ANTHROPIC'S IPO COULD REVEAL THE MOST EXPENSIVE AI BUSINESS MODEL YET..
Anthropic is growing at extraordinary speed. But its IPO filing reveals the other side of that growth: building a frontier AI company can require billions in computing costs before the economics of the business have fully matured.
THE NUMBERS ARE ALMOST DIFFICULT TO BELIEVE
Anthropic generated approximately $4.6 billion in revenue in 2025.
That was roughly twelve times its 2024 revenue.
But the company also recorded an operating loss of more than $8 billion.
And its reported net loss reached approximately $42 billion.
The $42 billion number needs an important qualification.
Around $34 billion of that loss came from an accounting charge related to financial instruments that could eventually convert into Anthropic shares.
In other words, it does not mean Anthropic literally spent $42 billion operating its business.
The operating economics are still extraordinary, though.
The company spent $12.65 billion on operating expenses, including roughly $7.33 billion on compute and infrastructure.
That is the real story.
ANTHROPIC'S BUSINESS HAS A MASSIVE PHYSICAL COST
A normal software company can sell another million subscriptions without constructing another million computers.
AI companies do not have that luxury.
Every time somebody asks Claude a question, the request requires computation.
Training a frontier model requires even more.
Running an increasingly powerful model for millions of customers requires enormous amounts of:
GPUs
TPUs
servers
electricity
data centers
networking
cooling
storage
So Anthropic's growth creates a strange equation:
More customers → more revenue
but also:
More customers → more inference → more compute
That is very different from traditional SaaS.
THE COST DOESN'T DISAPPEAR WHEN THE MODEL IS FINISHED
This is one of the biggest differences between AI and older software businesses.
A traditional software company might spend heavily developing a product.
Then it sells access to that product repeatedly.
The marginal cost of serving another customer can be relatively small.
With generative AI, the model keeps consuming computational resources every time it is used.
So Anthropic has two enormous categories of compute expenditure:
Training
Building and improving the models.
Inference
Actually running those models for customers.
And inference becomes increasingly important as usage grows.
CLAUDE'S SUCCESS CAN CREATE A BILL
That creates a fascinating business paradox.
Suppose Claude becomes extremely popular.
More people use it.
More companies integrate it.
More developers use Claude Code.
More agents run through Anthropic's models.
Revenue rises.
But the computational workload rises too.
So the company has to make sure:
Revenue per unit of compute > cost per unit of compute
Otherwise enormous user growth can actually create enormous costs.
That is why the AI industry's ultimate battle is not simply about who has the smartest model.
It is about who can operate intelligence economically.
ANTHROPIC'S COMPUTE SPENDING ALREADY TRIPLED
Anthropic's compute and infrastructure spending reached approximately $7.33 billion in 2025, around three times its 2024 level. It represented more than half of the company's operating expenses.
That tells investors something important.
Anthropic is not primarily spending its money like a conventional software startup.
A huge portion is going into the physical infrastructure required to make its software possible.
The business therefore sits somewhere between:
software company
and
technology infrastructure company
THEN COMES THE $518 BILLION NUMBER
The IPO prospectus reportedly says Anthropic expects to spend approximately $518 billion on cloud, computing and infrastructure obligations in coming years.
This does not mean Anthropic is writing a $518 billion check tomorrow.
It represents enormous future infrastructure commitments.
But the number reveals the scale of the bet.
Anthropic is essentially saying:
If we are going to build the AI systems we believe the market will demand, we need an extraordinary amount of computing capacity.
That creates an unusual capital structure.
ANTHROPIC NEEDS CLOUD COMPANIES TO SCALE
Anthropic is not building all of that infrastructure alone.
Amazon and Google are among its most important strategic partners.
Anthropic has agreements involving massive amounts of computing capacity, while Claude is available across AWS, Google Cloud and Microsoft Azure. Anthropic says AWS remains its primary cloud provider and training partner.
This creates a fascinating relationship.
The cloud companies are:
investors
and
suppliers
at the same time.
Anthropic needs their capital.
Anthropic also needs their computing infrastructure.
And the cloud companies benefit if Anthropic becomes one of the world's largest AI workloads.
AMAZON IS BOTH INVESTOR AND LANDLORD
That is one of the unusual characteristics of the AI economy.
Imagine a traditional startup.
An investor gives it money.
The startup uses the money to build its own servers.
With Anthropic, the relationship can be much more circular.
Amazon invests in Anthropic
↓
Anthropic buys enormous amounts of AWS compute
↓
AWS records cloud revenue
↓
Anthropic uses the compute to train and serve Claude
↓
Claude generates more revenue
↓
Anthropic needs more compute
The infrastructure provider and AI company therefore become financially intertwined.
