Hi Reader,
Earlier this year, people were talking about Anthropic going public at $2 trillion. This week it set a price range that values it closer to $1 trillion.
Most people are blaming the stock market. I think something more interesting is going on. Free AI models that anyone can download are getting good enough to do a lot of what Claude and ChatGPT can do, presenting a real problem for a company that has signed up for more than half a trillion dollars in compute bills.
In this week's piece, I explain what these free models are, what history says happens next, and who I think stands to benefit.
As always, would love to hear your thoughts!
Warmly,
Shruti
Anthropic Has Half a Trillion in Bills. Free Alternatives Could Make Them Harder to Pay.
The risk Anthropic flags in its own S-1, and what open source AI means for who benefits
By Shruti Shah, General Partner, Symphonic Capital. Part Five of an ongoing series.
When I started writing this, bankers were floating a valuation of up to $2 trillion for Anthropic's IPO. Then came last week's midterms and a record one-day drop in tech stocks. This morning, Anthropic set its IPO price range, valuing the company at roughly $1 trillion to $1.4 trillion, depending on how you count the shares.
Most of the coverage blames the market. I think something deeper is going on too. In my last piece, I wrote about the more than half a trillion dollars Anthropic has committed to spend on computing power. Those bills come due no matter what. Meanwhile, Anthropic faces a risk it flags in its own S-1: free, open source AI models.
If you haven't been following this closely, these are AI models anyone can download and use for free. Many people who use AI every day say they can do most of what Claude or ChatGPT can do at a fraction of the cost. They're usually a little behind the newest models, but they tend to catch up fast.
My hunch is that this will play out a lot like free software did in the early days of the internet. Open models will end up handling a huge share of everyday AI work. Anthropic and OpenAI will be fine, but they'll have to charge less. That's a real problem when your bills don't shrink.
To explain why, I want to start with the basics, since most people haven't had much reason to learn what an open model actually is.
So what is an open model?
Think about a big national restaurant chain, like Olive Garden or The Cheesecake Factory. Using Claude or ChatGPT, what people call "closed" models, is a bit like eating at one. The food is consistent, and you know what you are getting. You pay every time you order. The recipes are a closely guarded secret. In AI, there are really only a handful of these big chains: Anthropic, OpenAI, and Google.
An open model is what happens when a chain publishes its secret recipes. Anyone can make the same dishes at home for the cost of groceries. You can tweak them to your taste too. You do need your own kitchen, though, and the chain might still make it a little better.
In AI terms, the "kitchen" is your own computers or ones you rent. "Tweaking the recipe" means training the model further on your own data. The chain's edge is that the newest models from Anthropic and OpenAI are usually still a bit stronger on the hardest tasks.
The kitchen doesn't always have to be big, either. Many open models come in smaller versions that are cheap enough to run on a regular laptop or a modest server. Anthropic's S-1 names this as a risk too. It warns that customers may move to "smaller, cheaper, or open-source alternatives" that are good enough for the job.
Anthropic's own S-1 defines it simply enough. An open model is one where the finished version is released to the public, so anyone can download it, run it, and change it. What you usually can't see is the information it learned from. So the finished dish is shared, but not always how it was made in the kitchen.
You've probably heard some of the names of these open models. Meta's Llama models are the best known in the U.S. DeepSeek is the Chinese lab whose cheap, high-performing model rattled the stock market in early 2025. Alibaba's Qwen and Moonshot's Kimi are two more strong models out of China. Mistral is a French lab. Even Google and OpenAI have released open versions of their own, called Gemma and gpt-oss. When companies that make money selling closed models start giving some away, it's a sign they expect open models to be a big part of the future.
What this looks like in practice
Imagine a startup helping Medicaid patients manage diabetes. If it builds on Claude, every message a patient sends goes to Anthropic's servers. The startup pays for each one.
To be fair, Anthropic offers real privacy protections for businesses like this. Its S-1 describes tools designed to support HIPAA compliance for health care. Many of its business customers also have "zero data retention" terms, which means Anthropic doesn't keep their conversations. Still, the data has to leave the company's servers. The startup is relying on Anthropic's promises. The S-1 even notes that for its most powerful models, its safety rules may require keeping more data than those terms allow.
