Why is Anthropic pushing its partners to accelerate data center construction? The answer is simple: recent projections for AI data center projects under construction show that every server brought online earlier for product services could generate extremely high profits, including a data center Anthropic will lease outside Baton Rouge, Louisiana.

Last week, I traveled to rural St. Francisville for the project's groundbreaking ceremony, witnessing firsthand Anthropic's intense race to expand computing capacity. The Claude model developer will deploy Google AI server chips at a 330-megawatt campus there in the second quarter of 2027. Project developer Hut 8 plans to expand the campus to 1 gigawatt of computing capacity, requiring tens of billions of dollars in investment to build out and equip with chips. Fluidstack handles leasing and operations for the campus, while Anthropic supplies the Google chips. Like many of Anthropic's recent partners, Miami-based Hut 8 has shallow experience in AI infrastructure, originally focused on Bitcoin mining, with Donald Trump Jr. and Eric Trump both involved. The project site spans 627 acres of muddy land with some buildings still under construction, and the grand celebration was held inside a tent on this site, with Donald Trump Jr. in attendance. Louisiana's governor, 300 bankers, investors, power sector officials, and Anthropic's computing chief James Bradbury all attended. Anthropic plans to launch an IPO next month, and Hut 8 is just one of many new partners Anthropic has enlisted to rush data centers into production and grow revenue.

Over the coming years, Anthropic has committed to leasing up to 15 gigawatts of computing capacity from Amazon Web Services, Google, SpaceXAI, and more than a dozen other cloud providers and data center companies. People familiar with the matter revealed that Anthropic has also signed an undisclosed multi-year cloud services agreement with publicly listed IREN, which similarly transformed from a cryptocurrency mining company into an AI server supplier. AI infrastructure research firm Measured AI predicts that Anthropic's computing capacity will grow from 1.4 gigawatts in 2025 to 5 gigawatts this year, 9.5 gigawatts in 2027, and surpass 15 gigawatts by 2028. For reference, Microsoft Azure's cloud business had total computing capacity of about 5 gigawatts at the end of 2023, when that capacity served all customers, and large-scale AI expansion had not yet begun.

To beat OpenAI in the computing expansion race, Anthropic will increasingly rely on a group of former Bitcoin miners including Hut 8, TeraWulf, IREN, and Nscale. Some of these companies currently struggle to raise funds from public markets at their desired valuations.

Valuable Grid Access Rights

Despite fundraising pressure, these companies and Anthropic's other partners are highly valuable: they secured a batch of grid interconnection quotas in earlier years, often located at abandoned aluminum smelter sites, failed green hydrogen project parcels, renewable energy-rich areas, and overlooked locations that large cloud providers missed or abandoned. New projects face long interconnection cycles, making such grid access resources scarce and extremely valuable.

(From left to right): Santanu Biswas, head of AI infrastructure at Broadcom; Giordano Albertazzi, CEO of Vertiv; Asher Genoot, CEO of Hut 8; Jeff Landry, Governor of Louisiana; James Bradbury, head of computing at Anthropic; Bob Pragada, CEO of Jacobs Engineering. Photographed last week at the Hut 8 St. Francisville AI data center campus.

For example, Apollo Capital-backed Stream Data Centers acquired two shelved hydrogen project parcels from former star company Plug Power, whose grid access resources are highly attractive. Anthropic is in talks to lease up to 1 gigawatt of computing capacity from Stream in the future. Anthropic demands that partners push construction at full speed. For instance, leading cooling and electrical equipment company Vertiv is redesigning its products for rapid deployment at data centers like the Hut 8 Louisiana campus. Vertiv CEO Albertazzi explained at the Hut 8 event that they pre-assemble power modules, batteries, switchboards, cooling, and fire suppression systems into prefabricated kits at the factory, cutting conventional construction installation time in half. "The data center construction model is undergoing a complete transformation, with the goal of rapid deployment," he said.

In an era of server supply shortages, speed is money. The Louisiana campus is planned to consume 330 megawatts, producing 245 megawatts of usable computing capacity. Investors and analysts estimate that even if this capacity comes online just one month earlier than originally planned, Anthropic could add at least $1.3 billion in revenue, assuming all capacity is used for inference services rather than model training.

