Combined average of Moat (AI Resilience), Growth, and Valuation scores.
Moat ScoreWidening
Anthropic leads enterprise and coding AI on model quality, not on a structural lock: 80% of revenue is usage-based API spend that customers can cut with limited notice, 48% runs through the clouds of three rivals, and the compute underneath is rented from those same rivals. Every pillar is real and none is yet category-defining, so the moat rests on staying at the frontier.
Anthropic's edge is being the model enterprises and developers pick for work that matters, which shows in the revenue, not yet in contracts:
- Enterprise and Developer Pull: Revenue went from $386M in 2024 to $4.59B in 2025 and $16.3B in H1 2026, with more than 1,900 customers above $1M of trailing revenue and about 250 above $10M (S-1, September 30, 2026). Net revenue retention of about 900% says existing customers multiplied their spend; it measures demand, not lock-in.
- Claude Code and the Agent Surface: Claude Code, the Agent SDK, MCP and the Claude apps are where teams encode their rules, memories and skills, and Claude Code is the main driver of token consumption. That is real switching friction, the same class SpaceX's Cursor is rated intact on, but a model API is the easiest dependency in software to re-point.
- Rented Scale, Owned Efficiency: Anthropic says it earns more than $2 of revenue per $1 of compute and reported an 85% adjusted gross margin on inference in Q3. Training cost is fixed and spread over one of the two or three largest token volumes in the industry, but the chips are Google TPUs, Amazon Trainium and NVIDIA GPUs bought from rivals, so the unit-cost lead belongs to Alphabet, not Anthropic.
- What the S-1 Says Can Walk: Most revenue carries no long-term commitment, about 25% comes from two customers, and remaining performance obligations were only $10.8B at June 30 against $16.3B of half-year revenue. Anthropic is now trading discounted token rates for multi-year minimum-spend contracts; that is the step that would turn demand into embedding.
Moat Verdict
Anthropic is the purest AI beneficiary in the book: its revenue is AI demand, up from $386M in 2024 to $16.3B in H1 2026. The resilient pillars, embedding in enterprise workflows, usage scale and certifications, are all real and all intact, because the S-1 shows revenue that can walk, a channel run by three rivals and compute rented from them. The AI-vulnerable legs, the Claude Code interface and the model logic, are exactly what rival labs attack every release cycle. Durable while Claude stays at the frontier, but not yet locked in by contracts, data or owned silicon.
65.0 resilient · 65.0 vulnerable · 80/20 = 65.0 · = 65 · blended with 70 projected = 67
Open a moat to read its note.
Claude Code, CLAUDE.md files, skills, hooks and the Agent SDK are where developer teams encode how they work, and Claude Code is the main driver of token consumption. That is the same class of agent-IDE investment SpaceX's Cursor is rated intact on, and it is not strong for the same reason: the interface is a prompt and a model endpoint, Cursor, Copilot and Codex wrap rival models in the same workflow, and the S-1 says customers can reduce usage with limited notice.
The model weights and the training recipe are vendor-owned logic no customer can reproduce. They are not strong because rivals do reproduce the capability: Gemini, GPT and open-weight models have each matched a Claude release within months, and leadership on coding benchmarks has changed hands several times since 2024.
Anthropic does not monetise gated access to a public dataset; it trains on public data that every lab can crawl.
Frontier research talent is scarce, and Anthropic has held on to its founding research team through the 2025–26 pay war, but OpenAI, Google DeepMind, Meta Superintelligence Labs and SpaceXAI hire from the same pool. Same pool, same rating as Google's DeepMind (M8).
API, Claude apps, Claude Code and enterprise seats are sold together and Team and Enterprise plans bundle Claude Code, but the S-1 gives no attach or renewal figure, so the bundle is intact (M9). It is narrower than Alphabet's Workspace-Android-Gemini bundle or Microsoft's Office-Copilot bundle, both rated strong, because Anthropic owns no distribution surface of its own.
Usage feedback from coding agents and enterprise workloads improves the next model, but commercial API data is not used for training by default and belongs to the customer, so this is scale of usage, not a class of data no peer holds (M5). Alphabet's search and YouTube corpus and Meta's social graph are strong because the franchise owns the data; Anthropic does not.