GOOGLE HAS A SIMILAR ROLE
Google is also an investor and infrastructure partner.
Anthropic announced agreements involving up to five gigawatts of next-generation TPU capacity with Google and Broadcom, alongside additional capacity arrangements with Amazon and SpaceX.
This creates another layer of the business.
Anthropic is not simply buying generic computing.
It is securing access to specialized AI infrastructure.
The scarce resource is no longer only money.
It is compute capacity.
THAT IS WHY AI FUNDRAISING LOOKS DIFFERENT
Anthropic raised $65 billion in its Series H financing in May 2026, at a post-money valuation of approximately $965 billion.
That is an extraordinary amount of capital for a company founded only in 2021.
But the money has a very specific purpose.
Anthropic said the financing would help expand compute capacity, safety and interpretability research, and the products and partnerships supporting Claude.
In other words:
Capital → compute → models → customers → revenue
The infrastructure sits directly inside the growth strategy.
THE IPO IS ASKING PUBLIC INVESTORS TO FUND THE NEXT STEP
This is why the IPO matters.
Private investors have so far absorbed much of the risk.
Now public investors may eventually be asked to put a market value on the company.
Reuters reported that Anthropic's prospectus could support a valuation of more than $2 trillion, although that is not the same as a confirmed IPO price or final market capitalization.
The valuation question therefore becomes fascinating.
Investors are not simply asking:
How much revenue does Anthropic have today?
They are asking:
How much revenue can Anthropic generate after spending hundreds of billions on infrastructure?
WALL STREET IS LOOKING YEARS INTO THE FUTURE
Even before the latest prospectus details emerged, Reuters reported that investors were looking unusually far ahead when valuing Anthropic.
The company was reportedly projecting approximately $190 billion–$200 billion of revenue in 2028, compared with a run-rate figure of more than $47 billion disclosed in May.
That is an enormous assumption about future growth.
The logic is essentially:
Today's huge infrastructure spending
↓
much larger AI adoption
↓
much larger revenue
↓
eventually improving economics
The IPO will force investors to decide how much of that future should already be reflected in the valuation.
THIS IS NOT TRADITIONAL SOFTWARE MATH
A SaaS company is often analyzed through metrics such as:
ARR
gross margin
retention
customer acquisition cost
operating margin
AI companies still use these metrics.
But they add another critical question:
How expensive is intelligence to produce?
A company could have spectacular revenue growth while still having poor economics if every additional dollar of revenue requires too much compute.
That makes inference efficiency strategically important.
THE MODEL HAS TO GET CHEAPER
Anthropic's long-term economics therefore depend partly on technological progress.
If newer hardware becomes more efficient:
same AI workload → less electricity → less compute cost
If models become more efficient:
same customer task → fewer tokens/less computation → lower cost
If software optimization improves:
same hardware → more useful AI work
The company's economics can improve without simply raising prices.
BUT DEMAND COULD OUTRUN EFFICIENCY
There is another possibility.
AI becomes cheaper.
So people use much more of it.
The savings from efficiency are then partly consumed by higher usage.
This is one of the fundamental questions facing the industry.
If inference becomes 10 times cheaper but customers use 50 times more AI, total compute demand can still explode.
Anthropic's infrastructure commitments suggest management is preparing for exactly that possibility.
CLAUDE CODE CHANGES THE EQUATION
Anthropic is no longer selling only a chatbot.
Claude is increasingly being used for software development through Claude Code and for more autonomous work through its broader agent products.
That matters because an agent can generate substantially more model usage than a single question-and-answer interaction.
One human asking:
"Explain this code."
is one type of workload.
An AI agent repeatedly:
reads code → writes code → tests code → analyzes errors → rewrites code → runs tests again
creates a much larger computational workload.
The more useful Claude becomes as a worker, the more compute Anthropic may need to sell that usefulness.
THE BEST PRODUCT COULD ALSO BE THE MOST EXPENSIVE PRODUCT TO OPERATE
This creates a strange economic loop.
Better AI:
→ more useful
→ more customers
→ more tasks
→ more usage
→ more revenue
but also:
→ more inference
→ more compute
→ more infrastructure commitments
So Anthropic's objective cannot simply be maximize usage.
It needs to maximize economically valuable usage.
That is a much harder problem.
ENTERPRISE CUSTOMERS COULD BE THE KEY
Enterprise AI may help solve that equation.
Businesses can pay substantially more than individual consumers when AI becomes part of a revenue-generating or cost-saving workflow.
A company might pay for Claude because it helps:
write software
analyze documents
automate customer support
research markets
process contracts
operate internal workflows
The more deeply Claude becomes embedded in those processes, the more defensible the revenue can become.
