Translate this title to English: Musk dissolves xAI and founds SpaceXAI : Anthropic recovers Colossus 1 and 300 MW of compute

Musk dissolves xAI and founds SpaceXAI: Anthropic acquires Colossus 1 and 300 MW of compute

🔎 The day Musk handed his GPUs to his rival

On May 6, 2026, Elon Musk announced the outright dissolution of xAI, his artificial intelligence startup founded in 2023. The company merges into SpaceX under the name SpaceXAI. On the same day, Anthropic signed a contract to rent the entirety of Colossus 1, the Memphis supercomputer: 220,000 H100 GPUs and over 300 megawatts of power.

This is an unprecedented strategic reversal. Musk implicitly admits that xAI failed to leverage its compute advantage. Anthropic, excluded the day before from a federal AI testing agreement, overnight finds itself with the largest GPU farm dedicated to a single language model: Claude.

Why now? Because the compute race has entered a consolidation phase. The $12 billion raised by xAI was not enough to compete with the infrastructures of OpenAI and Google. One massive asset remains: Colossus 1. Rather than leaving it underutilized, Musk rents it to Anthropic. A cynical deal, but a financially logical one.

The essentials

• xAI is dissolved on May 7, 2026, and absorbed by SpaceX under the SpaceXAI brand. The engineering team joins SpaceX's workforce.
• Anthropic rents the entirety of Colossus 1: 220,000 H100s, 300+ MW, located in Memphis, Tennessee.
• Grok 4.1, xAI's latest model, will be retired on May 15, 2026. It no longer ranks in agentic benchmarks beyond 14th place (score 79), far behind GPT-5.5 (98.2).
• Anthropic doubles the usage limits of Claude Pro and Claude Max thanks to this additional compute.
• Dario Amodei, CEO of Anthropic, mentions a future partnership for "multi-gigawatt" orbital compute with SpaceX.

Recommended tools

The dissolution of xAI: an admission of failure

xAI no longer exists. The legal entity is dissolved, its employees transferred to SpaceX, its brand replaced by "SpaceXAI". This is the conclusion of an adventure lasting less than three years, launched in July 2023 with disproportionate ambitions.

Musk had justified the creation of xAI by the need for a "truth-seeking chatbot" that would correct the perceived biases of ChatGPT. Grok was supposed to be the anti-woke AI, integrated into X (formerly Twitter). The reality is more prosaic: $12 billion raised, a supercomputer built in record time, and results falling short of expectations.

The June 2025 agentic ranking speaks volumes. Grok 4.1 points to 14th place with a score of 79. For comparison, GPT-5.5 dominates at 98.2, Gemini 3 Pro Deep Think follows at 95.4, and Claude Opus 4.7 (Adaptive) places third at 94.3. Even in generalist terms, Grok 4.1 plateaus at 90, shared with Gemini 3 Pro Deep Think but well behind Gemini 3.1 Pro (92) and GPT-5.5 (91).

"Capital efficiency" is officially the cited reason. In plain terms: burning $12 billion for a mid-table model is over.

Grok 4.1: the numerical assessment

Grok 4.1 remains the latest offspring of the xAI lineage. Its general benchmark performance places it at the same level as Gemini 3 Pro Deep Think (90 points). But this score masks a depth problem: in agentic mode, where the model must plan, execute, and correct complex chains of actions, it drops to 79. Claude Opus 4.7 Adaptive reaches 94.3 in this same test, a gap of over 15 points.

The retirement of Grok 4.1 on May 15, 2026, marks the definitive end of xAI as a model player. SpaceXAI will likely continue to use third-party models for its internal needs (navigation, Starlink, launch optimization), without developing a proprietary LLM.

Colossus 1: 220,000 H100s handed over to Anthropic

Colossus 1 is the heart of the deal. This cluster of 220,000 NVIDIA H100 GPUs, built in Memphis, Tennessee, represents one of the densest compute concentrations in the world. The announced power exceeds 300 megawatts, equivalent to the consumption of a city of 200,000 inhabitants.

Anthropic is not renting a portion: it is taking the entire capacity. This is a crucial point. SpaceX retains no priority access to Colossus 1 for its own AI needs. The cluster becomes an exclusive tool at the service of Claude.

This contract was signed directly with SpaceX — and not with xAI, which no longer exists legally at the time of signing. Bloomberg confirms the May 6, 2026 announcement, and France24 specifies that Anthropic will use "the entire computing capacity" of the data center.

Why Anthropic rather than OpenAI or Google?

The question is legitimate. OpenAI has its own infrastructures via Microsoft. Google uses its internal TPUs. Anthropic, on the other hand, had until now depended on AWS and Google Cloud for its compute. Colossus 1 offers it a dedicated cluster, without sharing resources with other cloud clients.

