Anthropic Has No Moat: The Bear Case Before the IPO

I use Claude every day, mostly through Claude Code, and its models are the best I’ve used. I’m writing the bear case anyway, because a good product and a protected company are separate questions.

Anthropic, the company that makes Claude, is about to sell shares to the public for the first time. The amended S-1, the document a company must file first and where it has to list what could go wrong, sets a price of $70 to $78 a share for 825 million shares.

Trading on Nasdaq under the ticker ANT could start as soon as the week of November 16. The Motley Fool’s estimate puts the whole company at about $1.27 trillion to $1.41 trillion at that price.

My claim is that Anthropic has no moat, meaning nothing that keeps a customer once a rival is as good and cheaper. There are three reasons:

• The product is the model, and a customer can swap a model in an afternoon.
• The lead over cheaper models is about four months, and the price is falling from both sides.
• The bills are fixed for years, while customers pay as they go.

That leaves two questions I take at the end: whether a better model could crush Anthropic, even to a 90% fall in the stock, and whether any of this means you should short it.

I’m a developer, not an analyst, and none of this is investment advice. The filing’s numbers come through the breakdowns by Kepler, Mostly Metrics, Fortune and Digital Applied, and I name the one I used next to each figure.

The business is real, and I want that said first. Digital Applied’s read of the filing has revenue at $386 million in 2024, $4.6 billion in 2025 and $33.2 billion in the first nine months of 2026. That’s 13 times the same months of 2025.

What a moat is, and why Anthropic doesn’t have one

The word comes from castles, and investors usually credit it to Warren Buffett. A moat is the reason a customer stays when a rival offers the same thing for less.

Apple’s iPhone has one. Your photos, your messages and the apps you paid for live there, and an Apple Watch only works with an iPhone. A cheaper Android phone with a better camera still loses, because leaving costs you your stuff.

ASML has a different kind. It’s a Dutch company and the world’s only manufacturer of the EUV machines that print the most advanced chips, at more than $350 million for the newest ones. Intel, TSMC and Samsung can’t take their business elsewhere, because there is no elsewhere.

Visa has a third kind. Shops take it because shoppers carry it, and shoppers carry it because shops take it, so a new card network with lower fees starts with nobody on either side.

Those are three of the five kinds investors count. The other two are making the thing cheaper than anyone else can, and being so expensive to copy that nobody tries.

Anthropic against each one:

The product is the model, and a model can be swapped in an afternoon.

Anthropic sells Claude in several wrappers: chat apps, Claude Code for programmers, Cowork for office work, and direct access for other companies’ software. Every wrapper runs on the same few models. The model is the product.

The filing agrees. Rock Cyber quotes its strategy as one that “prioritizes model capability as a principal source of differentiation”.

A model is closer to a gas station you like than to an iPhone. You drive past a cheaper station because the fuel is better, and nothing about your car stops you from filling up somewhere else tomorrow.

A company uses a model by sending it text and getting text back. It pays by the token, a chunk of text about three-quarters of a word long. Everything the company built around Claude, its software, its instructions and its data, sits on its own computers.

Anthropic’s own documentation shows how small the change is. This is its example of calling Claude through the software kit OpenAI publishes for OpenAI’s models:

1
import OpenAI from "openai";
2

3
const openai = new OpenAI({
4
  apiKey: process.env.ANTHROPIC_API_KEY, // Your Claude API key
5
  baseURL: "https://api.anthropic.com/v1/" // Claude API endpoint
6
});
7

8
const response = await openai.chat.completions.create({
9
  messages: [
10
    { role: "system", content: "You are a helpful assistant." },
11
    { role: "user", content: "Who are you?" }
12
  ],
13
  model: "claude-opus-5-5" // Claude model name
14
});
15

16
console.log(response.choices[0].message.content);

Skip the code if you don’t write software. The line that matters is baseURL, the address the request goes to. Change it and the model name, and the same program talks to a different company.

Anthropic says this layer is “primarily intended to test and compare model capabilities”. Comparison shopping is a documented feature.