If it uses an open model instead, it can run everything on its own servers. Patient data stays on its servers. The cost per conversation drops a lot. Over time, the company can train the model on years of its own patient conversations so it gets better at exactly what those patients need.
For a lot of everyday work, like answering questions, sending reminders, or summarizing a visit, the open model is good enough. That's the part that should worry Anthropic and OpenAI.
How many people are actually doing this?
More than I expected. A September survey of 200 companies by Enterprise Technology Research found that open models now handle about a third of their AI usage, up from under a quarter a year earlier. Respondents expect that share to keep climbing. Among the companies already running open models, 60% had moved some work over from closed models like Claude or ChatGPT.
Some well-known companies have already made the switch. According to Mozilla's State of Open Source AI report, Stripe, the payments company, cut its AI costs by 73% after moving to open models.
Other data points the same way. OpenRouter is a popular service developers use to access hundreds of different AI models, so it can see which ones people actually use. Late last year, open models handled about a third of the work running through it. By September of this year, it was about 72%, mostly from Chinese labs. One caveat: OpenRouter's users skew toward developers watching their costs, so it probably overstates the trend. Open models also still bring in far less money than closed ones, even as more people use them.
The pattern most analysts describe is a split. Companies keep paying for the top models from Anthropic and OpenAI on their hardest, most valuable work. They move the high-volume, everyday work to open models.
We've seen this before
Something similar happened with software about 25 years ago.
Back then, most of the software that ran businesses was expensive and owned by companies like Microsoft. Then a free alternative called Linux came along. Linux is the basic software that tells a computer how to run, the same job Windows does on a PC. Anyone could download it, use it, and change it for free. A lot of people predicted it would put Microsoft out of business.
Some of that came true. Today, Linux runs most of the computers behind the websites and apps you use every day. Android phones are built on it too. Microsoft fought it hard at first. Its CEO even called Linux "a cancer" in 2001. By 2018, Microsoft had made peace with free software. It bought GitHub, the website where people share free code, for $7.5 billion.
Some of it didn't come true. Most people never switched their own laptops to Linux. They kept buying Windows and Macs because those were easier to use.
The biggest surprise was who made the money. It mostly wasn't the people giving the software away. It was Amazon, Google, and Microsoft.
This is where "the cloud" comes in. Instead of every business buying and maintaining its own computers, these companies built giant warehouses full of them. Businesses could rent computing power from them, the same way you pay a utility for electricity instead of building your own power plant. A lot of what ran on those rented computers was free software like Linux. The free software made the whole setup cheaper to run. The companies renting out the computers kept the profits. Amazon's cloud business is now one of the most profitable parts of the entire company.
Google and Microsoft got there in their own ways. Google ran its search engine on thousands of cheap computers using free software, which kept its costs low as it grew. Later, it gave away Android for free. That put Google's search engine and app store on billions of phones, where it makes money from ads and app sales. Microsoft, the company that once called Linux a cancer, built its own cloud business. Today, a lot of what its customers run on that cloud is Linux.
What I think happens next
Here's my hypothesis:
First, open models become a lot like Linux. They will run a big share of everyday AI work, often without people noticing. The survey data suggests that's already starting.
Second, Anthropic and OpenAI won't disappear. Like Windows and Macs, the polished products people use directly will hold on. So will the most powerful models for the hardest work. They'll likely have to charge less for everything else, though.
Third, the companies that make the most money will be the ones with something you can't download for free. For some, that's the data centers, the giant buildings full of computers that AI runs on. For most companies, though, the bigger advantage will be data. If everyone can download the same model for free, one of the biggest things that makes one company's AI better than another's is what it learns from. A health startup with years of conversations with its patients, used with their permission, has something no one else can download. So does a lender that understands how its borrowers actually pay their bills.
There's one big reason the Linux comparison could break down. Linux started as a volunteer project, so it cost relatively little to make. Open models still cost a fortune to build, so someone has to be willing to subsidize the cost. Right now that's Meta, a French lab, and several Chinese labs, each for its own strategic reasons. If they decide it's no longer worth it, the stream of free models could slow down. This is the part of my hypothesis I'm least sure about.
The catch
Before this starts to sound like a pitch for open source, there are real downsides.
The first is that you need your own kitchen. Running a model yourself means buying or renting powerful computers, hiring engineers, and having someone keep it all working. For a small startup, that can cost more in people and time than it saves on the AI bill. A recent Futurum Research report made the point in its title: businesses are embracing open models, but the companies renting out the computers to run them are making the most money.