Claude holds FedRAMP High and classified-cloud availability through AWS and Google Cloud, and its safety posture helps in regulated buyers, but Google, Microsoft and OpenAI hold the same authorisations, so it bars small entrants rather than peers (M4). The dispute with the US administration over government use of Claude is a headwind charged in growth keyRisk, not a lost licence.
There is no direct user-to-user network. There is a developer ecosystem around Claude Code, MCP servers and skills, and a usage loop where more agent traffic produces better coding models, but MCP is an open standard rivals also support, so the ecosystem does not lock users to Claude. Real and early, the same rating SpaceX's developer loop holds.
Claude sits inside customer products and workflows: more than 1,900 customers spend over $1M a year and about 250 over $10M. Not strong because the S-1 says most revenue carries no long-term commitment and can be cut with limited notice, 48% of sales runs through marketplaces where Gemini, Nova and GPT sit one click away, and remaining performance obligations were only $10.8B at June 30.
Widening over 5–10 years. Watch: Anthropic is trading discounted token rates for multi-year minimum-spend contracts. The trend is confirmed if remaining performance obligations grow faster than revenue for four quarters after the $10.8B of June 2026; it stops if RPO stays below one quarter of revenue.
A model is not the authoritative record for any business function; customer data, code and documents stay in the customer's own systems.
Training cost is fixed and spread over one of the two or three largest token volumes in the industry, and the company says it earns more than $2 of revenue per $1 of compute with an 85% adjusted inference gross margin, so the scale is real. It is not strong because the chips are rented from rivals: Alphabet is strong here on owned TPUs, and Anthropic pays Google, Amazon and Microsoft for the same capacity. It is not weakened like CoreWeave, whose product is the rented compute itself; Anthropic's product is the model the compute trains.
Most revenue is enterprise API spend chosen on model quality and cost, which the other pillars rate. Claude charges a per-token premium today, but that premium follows benchmark leadership rather than holding on the name alone (M10).
Moat in 5–10 years
Transaction embedding is the moat most likely to widen, because Anthropic is now converting usage into multi-year minimum-spend contracts and the S-1 gives a baseline ($10.8B of remaining performance obligations) to measure it against. Scale economics could follow if Anthropic Direct and the Broadcom TPU leases put a majority of compute on hardware it controls, but that is a plan, not yet a signal. The interface and model-logic legs hold at intact: rivals have matched each Claude release within months, and nothing in the filing suggests that cycle has stopped.
65 today → 70 if the trend completes · scored 67 (70/30)
Combined average of Moat (AI Resilience), Growth, and Valuation scores.
Moat ScoreWidening
Anthropic leads enterprise and coding AI on model quality, not on a structural lock: 80% of revenue is usage-based API spend that customers can cut with limited notice, 48% runs through the clouds of three rivals, and the compute underneath is rented from those same rivals. Every pillar is real and none is yet category-defining, so the moat rests on staying at the frontier.
Growth Score
Anthropic's public S-1 (filed October 26, 2026; Nasdaq ticker ANT; Goldman Sachs and Morgan Stanley lead 22 underwriters) shows revenue of $41M in 2023, $386M in 2024 and $4.59B in 2025, then $4.73B in Q1 2026 and $11.56B in Q2 (H1 $16.29B against $1.16B a year earlier). The company told investors on October 14 that Q3 revenue was about $16.8B (+45% QoQ), with an 85% adjusted gross margin, an adjusted operating margin above 11% and positive free cash flow; people at the meeting cited a ~$110B run rate by year-end, which Anthropic has not confirmed and which is not scored here. About 80% of H1 revenue was usage-based API spend and 20% seat subscriptions; 48% was sold through the AWS, Google Cloud and Azure marketplaces (47% in 2025, 32% in 2024). H1 adjusted operating income was $1.62B (10% margin) against $(1.85)B a year earlier; GAAP operating loss narrowed to $(693)M, with Q2 the first GAAP operating profit ($974M). Compute and infrastructure cost $8.95B in H1 (55% of revenue, from 201% a year earlier). The $(80.7)B H1 net loss is mostly a $72.4B non-cash fair-value charge on convertible notes, derivatives and SAFEs that ends when they convert at the IPO. Headcount was about 5,300 on September 30.