Claude Code is the wrapper I use all day, and even that isn’t tied to Claude. Three rival labs (Z.ai, Moonshot and DeepSeek) copied the way Anthropic’s servers talk, so Claude Code runs on their models.

This is the settings file from Z.ai’s guide:

1
{
2
  "env": {
3
    "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
4
    "ANTHROPIC_AUTH_TOKEN": "your_zai_api_key",
5
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5.3[1m]",
6
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-5.3[1m]",
7
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-5.3-flash[1m]"
8
  }
9
}

With that file in place Claude Code looks and works the same, and a Chinese model called GLM does the thinking. Anthropic’s gateway docs say it “doesn’t support routing Claude Code to non-Claude models through any gateway”.

ANTHROPIC_BASE_URL is documented on that same page, and Moonshot and DeepSeek publish the same kind of guide.

My own setup has nothing Claude-specific in it. My ~/.claude/CLAUDE.md is one line that imports a shared AGENTS.md, and my skills live in ~/.agents/skills with symlinks into ~/.claude/skills.

The plugs are shared too. MCP, the standard that lets a model use outside tools, was created by Anthropic and given to the Linux Foundation’s Agentic AI Foundation in December 2025. That was good for developers, and it means the tools built around Claude work with every other model.

Big customers have started making the move. At AI Engineer NYC on October 13, David Abitbol of the fintech Revolut said it now runs GLM 5.3 itself, on 56 Nvidia B200 chips, as the default for about 2,000 engineers. The blended cost is near $0.30 per million tokens.

Revolut keeps Claude Code and OpenAI’s Codex for the problems GLM can’t handle. The talk page also warns that the numbers “describe Revolut’s deployment and utilization, not a portable price for any self-hosted model”.

Cursor, the code editor, did the same thing as a product. Its Composer 2.5 model is built on Moonshot’s Kimi K2.5 and costs $0.50 per million tokens in and $2.50 out. By Cursor’s own benchmarks it scores 79.8% on SWE-Bench Multilingual against 80.5% for Claude Opus 4.7.

Cursor used to be one of Anthropic’s biggest customers. The S-1 says two customers were 12% of revenue each in 2025, per Digital Applied, and VentureBeat’s reporting, via PYMNTS, put Cursor and GitHub Copilot at about $1.2 billion of Anthropic’s $5 billion in revenue. SpaceX closed its $60 billion purchase of Cursor on August 14.

Meta moved too. PYMNTS reports it cut its Claude Code users from about 60,000 to 30,000, mostly by moving them to its own tools.

The customer base did get wider. No single customer was above 10% of revenue in the first nine months of 2026, and the top five were about 25%, per Digital Applied.

Nothing in how Anthropic charges slows a customer down, though. 79% of nine-month revenue was pay-as-you-go, billed by the token the way electricity is billed by the kilowatt-hour, so there’s no contract to wait out.

Morningstar, which values Anthropic at $1.7 trillion, declines to count switching costs as a moat for that reason.

The lead is about four months, and rivals learned from Claude itself.

Anthropic’s models are at the top right now. On Vals AI’s Terminal-Bench 4.0, a coding benchmark, Claude models held the top three places on November 9.

The trouble is what happens a few months later. Open-weight models are ones whose files anyone can download and run on their own machines, so they sell for close to what the chips cost to run.

Epoch AI measured how far behind they are. From January through May 2026 the best open-weight models trailed the best closed ones by about four months on its capability index, and a stricter version of the test stretches that to six.

Part of how they catch up is by asking Claude. In February 2026 Anthropic reported that three labs, DeepSeek, Moonshot and MiniMax, used about 24,000 fake accounts to run more than 16 million exchanges with Claude and train their own models on the answers.

I can’t tell you how much of any rival model came from Claude. I can tell you Moonshot’s Kimi K3.1 Pro sits fourth on that Terminal-Bench table, 1.5 points behind Claude Fable 5.2, and that an earlier Kimi is the base of Cursor’s Composer.

A recipe that can be learned by tasting the dish is hard to keep secret.

Then there are models you run on hardware you own. A guide to Mac models lists OpenAI’s own open-weight gpt-oss:20b running in 16GB of memory, and Qwen’s 27B model as a 17GB download with a licence that allows commercial use.