Extraordinary Profit Margins

Even if the project cannot come online early, the economic returns remain highly attractive, with a quarter-gigawatt (250 megawatts) of capacity generating about $1.3 billion in monthly revenue. Christian Okoye of infrastructure investment consultancy Occam Edge said that based on recently disclosed long-term AI contract pricing, Anthropic pays cloud providers about $200 million per month to lease chips. On this basis, if the Louisiana project is fully used for inference, gross margins could reach 85%. Anthropic is trying to bypass intermediaries like cloud service providers in some projects, which could further boost margins. (Gross margins exclude model training costs, personnel salaries, and other operating expenses.)

Okoye's analysis also covers data center projects equipped with Nvidia GPUs and alternative chips like Google TPUs. GPU lease prices are higher than Google TPUs, but Nvidia claims GPUs can output significantly more tokens per watt than other chips. Regardless, Anthropic's excellent inference business profitability explains the improvement in the company's operating costs: in spring 2025, Anthropic needed about $2.3 in operating costs for every $1 of revenue generated.

However, there are several risk points to watch: at the start of the year, Anthropic projected that model training costs would exceed inference costs by more than double; this year's massive server leasing from companies like SpaceX will certainly pressure margins. At the same time, Claude's large customers pose a risk of reducing usage, and the company faces intense competition from OpenAI and domestic open-source models. For now, though, Anthropic still maintains pricing power over most customers. For the entire industry and capital markets, a key question: Anthropic may be the only company with this kind of pricing power. Other generative AI service providers, including OpenAI, have inference business gross margins far below Anthropic's level.

Affordable Grid Costs

Infrastructure developers working with Anthropic say the Claude developer prioritizes rapid access to power, even at the cost of temporary curtailment during grid strain, switching to backup power, or building facilities farther from users, meaning AI inference results take longer to travel over fiber optics. Hut 8 CEO Asher Genoot said at the Louisiana event that he continues to communicate with power companies and various partners about "what we can do to speed up project construction."

The impressive revenue model of data center projects explains why Anthropic is willing to bear higher electricity prices, even as ordinary consumers are already sensitive to rising power bills. Energy analyst Hans Royal proposed a metric, the compute heat rate: the electricity price an AI inference data center can afford while remaining profitable. Royal calculated that by the third quarter of 2026, commercial AI application data centers could remain profitable even at electricity prices of $5,631 per megawatt-hour, more than a hundred times the conventional wholesale price of $50 per megawatt-hour. Royal also noted that for top AI companies like Anthropic, in extreme cases, they could remain profitable even at electricity prices as high as $12,791 per megawatt-hour. These calculations suggest power companies still have substantial room to raise rates for AI data centers.

Lowering Electricity Costs for Everyone

But there is a better approach than simply raising prices. Leveraging the strong returns from AI inference business, Anthropic and its data center partners can reach agreements with power companies: voluntarily disconnecting from the grid during a few peak demand periods each year in exchange for earlier partial grid access for data centers. According to analysis articles by Royal and Occam Edge, the revenue Anthropic loses during curtailment periods can be recovered through revenue in other periods. Meanwhile, several innovative companies are developing solutions: when the grid is under strain, AI computing loads can automatically switch to backup power or be dispatched to other computing locations.

This is also why new AI data centers could lower electricity costs for society as a whole. Power companies earn more revenue from existing assets, and fixed costs are spread across more electricity users. In Louisiana, local communities around the Hut 8 project look forward to improving their finances through this investment. West Feliciana Parish President Kenny Havard pushed a bill through the state legislature allowing local governments to cut residential property taxes by 25%, matching the tax incentives offered to Hut 8. After the project comes online, taxes paid by Hut 8 are expected to increase the parish's annual fiscal revenue from $40 million to $130 million, more than tripling it.

These prospects sound excellent, but they depend on continued adequate supplies of electricity, chips, engineering construction capacity, and debt financing in the coming years. This remains a massive unknown.