Elon Musk Praises Anthropic and Promises Not to Cut Off Its Compute, Reversing His Earlier Verdict

Elon Musk praised Anthropic and promised not to cut off its compute, reversing a public prediction that the AI company could never win.

The reversal matters because Anthropic now depends on infrastructure controlled by one of its most visible competitors. It signed an agreement in May to use the entire computing capacity at Colossus 1, a data center associated with SpaceX and xAI.

Musk is therefore playing two conflicting roles. He leads an AI rival through xAI, while SpaceX supplies infrastructure that helps Anthropic train models and serve Claude users. His reassurance addresses that conflict, but it does not remove it.

The immediate story began on X. Users questioned whether Musk might remove Anthropic from SpaceX infrastructure to weaken a competitor. Musk responded that doing so was “not my style,” according to the original compute dispute.

He also acknowledged that his previous assessment of Anthropic had been wrong. Musk had written in September 2025 that winning was never among the company’s possible outcomes.

Now he describes Anthropic as an AI leader and praises its Mythos and Fable models. That endorsement is striking because Claude competes directly with xAI’s Grok models for developers, businesses, and consumer attention.

This is more than an unexpected compliment between rivals. It is a test of whether infrastructure contracts can remain dependable when a supplier also wants to beat its customer.

Elon Musk Praises Anthropic and Promises Not to Cut Off Its Compute

The central change is Musk’s transition from dismissing Anthropic to publicly defending its access to infrastructure that his companies control.

Musk did not offer a neutral observation about a competitor. He admitted that his earlier forecast was wrong and described Anthropic’s latest models as leaders in the field.

That statement carries unusual weight because the relationship extends beyond social media. Anthropic announced its SpaceX agreement on May 6, linking the arrangement directly to higher Claude usage limits.

Under the agreement, Anthropic receives all computing capacity at Colossus 1. The company says that allocation provides more than 300 megawatts and over 220,000 Nvidia GPUs.

A GPU is a processor designed to execute many calculations in parallel. Frontier AI companies combine thousands of them to train models and answer user requests.

Anthropic said the new capacity would arrive within a month. It also connected the expansion to higher limits for Claude Code, its agentic software-development product, and the Claude API.

Those practical benefits make the arrangement harder to dismiss as a distant infrastructure plan. Developers can experience the result through greater availability, fewer capacity restrictions, and higher request limits.

Anthropic’s compute announcement said it doubled five-hour Claude Code rate limits across several subscriptions. It also removed certain peak-hour reductions and raised API limits for Claude Opus models.

The company did not publicly disclose the agreement’s commercial terms in that announcement. TechCrunch reported additional figures, but those terms have not been presented in Anthropic’s public statement.

The verified operational details are substantial on their own. Anthropic secured an entire data center’s capacity while continuing to use infrastructure from Amazon, Google, Microsoft, Nvidia, and other partners.

That diversification is important. Colossus 1 is a major addition, yet it is not Anthropic’s only source of computing hardware.

Anthropic says it trains and runs Claude across AWS Trainium, Google TPUs, and Nvidia GPUs. Different chips, cloud platforms, and regions reduce dependence on any single technical stack.

However, diversification at the company level does not make each workload portable overnight. Moving a large training run or inference fleet between providers requires engineering work, available hardware, networking, and compatible software.

Inference is the process that produces an answer from a trained model. At Anthropic’s scale, inference capacity must support large numbers of simultaneous requests with predictable response times.

That operational reality explains why users questioned Musk’s influence. A supplier does not need to control every server to create disruption. It only needs to control capacity that has become important to current demand.

Musk’s response attempted to settle that concern through personal reputation. He referenced his history of making Tesla patents available and opening the Supercharger network to other automakers.

Those examples support his claim that he does not automatically deny infrastructure to competitors. They are relevant precedents, but neither precisely matches the Anthropic relationship.

Tesla’s charging network serves an industry standardization role and earns revenue from outside vehicles. Anthropic’s agreement involves scarce computing resources used in a direct race between frontier AI developers.

The strongest protection is therefore not Musk’s public tone. It is the contract, combined with Anthropic’s ability to distribute workloads across other providers.

That distinction will define how seriously customers should take his promise. Public assurances can shape confidence, while enforceable obligations shape operational risk.

Anthropic’s Compute Deal Puts xAI in Two Businesses

xAI and SpaceX now benefit when Anthropic grows, even while Grok competes with Claude for many of the same customers.

The arrangement illustrates a widening split within the AI economy. Building models is one business, while supplying the infrastructure behind those models is another.

A company can lose a model benchmark and still earn value from the winner’s computing demand. It can also supply a rival while developing competing products.

Cloud providers already operate within this tension. Microsoft works with OpenAI while offering other models through Azure. Amazon supports Anthropic while providing infrastructure to a broader market.