None of them matches a current Claude model, and I found no test that says one does. The big ones are getting close to desk size, though. Meta’s Muse Spark Open, released October 15, needs about 1TB of memory, which Engadget prices at roughly $38,000 for four Mac Studios.

A local model doesn’t have to beat Claude. It has to be good enough for the easy work, and most work is easy.

AT&T is the example. PYMNTS reports it cut the cost of some AI tasks by as much as 56% by sending them to cheaper models, with quality down 2%. It wants open-source models handling 60% to 70% of its queries, up from 40%.

Menlo Ventures’ 2026 survey of nearly 400 enterprise buyers shows the same shift. Open-source models went from 11% to 19% of the tokens companies run in one year, while collecting 4% of the dollars.

Menlo lists Anthropic as a portfolio company, so read its numbers with that in mind.

The models don’t last either. Mostly Metrics found in the S-1 that models released in the previous six months made up over 80% of third-quarter revenue. Twice a year Anthropic has to win its revenue again with a new model.

The price can’t hold, because customers have started rationing.

A model bills like a taxi: the meter runs for as long as it works. In 2026 the coding tools got good enough to work for hours on their own, and the fares got big.

The Next Web reports that these tools “tripled enterprise AI bills even as per-token prices collapsed”. It says Microsoft found some staff spending $500 to $2,000 a month on Claude Code tokens alone.

So the biggest buyers capped the meter. Bloomberg reported, via 9to5Mac, that Apple cut some engineers’ daily AI allowance from $300 to $200 in October and is weighing $100. Claude, ChatGPT and Gemini are the services its staff can use.

Apple’s staff outside engineering already have a $100 cap, and the company has considered $25 for some teams. The same report says few workers were using their full $300, so this is a ceiling coming down, with the savings unclear.

Apple has company. PYMNTS reports Uber used up its 2026 AI budget by April and now limits each employee to $1,500 a month per coding tool, and that Tesla set a $200 weekly limit.

The Information reported, per PYMNTS, that Microsoft was on track to spend $1 billion on Anthropic’s tools internally this year and cut that by about a third. 9to5Mac, citing Forbes, has Accenture’s centrally funded Claude allowance at $29 per employee a month.

When the biggest customers cap what they spend, the seller either lowers the price or loses the usage. Anthropic’s own price list shows which way it went, in dollars per million tokens:

The Sonnet row carries a footnote on Anthropic’s page. The $2 and $10 price was “announced at launch as introductory pricing through August 31, 2026” and “is now the standard price”. I read that as a price rise that was planned and didn’t happen.

The rest of the market is public too. This command asks OpenRouter, a marketplace that resells every lab’s models, what each one costs:

1
curl -s https://openrouter.ai/api/v1/models | jq -r '
2
  .data[]
3
  | select(.id | test("^(anthropic/claude-(fable-5.2|opus-5.5|sonnet-5.5)|openai/gpt-6.1-(sol|terra)|google/gemini-4-argon|x-ai/grok-4.7|z-ai/glm-5.3|deepseek/deepseek-v4.1-flash)$"))
4
  | [.id, (.pricing.prompt | tonumber * 1e6), (.pricing.completion | tonumber * 1e6)]
5
  | @tsv' | sort -t$'\t' -k3 -nr | column -t

The output on November 10, in dollars per million tokens. The first number is for text you send and the second is for text the model writes:

1
anthropic/claude-fable-5.2    8       40
2
anthropic/claude-opus-5.5     4       20
3
openai/gpt-6.1-sol            2       10
4
google/gemini-4-argon         2       10
5
anthropic/claude-sonnet-5.5   2       10
6
x-ai/grok-4.7                 2       6
7
openai/gpt-6.1-terra          1       5
8
z-ai/glm-5.3                  0.0699  1.4
9
deepseek/deepseek-v4.1-flash  0.3     1.2

Price isn’t quality. Fable and Opus solve problems the cheap rows can’t, which is why I use them. GLM 5.3 is the model Revolut runs, though, and its output costs 7% of what Opus 5.5 charges.

