Anthropic’s IPO: What the numbers reveal

Key points:

• Massive demand comes with massive investment needs: Anthropic’s first-half revenue grew roughly 14x, alongside around $637bn in named multiyear computing arrangements.
• The operating picture is improving: Q2 delivered $974m in operating profit, while Claude Code and Cowork offer opportunities to expand customer usage and spending.
• The implications extend beyond the IPO: Cloud and chip partners stand to benefit from Anthropic’s expansion, but contract terms, delivery schedules and valuations matter.

---

Anthropic filed its S-1 IPO prospectus yesterday, October 26, putting numbers behind what we already know about AI: demand is enormous, and so is the investment needed to meet it. Revenue is growing rapidly as customers put Claude to work on more tasks, while the company is committing hundreds of billions of dollars to computing capacity.

Beneath those headline figures, the filing reveals how the business earns money, how its costs are changing, and what its expansion could mean for investors. Anthropic plans to list on Nasdaq under ANT, although the initial filing does not specify the offering price.

What the numbers say

Metric	What Anthropic disclosed
First-half revenue	$16.3bn, versus approximately $1.16bn a year earlier—roughly 14x growth
First-half operating result	A $693m loss, down from $2.14bn a year earlier
Q2 operating result	$974m profit on approximately $11.6bn of revenue—an operating margin of around 8.4%
First-half net loss	$80.7bn, including approximately $72bn in non-cash charges linked to financing instruments
First-half compute and infrastructure spending	Nearly $9bn, versus $2.3bn a year earlier
Cash and short-term investments	Approximately $132bn at June-end

Sources: Anthropic S-1

Making sense of the numbers

1. The operating business has improved more than the loss headline suggests

The $80.7bn net loss is heavily affected by non-cash accounting charges. It should not be read as $80.7bn of cash burned.

More revealing is the move to operating profitability in Q2, even though the first half overall remained loss-making. One profitable quarter does not establish a lasting trend, but it marks a significant improvement in the business.

2. Revenue has grown faster than compute spending

Across the first half, revenue grew roughly 14x, while compute spending rose about 4x. Revenue per dollar of compute spending improved from approximately $0.50 to $1.80.

That is encouraging, but it is not a profit margin or proof that every AI task costs less to deliver. Anthropic warns that new capacity can bring expenses before the revenue it supports, potentially squeezing margins.

3. The usage model helps explain customer spending

Approximately 83% of 2025 revenue came from usage-based customers, alongside subscriptions. Customers spend more as they use Claude for more work.

That helps explain the approximately 900% net revenue retention figure: the measured customer group generated roughly nine times its previous revenue, including expansion, reduced spending and customer losses. This measure excludes fixed-price Pro and Max subscriptions.

The large-customer numbers are striking too. More than 1,900 customers generated at least $1m over the preceding year, including around 250 that generated at least $10m each. That smaller group alone contributed at least $2.5bn, by calculation.

Many customers purchase usage through cloud marketplaces, which handled 48% of first-half 2026 revenue. This means customers buying Claude through platforms such as AWS, Google Cloud and Azure—not those platforms themselves consuming nearly half of Claude’s services.

4. Code opened the door. Cowork could widen it

Claude Code helps developers work with software. Cowork extends the opportunity into broader workplace tasks and teams outside software development.

Anthropic says Cowork adoption outpaced Code at the equivalent stage after their research previews. That is an adoption comparison, rather than evidence that Cowork already generates more revenue.

The opportunity is to expand from one team using Claude to multiple departments using it throughout the day. In a consumption business, that can increase spending within an existing customer account.

The next useful evidence will be whether Cowork becomes a meaningful second revenue engine alongside Code.

5. Meeting that demand requires substantial commitments

Anthropic’s approximately $132bn in cash and short-term investments provides considerable funding capacity. Much of that reflects fundraising.

Against that cash cushion sits a substantial infrastructure bill. Seven major arrangements total approximately $637bn, according to Bloomberg’s summary:

Partner	Amount	Duration
Microsoft	$44.9bn	Through May 2033
Google	$114.5bn	Through July 2033
Amazon	$110bn	Through April 2036
Broadcom	$161.2bn	Over the leases’ lives
Nscale	$44.6bn	Into 2033
Lambda	$41.3bn	Into 2034
SpaceX	$120.5bn	Into 2029

These amounts cover several years and should not be treated as one year’s spending or a single debt balance. Most agreements offer limited scope to cancel, although exceptions apply. SpaceX’s arrangement allows cancellation on 90 days’ notice once the initial fixed periods end.

Securing capacity can support growth and reduce supply uncertainty. But the payment schedules matter: how much must Anthropic pay, when must it pay, and how much flexibility does it retain if its needs change?

What investors can do now

Anthropic is not yet publicly traded, but its disclosures provide concrete information for assessing listed companies connected to its expansion.

Company	Connection to Anthropic	What to assess
Amazon (AMZN)	AWS capacity, Trainium chips and Claude distribution	Whether capacity deployment translates into AWS revenue and profit
Alphabet (GOOGL)	Google Cloud, TPUs and Claude distribution	Cloud growth and TPU adoption relative to infrastructure spending
Microsoft (MSFT)	Azure capacity and Claude distribution	Delivery schedules and conversion of commitments into revenue
Broadcom (AVGO)	TPU infrastructure partnership and equipment arrangements	Deployment timing, revenue recognition and customer concentration
Nvidia (NVDA)	GPUs used to train and run Claude	GPU demand alongside Anthropic’s use of alternative chips

These companies offer different exposures, and none is a direct substitute for Anthropic shares. A connection alone does not make a compelling investment: the relevant questions are how meaningful Anthropic is to the company’s earnings, what investment is required to serve it, and how much expected growth is already reflected in the share price.

Investors may use the prospectus to build a focused watchlist and sharpen their questions for upcoming earnings. A large multiyear agreement should prompt scrutiny of delivery and profitability, rather than being treated as immediate revenue.

What could challenge the story?

• Spending commitments arrive before guaranteed revenue. Much of Anthropic’s revenue depends on customer usage, while several computing agreements require minimum payments. If usage grows more slowly than expected, those bills still need funding.

• Q2 profitability may not be a straight-line trend. New capacity, model training and product development could push costs up before additional revenue arrives.

• Competition could affect pricing. Customers can choose competing models. Anthropic will need to keep improving Claude while managing the price it can charge for its services.

• Partners have competing interests. Amazon, Google and Microsoft help distribute Claude and supply computing capacity, but also compete in AI. Their decisions on pricing and capacity allocation matter to Anthropic.

• The IPO price could leave little room for disappointment. Strong business growth does not automatically make the shares attractive at any valuation. The offering terms will determine how much future success investors are being asked to pay for.

What to watch next

• The next quarter’s operating results: does Q2’s profitability persist as computing capacity expands?
• Cowork’s commercial contribution: does faster adoption translate into meaningful paid usage?
• Customer expansion: how does net revenue retention evolve as existing accounts become larger?
• Cash flow and payment schedules: how quickly do infrastructure arrangements draw on available funding?
• The IPO terms: what valuation, share count and dilution will public investors face? Compare actual revenue with clearly labelled forecasts and annualised run rates.