Google develops Gemini, supplies specialized hardware, and maintains a commercial relationship with Anthropic. Nvidia sells processors to nearly every major participant while developing its own AI software.

SpaceX and xAI bring a more personal competitive element. Musk publicly promotes Grok, criticizes rival laboratories, and influences the infrastructure organization serving Anthropic.

That combination makes the relationship feel less conventional than a standard cloud contract. Yet the underlying economic logic is familiar.

AI laboratories need vast amounts of electricity, chips, networking, storage, and cooling. Infrastructure owners need customers able to commit to sustained demand.

Anthropic’s model releases create exactly that demand. Fable 5 and Mythos 5 are designed for longer, more autonomous tasks than earlier Claude models, according to Anthropic.

Long-running agents consume more computation than short chatbot exchanges. A coding agent might inspect a repository, modify files, run tests, review failures, and revise its work.

A research agent may process millions of tokens, maintain notes, compare documents, and call external tools. Each additional action creates another inference workload.

This helps explain why model capability and infrastructure capacity must expand together. A highly capable model provides little commercial value when users constantly encounter restrictive limits.

Anthropic explicitly tied Colossus 1 to increased availability. That connection turns an abstract data-center agreement into a product decision affecting developers and enterprise buyers.

It also pressures xAI. Grok must compete against a Claude service whose capacity is partly supported by infrastructure within Musk’s corporate sphere.

If Claude usage grows, SpaceX benefits as a supplier. If Grok wins customers from Claude, xAI benefits as a model provider. Musk has economic exposure to both sides of the contest.

This structure can moderate destructive competitive behavior. Cutting off a paying infrastructure customer would sacrifice supplier benefits and potentially damage confidence among future customers.

However, the same structure introduces governance questions. Customers need confidence that resource allocation, maintenance, and contract enforcement remain separate from model competition.

Those questions become more serious during shortages. When both xAI and Anthropic want additional GPUs, networking capacity, or power, allocation decisions will reveal whose interests take priority.

Anthropic’s broader compute portfolio is its answer to that risk. Its announced agreements span multiple chip architectures and infrastructure providers rather than one exclusive environment.

The company has also expressed interest in orbital AI compute with SpaceX. Orbital systems would place computing equipment in space, where power, cooling, maintenance, and communications present difficult engineering constraints.

That project remains an expression of interest, not an operating production system. Readers should separate it from Colossus 1, which Anthropic says is already supporting increased capacity.

For developers, the most meaningful evidence appears in product behavior. Stable rate limits, predictable latency, and sustained access matter more than speculative infrastructure concepts.

Teams using AI coding tools also need records of prompts, outputs, decisions, and source material. A searchable AI knowledge base can preserve that context when models or providers change.

The infrastructure relationship therefore reaches beyond a rivalry between executives. It influences how businesses plan AI workflows, manage service risk, and decide whether a model can support production work.

Mythos and Fable Explain Musk’s Reversal

Musk changed his assessment after Anthropic presented models aimed at long-horizon coding, knowledge work, vision, and scientific research.

Anthropic describes Mythos-class models as a capability tier above its Opus line. Fable 5 and Mythos 5 use the same underlying model, but Anthropic applies different safeguards and access rules.

Fable is the broadly available configuration. Mythos offers fewer restrictions for vetted organizations working in selected areas, including cybersecurity.

This difference matters because Musk praised both names together. He was endorsing the underlying capability while acknowledging Anthropic’s attempt to separate general access from higher-risk use.

Anthropic says Fable 5 can work autonomously for longer periods than previous Claude models. The company also claims improvements across software engineering, document analysis, visual reasoning, memory, and life sciences.

Several examples come from customers with early access. Stripe reportedly tested Fable 5 on a migration across a 50-million-line Ruby codebase.

According to Anthropic, the model completed the migration in one day. Stripe estimated that a team would otherwise have needed more than two months.

That result is a customer report published by Anthropic, not an independent controlled evaluation. It still illustrates the kind of workload that drives demand for larger compute allocations.

Anthropic also says persistent file-based memory improved Fable’s performance in a long-running game test. Persistent memory lets an agent save notes outside its immediate prompt and consult them later.

The technique resembles how human teams preserve research across sessions. It also explains why knowledge workers increasingly care about personal knowledge management, rather than treating every model conversation as isolated.

The company reports ambitious scientific results for Mythos 5. Anthropic says its internal protein-design experts accelerated parts of their workflow by roughly tenfold.

It also says nine of 14 studied protein targets produced strong drug-design candidates under further investigation. Those findings remain company claims until external researchers validate them and published methods permit closer examination.

Anthropic has promised to publish additional genomics results. Until that happens, readers should treat the reported performance as preliminary evidence rather than settled scientific proof.

The company’s model release also details safety restrictions. Fable uses classifier models that detect requests involving cyberattacks, sensitive biology, chemistry, or model distillation.

