xAI leases core training cluster to Anthropic, sparking industry speculation

Article by Xiao Bing, Shenchao TechFlow

If three months ago you told a Silicon Valley investor that Musk would rent out xAI’s largest training cluster, Colossus 1, entirely to Anthropic, he would likely have laughed.

After all, in February, Musk was criticizing Anthropic on X for "hating Western civilization," and in March, he nicknamed the company "misanthropic." To Musk, this company was virtually the epitome of politically correct AI—a rival that, like OpenAI, needed to be taken down.

Then, on May 6, Anthropic and SpaceX jointly announced that Anthropic would receive the full computing capacity of Colossus 1—over 220,000 NVIDIA GPUs and 300 megawatts of power capacity—with delivery to be completed within one month. Anthropic explicitly stated that this computing power will be directly used to enhance the service experience for Claude Pro and Claude Max subscribers.

Musk posted a startling message on X: he said he had deep interactions with Anthropic’s executives over the past week, was "impressed," and noted that "they are all highly capable and genuinely working on the right things." He even said that Claude "will probably be good."

On the same day, he announced that xAI would be dissolved as an independent company and renamed SpaceXAI.

Major English media have framed this as "a landmark event in AI computing power sharing," but they've overlooked a key fact:

Colossus 1 is xAI's most critical training facility, not some "backup capacity."

Review the timeline. Colossus 1 was commissioned in Memphis in September 2024, going from groundbreaking to powered on in just 122 days—a miracle in data center construction history. It serves as the primary cluster for xAI to train Grok 3 and Grok 4, and embodies Musk’s narrative of “compute as power.” Equipped with over 220,000 GPUs, including H100s, H200s, and the latest GB200s, its scale ranked among the top three globally by the end of 2025.

Giving an entire training cluster of this scale to a direct competitor is equivalent to TSMC leasing all of its 5-nanometer production capacity to Samsung. This has never happened in the semiconductor industry. Anyone who understands the cycles knows that such a move only occurs under one circumstance: when you can’t use it all yourself.

According to SpaceXAI's official statement, Anthropic's computing power will "directly benefit Claude Pro and Claude Max subscribers." In other words, Anthropic is using this computing power for inference—running models for Claude's paying users, serving requests for the AI that Musk dislikes the most.

It is inaccurate to summarize this matter as "customer collaboration"; control of Colossus 1 has, in a sense, changed hands.

Why would it be "unlimited"?

The most direct answer lies in Grok's user data.

According to data released by Similarweb in April, Grok’s global mobile app daily active users (DAU) dropped from 13.9 million in March to 12.2 million in April, a monthly decline of 12.5%. In the U.S. market, the decline was even steeper, falling from 1.4 million to 1.1 million, a monthly decrease of 15.6%. Once the second-largest AI app globally after ChatGPT a year ago, Grok has now slipped to fifth place in April, overtaken by Claude, Gemini, and DeepSeek.

Claude's DAU increased from 16 million to 23 million, up 44% month-over-month.

This is a stark contrast: in 2026, as AI applications experience widespread and rapid growth, Grok is among the few leading products losing users. The reason is straightforward—Grok’s core use case has always been tied within the X (formerly Twitter) platform, functioning as a tool for “real-time search + sharp commentary.” Yet, it has never developed the same kind of “workflow stickiness” on its standalone app or web versions as Claude has. Many users on Reddit have complained that Grok has gradually moved its image and video generation features behind a paywall, compounded by regulatory investigations in multiple countries and threats of a ban from Apple—effectively bringing its growth engine to a near halt.

Even more critical is within xAI.

According to a April report by Fast Company, more than 80 employees have left xAI over the past few months, including several co-founders. A February report by the Financial Times noted that Musk has been exerting "unreasonable technical performance pressure" on the team in an attempt to catch up with competitors—a typical response from leaders during periods of consistent product setbacks.

When you consider these two points together, the reason Colossus 1 has excess capacity becomes clear: it was originally built to support a much larger Grok.

"Grok's insufficient demand" is merely superficial.

A deeper layer of logic is that Musk needs a new narrative to justify the $1.25 trillion valuation of SpaceXAI.

Recall what happened in February this year. SpaceX acquired xAI in an all-stock deal, with the merged entity valued at $1.25 trillion—the largest merger in history. Prior to the merger, xAI’s most recent funding round was its Series E in January, raising $20 billion at a $230 billion valuation. By embedding xAI into SpaceX, the company is essentially using the cash flow from SpaceX’s rocket business to sustain xAI’s money-burning black hole, which continues to lose $1.46 billion per quarter.

But even with SpaceX's financial support, SpaceXAI still faces a sharp question: what justifies its valuation?

OpenAI's most recent valuation is $852 billion, with an ARR of approximately $24–25 billion, resulting in a valuation-to-revenue ratio of about 35x. Anthropic is currently negotiating a $900 billion valuation with an ARR of $30 billion, yielding a valuation-to-revenue ratio of 30x.

What about xAI? In Q3 2025, it generated $107 million in revenue but incurred a net loss of $1.46 billion. Even if we optimistically project Grok’s 2026 revenue at $2 billion, the valuation-to-revenue ratio for SpaceXAI would still be far higher than that of OpenAI and Anthropic. In other words, Musk urgently needs to tell a new cash flow story for SpaceXAI—neither user growth nor enterprise API revenue can sustain it.