Valuation Score
Scored at $118 a share, the ~$2.0T valuation reported as the IPO target, divided by ~16.9B fully diluted shares (the S-1's Public Benefit Fund grant of 846.8M Class N shares is 5% of fully diluted shares at June 30, 2026). The S-1 does not yet carry a price range or share count, so this is a talked price, not a market price, and the file must be re-marked when the range is set and again at the first close. At $118 Anthropic sits between the base ($103) and the bull ($160), so the score is 60. The base of ~$1.75T is ~14× a 2027 revenue near $125B (Q4 2026 near $22B and roughly 15% sequential growth through 2027), with adjusted operating margin rising from H1's 10% toward the mid-20s as compute falls further as a share of revenue. That is above SpaceX's ~11× on its 2027 SI revenue because Anthropic's revenue is software-margin model access (85% adjusted inference gross margin) rather than GPU rental, and well above the neocloud ~5× because it does not own the capex. The bear of $65 (~$1.1T, near 1.15× the $965B May 2026 Series H) is a 2027 near $90B as price competition from Gemini, OpenAI and open models compresses token rates while the fixed compute bill keeps margins thin. The bull of $160 (~$2.7T) needs the ~$110B exit run rate to prove real and 2027 to approach $180B with margins heading to 30%+. At $118 the stock trades ~29× annualised Q3 revenue and ~32× H1 annualised. Earnings multiples are not meaningful: GAAP net income is distorted by the convertible marks, and adjusted net income was $2.67B in H1 (~$5.3B annualised, ~375× at $2.0T).
The Frontier-Quality Moat
Anthropic's edge is being the model enterprises and developers pick for work that matters, which shows in the revenue, not yet in contracts:
- Enterprise and Developer Pull: Revenue went from $386M in 2024 to $4.59B in 2025 and $16.3B in H1 2026, with more than 1,900 customers above $1M of trailing revenue and about 250 above $10M (S-1, September 30, 2026). Net revenue retention of about 900% says existing customers multiplied their spend; it measures demand, not lock-in.
- Claude Code and the Agent Surface: Claude Code, the Agent SDK, MCP and the Claude apps are where teams encode their rules, memories and skills, and Claude Code is the main driver of token consumption. That is real switching friction, the same class SpaceX's Cursor is rated intact on, but a model API is the easiest dependency in software to re-point.
- Rented Scale, Owned Efficiency: Anthropic says it earns more than $2 of revenue per $1 of compute and reported an 85% adjusted gross margin on inference in Q3. Training cost is fixed and spread over one of the two or three largest token volumes in the industry, but the chips are Google TPUs, Amazon Trainium and NVIDIA GPUs bought from rivals, so the unit-cost lead belongs to Alphabet, not Anthropic.
- What the S-1 Says Can Walk: Most revenue carries no long-term commitment, about 25% comes from two customers, and remaining performance obligations were only $10.8B at June 30 against $16.3B of half-year revenue. Anthropic is now trading discounted token rates for multi-year minimum-spend contracts; that is the step that would turn demand into embedding.
Moat Verdict
Anthropic is the purest AI beneficiary in the book: its revenue is AI demand, up from $386M in 2024 to $16.3B in H1 2026. The resilient pillars, embedding in enterprise workflows, usage scale and certifications, are all real and all intact, because the S-1 shows revenue that can walk, a channel run by three rivals and compute rented from them. The AI-vulnerable legs, the Claude Code interface and the model logic, are exactly what rival labs attack every release cycle. Durable while Claude stays at the frontier, but not yet locked in by contracts, data or owned silicon.
65.0 resilient · 65.0 vulnerable · 80/20 = 65.0 · = 65 · blended with 70 projected = 67
Open a moat to read its note.
Claude Code, CLAUDE.md files, skills, hooks and the Agent SDK are where developer teams encode how they work, and Claude Code is the main driver of token consumption. That is the same class of agent-IDE investment SpaceX's Cursor is rated intact on, and it is not strong for the same reason: the interface is a prompt and a model endpoint, Cursor, Copilot and Codex wrap rival models in the same workflow, and the S-1 says customers can reduce usage with limited notice.
The model weights and the training recipe are vendor-owned logic no customer can reproduce. They are not strong because rivals do reproduce the capability: Gemini, GPT and open-weight models have each matched a Claude release within months, and leadership on coding benchmarks has changed hands several times since 2024.
