Introduction

Elon Musk’s most dramatic moment isn’t his latest legal battle with OpenAI or his integration of xAI into SpaceX, but rather his decision to open his most dazzling AI asset—Colossus 1—to competitor Anthropic. SpaceXAI officially announced that Anthropic will gain access to Colossus 1, a supercomputer equipped with over 220,000 NVIDIA GPUs, which can be used for training, fine-tuning, inference, and high-performance computing.

This indicates that what was once packaged as xAI’s strategic weapon to catch up with OpenAI, Google, and Anthropic has now transformed into a rentable infrastructure. Reuters confirmed that SpaceX has signed an agreement with Anthropic, which will use the additional computing power to enhance the user capacity of Claude Pro and Claude Max. This isn’t just a typical business collaboration; it’s a reversal in the AI race: those who build models are now selling computing power to those who can better utilize the models.

The Colossus Story

The story of Colossus initially resembled an engineering myth. According to reports, this AI supercomputer located in Memphis, Tennessee, was built from scratch in just 122 days, with the first rack starting training just 19 days after installation. Jensen Huang noted that such speed typically requires four years. For Musk, this demonstrated his traditional advantages: extreme engineering, capital mobilization, supply chain compression, and organizational sprinting.

However, AI isn’t a rocket or an electric vehicle. The core challenges of rockets lie in physical limits, while those of cars focus on manufacturing systems. The core challenges of large models involve talent, data, algorithms, product feedback, and continuous iteration. A supercomputer can rise in 122 days, but model capabilities cannot be linearly stacked by the number of GPUs. Computing power is merely an entry ticket, not a victory in itself.

This is precisely where xAI finds itself in an awkward position. It possesses a sufficiently large hardware narrative but has not simultaneously established a stable software organization. Reports indicate that by the first quarter of 2026, all 12 co-founders of xAI had left. If this fact holds true, renting Colossus to Anthropic is not merely a resource optimization but rather a forced recalibration of its business model.

Musk’s Ambitions

Musk certainly lacks ambition. xAI’s initial slogan was “Understanding the true nature of the universe,” a typical Musk declaration: grand, abstract, and philosophical. However, the market ultimately asks simple questions: Where are the users? Where is the revenue? How does model performance stack up? How do marginal costs decrease, and how is capital expenditure recouped? When the answers are unclear, the most certain cash flow often comes from renting out computing power.

This reflects a deeper change occurring in the AI industry. Over the past year, the market’s focus has shifted from “who has the strongest model” to a split in competition: one line focuses on model capabilities and user entry points, while the other concerns infrastructure made up of power, land, chips, networks, and cooling. Anthropic lacks computing power, while SpaceXAI lacks efficiency in model commercialization; their transaction essentially meets each other’s needs.

Implications for Anthropic and SpaceXAI

For Anthropic, this deal alleviates expansion bottlenecks. Claude is gaining higher usage intensity in enterprise programming, knowledge work, and professional subscription markets, with inference demand rising rapidly. Computing power is no longer just a capital asset for the training phase but has become an operational bottleneck for everyday service capabilities. SpaceXAI stated that Anthropic will use this computing power to improve the user capacity of Claude Pro and Claude Max subscriptions.

For SpaceXAI, leasing Colossus represents a revaluation of assets. If its own models cannot fully utilize 220,000 GPUs in the short term, then idleness is waste, and renting reduces depreciation pressure. AI supercomputers are not one-time trophies; they are high-debt, high-depreciation, high-energy-consuming asset pools. They must be used frequently; otherwise, the larger the scale, the greater the financial pressure.

This explains why “computing power lessors” are not necessarily failures. Amazon Web Services, Microsoft Azure, and Google Cloud essentially rent infrastructure to others, thus obtaining more stable commercial value than many application companies. The real question isn’t about renting but whether it is an active strategy or a passive loss mitigation; whether it reflects platform capabilities or a patch after model competition falters.

Musk’s Unique Approach

Musk’s uniqueness lies in his attempt to recombine computing power, rockets, satellites, electricity, and AI. SpaceXAI even mentioned that Anthropic is interested in collaborating to develop “multi-gigawatt” orbital AI computing power, stating that the computational needs of next-generation AI systems are outpacing the supply of terrestrial power, land, and cooling conditions. This sounds like science fiction, but it aligns with Musk’s consistent logic: when Earth’s resources become bottlenecks, industries should be moved to space.

However, capital markets will not pay solely for imagination. The higher the threshold for AI infrastructure, the more important commercial discipline becomes. 220,000 GPUs can create a sensation, but they consume power, depreciate, and incur capital costs daily. Those who can convert computing power into high-margin revenue will gain a true competitive moat in the AI war. Otherwise, the largest supercomputer may merely become the largest cost center.

Industry Paradox

This transaction also reveals a paradox in the AI industry: it is increasingly difficult to completely separate competitors. OpenAI relies on Microsoft, Anthropic on Amazon and Google, and now turns to SpaceXAI for computing power; NVIDIA sells chips to everyone, and cloud vendors are both suppliers and competitors. The AI industry does not have the clear boundaries of traditional internet companies; it resembles the early electricity industry, where power plants, transmission networks, and large consumers are intertwined.

Thus, “the builder of the world’s largest supercomputer has become a computing power lessor” appears to be a reversal for Musk, but at a deeper level, it signifies the AI industry moving from a hero narrative to a balance sheet narrative. Who can build quickly only solves the first problem; who can utilize it fully, for a long time, and profitably addresses the second problem. Musk has proven the limits of engineering speed, while Anthropic reminds the market that in the AI era, computing power must ultimately serve the product.

What truly deserves attention is not whether Colossus still belongs to Musk, but whether its fate foreshadows a new division of labor in the AI industry. In the future, not every model company may be able to build supercomputing power, nor can every computing giant necessarily train the strongest models. Hardware, energy, models, and applications may be re-split, forming a new cloud computing order. At that point, today’s transaction will not just be a lease but the beginning of a restructuring of the AI industry’s business model.