Renting Colossus to Anthropic was the beginning of the story.

It instantly repositioned SpaceXAI from a "model company" to an "AI cloud infrastructure provider"—a player somewhat like CoreWeave, but larger and with greater power capacity. In the world of narrative valuations, cloud providers are worth more than model companies, as they can offer long-term contracts and predictable cash flows—something pure model companies struggle to deliver.

Add the vague memo about Anthropic and SpaceX’s “orbital compute hub,” where both parties agreed to “explore” deploying multi-gigawatt AI data centers in space, and you understand the full intent: this is a new balance sheet being prepared for SpaceX’s IPO. Rockets, Starlink, ground-based data centers, and orbital compute—all bundled into one massive infrastructure narrative. Grok itself is irrelevant; what matters are the GPUs, power, and launch pads that Musk controls.

Within this framework, Musk's dramatic shift in attitude toward Anthropic takes on another interpretation.

It is a transaction.

Anthropic has provided SpaceXAI with more than just rental space—it has offered a credibility endorsement. By publicly endorsing the availability, scalability, and operational quality of Colossus 1, Anthropic has effectively granted SpaceXAI an invitation to join the “compute infrastructure club,” whose members include AWS, GCP, Azure, and CoreWeave. Prior to this, xAI had virtually no reputation in the cloud services market; it had only used compute resources to train its own models and had never engaged in any external commercial operations.

For Anthropic, the deal is also extremely advantageous. It is raising funds at a $900 billion valuation and is likely to go public in October. It has publicly disclosed a demand for 5 gigawatts of training compute; SpaceX’s offer of 300 megawatts may seem modest, but its value lies in “immediate delivery”: power availability within a month, directly alleviating Claude’s current inference pressures. In April, Anthropic publicly acknowledged that Claude’s reliability and performance had been impacted during peak times due to “infrastructure strain.” The emergency capacity of 300 megawatts is worth far more than its face value.

This is a two-way narrative trade: Anthropic received service stability, while SpaceXAI received a valuation story.

Who gave in? Musk himself indeed stepped back—he did business with his longtime rival and even spoke well of him. But on a deeper level, it was Grok that gave in. As a product, as a model company, and as Musk’s flagship weapon against OpenAI/Anthropic, Grok is being quietly downgraded to just another ordinary offering within SpaceX’s AI portfolio. The fact that Colossus, a core strategic asset, has been made available to customers signals that Musk no longer views “in-house models” as the primary battleground.

In this sense, May 6 marks the end of Grok’s era as a "cutting-edge model company."

Zoom out a bit further—this event’s industry significance may be greater than what we can currently see.

Throughout 2024 and 2025, the AI computing power market has been characterized by intense competition across the entire industry. OpenAI is competing, Anthropic is competing, xAI is competing, Mistral is competing, and sovereign wealth funds from around the world are competing. GPUs are hard currency, data center locations are geopolitical issues, and power supply is a matter of national strategy. In this environment of universal scarcity, no one will rent their training clusters to competitors, because every GPU hour you lease out today could be the critical computing power your rival needs to catch up tomorrow.

And now, xAI has done it.

This means the AI computing power market is experiencing its first major divergence: demand from leading model companies (OpenAI, Anthropic, Google DeepMind) continues to grow exponentially, while second-tier and smaller model companies are beginning to see excess capacity. This divergence occurs in the middle to later stages of every capacity expansion cycle—from solar panels to electric vehicle batteries to Bitcoin miners—with nearly identical patterns. In the early stage, everyone faces shortages; in the middle stage, capacity begins to overflow to second-tier players; in the late stage, top players consolidate upstream and downstream supply chains, while second-tier players either transition into infrastructure service providers, get acquired, or exit the market.

CoreWeave is the best benchmark. Originally an Ethereum mining farm, it seized the window of excess GPU capacity in 2018 to pivot into AI cloud services, reaching a $60 billion market cap at its 2024 IPO. Its very existence proves that shifting from models to compute is a viable path. SpaceXAI is retracing this route with an even more aggressive version—beyond selling ground-based compute, Musk aims to sell compute capabilities into space.

A true signal of the AI bubble's peak may be when second-tier model companies纷纷 transition into cloud service providers. When the core narrative of an industry shifts from "I have the best model" to "I have the most GPUs," it typically means differentiated competition has reached its end.

One noteworthy detail is that in Memphis, where Colossus 1 is located, xAI deployed dozens of natural gas-powered turbines to power construction efforts, claiming the use was "temporary" and therefore exempt from federal permits. Local residents have continued to protest over air pollution concerns, and the issue remains unresolved to this day.

Now, these GPUs powered by gas turbines will be used to run Anthropic’s Claude, one of the most rigorous labs in terms of AI ethics and climate issues.

More absurdly, Anthropic and SpaceX "expressed interest" in deploying gigawatt-scale AI computing power in orbit. Musk’s logic is that Earth’s power and cooling capacity will eventually be insufficient—the future of AI lies in space.

Between Memphis’s gas turbines and the orbital solar panels in Musk’s PowerPoint lies a massive valuation expectation. Colossus 1 leased to Anthropic is the first new story Musk told to justify this valuation expectation.

In just three months, xAI went from being an adversary to a supplier. Who will be next to be revalued?