Anthropic does not monetise gated access to a public dataset; it trains on public data that every lab can crawl.
Frontier research talent is scarce, and Anthropic has held on to its founding research team through the 2025–26 pay war, but OpenAI, Google DeepMind, Meta Superintelligence Labs and SpaceXAI hire from the same pool. Same pool, same rating as Google's DeepMind (M8).
API, Claude apps, Claude Code and enterprise seats are sold together and Team and Enterprise plans bundle Claude Code, but the S-1 gives no attach or renewal figure, so the bundle is intact (M9). It is narrower than Alphabet's Workspace-Android-Gemini bundle or Microsoft's Office-Copilot bundle, both rated strong, because Anthropic owns no distribution surface of its own.
Usage feedback from coding agents and enterprise workloads improves the next model, but commercial API data is not used for training by default and belongs to the customer, so this is scale of usage, not a class of data no peer holds (M5). Alphabet's search and YouTube corpus and Meta's social graph are strong because the franchise owns the data; Anthropic does not.
Claude holds FedRAMP High and classified-cloud availability through AWS and Google Cloud, and its safety posture helps in regulated buyers, but Google, Microsoft and OpenAI hold the same authorisations, so it bars small entrants rather than peers (M4). The dispute with the US administration over government use of Claude is a headwind charged in growth keyRisk, not a lost licence.
There is no direct user-to-user network. There is a developer ecosystem around Claude Code, MCP servers and skills, and a usage loop where more agent traffic produces better coding models, but MCP is an open standard rivals also support, so the ecosystem does not lock users to Claude. Real and early, the same rating SpaceX's developer loop holds.
Claude sits inside customer products and workflows: more than 1,900 customers spend over $1M a year and about 250 over $10M. Not strong because the S-1 says most revenue carries no long-term commitment and can be cut with limited notice, 48% of sales runs through marketplaces where Gemini, Nova and GPT sit one click away, and remaining performance obligations were only $10.8B at June 30.
Widening over 5–10 years. Watch: Anthropic is trading discounted token rates for multi-year minimum-spend contracts. The trend is confirmed if remaining performance obligations grow faster than revenue for four quarters after the $10.8B of June 2026; it stops if RPO stays below one quarter of revenue.
A model is not the authoritative record for any business function; customer data, code and documents stay in the customer's own systems.
Training cost is fixed and spread over one of the two or three largest token volumes in the industry, and the company says it earns more than $2 of revenue per $1 of compute with an 85% adjusted inference gross margin, so the scale is real. It is not strong because the chips are rented from rivals: Alphabet is strong here on owned TPUs, and Anthropic pays Google, Amazon and Microsoft for the same capacity. It is not weakened like CoreWeave, whose product is the rented compute itself; Anthropic's product is the model the compute trains.
Most revenue is enterprise API spend chosen on model quality and cost, which the other pillars rate. Claude charges a per-token premium today, but that premium follows benchmark leadership rather than holding on the name alone (M10).
Moat in 5–10 years
Transaction embedding is the moat most likely to widen, because Anthropic is now converting usage into multi-year minimum-spend contracts and the S-1 gives a baseline ($10.8B of remaining performance obligations) to measure it against. Scale economics could follow if Anthropic Direct and the Broadcom TPU leases put a majority of compute on hardware it controls, but that is a plan, not yet a signal. The interface and model-logic legs hold at intact: rivals have matched each Claude release within months, and nothing in the filing suggests that cycle has stopped.
65 today → 70 if the trend completes · scored 67 (70/30)
Growth Analysis
Growth Drivers
Key Risk
Anthropic carries at least $430B of non-cancellable compute commitments (Broadcom, Google, Amazon, Microsoft) and up to ~$637B with SpaceXAI, Nscale and Lambda, while most of its revenue can be cut with limited notice, about 25% comes from two customers, and 48% is sold through the clouds of Google, Amazon and Microsoft, each of which ships a competing model. If OpenAI, Gemini or a cheaper open model takes the coding and agent workloads, or a top customer moves in-house, before the multi-year minimum-spend deals cover the commitments, the fixed cost lands on a shrinking base. Graded high, not CoreWeave's severe: $131.7B of cash and short-term investments covers about two years of commitments, the convertibles that create the GAAP loss convert into equity at the IPO rather than needing repayment, the company is profitable on an adjusted basis, and the SpaceXAI tranche can be cancelled on 90 days' notice after early 2027. The dispute with the US administration over government use of Claude is a further headwind to public-sector demand.