Anthropic's exclusion from the federal AI testing agreement of May 5, 2026 (which concerned Microsoft, Google, and xAI) likely accelerated the negotiations. Anthropic needed a media coup and a massive compute lever to stay in the race. Musk, for his part, needed to monetize an expensive asset that was losing value every month.

Claude doubles its limits: what changes concretely

The immediate impact for users is clear: Anthropic announced that the compute from Colossus 1 would be used to "improve capacity for Claude Pro and Claude Max users". In practice, this translates to a doubling of usage limits.

For a Claude Pro user, this means more requests per day, longer sessions with Claude Opus 4.7, and reduced waiting times during peak hours. For Claude Max subscribers (enterprise plan), priority access to the cluster allows for chaining complex agentic workflows without interruption.

Claude Opus 4.7 (Adaptive) is the model that benefits the most from this influx. Its adaptive architecture dynamically adjusts its compute usage based on the complexity of the task. With an additional 220,000 H100s, the model can switch to its "high intensity" mode more often without saturating the infrastructure.

Comparison of models impacted by the deal

Orbital compute: the real ambition behind the deal

The most fascinating detail of this announcement is Dario Amodei's mention of scaling to "many terawatts in orbit." This is where the partnership with SpaceX makes complete sense.

The fundamental problem with AI compute is energy. A 300 MW cluster like Colossus 1 requires a massive power supply, industrial cooling, and very high-bandwidth network connections. On Earth, these physical constraints limit the maximum size of a data center.

In orbit, solar panels provide quasi-unlimited energy 24/7. The vacuum of space allows for passive cooling through thermal radiation. And Starlink constellations (also managed by SpaceX) can ensure connectivity.

The scenario is as follows: Anthropic provides the AI expertise, SpaceX provides the launcher (Starship), the satellites, and the connectivity. Together, they could deploy orbital data centers of several gigawatts, or even terawatts. It is science fiction becoming a business plan.

The challenges of orbital compute

Nothing is simple, however. Launching computer hardware into orbit is still expensive, even with Starship. Maintenance is impossible without a return to Earth. Cosmic radiation degrades electronic components. And the latency between low orbit and the ground remains a problem for real-time conversational applications.

But for model training (which is done in batch, without latency requirements), orbital compute becomes credible by the 2028-2030 horizon. Anthropic is positioning itself right now as the AI partner for this future infrastructure.

Federal exclusion: the political context of the deal

One element often omitted in media coverage: on May 5, 2026, the day before the deal was announced, Anthropic was conspicuously absent from a federal agreement on AI testing. This agreement, negotiated between the US government, Microsoft, Google, and xAI, aimed to establish safety protocols for frontier AI models.

Anthropic's exclusion was perceived as a political signal. The company founded by Dario and Daniela Amodei had distinguished itself through its pro-safety stances and its alignment with Democrats. The change in administration had weakened its institutional position.

The very next day, Anthropic signs with SpaceX — a company whose CEO is the president's official technology advisor. The political reading is tempting: a compute deal in exchange for reintegration into the institutional fold. No source confirms this transaction, but the timing is troubling to say the least.

What this means for the AI model benchmark

Access to Colossus 1 changes the dynamics of the AI compute race. Anthropic shifts from a position of cloud dependency (AWS, GCP) to a player with its own dedicated infrastructure. This could translate into significant improvements in upcoming benchmark cycles, particularly for agentic tasks where the volume of compute determines the quality of reasoning.

SpaceXAI: what happens to Grok and the team?

The merger of xAI into SpaceX creates a new division: SpaceXAI. Its role is not to compete with OpenAI or Anthropic in the consumer LLM market. It is about integrating AI capabilities into SpaceX's space operations.

Grok 4.1 continues to exist as a SpaceX product, but its future is uncertain. Without a research team dedicated to developing foundational models, it is hard to imagine how SpaceXAI could maintain a competitive pace of innovation. LLM models evolve every 3-6 months. A model frozen at the level of Grok 4.1 (agentic score 79) will be obsolete before the end of 2026.

xAI's engineering team joins SpaceX with SpaceX employment contracts. Some key talent might leave for other labs. Brain drain is a real risk: why would an LLM researcher stay in a space division that no longer develops models?

What are the $12 billion raised by xAI worth?

xAI had raised a total of $12 billion. Part of it was invested in Colossus 1 (estimated between $2 and $4 billion for the H100 cluster). The rest funded salaries, R&D, and operations. With the dissolution, these assets revert to SpaceX.

The Colossus 1 lease to Anthropic generates recurring revenue for SpaceX. At 300 MW, even with a preferential rate, the contract likely represents several hundred million dollars per year. Not enough to pay back $12 billion, but sufficient to make the infrastructure profitable.

