September has offered one of the clearest pictures yet of what the frontier AI race looks like when translated from models and tokens into physical infrastructure.

Anthropic has moved aggressively to lock down dedicated compute in the United States while establishing its first major data center foothold in Australia. OpenAI, meanwhile, has expanded into Malaysia through Nvidia-backed Firmus as the financing behind its much larger infrastructure ambitions continues to grow more complicated.

All in all, the developments suggest that competition between the leading AI labs is entering another phase. Securing GPUs remains essential, but the harder problem is increasingly assembling the entire chain around them: land, power, cooling, networks, project finance and counterparties capable of delivering capacity measured in hundreds of megawatts — and increasingly gigawatts.

That distinction is important for the data center industry. The AI infrastructure story is no longer simply about projected demand. It is increasingly about which commitments can actually become operating megawatts.

Anthropic’s $45 Billion Bet Gets More Concrete

The most revealing new detail came not from Anthropic itself, but from Nscale. The Nvidia-backed AI infrastructure provider filed for a U.S. initial public offering on Sept. 18, providing new financial and technical detail around a massive compute agreement first reported in August.

Nscale’s SEC filing says it entered four GPU services agreements with Anthropic on Aug. 25 that could generate approximately $44.6 billion in aggregate payments. The agreements call for Nscale to provide Anthropic with dedicated infrastructure built around Nvidia Vera Rubin NVL72 systems at the company’s planned Monarch Compute Campus in Mason County, West Virginia. The deployments are structured in four tranches with multiyear service terms.

Reuters previously reported the agreement at roughly $45 billion over six years, covering about 460 MW of compute capacity at Monarch. (The agreement follows Anthropic’s $19 billion, 401-MW lease with TeraWulf in Kentucky, another enormous commitment that illustrates how aggressively the AI developer is securing dedicated infrastructure ahead of demand). That alone would make the project notable. But Nscale’s IPO prospectus reveals the scale of the larger campus around it.

Monarch sits on approximately 2,250 acres and is planned with more than 8 GW of gross power capacity, including an initial 2 GW of gross generation supporting roughly 1.37 GW of IT load targeted for the first half of 2028. Nscale has increasingly described itself as a vertically integrated infrastructure platform stretching from behind-the-meter generation through data centers, GPUs and cloud software.

There is, however, an important caveat. Nscale disclosed that it must obtain qualifying financing for the GPU equipment and data center infrastructure required under the Anthropic agreements — and that as of the IPO filing it had not secured binding commitments for that financing.

The disclosure reinforces a challenge DCF has been tracking as the gigawatt AI buildout faces its execution test: enormous contracted demand does not by itself secure the financing, power, equipment or construction capacity needed to deliver operating megawatts. The agreements also remain subject to specified delivery and service-availability requirements.

That qualification may be as significant for the data center market as the $44.6 billion headline. The demand signal is extraordinary. But contracts of this magnitude increasingly represent the beginning of the infrastructure challenge rather than its conclusion.

Nscale Illustrates the New AI Infrastructure Math

Nscale itself offers a useful snapshot of how dramatically that equation has changed. The company reported $140.6 million in revenue for the first six months of 2026, up from $10.4 million a year earlier, while posting a net loss of approximately $1.02 billion.

It now claims more than $103 billion in active and contracted total contract value and a power pipeline exceeding 10 GW. Reuters noted that its largest customer represented 52% of first-half revenue, while Microsoft and Anthropic are expected to become major customers going forward.

This is increasingly characteristic of the AI infrastructure market: enormous contracted demand sitting against businesses that must spend staggering sums before much of the associated capacity can generate revenue.

Nscale is hardly alone. What makes its filing valuable is that it exposes the machinery behind the boom unusually clearly.

Another $35 Billion Compute Commitment

Nscale is not Anthropic’s only enormous infrastructure commitment. At the beginning of September, Reuters reported that Anthropic had signed a roughly $35 billion cloud computing agreement with Nvidia-backed Lambda, securing access to infrastructure at a roughly 350-MW data center being developed by Hut 8 in Nueces County, Texas.

The structure of that agreement is particularly indicative of where AI infrastructure financing is heading. Nvidia is an investor in Lambda and the dominant supplier of the GPUs powering these systems; reports also indicate that Nvidia holds the lease on the Hut 8 facility supporting the Anthropic deployment.

The arrangement extends a trend DCF has already examined around Nvidia’s expanding role in AI data center development and infrastructure financing, where the chipmaker increasingly appears not only as an equipment supplier but as a financial and contractual participant in the capacity being built around its hardware.

The Lambda structure shows how deeply the semiconductor, cloud and data center layers of the AI infrastructure stack are becoming intertwined.

For Anthropic, the Lambda commitment also reinforces the scale of its diversification strategy. Coming alongside the company’s approximately $45 billion Nscale agreement, it suggests that Anthropic is not relying on a single hyperscaler, neocloud or data center platform to satisfy its rapidly growing compute requirements.

Anthropic Also Goes Global — and Inference Goes Gigawatt-Scale

Anthropic’s infrastructure strategy is also extending well beyond its Nscale relationship — across providers, architectures and geographies.

On Sept. 16, Reuters reported that the company had signed its first Australian data center lease, covering a proposed 2.16-GW campus being developed by Zerra DC roughly 250 kilometers from Brisbane. The facility is expected to begin coming online in 2027 and, importantly, is intended for inference rather than model training.

