xAI Trained on Claude Outputs for Months Before Anthropic Cut Access

Key Takeaways

• xAI trained its coding models directly on Anthropic's Claude outputs before access was revoked in January 2026
• xAI's pretraining team has shrunk from 20+ engineers to under 5, with four Grok code leads departing within months
• xAI now earns $1.25 billion monthly leasing its Colossus infrastructure to Anthropic instead of training its own models

Elon Musk's xAI spent months training its Grok coding models on outputs from Anthropic's Claude before getting cut off. The company continued the practice through workarounds even after Anthropic revoked official API access in January, The Information reports.

The revelation adds another chapter to AI's ongoing data practices controversy. It also highlights the irony of xAI's current business situation: the company now earns $1.25 billion monthly leasing its GPU infrastructure to the same competitor it was allegedly scraping.

How xAI Used Claude to Train Grok

According to The Information, xAI trained its coding model directly on Claude's outputs. This process, known as model distillation, involves using a more capable model's responses to train a smaller or newer model. It's faster and cheaper than training from scratch on raw data.

When Anthropic discovered the activity and revoked xAI's official API access in January, the scraping didn't stop. xAI engineers kept going through personal accounts and the intermediary service Blackbox AI. The workarounds continued until May 2026.

“Partially [using competitors' models to train] is industry standard.”

— Elon Musk, CEO of xAI, in court testimony regarding the training practices of his models

Musk made that statement during litigation when he admitted xAI "partially" used OpenAI models to train Grok. The same defense now applies to the Claude situation. Whether the practice is truly standard or simply common depends on who you ask. Most AI companies prohibit it in their terms of service.

Internal Turmoil at xAI

The distillation controversy comes as xAI faces significant internal challenges. The company's pretraining team has shrunk dramatically. What was once a team of over 20 engineers is now under five people.

Four Grok code leads left within months. Many co-founders have also departed. In one incident, an employee accidentally deleted critical training data. The mistake cost two to three weeks of work, according to The Information.

The staffing problems suggest xAI may be stepping back from frontier model competition. The company's pivot to infrastructure leasing supports that interpretation.

From AI Developer to GPU Landlord

All that compute Musk bought? He's now renting it out instead of using it to train his own models.

xAI operates Colossus, a massive data center in Memphis housing 220,000 NVIDIA GPUs. The facility was built to power Grok's training runs. Now it generates revenue from competitors.

Anthropic agreed to pay xAI $1.25 billion monthly to lease Colossus 1 infrastructure. Google has a separate deal through SpaceX worth $920 million monthly for 110,000 chips. The combined revenue exceeds what most AI companies spend on their entire operations.

xAI reportedly calls this a stopgap measure. The company claims it will return to training its own models once it resolves internal issues. But the economics of GPU leasing may prove hard to abandon. Compute revenue is predictable. Frontier AI research is expensive and uncertain.

The Hypocrisy Question

Online discussion has focused heavily on the perceived hypocrisy. Musk has positioned himself as an AI safety advocate. He co-founded OpenAI (then left) and has criticized data scraping practices. Now his company stands accused of the same behavior.

The irony deepens when you consider the business relationship. xAI is essentially landlording for the same company it allegedly scraped. Anthropic pays xAI over a billion dollars monthly while xAI's engineers were using workarounds to access Claude after being explicitly cut off.

Comments on Hacker News and Reddit have been unsparing. Many users noted that Musk's public stance on data theft doesn't match xAI's reported practices. Others see the compute deals as evidence that xAI has given up on competing at the frontier and settled for infrastructure utility.

What Happens Next

Anthropic hasn't commented publicly on the distillation accusations. The company's decision to lease compute from xAI despite the controversy suggests the business relationship remains intact. Compute is compute.

For xAI, the path forward is unclear. The talent exodus makes a quick return to frontier model training difficult. The compute leasing business is lucrative but transforms the company from an AI competitor into an infrastructure provider.

Musk's prediction from fall 2025 came true, as he noted on X. Just probably not the way he pictured it.

Frequently Asked Questions

What is model distillation in AI?

Model distillation is training a new AI model using outputs from an existing, often more capable model. Instead of training on raw data, you train on the responses generated by another AI. It's faster and cheaper but typically violates the source model's terms of service.

Did xAI break any laws by using Claude outputs?

The legality is uncertain. Using another company's AI outputs to train your own model likely violates terms of service, but whether it constitutes theft or copyright infringement remains legally untested. Musk has called the practice "industry standard" in court.

Why is Anthropic leasing GPUs from xAI after being scraped?

Compute access is scarce and expensive. Anthropic needs GPUs to train its models, and xAI's Colossus facility has 220,000 available. Business pragmatism often trumps grievances when compute is on the line.

How many GPUs does xAI have?

xAI's Colossus facility in Memphis houses 220,000 NVIDIA GPUs. The company leases portions to Anthropic and Google, generating over $2 billion monthly in compute revenue.

Is xAI still developing Grok?

Unclear. The pretraining team has shrunk to under five people, four code leads have left, and the company is leasing its compute to competitors. xAI calls this a stopgap, but the talent exodus suggests the company may be shifting away from frontier model development.

Source: The Decoder / Matthias Bastian