More from this creator
Other episodes by Kitty Cat.
More like this
If you liked this, try these.
Transcript
The full episode, in writing.
Elon Musk launched xAI in 2023 with a clear goal: to take on OpenAI and Anthropic at the top of the artificial intelligence world. He raised about $42 billion and built up a stockpile of GPUs, preparing xAI to compete in the high-stakes race to develop advanced AI models. This investment laid the groundwork for xAI’s first offering, the Grok chatbot, which came out in 2023. But despite the war chest and technical ambition, xAI was behind. It didn’t release its first coding-focused model until August 2023. By that time, both Anthropic and OpenAI had not only released their own coding tools, but were already selling versions widely considered more advanced and more robust.
Anthropic moved quickly to protect its technical advantage. In August 2023, it began restricting direct access to its Claude API for rivals, citing concerns about violations of its terms of service. OpenAI was the first to be cut off, after technical staff were found using Anthropic's coding tools in ways Anthropic said weren’t allowed. Anthropic said this kind of access, even for benchmarking or evaluation, could be abused in ways that threatened its intellectual property and model safety. xAI soon faced the same fate. By the end of that summer, Anthropic had also removed xAI’s corporate access to its flagship Claude models, locking one of its most direct competitors out of the Claude ecosystem.
But the cutoff did not end xAI’s interest in Anthropic’s technology. For months after losing formal access, xAI continued to get Claude outputs through other channels. Engineers reportedly used personal accounts and intermediaries, including a third-party service called Blackbox AI, to circumvent restrictions. As of May 2024, xAI was still accessing Anthropic’s models using these workarounds. Blackbox’s encrypted service allowed users to route requests to rival AI labs, masking the true origin and intent of the queries. This ongoing access enabled xAI to continue a multimonth distillation project, training its own coding model directly on Anthropic’s responses. Two people familiar with the work said outputs from Claude were fed into training datasets, giving xAI a shortcut to capabilities it was struggling to match internally.
Inside xAI, these backdoor strategies unfolded against a backdrop of turmoil. In 2024, the lab cycled through at least four Grok Code leads in a matter of months. Names like Tony Wu and Beibin Li were associated with the leadership churn. By mid-May 2024, the pretraining team had shrunk from more than two dozen to fewer than five people. Staff departures accelerated after SpaceX acquired xAI, with eight of the startup’s co-founders leaving in a wave of exits. Budget cuts and engineering layoffs followed, leaving the core development group understaffed during a critical phase of competition.
Technical problems compounded the instability. In early 2024, a staffer handling data migration accidentally deleted two to three weeks of core training data for xAI’s coding product. That lost data represented weeks of progress and forced the team to rebuild part of their training pipeline.
Despite these setbacks, xAI kept moving. In January 2024, Tesla poured $2 billion into xAI, deepening the financial ties between Musk’s companies. Days later, SpaceX acquired xAI at a valuation of $250 billion. These moves gave xAI both capital and access to wider infrastructure, but also triggered more staff departures as the original team gave way to new management and priorities.
As technical and organizational problems mounted, Elon Musk began seeking help from outside companies. In March 2024, he announced a collaboration between Tesla and xAI on the Macrohard project, designed as a computer use agent with the ambitious goal of replacing entire categories of white-collar work. Members of Tesla’s Autopilot, AI infrastructure, and Dojo supercomputer teams started working at xAI, adding more technical muscle but also blurring the boundaries between the Musk companies. Around this time, xAI initiated partnerships with Cursor and Mistral AI. Cursor, a coding assistant startup led by Michael Truell, received access to xAI’s infrastructure to train its own Composer model. SpaceX secured an option to buy Cursor for $60 billion or pay a $10 billion breakup fee if the deal didn’t go through. Cursor’s CEO was seen regularly at xAI’s offices, and the two companies held early meetings to share technical updates and proprietary details.
Mistral AI executives visited xAI headquarters in mid-April 2024, with hopes of leveraging Mistral’s customizable models for pretraining and post-training work, but no formal partnership emerged. Mistral founding engineer Devendra Chaplot, who joined xAI as a pretraining lead, left the company by May, ending the brief collaboration.
While these partnerships were forming, xAI pivoted to focus on infrastructure deals. In April, Cursor gained access to xAI’s compute resources, enabling it to accelerate work on Composer. The infrastructure strategy took a dramatic turn in May 2024, when xAI and Anthropic announced a deal for Anthropic to lease a significant portion of xAI’s data center capacity. The terms were stark: $1.25 billion per month, with an initial six-month commitment and month-to-month flexibility after that. This deal gave Anthropic access to Colossus 1, xAI’s flagship data center, equipped with over 220,000 NVIDIA GPUs—including H100, H200, and next-generation GB200 accelerators. That GPU count is larger than the population of Des Moines, Iowa, and represents one of the largest private AI compute clusters in the world.
The strategy behind the deal was complex. On one hand, the payments from Anthropic immediately boosted xAI’s and SpaceX’s financials, bringing in $7.5 billion over the initial term—more than the GDP of Liechtenstein. On the other hand, it raised questions about whether xAI was pivoting away from developing its own breakthrough models toward becoming a cloud-like provider of compute capacity. This was not lost on industry analysts. TechCrunch speculated that xAI’s willingness to lease so much capacity to a direct rival might signal it had already shifted its own training efforts to a newer facility, Colossus 2, and didn’t need both centers running at full tilt.
