New disclosures from SpaceX's IPO filing indicate that frontier AI firms are beginning to treat compute infrastructure as a standalone commercial business. Elon Musk’s xAI has agreed to provide large-scale AI capacity to its competitor Anthropic, marking a significant shift in the AI infrastructure landscape.
The filing reveals that Anthropic has committed to purchasing compute services delivered through xAI’s Colossus and Colossus II AI infrastructure clusters through May 2029. The agreement is valued at approximately $1.25 billion per month, making it one of the largest publicly known infrastructure deals in the AI industry.
This arrangement is notable because Anthropic competes directly with xAI in the market for frontier AI models and enterprise AI services. The deal suggests that at least some AI developers are increasingly willing to buy large-scale compute capacity from rival infrastructure operators rather than rely exclusively on internally owned GPU fleets or traditional hyperscaler cloud platforms such as AWS, Microsoft Azure, or Google Cloud.
SpaceX also stated in the filing that it “may enter into additional compute capacity agreements with third parties in the future,” indicating the Anthropic deal may not remain an isolated arrangement. This signals that the infrastructure arms race among AI companies is evolving into a more open market where compute capacity is becoming a tradable commodity.
Compute as a strategic asset
Analysts said the disclosures point to a broader structural shift underway in the AI industry, where excess compute infrastructure itself is emerging as a monetizable strategic asset independent of the AI models running on top of it. The era when AI models and the hardware they run on were inextricably linked is fading, as companies recognize that the value of compute can be separated from the intellectual property of the models.
Sameh Boujelbene, vice president at Dell’Oro Group, said: “This is less about excess capacity and more about compute becoming its own strategic asset class. Frontier AI companies are building at a scale where infrastructure can be used both internally and commercially.” This perspective underscores a fundamental change in how AI infrastructure is financed and deployed. Rather than being a cost center tied to model development, GPU clusters and associated networking and power systems are becoming profit centers in their own right.
The scale of the deal—$1.25 billion per month or roughly $45 billion over the full three-year term—provides a rare public price signal for frontier compute. Until now, the cost of training and inference at the highest tier has been opaque, hidden inside cloud bills and private contracts. This agreement gives enterprises and investors a tangible reference point for what premium AI infrastructure is worth.
More compute options for CIOs
For CIOs and enterprise infrastructure leaders, the disclosures signal that AI infrastructure sourcing is becoming strategically more complex as the market evolves beyond traditional hyperscaler cloud consumption models. The days when an enterprise could simply sign a contract with a single cloud provider for all its AI needs are giving way to a more fragmented and specialized supply chain.
Shay Boloor, chief market strategist at Futurum Group, said that enterprises may increasingly source AI infrastructure from a broader mix of providers, including hyperscalers, neocloud operators, specialized infrastructure vendors, and even frontier AI labs themselves. “The old assumption was that enterprises would simply buy AI capacity from the major hyperscalers. This filing suggests the market is moving toward a more complex supply chain where compute can come from hyperscalers, neoclouds, frontier labs, vertically integrated AI platforms and specialized infrastructure providers,” Boloor explained.
Boujelbene added that enterprises should increasingly think of GPU infrastructure as both a sourcing and utilization challenge rather than simply a cloud procurement decision. “The key questions are no longer only ‘which model should we use?’ but ‘where should workloads run, at what cost, and with what level of utilization?’” He emphasized that optimizing GPU utilization across multiple providers could become a core competency for large enterprises, especially those running both training and inference workloads at scale.
Arnal Dayaratna, research VP for software development at IDC, highlighted that the real challenge in AI deployments has been about accessing GPUs and managing them at scale affordably. “Putting public price tags on these arrangements gives enterprises a clearer signal of what frontier-scale infrastructure actually costs, which is essential context for building realistic AI ROI models and understanding why inference costs, usage limits, and API pricing look the way they do. For CIOs, it also clarifies that the economics of AI services are set upstream of the software layer, largely before a vendor ever writes a line of product code.”
This newfound transparency is crucial as enterprises build internal business cases for AI adoption. Knowing that top-tier compute commands $1.25 billion per month provides a benchmark for negotiation with cloud providers and alternative vendors.
