How xAI Allegedly Distilled Claude, Got Blocked, and Still Landed a $45 Billion Deal With Anthropic

Months before Anthropic and xAI inked a massive $45 billion compute partnership, a quieter fight was already underway. In June 2026, media reports revealed that Elon Musk's AI company, xAI, had been systematically extracting outputs from Anthropic's Claude to train its own Grok models—a contentious technique known as model distillation. When Anthropic caught on and shut off access, xAI didn't stop. Instead, its engineers allegedly turned to personal accounts and an encrypted relay service to keep the pipeline running .

This episode places xAI at the center of a widening crackdown on competitive distillation, one that the rest of the US AI industry has mainly focused on Chinese state-linked labs. The fact that these same two companies would, within months, sign a landmark compute deal reveals just how tangled competition and dependency have become at the AI frontier.

The Core Allegation: Months of Distilling Claude for Grok

Reports published on June 7, 2026, by outlets including The Decoder and Sina Finance, citing project sources, claim xAI ran a months-long model distillation project specifically designed to improve its own coding models using Claude's outputs . Model distillation is a legitimate machine learning technique, often used within a single company to train a smaller model on the outputs of a larger one. The controversy, however, arises when one company does this using another company's model, effectively extracting its capabilities at a fraction of the original training cost and in violation of most terms of service .

The timing of the revelations about Claude was especially awkward for Musk: on April 30, 2026, he testified under oath that his company had partially used OpenAI models through distillation to train Grok, dismissing it as a common practice . The Claude allegations suggest the practice was wider than previously known, making it a deliberate strategy rather than an isolated incident.

How Anthropic Cut Access and xAI Fought Back

The tensions began to surface in early January 2026 when Anthropic discovered xAI was accessing its models through Cursor, a popular developer IDE. The company saw this as a direct violation of its terms, which forbid using Claude to train competing systems. It revoked xAI's official enterprise API access within days and tightened its technical safeguards to prevent third-party tools from masquerading as legitimate clients . xAI co-founder Tony Wu confirmed the block in an internal company memo .

But the cut-off did not end the effort. According to sources cited in a Sina Finance report, at least two workarounds were immediately deployed :

• Personal Claude accounts: xAI engineers began using individual subscriptions to keep querying the model. Anthropic detected and deactivated large batches of linked personal accounts in response.
• Blackbox AI relay service: Queries were routed through Blackbox AI, a third-party platform that encrypts traffic to hide which model is ultimately being called. Two industry insiders claim that as of mid-May 2026, xAI was still accessing Anthropic models through this method for benchmarking and other R&D work.

Neither xAI nor Musk has disclosed the full extent of Claude usage or how the ongoing workarounds were sanctioned internally . Anthropic did not comment on the ongoing use through Blackbox AI.

The Bigger Picture: xAI's Playbook Fits a Global Pattern

The xAI allegations surfaced alongside a much larger, geopolitically charged backlash against adversarial distillation. U.S. AI companies have been mobilizing largely around one target: Chinese AI labs.

In February 2026, Anthropic published a detailed report documenting industrial-scale campaigns by DeepSeek, Moonshot, and MiniMax to illicitly extract Claude's capabilities. The three companies generated over 16 million exchanges using approximately 24,000 fraudulent accounts, often routed through proxy services to hide their origins . OpenAI sent its own memo to Congress the same month, accusing DeepSeek of "ongoing efforts to free-ride" on US frontier models and using increasingly sophisticated obfuscation techniques . By April 2026, OpenAI, Anthropic, and Google had formalized a joint mechanism through their Frontier Model Forum to share threat intelligence and detect adversarial distillation attacks .

The structural similarity is inescapable: xAI's alleged tactics—systematic querying, using intermediaries to circumvent blocks after a ban—are identical to the playbook US companies condemn when used by actors in Beijing. The difference is that xAI is an American competitor, not a foreign lab, and this distinction has generated considerable awkwardness within the industry . Musk has not publicly addressed the apparent hypocrisy, though he has separately railed against Anthropic for its own alleged data scraping practices .

From Blocking Models to Buying Compute: A $45 Billion Paradox

The most dizzying twist came in May 2026. Despite the ongoing access war, Anthropic signed a three-year agreement to lease the entire 300-megawatt output of xAI's Colossus 1 supercomputer near Memphis, Tennessee. The terms are staggering: $1.25 billion per month, totaling roughly $45 billion over the life of the contract through May 2029, with discounted rates during the first two ramp-up months .

The deal was technically executed through SpaceX, the parent entity that built Colossus in 2024 with more than 220,000 Nvidia GPUs. It essentially transforms xAI into a cloud infrastructure provider—at least for this one tenant. In a statement, xAI said the arrangement "allows us to monetize unused compute capacity in our infrastructure" . Industry analysts have noted the paradox sharply: the same company that actively blocked xAI from using Claude's API is now its single biggest compute customer .

The contradiction is a defining feature of the current AI landscape. Frontier labs need each other's computing muscle as desperately as they need to protect their own model intellectual property. Anthropic itself is the world's largest contracted AI compute buyer, with agreements spanning AWS, Google Cloud, CoreWeave, and now xAI . For xAI, the deal offers a revenue stream approaching SpaceX's total annual reported income from a single contract, while simultaneously underscoring how quickly competitive lines can blur when hardware scale becomes synonymous with survival .

The relationship, as one analysis put it, "is deeply paradoxical"—and it's not likely to become simpler anytime soon .