Anthropic is reportedly preparing to list at a valuation of around US$2 trillion, only a few years after it began selling Claude commercially. A leaked draft of its S-1 filing, reported by Reuters, gives a first look at what sits behind that figure: rapid revenue growth, and a very large forward bill for the computing power needed to deliver it.

Anthropic is not alone: OpenAI shipped three GPT-6 models in September and used DevDay to push into always-on agents, while Meta's Muse took agents to mainstream consumers. Together they point to three forces behind AI infrastructure demand: the model race, the rise of agentic AI, the scramble for power and capex from hyperscalers.

The model race is an infrastructure race

The excitement rests on growth. Reported run-rate revenue climbed from roughly US$1 billion at the end of 2024 to around US$9 billion at the end of 2025, and to more than US$65 billion by July 2026. The leaked S-1 also discloses around US$518 billion of contractual commitments for cloud, compute and data centres over roughly the next decade¹, a reminder of how much capital sits behind that growth.

Chart 1: Anthropic’s reported annualised run-rate revenue

Source: Anthropic, Reuters, Bloomberg. Run rate annualises recent sales and is not booked revenue. Anthropic reports it gross of cloud-partner sales

OpenAI is racing just as hard, and now cites 1.2 billion weekly users². Its September releases form a tiered range, from a flagship (Astra) to lower-cost models (Sol, Luna), reflecting a product strategy of matching capability and price to the use case.

The race is fought with compute, and the commitments are now measured in gigawatts. Anthropic has announced up to 5GW from Amazon, 5GW of next-generation TPU capacity with Google and Broadcom, and access to SpaceX’s data centres³. Such willingness to sign obligations of this size shows how much capacity the leading labs expect to need.

AI agents could broaden compute demand

Personal AI agents are not new. OpenClaw, the open-source agent that went viral earlier this year, remained largely the preserve of technical users. Muse lowers that barrier sharply: users simply install the app and sign in with a Meta account. The response has been striking. Launched on 8 September, Muse reached number one on the US App Store within ten days notes⁴. Its early US downloads surpassed those of ChatGPT, helped by Meta’s installed base across Instagram, WhatsApp and Facebook.

That lower bar is drawing in small businesses too. Nearly a third of early users have already linked a business account⁵, and Muse for Small Business connects to production tools. OpenAI’s Dots, launched at DevDay, can be read as a response: always-on agents that run on their own cloud computers and connect to thousands of apps.

For AI infrastructure, the implications lie in how agents change the workload. Agents run continuously, call other tools and work through multi-step tasks. Each step involves further model calls, so a single task can require far more inference than a one off chat prompt. Training models to act reliably also requires heavy post-training. The effect should extend beyond accelerators to the CPUs that host agent environments, the memory that holds context and the networking that connects them.

Chart 2: Chatbot (upper) and agentic (bottom) inference workflows compared

Source: WisdomTree

Power is becoming part of the AI stack

All of this compute needs electricity, and demand is rising steeply. Global data centre electricity consumption is estimated to roughly double by 2030, to close to 3% of global electricity demand⁶. Securing a grid connection means a long queue, and local opposition adds friction in some places.

Data centre operators are therefore looking beyond waiting for grid supply. Hyperscalers are signing long-term agreements with alternative solutions including nuclear and renewables. Amazon and Google both announced deals with Constellation in recent days, partly to fund new capacity at those plants. They are also backing small modular reactors (SMRs), such as Meta’s agreements with Oklo earlier this year. For nearer-term needs, some are turning to behind-the-meter generation that bypasses the grid, such as Oracle’s agreement for Bloom Energy fuel cells, indicating on-site power to supply a growing share of new capacity. Solar, storage and grid upgrades also have a role, particularly where they can be deployed faster than new reactors. The AI infrastructure opportunity therefore extends beyond semiconductors and data centres into power generation and energy infrastructure.

Chart 3: US data centre power shortfall through 2029

Source: Morgan Stanley, Living on the Edge: The Rise of Distributed Power. Based on estimation of midpoint scenario.

Hyperscaler capex funds the build-out of AI Infrastructure

Hyperscaler capex is the fuel for the AI infrastructure build-out. The four largest US hyperscalers have raised their spending guidance through the year. That spending builds their own cloud and AI capacity, and in doing so becomes revenue for chip designers, memory makers, networking suppliers, power equipment providers and other parts of the chain. Micron’s latest quarter offers a recent example, with data centre revenue up 56% on the previous quarter, as cloud customers expanded AI capacity.⁷

Chart 4: Hyperscaler capex funds the AI infrastructure ecosystem

Source: Sources: Microsoft - Microsoft Corporation. (2026, April 29). Microsoft cloud and AI strength fuels third quarter results [Press release]; Alphabet - Alphabet Inc. (2026, April 29). Alphabet announces first quarter 2026 results [Press release]; Amazon - Amazon.com, Inc. (2026, April 29). Amazon.com announces first quarter results [Press release]; Meta - Meta Platforms, Inc. (2026, April 29). Meta reports first quarter 2026 results [Press release].

This spending also looks more sustainable than past infrastructure booms. In the dot-com era, telecom carriers spent heavily on fibre, many went bankrupt, and internet companies such as Amazon and Google later reaped the benefits. This time, the biggest spenders are also the main beneficiaries, earning revenue directly from the capacity they build. Returns are meaningful too: return on invested capital is estimated at about 31% when hyperscalers rent out GPU capacity to customers, and about 46% when they sell access to their own AI models running in their own data centres, suggesting the spending rests on economics rather than speculation.

Conclusion

Anthropic's reported IPO is a useful marker, but the more lasting signal lies in what sits beneath it. The leading labs are committing to compute at gigawatt scale, and agentic AI is broadening the workload beyond accelerators to CPUs, memory and networking. Power has become part of the infrastructure itself, with hyperscalers turning to existing nuclear plants, future SMRs and on-site generation alongside the grid. The direction of travel is consistent across models, agents and energy: AI is increasingly an infrastructure build-out, and its beneficiaries extend across the whole AI infrastructure ecosystem.

¹ Reuters: Anthropic's IPO prospectus shows sweeping AI vision, surging costs
² OpenAI: DevDay 2026 Recap
³ Source: Company press https://www.anthropic.com/news/series-h
⁴ Source: BNP Paribas “The Muse(ic) Keeps on Playing, Week 3” 02 Oct 2026
⁵ Source: Axios Meta courts small business with new Muse agent
⁶ Source: IEA Energy demand from AI
⁷ Micron, company press