Elon Musk publicly characterized the Anthropic deal as short-term, writing on X that the rationale was, “I thought we might need the compute back at some point.” He positioned xAI as still focused on building its next generation of models, but the lease agreement with Anthropic brought new scrutiny to xAI’s long-term direction—was it primarily an AI developer, or was it morphing into a provider of industrial-scale infrastructure to its own competitors?
Meanwhile, Anthropic’s commitment to responsible AI development remained a central point of friction. The company had recently tangled with the Pentagon, resisting demands to loosen safety measures or provide broader access for military purposes. Anthropic’s ethical stance—advocating for careful oversight and restraint in AI deployment—set it apart in an industry obsessed with speed and competitive edge.
The Colossus 1 data center, now central to Anthropic’s compute lease, brought its own challenges. The facility has faced environmental lawsuits, complicating both xAI’s and Anthropic’s operations. The scale of the data center—housing more GPUs than most national supercomputers—has drawn attention from regulators and environmental groups concerned about energy use and local impact.
The rivalry between xAI and Anthropic, fought both on technical and strategic fronts, has shaped every stage of xAI’s development. The decision to lease compute to Anthropic was born out of both financial need and an evolving view of xAI’s place in the AI ecosystem. Steep first-quarter losses at SpaceX highlighted the unprofitable nature of xAI’s AI unit, and leasing capacity to Anthropic provided much-needed revenue. At the same time, the deal let xAI and SpaceX pitch themselves to investors as AI growth stories ahead of their IPOs, emphasizing enterprise and consumer offerings as a key part of the overall business.
Anthropic’s reliance on xAI’s infrastructure came at a time when it was preparing to go public. In early June 2026, Anthropic confidentially filed for an initial public offering with the U.S. Securities and Exchange Commission, seeking to raise capital and broaden its investor base. The timing put Anthropic in direct competition with both OpenAI and SpaceX, as all three raced to list on public markets.
For xAI, the months leading up to the partnership were marked by missed targets and course corrections. The original plan was to launch a new version of its coding model alongside a general-purpose upgrade by the end of 2023, but delays forced a pivot. xAI instead rolled the model into Grok Build, a new developer platform released in May 2024. Tesla engineers received early access to Grok Build, offering feedback and helping xAI refine its coding assistant despite internal upheaval.
SpaceX’s IPO paperwork, filed ahead of its record-shattering offering, played up ongoing training of the next major Grok model and recent releases of video and voice-generation features. Yet, no launch timeline for the next Grok release was set, and the public documentation focused as much on infrastructure growth as on new AI breakthroughs.
The pattern repeated: technical ambition, organizational tumult, and a constant recalibration of strategy. As of mid-May 2024, xAI’s core pretraining team was down to fewer than five people, while the company had cycled through four Grok Code leads in just a few months. Beibin Li, the most recent lead, left in May, joining a long list of departures.
Even as xAI’s access to Anthropic’s Claude models was being shut down, engineers continued to seek out ways to benchmark and train against them—using personal accounts and intermediaries to evade API restrictions for as long as possible. These cat-and-mouse tactics persisted into May 2024, months after the formal cutoff. The drive to match Anthropic’s technical lead remained relentless, even as xAI’s own internal resources dwindled.
The most recent development in the rivalry is the sheer magnitude of the compute deal. Anthropic’s $1.25 billion per month payment for a substantial portion of Colossus 1 represents one of the largest infrastructure leases in AI history. The initial six-month commitment, with month-to-month extension options, gives both companies flexibility—but also keeps the pressure on xAI to decide whether it will reclaim the compute for its own next-generation model training.
The move to lease such a large slice of capacity was partly enabled by xAI’s own migration to Colossus 2, a newer data center reportedly designed to support the company’s long-term AI ambitions. Analysts questioned whether xAI’s willingness to rent out Colossus 1 meant it was deprioritizing AI research in favor of becoming a neocloud—a next-generation infrastructure provider for the industry. The deal’s sheer scale dwarfs most prior agreements: $7.5 billion in six months is more than the market capitalization of some entire tech firms.
Cursor’s involvement further illustrated the shift in strategy. With access to xAI’s infrastructure, Cursor is pushing its Composer model forward, leveraging the same GPU clusters that power Anthropic’s Claude runs. SpaceX’s $60 billion option to acquire Cursor, or pay a $10 billion breakup fee, signals the stakes involved in controlling the next wave of coding assistants and developer tools.
Amid all these maneuvers, the boundaries between partnership and rivalry are blurring. As xAI provides the hardware backbone for Anthropic’s training, both companies are racing to establish their models and platforms as indispensable in a crowded market. Market watchers are closely tracking whether xAI will reclaim its compute, double down on model research, or pivot even further toward offering infrastructure as its main product.
Anthropic’s ethical positioning continues to be a point of differentiation. The company has called for industrywide pauses in development to manage AI risks and has resisted government and military pressure to relax safety protocols. Its decision to lease compute from a direct rival is both a pragmatic response to scaling demands and an acknowledgment of the limited options for high-density AI infrastructure at the cutting edge of the field.
What comes next is uncertain. xAI is actively training the next major Grok model, deploying new video and voice-generation features, and fielding a developer community around Grok Build. But the company’s public messaging, internal headcount, and willingness to lease out its most advanced infrastructure all point to a strategic crossroads. The latest available figure: Colossus 1 is currently running more than 220,000 NVIDIA GPUs—more than the combined total in most national data centers—primarily to power a competitor’s AI model.