As Anthropic, the company behind the AI model Claude, prepares for its blockbuster October 2026 IPO, its financial filings are set to pull back the curtain on one of the biggest questions in tech: what does it actually cost to build a frontier AI?
A Gold Rush for Silicon
At the
heart of the AI revolution is a sliver of silicon: the Graphics Processing Unit, or GPU. Originally designed for video games, these chips are uniquely suited for the parallel processing required to train large language models. This has created a massive market dominated by companies like NVIDIA. A single NVIDIA H100 GPU can cost between $25,000 and $40,000. Training a model like Anthropic's Claude requires not just a few, but hundreds of thousands of these chips working in concert for months. Filings reveal that Anthropic has committed to spending hundreds of billions of dollars over the next decade on infrastructure, a significant portion of which is for these specialized processors. While the one-time training cost for a model like GPT-4 was estimated to be over $100 million, the ongoing costs of running the model for users, known as inference, represent a continuous and punishing operational expense.
The Cloud Partnership Equation
To manage these astronomical hardware costs, AI labs like Anthropic have forged deep, and expensive, partnerships with cloud computing giants. Anthropic has committed to spending over $110 billion with Amazon Web Services (AWS) and another $111 billion with Google Cloud over the next decade. These are not just simple customer-vendor relationships; they are complex, multi-faceted deals that often involve the cloud providers also taking significant equity stakes in Anthropic. These agreements provide Anthropic with access to massive, pre-built infrastructure, including Amazon's custom Trainium AI chips and Google's Tensor Processing Units (TPUs). In return, Anthropic locks itself into long-term, often non-cancellable spending commitments, where it may have to pay for the computing power regardless of whether it uses it. This strategy of relying on multiple providers allows Anthropic to leverage the best technology from each, but it also underscores a critical dependency that potential investors will be scrutinizing.
Beyond Chips: Power and Real Estate
The costs don't end with the chips themselves. The infrastructure required to house and run these massive GPU clusters is another enormous expense. AI data centers are highly specialized buildings requiring immense amounts of electricity and sophisticated cooling systems to prevent the processors from overheating. The energy consumption of AI is staggering; by some estimates, data centers could account for a significant percentage of global energy demand by 2030. This demand strains local power grids, and AI companies are increasingly involved in sourcing their own power, sometimes by investing in energy infrastructure or repurposing old power plants. For Anthropic, this translates into commitments for gigawatts of power capacity—enough to power a small city—just to keep its models running. These physical-world costs are a crucial, and often overlooked, part of the AI infrastructure bill.
The Path to Profitability
With such colossal expenses, the central question for investors is how Anthropic plans to turn a profit. The company's filings show a dramatic rise in compute costs, which at times have grown faster than revenue. For a period in 2025, the company reportedly spent around $1.60 on computing for every $1 of revenue it generated, leading to significant operating losses. The road to profitability depends on selling access to its Claude models to enterprise customers through subscriptions and API usage. While revenue has been growing at an explosive rate, much of it has been tied to its cloud partners, who are both investors and customers. The challenge for Anthropic, and the AI industry as a whole, is to prove that the value generated by its models can eventually outpace the immense and ongoing cost of the infrastructure required to create and operate them. This balancing act will be the core narrative as it steps into the public market.