The filing names the risk. Digital Applied quotes its warning of “pricing pressure as a result of competition or otherwise”, and Fortune quotes it saying price competition could “adversely affect our ability to achieve or sustain profitability”.

The bills are fixed for years, and the customers pay as they go.

Picture a restaurant that signs a ten-year lease on the biggest building in town. Its customers pay per meal and can eat next door tomorrow. If they do, the rent is still due.

Anthropic’s rent is for chips and data centers, which the industry calls compute. The November 2 amendment, as Digital Applied reads it, commits Anthropic to $895.4 billion of spending, nearly all of it compute.

$447.5 billion of that is expected before the end of 2030. By my arithmetic the total is 27 times the revenue of the last nine months, or 13 times the third quarter annualized.

The landlords are the awkward part. These are the biggest contracts in Kepler’s breakdown, with what each supplier also does:

The SpaceXAI deal is the flexible one: either side can cancel with 90 days’ notice once the initial periods end in January and March 2027. The rest are minimums.

Three of those suppliers are also the store. 44% of nine-month revenue went through the AWS, Google Cloud and Azure marketplaces, per Digital Applied.

The S-1 is blunt about the conflict. Mostly Metrics quotes it saying the cloud companies, which it calls CSPs, hold “significant equity investments in our company”, that “These CSPs are also our customers”, and that they have “visibility into our pricing and commercial terms”.

So Google rents Anthropic the chips, sees its commercial terms, sells Gemini 4 Argon at Sonnet’s price and takes a fee when a customer buys Claude through Google Cloud.

Those fees are why the profit on each sale is lower than it looks. Gross margin is what’s left of a dollar of sales after the direct cost of delivering it, and Mostly Metrics found the inputs in Note 8: $2.4 billion of cost against $16.3 billion of first-half revenue, an 85% margin.

The marketplace fees sit in a different line, sales and marketing, where Kepler shows channel partner fees rising by $2.7 billion. Les Barclays adds them back and gets about 65%, and Mostly Metrics lands in the low 60s.

The mismatch is in the dates. The leases run five years and the cloud contracts run to 2033 and 2036, while four-fifths of the revenue that pays for them is billed by the token. If a cheaper model takes the bulk of the work, the bill stays.

Growth has gone from absurd to fast. Second-quarter revenue rose 144% over the first, from $4.7 billion to $11.6 billion in Kepler’s tables, and the Fool has the third quarter up 47%, to $17.0 billion.

What it means for the stock

Can a better model crush Anthropic, and could the stock fall 90%?

A better model can take the customers, and Anthropic is the proof. In Menlo’s surveys, OpenAI held 50% of enterprise LLM spending in 2023 and 27% by 2025. Anthropic went from 12% to 40% over the same two years, and OpenAI’s customers had nothing holding them either.

It can happen in a single year. In the same surveys Google fell from 21% in 2025 to 9% in 2026.

The field is crowded with companies that can afford to try. OpenAI, Google, Meta and SpaceXAI all train their own models, and Z.ai, Moonshot and DeepSeek sell theirs for a fraction of the price.

Today the better model is Anthropic’s. Menlo has it at 55% of enterprise LLM spending in 2026, and OpenAI at 31%.

The margin at the top is thin, though. Fable 5.2 scores 65.15% on that Terminal-Bench test and Kimi K3.1 Pro scores 63.64%. OpenRouter lists Kimi K3.1 Pro at $2 in and $10 out, a quarter of Fable’s price.

Now the 90%. That fall would take the company from about $1.3 trillion to about $130 billion, which is roughly the cash and investments it holds today: Digital Applied counts $95.6 billion in cash and $36.3 billion in short-term investments.

So the market would have to value the business itself at nothing, and one better model from a rival doesn’t do that. OpenAI lost almost half its share in two years and still holds 31%.

What could do it is the combination. Customers ration and swap, growth stalls, and the $895 billion of bills keeps arriving.

Aswath Damodaran, the NYU professor who values companies in public, treats those commitments as debt. On October 28 he valued Anthropic’s shares at $1.475 trillion in total, assuming revenue of about $1 trillion by 2036 at a 27.5% operating margin.

