Three AI Companies Are Going Public This Year. Their Combined Compute Bill Explains Why.

SpaceX, OpenAI, and Anthropic are all expected to go public in 2026. Combined, they could raise close to $200 billion in the largest IPO wave since the dot-com era. But the numbers buried in their filings and investor disclosures reveal something more consequential than fundraising totals: these three companies are collectively betting hundreds of billions on the premise that AI workloads will consume compute at a scale that dwarfs current demand.

The filings put hard numbers on what had been speculation. Anthropic is paying SpaceX $1.25 billion per month through May 2029 for access to Colossus and Colossus II data center clusters, according to WIRED and confirmed by Reuters. That is $15 billion per year, locked in for three years. An Anthropic spokesperson confirmed the figure to Business Insider. The company’s compute chief, Tom Brown, wrote earlier in May that the Colossus capacity would be used for inference, not just training.

OpenAI has reportedly told investors it plans to spend $600 billion on computing power by 2030, according to The Hill. The company also serves as operational lead for Stargate, a $500 billion joint venture targeting 10 gigawatts of AI data center capacity across the US and internationally by 2029, per IG International.

The Inference Problem

The distinction between training compute and inference compute matters here. Training a frontier model is expensive, but it is a fixed cost paid once per model generation. Inference is a variable cost that scales with usage. Every API call, every agent action, every tool invocation burns tokens.

Brown’s clarification that Anthropic’s SpaceX capacity is earmarked for inference, not training, is the tell. Anthropic is not paying $15 billion a year to build the next Claude. It is paying $15 billion a year to run the current one at scale.

The math changes when agents enter the picture. A human using ChatGPT might generate a few thousand tokens per session. An autonomous agent running a multi-step workflow, calling tools, evaluating results, and retrying failures can burn hundreds of thousands of tokens per task execution. Multiply that by continuous operation (agents do not sleep, do not take breaks, do not log off at 5 PM) and per-user compute consumption jumps by orders of magnitude.

What the Valuations Imply

SpaceX is expected to price at $1.75 trillion to $2.3 trillion, according to IG International. OpenAI is estimated at $850 billion to $1.1 trillion. Anthropic is reportedly raising at $900 billion and targeting an October listing.

These are not revenue multiples. OpenAI reported annualized recurring revenue exceeding $20 billion for 2025, per IG International. SpaceX posted $18.7 billion in 2025 revenue. At a $2 trillion valuation, SpaceX trades at roughly 107x revenue. OpenAI at a $1 trillion valuation trades at 50x.

Investors pricing these companies are not buying current earnings. They are buying the assumption that AI agent workloads will grow compute demand faster than any technology cycle before them, and that the companies controlling model development and compute infrastructure will capture that value.

As Marketplace reported, Cornell finance professor Minmo Gahng noted these companies are “not likely to be profitable in the near future because they’re spending so much on hardware.” The IPO market is being asked to fund infrastructure buildout, not reward earnings.

The Agent Compute Thesis

The unstated assumption connecting all three IPOs is that autonomous agents will be the primary consumers of AI compute within the next 3 to 5 years. Not chatbots. Not search augmentation. Not content generation tools. Agents: persistent, tool-using, multi-step autonomous systems that run continuously and make API calls at machine speed.

This is why Anthropic’s Tom Brown specified inference capacity. Training Claude costs billions, once. Running millions of Claude-powered agents costs billions, every month, indefinitely. The $1.25 billion monthly payment to SpaceX is not a research budget. It is an operating expense for a product category that barely existed 18 months ago.

OpenAI’s $600 billion target tells the same story from a different angle. The company is not planning to train 600 models. It is planning to serve a world where its models power autonomous workflows at enterprise scale, consumer scale, and developer scale simultaneously.

The Risk Nobody Prices

The risk is simple: what if agent adoption plateaus? ChatGPT stalled at approximately 900 million weekly active users in early 2026, falling short of internal targets, according to IG International. Monthly revenue milestones have been missed on several occasions as competition from Google and Anthropic intensifies.

If agent workloads do not grow fast enough to justify $15 billion annual compute contracts, Anthropic is locked into payments that could consume its cash reserves regardless of revenue trajectory. If OpenAI’s $600 billion compute plan runs ahead of demand, the company builds infrastructure nobody uses.

These are infrastructure bets. They pay off spectacularly if agent adoption follows the trajectory these companies project. They become crippling liabilities if it does not.

The IPO wave will test whether public markets share the conviction that autonomous agents are the next computing platform. The S-1 filings have made the bet explicit. Now the market has to price it.