Hi Reader,
As you've likely heard, Anthropic filed to go public on Monday. I've spent the week digesting the S-1.
You'll see the revenue numbers everywhere, but there is another number I haven't stopped thinking about: Anthropic has committed to spend more than half a trillion dollars on computing power. Regular people are already helping pay for that through their retirement accounts, their utility bills, and their water.
This week's piece covers what's actually in the filing, where I got it wrong in an earlier piece, what the filing says (and doesn't say) about the communities hosting these data centers, and who I think should get the benefit.
It's Part Four of my series on the AI IPOs. As always, I welcome your comments and thoughts!
Best,
Shruti
Anthropic Filed Its S-1. Here's Who's Paying for It.
What Anthropic's S-1 says about who pays for AI. And who it's for.
By Shruti Shah, General Partner, Symphonic Capital.
Anthropic filed its S-1 on Monday.
I spent the last few days reading it. The headline numbers are pretty wild. Revenue went from $41 million in 2023 to $16.3 billion in just the first half of this year. About 1,900 customers now pay Anthropic more than $1 million a year. You'll likely see those numbers everywhere this week, so I won't spend much time on them here.
The number that I think is really worth paying attention to is a different one. Anthropic has committed to spend more than half a trillion dollars on computing power over the coming years. Here's what's peculiar about this: that money pays for compute in the data centers Anthropic rents to train its models and run them for customers. Yet every new model Anthropic releases does the same work for less money than the one before it. Newer models finish tasks more efficiently, and anthropic often cuts its prices when it launches them too. Its newest model cut what customers pay per task by about 40%.
Why would a company make its own product cheaper? Anthropic is betting that as AI gets cheaper, people will find so many new uses for it that they end up spending more overall. That doesn't always happen with cheaper products, but so far, it seems to be working here. Anthropic cut prices with a new model last November. Over the next quarter its revenue growth jumped from 48% to 132%. The filing credits a mix of things: lower prices, better capabilities that opened up new uses, new products, and more customers signing on. So the company is spending more than almost anyone ever has on a product it keeps making cheaper.
But after reading the filing, I can't help but reflect on the idea that regular people are already helping pay for that half trillion in spend, through their retirement accounts, their utility bills, and their water. The question that I continue to sit with is whether they'll ever see the benefit.
This is part of the same line of questioning I began when discussing the SpaceX IPO. When AI makes money, who gets the returns? When it costs money, who gets the bill?
Who's paying for it
In a previous essay about the circular spending about AI, I called Anthropic the most circular of the AI companies. The money it spends flows right back to the same companies that have invested in it. This filing shows just how circular.
Google invested $10 billion in May. It has up to $20 billion more lined up through notes that turn into stock. Amazon has up to $15 billion lined up the same way.
Then, Anthropic turns around and buys computing power from them. At least $114.5 billion to Google. At least $110 billion to Amazon. Most of it owed "regardless of actual usage," according to the phrase in the filing. Anthropic pays whether it uses the compute or not. So the investors are also the suppliers. The money flows right back to them.
SpaceX is in the S-1 too. Between May and October, Anthropic agreed to rent up to $120.5 billion of computing power from SpaceXAI, the old xAI. Once it's fully up in December, that's about $3.1 billion a month. More than ten times what SpaceX's entire AI segment brought in last year. Four days after its IPO, SpaceX paid $60 billion for Cursor to compete with Anthropic in AI coding. So SpaceX is fighting Anthropic for customers while Anthropic pays its rent.
In that same piece, I said Anthropic had committed $200 billion to Google. The real number is much higher. The filing shows at least $114.5 billion to Google through 2033, plus an estimated $125 to $150 billion more for Google's chips between 2028 and 2031. That's roughly $240 to $265 billion to Google alone. Across all of its suppliers, including Amazon, SpaceX, and Microsoft, Anthropic's commitments come to more than half a trillion dollars.
This is what I mean when I say the AI boom runs on a loop. A handful of companies are each other's investors, suppliers, and customers. Google, Amazon, Microsoft, and Broadcom are all among the ten biggest companies in the S&P 500. If your 401k holds an index fund, you own a lot of them.
Where I got it wrong
I previously stated that I thought Anthropic would head into its IPO with the same profitability problems as OpenAI.
But wow, I was wrong about that.
You've probably seen a scary headline this week. An $80.7 billion net loss in the first half of 2026. Almost all of it, $72.4 billion, isn't money Anthropic spent. Here's what happened: Anthropic raised a lot of money through convertible notes. Those are investments that turn into shares later, at a price set back when the company was worth less. Anthropic's value has shot up since. So the shares those investors will get are now worth far more than what they put in. Accounting rules count that gap as a loss. No cash leaves the company. The real cost is that those investors end up owning a bigger slice of it.
Take that accounting entry out and the picture looks very different. Running the business cost Anthropic $693 million more than it made in the first half of this year, on $16.3 billion in revenue. For a company growing this fast, that's pretty close to breaking even. It's actually generating cash too. Anthropic ended those six months with $3.9 billion more cash from operations than it started with, partly because customers are paying it faster than it pays some of its own bills.
OpenAI's leaked numbers showed it losing about $1.60 for every dollar it made in 2025. Anthropic only lost about four cents.
That $16.3 billion in revenue comes from businesses paying Anthropic to get work done for less. Mostly that means companies using Claude to write code, answer customers, or get through contracts faster than their teams could on their own.
So the business seems to be working but the question is will it keep working?
The bills are fixed, but the revenue isn't.