His worst case, $200 billion of revenue at a 10% margin, comes out at negative $373 billion. That’s worse than a 90% fall, and he builds in a one-in-ten chance that the company fails.

Morningstar’s bear case is $500 billion, more than 60% below the IPO range.

The price has already slipped once. The Fool notes that October reports had the IPO at $1.5 trillion to $2 trillion, and the range filed on November 10 works out to less than that before a single share has traded.

Markets also overreact to this exact news. On January 27, 2025, after the Chinese lab DeepSeek released a cheap model, Nvidia fell 17% and lost about $589 billion of value in one day. Nvidia sells the chips and doesn’t make a competing model.

One thing does slow a rival down: chips. A lab with a better model still needs data centers to serve it, and Mostly Metrics quotes the S-1 admitting that Anthropic itself “sought quantities of compute from CSPs in excess of what those partners were willing to provide”.

So the $895 billion of commitments is also the closest thing Anthropic has to a moat. It’s a rented one.

Should you short it?

Shorting means borrowing shares, selling them today and buying them back later. If the price falls you keep the difference. If it rises there’s no ceiling on what you lose, and you pay a fee for every day you hold the loan.

I wouldn’t. Everything above is about the business over years, and a short is a bet on the share price over weeks.

Start with how few shares will trade. 825 million shares out of about 16.9 billion is roughly 5% of the company, and you can only borrow shares that trade. Bloomberg reported on November 10 that orders covered the whole deal on the day it launched.

Then there are buyers with no choice. Nasdaq lets a new listing join the Nasdaq-100 index after about 15 trading days if it ranks in the top 40 by value, and every fund that tracks the index then has to buy it.

SpaceX is the closest precedent, and Leverage Shares kept a record of its first weeks. It priced at $135 on June 11, 2026, opened at $150 the next day and closed at a record $201.80 on June 16. Someone who shorted the first trade was down 35% in three trading days.

SpaceX joined the Nasdaq-100 on July 7. On July 15 it closed at $135.27, and that short was up 10%.

On November 10 the Fool’s quote had it around $141. The bears were right about the price and still had to survive June.

Nothing in my case has a date on it either. Anthropic can pay its bills for years, and a short pays to wait.

The dates a bear would watch are in the filing. Damodaran reads the lock-up, the rule that stops insiders from selling straight away, as letting them sell up to 10% of their shares on day one, the rest in stages over 180 days, and the founders after a year.

There are other ways to bet against a stock, and each has a catch. Options can list from the second business day after an IPO this size, by optionstrading.org’s reading of Cboe’s rules, and options on a new listing are expensive.

Leverage Shares has short Anthropic funds marked “coming soon”. Daily funds like those are built for one-day bets and drift away from the stock over weeks.

The cheapest way to act on a bear case is to not buy.

A buyer should also know they don’t get a vote that counts. Fortune reports that the five members of the Long-Term Benefit Trust can appoint a majority of the board, and Mostly Metrics quotes the filing saying the structure “may effectively eliminate the ability of holders of our Class A common stock to influence important decisions”.

That structure exists to protect the safety mission. It also means public shareholders can’t change course if the economics turn.

The numbers that argue against me are strong. Kepler has net revenue retention at about 900% at September 30, meaning last year’s customers spend roughly nine times what they did, and more than 11,900 customers spending over $100,000 a year, up from just over 950. Prices fell all year and revenue still grew 13 times.

The losses are smaller than the headline. Digital Applied puts the nine-month net loss at $104.0 billion, and $94.3 billion of it is the accounting cost of convertible notes and SAFEs rising with the valuation.

The operating loss for those nine months was $1.3 billion. Kepler shows a $974 million operating profit in the second quarter, and the Fool reports a $567 million loss in the third.

Morningstar also disagrees with my headline. It rates the moat “narrow”, on Anthropic’s technology and cost position.

I’d change my mind if the Sonnet price holds for a year, or if pay-as-you-go revenue turns into multi-year contracts.

I’ll keep using Claude after the IPO. The day a cheaper model does my work as well, I’ll change one line in settings.json, and so can every customer in that S-1.