Resemblance to cloud economics
Until recently, frontier AI companies largely treated compute infrastructure as a tightly controlled internal capability closely tied to proprietary model development. OpenAI, for example, built its own supercomputers with Microsoft’s help, while Google used its internal TPU infrastructure. The idea of selling compute to a direct competitor was almost unheard of.
The SpaceX filing, however, suggests the economics of AI infrastructure may be evolving toward something more closely resembling cloud infrastructure markets, where compute capacity itself becomes commercially tradable. Just as Amazon Web Services emerged from Amazon’s internal infrastructure, xAI’s willingness to sell to Anthropic hints that many AI labs may eventually spin off their compute divisions.
Boujelbene said the arrangement points to “more fluid compute-sharing models” emerging across the industry as infrastructure spending continues accelerating and AI demand remains high. The scale of investment required for frontier training clusters—often exceeding $10 billion per facility—makes it rational for companies to seek revenue from spare capacity, even if that means selling to rivals.
The filing repeatedly emphasizes the scale of xAI’s infrastructure ambitions, referencing continued investment in “AI infrastructure, compute capacity, and power systems” needed to support expanding training and inference workloads. It also provides one of the clearest public reference points yet for the economics underpinning frontier-scale AI compute infrastructure, an area where pricing, utilization rates, and long-term return models have largely remained opaque despite the industry’s aggressive datacenter expansion.
Boloor said the agreement effectively places one of the first meaningful public market values on frontier AI compute capacity. “The $45B Anthropic/SpaceX agreement shows that scarce, high-quality AI compute has become valuable enough that one frontier AI company is willing to pay another infrastructure operator tens of billions of dollars to access it,” Boloor said. The disclosures, he added, begin putting “a dollar value around frontier compute capacity” while offering insight into “the pricing power of scarce GPU clusters and ROI for companies building these systems.”
The deal also has implications for the broader GPU supply chain. Nvidia’s flagship chips, such as the H100 and B200, are in short supply, and the ability to secure large clusters has become a competitive advantage. By selling compute to Anthropic, xAI effectively monetizes its privileged access to hardware.
Analysts reject simplistic 'oversupply' interpretation
The filing has also fueled debate over whether the AI industry’s aggressive datacenter buildout could eventually outpace enterprise demand for frontier AI services. Some market observers worry that the tens of billions of dollars being poured into new GPU clusters may lead to a glut, similar to the dot-com boom’s fiber-optic overbuild.
But analysts cautioned against interpreting the Anthropic arrangement as evidence that major AI companies are sitting on large amounts of idle infrastructure. “I wouldn’t frame this as clear evidence that frontier AI firms are overbuilding GPU capacity. This is more of the natural evolution of AI compute becoming its own monetizable infrastructure layer,” said Boloor.
He said frontier AI companies are effectively forced to build infrastructure ahead of demand because “training runs, inference demand and agentic workloads don’t scale in a perfectly smooth line,” while procurement lead times for GPUs, networking systems, memory, and power infrastructure remain lengthy. A single large cluster can take 18-24 months to design, build, and bring online, so companies must place orders years in advance.
Alvin Nguyen, senior analyst at Forrester, similarly said the arrangement is likely to reflect the evolving workload dynamics rather than simple excess capacity. “There is enough demand for AI overall that all AI infrastructure is finding use. This is the natural evolution toward compute sharing and infrastructure monetization.” Nguyen compared the trend to the early days of cloud computing, when companies like Amazon started selling unused server capacity to external customers.
The deal also highlights a growing divergence in the AI market: while hyperscalers like Microsoft, Google, and Amazon continue to build massive clouds, frontier labs are creating their own private infrastructure networks. This dual-track approach gives enterprises more choice but also more complexity when planning their AI strategies.
For now, the xAI-Anthropic agreement stands as a landmark transaction that confirms the rise of AI compute as a standalone business. Whether it remains an outlier or becomes the template for the industry depends on how quickly other labs follow suit, but the filing has already changed how investors and CIOs evaluate AI infrastructure.
Source: Network World News