Score Derivation
95.0 base − 2.0 trajectory + 4 margin − 10 risk = 87
Base 95 (40–60% CAGR, saturated above 30% by design, G6) − 2 trajectory (API usage decelerating sequentially from +144% to +45% QoQ; subscriptions held stable with no separate series) + 4 margin expanding (adj. operating margin from −159% in H1 2025 to +10% in H1 2026; compute from 201% to 55% of revenue) − 10 high keyRisk (fixed compute commitments against revenue that can walk; see keyRisk) = 87.
Price Scenarios (12–24 Months)
Price vs Targets
Dashed lines mark the 12–24 month bear, base and bull targets. Daily closes.
Valuation Analysis
P/E is omitted: H1 2026 GAAP net loss was $(80.7)B, almost all of it a $72.4B non-cash fair-value charge on convertible notes, derivatives and SAFEs, while adjusted net income was $2.67B. Valuation is anchored on price/sales: at the talked ~$2.0T, ~29× annualised Q3 revenue (~$67B) and ~16× the base case 2027 revenue of ~$125B. Per-share figures use ~16.9B fully diluted shares implied by the S-1's 5% Public Benefit Fund grant and will move with the IPO's primary share count. $103 (base): the talked IPO price sits ~15% above fair value.
Where We Are vs Targets
Loading live price…
Frontier parity arrives: Gemini, OpenAI and open models match Claude on coding and agents, token prices fall faster than volume grows, and the fixed compute bill keeps margins thin; the market pays ~12× a 2027 revenue near $90B, ~$1.1T.
- Sequential revenue growth drops into single digits by mid-2027 as a top-two customer builds or switches models, and cloud-marketplace sales shift toward Gemini on Google Cloud and Nova on AWS
- Compute stays above 50% of revenue because committed TPU, Trainium and GPU capacity arrives ahead of demand, so adjusted operating margin stalls near 10%
- Lock-up windows on days 27, 48, 62 and 76 add supply into a thin float while the government-use dispute keeps federal demand out
Claude keeps its enterprise and coding lead without widening it, 2027 revenue lands near $125B, margins climb toward the mid-20s, and the stock settles near ~$1.75T after the IPO pop fades.
- Q4 2026 revenue near $22B and roughly 15% sequential growth through 2027, led by Claude Code and agent workloads
- Multi-year minimum-spend contracts lift remaining performance obligations well above the $10.8B of June 2026, reducing the share of revenue that can walk
- Anthropic Direct and the Broadcom TPU leases bring a growing share of compute onto leased hardware, holding inference gross margin in the mid-80s
The ~$110B exit run rate proves real, agents move from coding into every enterprise function, and 2027 revenue approaches $180B at 30%+ margins, re-rating Anthropic toward ~$2.7T.
- Q4 2026 revenue tracks the ~$110B run rate investors were told about, and 2027 approaches $180B
- Claude becomes the default agent model across Microsoft 365, AWS and Google Cloud marketplaces while direct sales grow faster than the 48% marketplace channel
- Owned or directly leased compute reaches the majority of capacity by 2028, giving Anthropic its own unit-cost position
ANT: common questions
Is Anthropic, PBC stock a buy?
We rate Anthropic, PBC Hold, with a composite score of 71/100: moat 67, growth 87 and valuation 60. The valuation score is struck against our bear, base and bull targets and moves with the price.
What is Anthropic, PBC's moat?
Anthropic, PBC scores 67/100 on moat. Anthropic leads enterprise and coding AI on model quality, not on a structural lock: 80% of revenue is usage-based API spend that customers can cut with limited notice, 48% runs through the clouds of three rivals, and the compute underneath is rented from those same rivals.
How fast is Anthropic, PBC expected to grow?
We estimate 40–60% a year over the next three to five years, for a growth score of 87/100.
What is Anthropic, PBC's price target?
Our base case is $103, with $65 in the bear case and $160 in the bull case.