I think it’s uncontroversial to say that Anthropic’s plans to go public offer investors an unusual proposition in financing artificial intelligence’s coding darling.
On one hand, the company is growing at extraordinary speed, its revenue is increasing rapidly and its artificial intelligence models are being adopted by businesses around the world, yet it is also losing billions of dollars and committing itself to hundreds of billions more in future computing and infrastructure costs.
According to Anthropic’s IPO prospectus, the company generated almost US$4.6 billion in revenue during 2025, roughly twelve times what it generated a year earlier. It nevertheless recorded a net loss of about US$42 billion.
That headline figure requires some qualification because roughly US$34 billion of the loss came from an accounting charge associated with financing instruments that could eventually convert into shares, rather than from money spent running the business. The company’s operating loss was approximately US$8 billion, more than most companies could ever possibly earn.
Even after making that distinction, I think we can agree that the numbers are difficult to ignore. Anthropic spent US$7.33 billion on computing and infrastructure during the year, more than half of its total operating expenses, while making long-term commitments for cloud computing, chips and other infrastructure that Reuters has estimated at around US$518 billion over the coming years.
The proposed IPO could value the company at more than US$2 trillion
That combination of enormous growth, enormous losses and an even more enormous appetite for capital goes a long way towards explaining why the finances of the AI industry have become so unusual.
Anthropic is not simply trying to build a software company that can eventually fund itself from subscriptions and enterprise contracts. Rather, it is trying to build one of the most computationally intensive businesses in the world, and it needs investors, cloud providers, chip manufacturers and other technology companies to finance that expansion long before the economics of the finished business are certain.
The same problem can be seen at OpenAI. Financial documents reported by journalist Ed Zitron and independently verified by the Financial Times indicate that OpenAI generated US$13.07 billion in revenue during 2025 while recording around US$34 billion in costs and expenses. Its reported net loss attributable to the company was US$38.53 billion, although that figure was also affected by substantial accounting adjustments associated with its corporate restructuring and financing arrangements.
The operating numbers are more straightforward, but they tell much the same story. OpenAI’s costs are rising extraordinarily quickly as it spends on research, computing and the infrastructure required to train and operate its models. The company has been able to raise enormous amounts of capital to support that spending, just as Anthropic has, but neither company has yet reached the point where its own revenues can comfortably finance the scale of expansion its ambitions require.
That is what makes the forthcoming wave of AI IPOs so interesting, even for those of us that aren’t particularly invested in the AI boom’s success. The important question is not simply whether Anthropic can command a valuation of US$2 trillion, or whether OpenAI can eventually achieve an even larger one. It is how long investors will remain willing to finance companies whose growth depends upon spending extraordinary amounts of money before they have demonstrated that those investments can produce conventional profits.
AI is an unusually expensive business to scale
The economics of artificial intelligence are different from those of much of the software industry that preceded it.
A conventional software company can spend heavily developing a product and then sell that product repeatedly without necessarily increasing its underlying costs by anything like the same amount. Once the software has been written, adding another customer might require very little additional infrastructure.
Generative AI does not have that advantage. Training advanced models requires huge quantities of computing power, while operating those models for customers requires further computing capacity every time someone sends a request. The more successful an AI service becomes, the more infrastructure its operator needs to keep it running.
This creates a peculiar situation in which rapid growth can actually increase a company’s need for capital, rather than reducing it.
Anthropic’s financial results demonstrate the problem. The company managed to increase revenue from hundreds of millions of dollars to almost US$4.6 billion in a single year, but the cost of supporting that expansion also increased dramatically. Computing and infrastructure alone accounted for US$7.33 billion of expenditure in 2025.
The company is effectively betting that today’s infrastructure expenditure will allow it to capture tomorrow’s demand. If customers continue adopting Claude and other AI services at the rate Anthropic expects, the computing capacity it is securing now could become a valuable competitive advantage. If demand grows more slowly, however, the company will still have to deal with infrastructure commitments that were made on the assumption of much higher future usage.
That is why the US$518 billion figure in the prospectus deserves attention. It is not a statement that Anthropic will simply spend US$518 billion in the next twelve months. It represents long-term commitments for computing, cloud services and infrastructure stretching across several years, with a substantial proportion reportedly difficult or impossible to cancel.
The company is therefore making a very large financial bet on the future size of the AI market, and for a winner-takes-all industry.
The investors are also the suppliers
One of the more unusual features of that bet is the identity of the companies involved.
Amazon is an important investor in Anthropic and also provides much of its cloud infrastructure. Google is both an investor and a supplier of computing capacity. Microsoft has an enormous commercial relationship with OpenAI while also developing its own AI products and services.