Getting it to actually work day to day is harder too. A Mozilla survey found that only about half of developers trying open models got them running for real customers. For closed models, it was closer to two-thirds.
Open models are also still a step behind on the most difficult work. Even big fans of open models admit the gap can be weeks or months. For complex reasoning or high-stakes decisions, that gap can be consequential.
Then, there's customer support. If something breaks with Claude, Anthropic is responsible for fixing it. With an open model, that's on you. "Open" doesn't always mean you can use it however you like, either. Some come with rules about how businesses are allowed to use them.
There are risks for everyone else as well. Once someone downloads a model, they can strip out the safety guardrails built into it. That's one of the strongest arguments the frontier labs make, and one that I think is very valid. Where a model comes from matters too. Many of the best open models come from Chinese labs, which makes some companies and governments nervous about hidden behavior or compliance problems.
Is it secure?
It depends on what you mean by secure.
In one way, an open model can be more secure because your data stays with you. When you use Claude or ChatGPT, your data goes to their servers. Even with strong privacy promises, you're trusting another company with it. Run an open model on your own computers and nothing leaves, which is why hospitals, banks, and law firms find them so appealing.
The flip side is that you're the security team now. Anthropic and OpenAI have large teams protecting their systems. If you run the model yourself, keeping it locked down is your job. A small team with a weak setup can easily end up less secure than if they had just used Claude.
OpenClaw is a good example of how this can go wrong. OpenClaw is a free, open source AI assistant that lives on your computer and takes actions for you. It can read your email, open your files, and move around websites for you. It got hugely popular this year. It also became a bit of a security mess. People can install add-ons called skills to give it new abilities. Security researchers found that many of those add-ons were actually disguised programs built to steal people's data. Cisco's team tested one and found it was sending user data to an outside server. The Dutch data protection authority went as far as warning organizations not to use agents like this on systems with sensitive data. OpenClaw has since started scanning skills for malware, though its own team says that won't catch everything.
The lesson I take from it is that open doesn't automatically mean safe or unsafe. It depends on how much access you hand over and who you're trusting along the way.
What Anthropic's S-1 says about it
In its risk factors, Anthropic says open models could take part of the market and push down its prices. It says customers may decide the newest frontier models aren't worth the extra cost, then switch to smaller, cheaper, or open alternatives that are good enough. It also says competitors have tried to copy its models and will likely keep trying.
Anthropic's S-1 also states that if the ability to build highly capable models spreads widely, competition could come down to who has the most computing power rather than who builds the best models.
This connects back to the half trillion dollars from Part Four. In investing, a moat is whatever keeps competitors from taking your customers, the way a moat protects a castle. If anyone can download a good model for free, the model itself stops being much of a moat. Anthropic seems to be betting that computing power will be its moat instead. The company with the most chips and data centers can train new models fastest and run them at the lowest cost. It's the same playbook Amazon used years ago, when it made a fortune renting out computers, many of them running free software like Linux.
The fixed costs problem
This is the part I think matters most.
In Part Four, I wrote about the more than half a trillion dollars Anthropic has committed to spend on computing power. Most of those commitments are owed "regardless of actual usage," in the filing's words. Many run well into the 2030s. Anthropic has to pay whether it has the customer base or not.
Think about the restaurant chain again. Imagine it just signed ten-year leases on hundreds of new locations, sized for a future where more and more people eat out. Then the recipes get published, and a lot of people start cooking at home. Or, as happened a few years ago, a global pandemic keeps everyone home. The chain probably doesn't go out of business. Its best customers may still come in for the dishes that are hard to make at home. But it likely has to lower its prices to compete, and fewer people are coming in. Meanwhile, the rent on every one of those locations is still due.
This is the risk for Anthropic. Its revenue depends on customers choosing its models over and over. The S-1 says over 80% of last quarter's revenue came from models released in the previous six months, so it has to keep winning those customers. So far, it has. The number of customers spending more than $100,000 a year grew from about 1,500 at the end of 2025 to more than 11,900 by September. The open question is whether that keeps up as the free alternatives get better. Its costs, on the other hand, are fixed. They stay the same no matter how many customers it has. If open models take more of the everyday work and prices fall faster than Anthropic grows, Anthropic could find itself bringing in a lot less revenue while still owing the same giant bills. Reuters reports that some investors questioned Anthropic's earlier, higher valuation partly because of these commitments. A company with fixed bills this large is a riskier bet when the future of its prices is uncertain.