BUT CUSTOMER CONCENTRATION IS A RISK
Anthropic's prospectus also reportedly shows that nearly one-quarter of revenue came from just two customers in 2025.
That creates another important IPO issue.
Rapid revenue growth is valuable.
But investors also need to know where the revenue comes from.
If a small number of customers account for a large percentage of revenue, losing one major contract could have an outsized effect.
Anthropic itself reportedly warns that many major customers are not locked into long-term contracts and could reduce or stop spending.
CASH PROVIDES SOME BREATHING ROOM
Anthropic had approximately $20.28 billion in cash, cash equivalents and short-term investments at the end of 2025, according to the reported prospectus.
That is an enormous liquidity position.
But against hundreds of billions of future infrastructure commitments, the number looks different.
It is not necessarily a sign that Anthropic lacks money.
It is a reminder of the scale of the capital requirements involved.
THE AI COMPANY IS STARTING TO LOOK LIKE AN INFRASTRUCTURE PROJECT
This may be the most unusual aspect of the IPO.
A software startup traditionally raises money to:
hire engineers
build product
sell software
Anthropic needs to do all of that.
But it also needs to secure enormous amounts of computing infrastructure.
That makes its business resemble a combination of:
software
semiconductor demand
cloud computing
data-center infrastructure
energy consumption
The AI application may be digital.
Its cost base is increasingly physical.
THAT CHANGES THE INVESTMENT QUESTION
The public-market question is therefore not simply:
"Will people use Claude?"
The evidence already shows strong demand.
The more difficult question is:
"Can Anthropic turn that demand into sustainable margins after paying for the infrastructure required to deliver it?"
That is the economic problem hidden underneath the AI race.
ANTHROPIC'S IPO COULD BECOME A BENCHMARK FOR THE ENTIRE INDUSTRY
If Anthropic goes public at a valuation around or above the levels being discussed, investors will receive one of the first large-scale public-market windows into a frontier AI laboratory.
That means analysts will be able to examine:
revenue growth
compute costs
gross margins
infrastructure commitments
customer concentration
cash requirements
future capital needs
Instead of relying primarily on private funding rounds, the market will eventually have a continuously updated price for the business.
That changes the game.
PRIVATE AI VALUATIONS WILL MEET PUBLIC ACCOUNTING
Private markets can tolerate enormous uncertainty.
A venture investor can say:
"This company could become worth $2 trillion."
A public company eventually has to publish financial statements and explain how it gets there.
The IPO therefore turns the AI story from:
venture-capital thesis
into:
public-company financial model
That is why Anthropic's filing is so interesting even before the first share trades.
THE $42 BILLION LOSS CAN BE MISLEADING — BUT THE $8 BILLION OPERATING LOSS ISN'T
This distinction deserves repeating.
The headline $42 billion loss includes a huge accounting charge related to financing instruments.
It should not be interpreted as $42 billion of cash disappearing through ordinary operations.
But the company still recorded an operating loss of more than $8 billion in 2025.
And it spent approximately $7.33 billion on compute and infrastructure alone.
So the underlying economic challenge is very real even after stripping away the unusual accounting item.
ANTHROPIC'S BUSINESS MODEL IS EXPENSIVE BECAUSE THE PRODUCT IS COMPUTATION
This is ultimately the simplest way to understand the story.
Netflix needs servers.
Spotify needs servers.
Salesforce needs servers.
But Anthropic's product itself is computation.
When the customer buys Claude, they are effectively buying access to intelligence generated by enormous amounts of computing power.
That means the cost of producing the product rises directly with usage.
The company therefore has to continuously improve the ratio:
AI value delivered / compute required
That ratio may become one of the most important metrics in the entire AI industry.
MAACAT PERSPECTIVE
Anthropic's IPO could reveal something much bigger than how much Claude is worth.
It could reveal what a frontier AI company actually costs to operate once the numbers are exposed to public markets.
The basic machine looks like this:
Capital → compute → better models → more customers → more usage → more revenue
But there is another arrow running in the opposite direction:
More usage → more inference → more compute → more infrastructure spending
That is what makes the AI business model so unusual.
Anthropic generated nearly $4.6 billion in 2025 revenue, yet spent billions on infrastructure and still recorded an operating loss of more than $8 billion. At the same time, it is preparing for infrastructure commitments measured in the hundreds of billions.
The ultimate question is therefore not whether AI can generate enormous revenue.
It already can.
The question is whether intelligence can eventually become cheap enough to produce that the revenue growth outruns the cost of the machines producing it.
If Anthropic's IPO gives public investors the first detailed answer to that question, its financial statements could become one of the most important documents in the AI economy.