There is nothing inherently strange about a technology company investing in a customer or strategic partner. What is unusual is the scale of the relationships now developing around AI.
Anthropic has announced an agreement with Amazon involving up to five gigawatts of computing capacity and more than US$100 billion in planned expenditure on AWS technologies over ten years. Amazon has also committed another US$5 billion to Anthropic, with the possibility of investing a further US$20 billion.
Google has similarly invested billions of dollars in Anthropic while supplying the company with access to its computing infrastructure.
The result is an industry in which the boundaries between capital provider, infrastructure supplier and customer are increasingly difficult to separate. The companies building AI models need vast quantities of computing power, while the companies selling that computing power have a direct financial interest in the success of the businesses consuming it.
That relationship has now extended into the financing of the hardware itself. Reuters reported in October that Broadcom had agreed to lend Anthropic up to US$42 billion to help finance the leasing of its chip technology. The arrangement forms part of Anthropic’s much larger commitment to computing capacity involving Broadcom and Google, with some of the financing potentially convertible into Anthropic shares.
It is an unusual arrangement, but perhaps an increasingly logical one for an industry in which the cost of the machinery required to operate the product is itself one of the largest expenses.
OpenAI’s finances show how quickly the numbers can grow
OpenAI has been following a similar trajectory, although its corporate structure and financing arrangements are different.
The financial documents reported by Ed Zitron show that OpenAI had US$3.7 billion in revenue in 2024 against US$12.48 billion in costs and expenses. In 2025, revenue increased to US$13.07 billion while costs and expenses climbed to approximately US$34 billion.
Research and development accounted for US$19.18 billion of those 2025 costs, while cost of revenue was approximately US$7.5 billion. The figures provide an indication of just how expensive it has become to operate at the frontier of AI development.
The company’s relationship with Microsoft is particularly significant. According to the same documents, OpenAI paid Microsoft approximately US$17.2 billion during 2025 across research and development, cost of revenue, sales and marketing and general administrative expenses. About US$10.59 billion was recorded under research and development.
Microsoft is also one of OpenAI’s major investors and strategic partners. That means one of the largest sources of OpenAI’s expenditure is also a company with a direct interest in OpenAI’s continued growth.
Again, there is nothing inherently unusual about such an arrangement. Large technology businesses have always formed partnerships with suppliers and invested in companies whose success could benefit their own businesses. What makes the current situation remarkable is the amount of money involved and the speed at which the relationships are expanding.
OpenAI needs computing capacity to train and operate its models. Microsoft wants that computing demand to run through its infrastructure. OpenAI’s investors want the company to become enormously valuable. Microsoft itself wants to capture a substantial share of the economic value created by AI.
The interests overlap, which helps explain how companies that are still losing billions of dollars can continue to secure the capital necessary to expand.
Venture capital is only one part of financing artificial intelligence
The AI industry is often described as though venture capital is simply providing the money needed to keep startups alive until they become profitable.
At this scale, that description no longer really works.
Anthropic has raised enormous private funding rounds, including US$30 billion at a valuation of US$380 billion and another US$65 billion round at a valuation of US$965 billion. It also has strategic investment from major technology companies, long-term infrastructure agreements, equipment financing and access to debt.
OpenAI has developed an equally complicated network of investors and commercial partners, while its enormous computing requirements have tied its finances closely to Microsoft and the wider infrastructure industry.
The eventual move into public markets adds another source of capital and, perhaps more importantly, another group of people who will have to decide whether the economics make sense.
A private investor can afford to wait for a company to become profitable if they believe the eventual return will justify the patience. Public investors have more immediate information about the company’s financial position and can respond much more quickly if expectations change.
Anthropic’s IPO will therefore provide a particularly interesting test of the market’s appetite for the AI business model. Investors will not simply be deciding how much an AI company is worth. They will be deciding how much they are willing to pay for a company that expects to spend vast sums on infrastructure for years before the ultimate economics of that infrastructure are known.
The biggest risk may not be the losses themselves
It would be easy to look at Anthropic’s US$42 billion loss or OpenAI’s reported US$38.5 billion loss and conclude that the companies are simply burning through money at an unsustainable rate.
The financial statements are more complicated than that.
Anthropic’s enormous accounting loss was largely driven by the changing value of financing instruments rather than equivalent cash expenditure. OpenAI’s headline loss was also affected by accounting changes associated with its restructuring.
For both companies, the more useful figures are the relationship between revenue and operating expenditure, the amount being spent on computing and the commitments being made against expected future demand.
Those numbers are still enormous, but they tell a more useful story.