As I wrote in Part Four, those bills are also revenue for Google, Amazon, Microsoft, and SpaceX, some of the biggest companies in the stock market. Seven of Anthropic's suppliers and partners, including Google, Microsoft, and Nvidia, have also agreed to buy $26 billion of its stock at the IPO price. The money keeps going in circles. If Anthropic ever struggles to pay those bills, the pain won't stay inside Anthropic.
The regulation question
There's one more debate here I want to be careful with. Some people believe the push for AI regulation from Dario Amodei and Sam Altman is mostly about protecting their businesses from open source competition.
Critics argue the biggest labs want rules that are easy for them to follow and hard for smaller or open competitors to meet. The labs say their concerns are about safety and national security, including foreign labs copying their models. Anthropic's S-1 even describes that copying as something that could raise national security concerns.
I don't think it has to be either safety or self-interest. The labs can worry about safety while also benefiting from rules that make it harder for open models to keep up. What I want to pay attention to is whether the rules that actually get written make AI safer or mostly make it more expensive to compete.
What this means for the rest of us
For the founders we work with, this is mostly good news.
The companies we back at Symphonic serve people the market has overlooked: families on Medicaid, hourly workers, communities getting ready for the next flood. That work tends to be high-volume and cost-sensitive. It usually involves sensitive health or financial data. Cheap open models that can run privately are a big deal for those businesses.
It also sharpens something we already believe. Data is what makes a company hard to copy. Open models let a startup turn that data into something it actually owns.
Think about a startup that helps hourly workers manage their money. Over a few years, it learns things no general AI model knows: which week of the month people tend to run short, which bills get paid first, what actually helps someone avoid an overdraft fee. With an open model, the company can train the model on all of that and keep the result. The improved model lives on its own servers. It gets better every year, and no one else has it.
On Claude, the company can still use its data, but it's building on a model it rents. Anthropic sets the price and decides when older models get retired. The S-1 says over 80% of its revenue comes from models less than six months old, so the ground keeps shifting under the companies that build on them. For a startup serving people where every dollar of cost matters, owning the model means having more control over its costs too.
Open models also let a startup go places the big models can't easily follow. A clinic with spotty internet can run one on a laptop. A health company can keep every patient record inside its own walls.
None of this means every company should switch. The newest models from Anthropic and OpenAI are still stronger on the hardest problems. But for the founders we back, the combination of their own data and a model they own is starting to look like a real advantage.
The responsibility shifts too. A startup that runs its own model is on the hook for keeping it safe and secure. For a small team serving vulnerable people, that has to be taken seriously from day one.
In Part Four, I asked whether the regular people helping pay for all this AI spending would ever see the benefit. Open models give me a little more hope on that front.
Anthropic and OpenAI are the big chains in this story. Other labs keep publishing recipes that are almost as good, for free. If history is any guide, the chains will survive. They’ll just have to charge less to keep getting customers or they will operate with one specific set of customers that can afford to pay their prices.
The bigger change is who else gets to cook. When something expensive suddenly gets cheap, the benefits rarely stay only with the companies that built it. Free software didn’t just make Amazon rich. It let a whole generation of small companies get started without buying their own servers. Open models could do the same for AI. A startup no longer has to pay Anthropic or OpenAI every time it needs to answer a question. That makes it possible for a small team to build AI tools in markets where every dollar really counts.
Whatever valuation Anthropic ends up with, I don't think it's the number that matters most. The one I think really matters is how much of the world's AI work runs on open models. Depending on who's counting, it's already somewhere between a third and half. Whether that keeps climbing depends on a few things. Do open models keep getting better? Do Meta and the Chinese labs keep giving them away? Do new rules make them harder to use?
These answers will determine how much Anthropic can charge to cover its fixed costs. They'll also determine who gets to use AI at all. If open models keep improving, small teams can build great products at affordable prices. If they don't, AI's benefits will mostly go to the people and businesses that can already afford it.
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 Five 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. Part Four: Anthropic filed its S-1. Here's who is paying for it.