turns every `` figure into a filed fact — the single biggest information event); (2) a revenue print that clears the bar management implied (~$30B+ exiting 2026 — revenue is the variable that decides Oct-2026 vs 2027 timing ); (3) market-window conditions for a mega-cap tech IPO.
• The likely path-to-tradeable: an IPO in the Oct-2026 → 2027 window at a valuation anchored on the ~$965B last round and ~$1T secondaries — i.e. the public market is being asked to underwrite a ~$1T entry on a ~$30–47B run rate (~20–32x ARR) and negative-to-thin margins until 2028.
• Brier-scoreable forecast to log (not run in this unattended breadth loop): "Anthropic files a public (non-confidential) S-1 or completes its IPO by 2027-06-30," p≈0.70. And a second: "Anthropic IPO prices at ≥ $965B (last primary) if it lists by end-2026," p≈0.55.
• Write-back (deferred per wave rules): register in _index.json and seed private-watch.json at readiness 5 with dossier → this file; both are out of scope for this unattended wave (no index/watchlist edits) — noted for the post-wave pass.

Lens 12 · Bull vs Bear

Bull. Anthropic is, on the merits, the best-positioned company in applied AI: enterprise-led revenue (80%, the stickiest kind), the coding category leader (~54% share), the only frontier model on all three hyperscalers, a structurally diversified compute supply that no rival matches, superior unit economics ($2.10/compute-$ vs OpenAI's $1.60 target), the most credible safety brand (a real GTM moat in regulated buyers), and the deepest founding bench in the field. Revenue is doubling on a cadence of weeks and the >$1M-ARR cohort is deepening. If the margin curve bends as guided (40%→77% by 2028) and Claude embeds into enterprise workflows as durable infrastructure, ~$965B is a floor, not a ceiling — this is a future trillion-dollar public company and the cleanest way to own the enterprise-AI layer.

Bear. The price already assumes all of that. ~$965B / ~$1T on a $30–47B run rate is a ~20–32x ARR mark on a company that won't break even until 2028 and just lowered its margin projection. The model itself is commoditizing (Gemini 3.1 within 4 Elo, GPT-5.5 coming) — frontier quality is a depreciating moat, and the revenue is usage-based and swappable, not contracted seats. The compute is funded by a circular ring of vendor-investors ($180B+ committed, ~1.7x the largest round) that flatters both sides of the ledger. And the governance (LTBT can subordinate shareholder return to mission) is a structural discount a public investor must price. Pre-mortem (18 months out, thesis broke): Gemini/GPT close the gap to parity, enterprises route agentic coding to whoever is cheapest that quarter, usage-based revenue decelerates from "doubling" to merely "fast," the margin curve stalls below 50% as compute commitments come due, the IPO prices below the $965B last round, and the ~$1T secondary marks unwind — the classic late-stage-private-down-round-into-the-IPO. Contrarian read the market is refusing to see: the bull and bear are both right — Anthropic is simultaneously the best business in AI and a valuation that already prices a decade of flawless, margin-expanding, lead-retaining execution. The market is conflating "best company" with "best entry price." Those are different claims.

Lens 13 · Devil's advocate (short-seller)

Dismantling the bull case. The short isn't "Anthropic is bad" — it's that you are underwriting a $1T entry on three things that all have to break right at once.

1. The moat is the most perishable kind. "Best model" lasted days against Gemini 3.1 last cycle. A frontier lead that resets every quarter is not a moat; it's a treadmill funded by ever-larger compute bills.
2. Revenue concentration by workload, not by customer. The growth story is overwhelmingly agentic coding (Claude Code + the coding API). Coding is the single most price-sensitive, most-benchmarked, fastest-commoditizing AI workload — exactly where a cheaper Gemini/GPT/open-weight model bleeds share fastest. Strip aggressive coding growth and the $30B→$47B curve flattens.
3. The circular financing is the dangerous part. Amazon/Google/NVIDIA/Microsoft are investor + supplier + (via Bedrock/Vertex) competitor. Their dollars in and Anthropic's compute dollars out inflate both sides; if any one of them tightens capacity or pricing — or decides its own model is good enough — the supply story and the demand story both wobble. ~$180B in compute commitments against a pre-break-even P&L is a take-or-pay-style obligation the revenue must grow into or the margin math implodes.
4. Worst capital-allocation / governance flag: the LTBT can legally override shareholder economics for mission. A public minority investor at ~$1T has the least control of any comparable mega-cap. What permanently impairs it: frontier-model capability converges to "good enough and roughly equal" across 3–4 labs + open weights, coding margins compress to commodity infra economics, and Anthropic is left servicing $180B of compute with software-multiple expectations and utility-multiple gross margins. Plausibility: moderate — not the base case, but far more plausible than a ~32x-ARR mark admits. Valuation if growth disappoints 20–30%: at ~$22–35B run rate with a stalled margin, a sane public market re-rates toward 10–15x ARR ($220–525B) — i.e. a 45–75% drawdown from the ~$1T secondary mark.

Lens 14 · Management questions (15, ordered by information value)

1. Of the ~$30B run rate, what % is usage-based coding (Claude Code + coding API), and what is the net revenue retention of your top 50 accounts? (Tests the "infrastructure vs discretionary compute" thesis directly.)
2. Reconcile the $30B management figure with the $44–47B aggregator reports — what is the exact GAAP-comparable run rate and how is it measured?
3. What are the contractual terms of the $180B+ compute commitments — take-or-pay, minimums, exclusivity — and how are they carried (purchase obligation vs recognized)?
4. Why was the gross-margin projection lowered, and what specifically bends the curve from ~40% to 77% by 2028?
5. With Amazon/Google/NVIDIA/Microsoft as investor + supplier + marketplace competitor, how is compute priced at arm's length, and what's the related-party revenue %?
6. What is the true 2026 cash burn in dollars (not % of revenue), and what's the bridge to break-even in 2028?
7. As models converge, what is the defensible moat in 24 months — and what's the evidence it's brand/embedding rather than raw capability?
8. How concentrated is revenue across the cloud marketplaces (Bedrock/Vertex), and what happens to economics if a hyperscaler favors its own model?
9. What governance protections will public shareholders have, given the LTBT can elect a board majority and subordinate returns to mission?
10. Total provisioned cost of the copyright exposure (Bartz $1.5B + the $3B music suit + future claims) and the go-forward training-data licensing cost structure?
11. What is consumer/first-party mix and is there any intent to chase it, or is enterprise the permanent strategy?
12. What's the plan if the IPO window forces a price below the $965B last round — down-round into the public market, or wait?
13. Capex/data-center strategy ($50B US DCs): owned vs leased, and how does it interact with the hyperscaler compute deals?
14. Safety/RSP as the lead toward ASL-4 — does the scaling-policy gate ever delay a revenue-driving release, and has it?
15. What single competitive development (a specific rival model/price move) would most threaten the 2026 plan, and what's the contingency?