Anthropic’s revenue is growing quickly. Its gross economics have also improved substantially. Reuters reported that its implied gross margin rose to roughly 40% in 2025, compared with negative 95% in the previous year.
That improvement is important because it suggests the underlying service may be becoming more economically viable as the company scales.
The problem is that improving the economics of each unit of AI usage does not necessarily mean the company needs less money. If the number of customers and the amount of computing they consume grow quickly enough, total infrastructure spending can continue increasing even while the economics of the underlying service improve.
That is the central financial gamble being made across the industry.
Someone eventually has to pay for the infrastructure
There is a simple reality underneath all of these complicated financing arrangements. The infrastructure being built for AI is real, and someone ultimately has to pay for it.
The chips have to be manufactured. Data centres have to be constructed. Electricity has to be generated. Cloud providers have to maintain the infrastructure. Engineers and researchers have to be paid. Debt has to be serviced and infrastructure leases have to be honoured.
At present, the industry is relying heavily on investors’ willingness to fund those costs on the expectation that AI will become sufficiently valuable to justify them.
That expectation is not unreasonable. AI is already generating billions of dollars in revenue, and businesses are adopting the technology across software development, customer service, research, marketing, finance and a growing range of other areas.
What remains uncertain is how large the eventual market will become and how much of the resulting revenue will remain with the companies developing the models after infrastructure costs, employee compensation and the enormous cost of ongoing research are taken into account.
That distinction will become increasingly important as the industry matures.
The IPO changes the question
For years, investors could largely treat frontier AI as a private-market experiment. Venture capital firms, technology companies and sovereign wealth funds were prepared to provide enormous amounts of capital because they believed that owning a stake in one of the companies that eventually dominated AI could be worth far more than the money invested today.
An IPO changes the audience.
Anthropic will have to explain its economics to a much broader group of investors, many of whom will be assessing the company through the familiar measures of revenue growth, margins, cash flow and capital expenditure.
The prospectus already provides some reasons for caution. Nearly a quarter of Anthropic’s 2025 revenue came from two customers, while many of its largest customers were not locked into long-term contracts. The company has also made infrastructure commitments that stretch far into the future.
That combination creates an unusual risk. Anthropic is making long-term commitments on the assumption that demand will continue to rise, while a significant portion of its revenue comes from customers who retain the ability to reduce their spending.
The company’s argument is that AI demand is still at an early stage and that the capacity being secured now will be needed as adoption accelerates.
Public investors will ultimately have to decide how much confidence they have in that argument.
The AI economy is becoming a financing ecosystem
This is perhaps the most interesting aspect of the entire story.
The AI industry is no longer simply a collection of startups raising money from venture capital firms. It is becoming an interconnected financing ecosystem in which the same companies can occupy several positions at once.
An investor can also be a cloud provider. A cloud provider can also be a customer. A chip manufacturer can provide both hardware and financing. A technology company can invest in an AI laboratory while competing with it in other parts of the market.
Capital moves into AI companies, those companies spend the capital on infrastructure, infrastructure companies receive the resulting revenue and some of them reinvest in the AI companies. Rising valuations make it easier to raise additional capital, which allows the cycle to continue.
The system can function for a long time if the underlying economic proposition continues to strengthen.
The difficult part comes if growth slows before the companies have reached the scale necessary to support the commitments they have already made.
Anthropic’s prospectus makes that tension unusually visible. The company is growing rapidly enough to attract a potential US$2 trillion valuation, but it is also committing hundreds of billions of dollars to the infrastructure required to support the growth on which that valuation depends.
That is the strange financial loop at the heart of the AI boom.
The companies need enormous amounts of capital because AI requires enormous amounts of infrastructure. Investors provide the capital because they believe AI will eventually generate enormous amounts of economic value. Infrastructure companies benefit from the spending and therefore have an incentive to finance the customers generating it. The AI companies then need to grow quickly enough to justify the next round of investment.
For the moment, the money keeps coming.
The more interesting question is what happens when investors begin asking whether the industry can eventually generate enough cash to replace the capital that has been poured into is one that OpenAI will eventually have to confront as well. The future of AI may depend on how capable its models become, but the future of the companies building those models will depend on something considerably less glamorous: whether the financial machinery supporting the AI boom it.
That is the question Anthropic’s IPO will begin to answer, and it is one that OpenAI will eventually have to confront as well. The future of AI may depend on how capable its models become, but the future of the companies building those models will depend on something considerably less glamorous: whether the financial machinery supporting the AI boom can keep running long enough for the promised returns to arrive.
Want to read more? Explore the Trafficon Digital blog for more exciting articles on everything from marketing and AI to complaints I have about being on the internet in my 30s.