Impact on the AI race: the new landscape

The withdrawal of xAI and the strengthening of Anthropic are redrawing the AI map in May 2026. The LLM market is structuring itself around three dominant poles.

OpenAI remains the undisputed leader with GPT-5.5 (98.2 agentic, 91 general). The company has Microsoft's infrastructure, a massive user base, and a sustained release pace (GPT-5.4, GPT-5.4 Pro, GPT-5.3 Codex in a few months).

Google offers the most balanced lineup: Gemini 3.1 Pro (92 general), Gemini 3 Pro Deep Think (95.4 agentic, 90 general). Google's advantage is vertical integration with Search, Workspace, and Android, coupled with proprietary TPUs that reduce dependence on NVIDIA.

Anthropic was the weakened third pillar. With Colossus 1, it becomes the best-armed challenger in pure compute. Claude Opus 4.7 Adaptive (94.3 agentic, 90 general) and Claude Sonnet 4.6 (81.4 agentic, 83 general) benefit from a dedicated infrastructure that neither OpenAI nor Google have in this form.

Secondary players — DeepSeek (V4 Pro Max at 88 general), Moonshot AI (Kimi K2.6 at 88.1 agentic), Z.AI (GLM-5 at 82 agentic) — remain relevant in their respective markets but lack the compute scale to compete on frontier models.

The Musk paradox

The most ironic aspect of this affair: Elon Musk, who founded xAI to create an alternative to the "corrupted AI" of OpenAI and Anthropic, ends up providing the compute that strengthens Anthropic. His direct rival inherits the supercomputer he himself had built.

Musk doesn't seem bothered by this. The SpaceXian logic prevails: an asset must generate revenue or strategic value. Colossus 1 wasn't generating enough value via Grok. It does so via Anthropic. The rest is entrepreneurial sentimentality, and Musk has never had much of that.

❌ Common mistakes

Mistake 1: Confusing SpaceXAI and xAI

SpaceXAI is not xAI renamed. It is a SpaceX division with a different mandate: embedded AI, launch optimization, autonomous navigation. There is no indication that SpaceXAI will continue to develop consumer LLMs. Claude and Grok are not competitors in the same space.

Mistake 2: Thinking Colossus 1 is enough to catch up with OpenAI

220,000 H100s is massive. But OpenAI has equivalent clusters via Microsoft, plus access to next-generation GPUs. Anthropic gains autonomy, not supremacy. The agentic leaderboard shows that Claude Opus 4.7 (94.3) remains behind GPT-5.5 (98.2) despite this reinforcement.

Mistake 3: Believing orbital compute is imminent

Amodei speaks of "terawatts in orbit," but that is a long-term vision. No orbital data center is operational in 2026. Launch constraints, maintenance, and latency remain major obstacles. The current deal concerns Colossus 1 in Memphis, not a satellite.

Mistake 4: Underestimating the impact of Grok 4.1's withdrawal

X users who used Grok daily will lose their tool on May 15. This is a functional void for the platform. SpaceXAI will either have to integrate a third-party model or offer an internal alternative. Neither option is simple.

❓ Frequently asked questions

Why did Musk dissolve xAI rather than sell it?

xAI did not have a sufficiently differentiated product to interest an acquirer. Grok 4.1 is mid-table, the team is individually recruitable, and the only unique asset is Colossus 1 — which already belongs to SpaceX. Dissolution was the cleanest exit.

Does Anthropic own Colossus 1?

No, Anthropic leases the compute capacity. SpaceX remains the owner of the physical infrastructure. It is a long-term lease, not an acquisition. Anthropic did not buy the GPUs.

What happens to Grok subscribers?

Grok 4.1 will be retired on May 15, 2026. X will need to inform its users and offer an alternative. No migration plan has been publicly announced at this stage.

Does the Anthropic-SpaceX deal affect Claude's pricing?

Anthropic has indicated that the additional compute would be used to improve usage limits, not to lower prices. Claude Pro and Claude Max rates remain unchanged for now. The lower cost per request (thanks to dedicated compute) could, however, translate into future adjustments.

Will Claude Opus 4.7 surpass GPT-5.5 thanks to Colossus 1?

Not automatically. The agentic score depends on the model's architecture, the quality of training data, and reasoning techniques, not just the volume of compute. Colossus 1 allows for more frequent deployment of Opus 4.7's high-intensity mode, but closing the 3.9-point gap with GPT-5.5 will require a new training cycle.

✅ Conclusion

Musk sacrifices xAI to make Colossus 1 profitable, Anthropic lands the largest dedicated GPU cluster in history, and the AI race loses one player to strengthen another. The real signal is orbital compute: Anthropic and SpaceX are preparing the post-Earth era for model training. To follow the evolution of Claude with this new infrastructure, check out our detailed analysis of the Anthropic-SpaceX partnership for Colossus 1.