Over 80% of Anthropic's revenue last quarter came from models released in the six months before. Customers move to the newest model almost as soon as it ships. They expect each one to be better than the last. The filing admits that even a small slip in performance could send them to a competitor. Anthropic says that it expects growth to slow and admits it has struggled to forecast its own revenue.
Meanwhile the commitments it has made run into the 2030s, and most can't be canceled.
This is where the risk for this business lives now. As long as Anthropic keeps growing, the loop holds. If growth slows faster than the bills come due, the loop runs in reverse. Google, Amazon, Microsoft, and SpaceX all lose a customer they've counted on, and their stock prices will reflect it.
Many of those stocks probably sit in your 401k, and that's is first way many of us are tied to all this, but it isn't the only one.
Who else pays
So the first thing I searched for in Anthropic's filing was mention of the communities that might be impacted.
And sure enough, they discuss it. Anthropic says it has publicly committed to cover any increase in electricity prices its data centers cause for regular customers. It lists other promises too, like paying for grid upgrades, adding new power to match what it uses, and cutting back during peak demand. Water, however, seems to be a different story. The filing only treats it as a risk to Anthropic. Its data centers need water for cooling, which is scarce or getting expensive in some places. It even acknowledges public controversy over water use in some of the regions where it has signed compute deals. But there's no promise about water like the one about electricity.
The way the filing talks about climate is also telling. Climate change shows up as a risk to Anthropic. Hotter temperatures will make it more expensive to cool its servers. Fires, floods, and storms could knock data centers offline. The filing even notes that its Bay Area headquarters faces a growing threat of fires.
To be fair, that's what risk factors are for in an S-1. They tell investors what could hurt the company. But it means the filing never asks the reverse question. What does all this computing power mean for the climate and for the communities living next to it? I couldn't find an emissions number or a clean energy goal anywhere in it. Anthropic says it will add new power to match what its data centers use but it doesn't say what kind.
The communities hosting these data centers face the same heat, fires, and floods. Most of them have far less money to deal with it.
But, it's worth also reading closely here. I talked earlier in this piece about how a lot of Anthropic's compute runs in data centers it rents from Google, Amazon, SpaceX, and others. The filing doesn't say explicitly whether the electricity promise applies to those rented data centers, or even where they are. So if some of that computing power runs in Memphis, it's not clear that Anthropic has promised the people there anything.
Communities near data centers aren't the only ones carrying costs they didn't sign up for. Anthropic's models learned from huge numbers of books. Authors sued, saying their work was used without permission. Anthropic settled with them for $1.5 billion. It's one of the only places in the filing where the people whose work made all this possible actually got paid.
To Anthropic's credit, it's also trying to give some of the upside back. It set aside 5% of its shares for a fund dedicated to public benefit. All seven co-founders have pledged 80% of their own stakes to philanthropy. At these valuations that's a lot of money. But Anthropic still decides how that fund gets spent. I'd love to see the communities paying some of these costs get a real say in where that money goes.
Anthropic is also a public benefit corporation, or PBC. Most companies have one legal job: make money for their shareholders. A PBC has to name a public benefit in its charter. Its board then has to balance shareholder returns against that benefit and the interests of the people affected by its work. Anthropic's stated purpose is "the responsible development and maintenance of advanced AI for the long-term benefit of humanity."
While that is a real commitment, there is a catch. Under Delaware law, only shareholders can sue to enforce it. Even then, they need to own at least 2% of the company or $2 million worth of stock. The people living next to the data centers are exactly who the law tells the board to consider. They just have no way to hold the company to it.
Who gets the benefit
This is the part of the filing that matters most to me.
Anthropic is spending half a trillion dollars on the computing power to build and run its AI. Only a handful of balance sheets on earth can play at that level.
But look at what all that spending produces. Anthropic's newest model cut the cost per task for customers by about 40%. New models now ship every few months. Every one of them makes intelligence cheaper for everyone who builds on top of it.
Right now, the people getting the most out of cheaper AI are businesses that already had money to spend. Think of software companies using Claude to write code, or companies using it to review contracts. They could already afford engineers and lawyers. AI just makes that work faster and cheaper for them.
We think the bigger opportunity is everyone else: families on Medicaid, hourly workers who need real financial tools, towns getting ready for the next flood. These are huge markets that most investors haven't paid much attention to. Cheaper AI makes it a lot easier to build real businesses that serve them well. That's the opportunity we see.
Many of them are the same people already paying some of the costs of this boom, through the data centers in their towns and the stocks in their retirement accounts. They're also the biggest market AI hasn't reached yet.
The companies that reach them won't win because of which AI model they build on. Every startup can use the same models from Anthropic, OpenAI, or Google. They all get each upgrade on the same day. They'll win on the data they earn by serving people well and on how close they are to the people they serve.
Earlier this week I wrote about who holds the pen.
An S-1 is a company holding the pen on its own story. Anthropic's is more honest than most. It tells you what the company is spending to make AI cheap. What it can't tell you is whether that cheap AI will ever reach the people already helping pay for it. That part is up to the rest of us. It's what we're betting on at Symphonic.
OpenAI is a public benefit corporation too. When it files, I'll be asking the same question.
Shruti Shah is General Partner at Symphonic Capital, an early-stage venture fund investing in AI as infrastructure for underserved communities across health, wealth, and climate resilience.
This is Part Four of an ongoing series. Part One: The AI IPO Paradox. Part Two: Everyone Is Reading the SpaceX S-1. Here's What They're Missing. Part Three: SpaceX's First Week. Here's What Actually Happened.