DXYN

Investors got a preview of Anthropic's S-1 filing ahead of its upcoming IPO when The Wall Street Journal recently reported details from a draft of the filing.

And while it came as no surprise that Anthropic spends a lot of money on artificial intelligence (AI) compute capacity -- essentially, the computing power it needs to run and train Claude -- it was shocking to learn that spending could reach $750 billion over the next decade.

Here's where the money is going and why it's a major hurdle on Anthropic's path to profitability.

Where is that $750 billion going?

Anthropic's costs are so high because the company has to invest in compute capacity to run AI queries and training. Every time you type a request into Claude, it takes a certain number of "tokens" to process it and return an answer.

And those tokens are very expensive. Recent data from SemiAnalysis shows that while some Claude Max users pay $200 per month for their subscriptions, if they actually used all of their tokens, it would cost Anthropic about $8,000.

Anthropic said it could spend up to $750 billion in compute spending over the next decade or so, with the bulk of the money roughly going to these companies:

• $160 billion to lease Alphabet processors in a deal with Broadcom
• $111 billion for Google Cloud services
• $110 billion to Amazon for its Amazon Web Services (AWS)
• $86 billion to rent Nvidia semiconductors from third-party companies
• $31 billion to Microsoft for its Azure cloud services

That's not all of the company's total spending, but it represents most of it. Also worth mentioning is that $450 billion in cloud compute spending contracts are "noncancelable," according to WSJ.

That means that over the next seven to 10 years, Anthropic has to pay that amount to cloud providers, even if it doesn't actually use the computing power. You could see why that might be a problem for Anthropic (and its future shareholders) in the coming years if, say, Claude usage slows down or rising competition from other chatbots and AI agents stifles the company's growth.

To put this all into context, Anthropic spent more on AI compute in the first half of this year than it did for all of 2025 -- and it still needs more.

Its $80 billion loss isn't as bad as it looks, but Anthropic is still risky

If there's some good news in the draft prospectus, it's that revenue is accelerating very quickly. The company generated $16.3 billion in sales in the first half of this year, compared to just $1.1 billion in the comparison period of 2025.

As for its losses, they reached a staggering $80 billion in the first half of the year, of which $72 billion was due to an accounting adjustment primarily tied to Anthropic's convertible notes.

Still, Anthropic's losses are significant despite its accelerating sales, and the company's long-term computing commitments mean that it could have a difficult time reaching profitability any time soon on a generally accepted accounting principles (GAAP) basis. The company did say in the prospectus that it was slightly profitable on a non-GAAP (adjusted) basis in the third quarter.

Given its current spending, GAAP profitability doesn't appear to be in the cards anytime soon. And even profitability on an adjusted basis could come under pressure if Anthropic loses a major customer (two of its customers account for 20% of total revenue), or if revenue growth decelerates.

All of this means that potential Anthropic investors will have to assume a fair amount of risk if they buy this AI stock after its upcoming IPO.