That distinction fits a broader infrastructure shift DCF has been tracking as AI infrastructure scales both outward and upward, with inference pushing compute closer to regional demand even as individual campuses grow toward gigawatt scale. The planned project also provides clues about the infrastructure requirements emerging around increasingly distributed inference.

According to Reuters, the developer would obtain renewable power through power purchase agreements and bear the project's grid-connection costs. The site is also expected to use a closed-loop air-cooled system designed to reduce clean-water requirements. The agreement remains subject to approval by Australia’s Foreign Investment Review Board.

A 2.16-GW campus devoted to inference is a notable marker in itself. Training clusters may still command the biggest headlines, but inference is increasingly becoming a large-scale infrastructure planning problem as frontier AI moves deeper into enterprise and consumer workloads.

Anthropic is simultaneously pursuing one of the industry’s most diversified compute strategies. Earlier this year it committed to up to 5 GW of additional AWS capacity while expanding its use of Google TPUs and Broadcom infrastructure, including approximately 3.5 GW expected through the Broadcom portion of that partnership beginning in 2027.

Anthropic says it continues to run workloads across AWS Trainium, Google TPUs and Nvidia GPUs. That diversification now extends increasingly to geography as well as silicon.

OpenAI Makes Its Own Asia-Pacific Move

OpenAI is following a similar path. On Sept. 8, Australian AI infrastructure provider Firmus announced a multiyear agreement making OpenAI an anchor customer for dedicated AI compute capacity at two data centers in Malaysia.

Firmus said the deal pushed its contracted capacity across customers above 900 MW. The company operates or is developing seven AI facilities across Australia, Singapore, Indonesia and Malaysia and plans to deploy Nvidia Vera Rubin systems across its Asia-Pacific footprint.

The timing is notable. OpenAI released GPT-6 Astra on Sept. 3, pushing its frontier model deeper into computer use, software engineering and professional workflows while beginning a broader enterprise rollout.

That relationship between model capability and infrastructure consumption remains central to the sector. More capable models may become more efficient on a per-task basis, but broader adoption, agentic workloads and much higher utilization can still translate into greater aggregate demand for compute.

The Financing Layer Moves to Center Stage

The scale of that demand is now putting increasing attention on who finances the infrastructure beneath it.

Reuters reported Sept. 18 that OpenAI expects to burn nearly $280 billion in cash through 2030, based on Financial Times reporting, as spending on compute and infrastructure accelerates.

And on Thursday, Sept. 24, SoftBank raised $11.1 billion in high-yield bonds, described by Reuters as the largest high-yield corporate bond sale on record, to fund investments in OpenAI and other AI companies and infrastructure. The financing extends the infrastructure strategy DCF examined earlier this year around SoftBank, DigitalBridge and the next phase of OpenAI’s Stargate buildout, as increasingly large pools of private capital are assembled around OpenAI’s expanding compute requirements. SoftBank has committed an additional $30 billion to OpenAI during 2026 after investing roughly $30 billion during 2025.

OpenAI is simultaneously pursuing other enormous infrastructure structures, including the 8-GW PORTS-Pike development examined by DCF in August, where long-duration compute commitments and Nvidia-backed credit support similarly illustrate the financial engineering increasingly required behind gigawatt-scale AI campuses.

Meanwhile, another OpenAI-linked project provided an unusually timely reminder that even enormous financing packages cannot remove development risk.

Reuters reported Thursday that Oracle sent a force-majeure notice associated with Project Jupiter, the 1,400-acre New Mexico data center campus connected to Oracle’s infrastructure agreement with OpenAI. The campus has reportedly secured about $18 billion in loans but has encountered issues surrounding a planned natural-gas pipeline along with challenges related to water and air-quality permitting.

Oracle stressed that Project Jupiter remains on schedule and said force-majeure notices are common on projects of this scale and do not themselves establish a delay. Blue Owl, whose Stack Infrastructure unit is developing the project, said the notice does not alter financial commitments to the development.

From Compute Scarcity to Execution Risk

The September developments around Anthropic and OpenAI therefore point to something larger than another round of AI capacity announcements. The infrastructure race is becoming both more geographically distributed and more financially interconnected.

Frontier labs are securing compute through hyperscale clouds, neocloud providers and dedicated campuses; spreading inference into new regions; mixing Nvidia GPUs with custom silicon; and signing contracts whose values can exceed the capitalization of many established data center companies.

But the bottleneck is also moving. Finding demand is not the problem. Converting that demand into powered, financed, permitted, cooled, connected and operational infrastructure — on the schedules embedded in these enormous contracts — increasingly is.

For data center developers, utilities, equipment suppliers and investors, that may be the most important signal coming from the September AI news cycle. The next phase of the AI buildout will not be determined solely by which lab builds the strongest model. It will also be determined by which infrastructure ecosystem can actually deliver the megawatts.

At Data Center Frontier, we talk the industry talk and walk the industry walk. In that spirit, DCF Staff members may occasionally use AI tools to assist with content. 

Keep pace with the fast-moving world of data centers and cloud computing by connecting with Data Center Frontier on LinkedIn, following us on X/Twitter and Facebook, as well as on BlueSky, and signing up for our weekly newsletters using the form below.