A classifier is a separate AI system that categorizes a request before the main model answers. When a request triggers certain restrictions, Fable may reject it or route it to Claude Opus 4.8.

Anthropic acknowledges that benign requests can trigger these controls. That tradeoff makes Fable safer by the company’s standards, but potentially less useful for legitimate researchers.

Mythos provides vetted partners with fewer restrictions in approved domains. This creates a two-level access model based on user identity, institutional controls, and intended work.

That access strategy distinguishes Anthropic from laboratories focused on broad releases. It also creates commercial and political tension around who receives the strongest capabilities.

Musk’s praise does not validate every benchmark or safety claim. It shows that a highly interested competitor considers Anthropic’s progress significant enough to reverse his earlier judgment publicly.

The competitor’s identity gives the statement news value. It does not give the statement scientific authority.

Independent comparisons remain essential because benchmark results can depend on prompting, tool access, evaluation design, and reasoning budgets. A model can lead one coding test and trail another.

Real-world reliability also involves more than task completion. Production agents must handle ambiguous instructions, recover from errors, protect data, and make their actions auditable.

Anthropic’s early-customer examples suggest progress on difficult workloads. They do not establish that Fable or Mythos is universally superior across every business application.

Still, the direction is clear. Anthropic is pushing Claude from conversational assistance toward sustained execution, and sustained execution requires far more infrastructure.

That connection links the model story to the SpaceX agreement. Musk is not merely praising an impressive demonstration. His infrastructure helps support the workload created by those capabilities.

The Promise Does Not Eliminate Anthropic’s Supplier Risk

Musk’s assurance lowers the temperature of the dispute, but it cannot substitute for technical redundancy, enforceable contracts, and transparent service performance.

The skeptical case starts with the source of the promise. Musk made it during a conversation on X rather than through a joint corporate statement detailing governance protections.

His answer addressed the most dramatic scenario, an intentional cutoff designed to hurt Anthropic. Operational risk includes many less dramatic possibilities.

Capacity can become constrained. Hardware can fail. Network performance can deteriorate. Expansion schedules can slip, and resource priorities can change.

A supplier can also comply with a contract while delivering a relationship that becomes commercially unattractive. Service-level terms determine what happens when performance or availability falls short.

Those details have not been publicly disclosed. Readers therefore cannot independently assess termination rights, minimum capacity commitments, remedies, or allocation protections.

TechCrunch reported that contract provisions provide safeguards. Even so, the public record does not reveal enough detail to measure their strength.

Anthropic’s multi-provider strategy offers the clearest visible defense. The company has announced large compute relationships with Amazon, Google, Microsoft, Nvidia, and Fluidstack.

Multiple providers reduce concentration, but they do not create instant interchangeability. Trainium, TPUs, and Nvidia GPUs use different software environments and optimization paths.

Training a frontier model also involves tightly coordinated clusters. Engineers cannot simply move an interrupted run to another provider as easily as transferring an ordinary web application.

Inference can be more portable, especially when deployment planning begins early. Yet moving heavy traffic still requires spare capacity, tested systems, and reliable data connections.

The relationship presents another risk for SpaceX. Other AI companies might hesitate to place strategic workloads on infrastructure connected to a direct model competitor.

Musk’s promise therefore protects more than Anthropic. It supports SpaceX’s credibility as an infrastructure supplier to outside AI laboratories.

If SpaceX consistently serves rivals without interference, it can make a stronger case for becoming a neutral compute provider. If disputes affect service, potential customers will demand stricter protections or go elsewhere.

Questions also surround Anthropic’s own model claims. Its safeguards are central to the distinction between Fable and Mythos, but false positives can block legitimate work.

Anthropic says its cyber classifiers rejected harmful requests in internal and partner testing. Attackers, however, continuously adapt their methods after public deployment.

The company also retains Mythos-class traffic for 30 days to investigate complex attacks and reduce false positives. That policy creates a separate concern for businesses handling sensitive information.

Anthropic says retained data will not train new Claude models and will support safety purposes. Enterprise buyers must still evaluate whether the policy matches their compliance duties.

Government intervention adds another layer of uncertainty. Access to the most capable models can change because of export controls, security directives, or rules concerning foreign nationals.

In June, Anthropic temporarily took Fable and Mythos offline in response to a United States government directive, according to an access restrictions report. That episode demonstrated that providers are not the only source of cutoff risk.

An infrastructure contract cannot override government restrictions. Nor can a personal promise guarantee access when legal requirements change.

This broader context makes the X exchange slightly misleading. The relevant question is not simply whether Musk intends to disconnect a rival.

The better question is whether Anthropic can keep Claude available across commercial disputes, technical failures, regulatory interventions, and sudden increases in demand.

For enterprise buyers, resilience should be assessed through evidence. They should watch uptime, rate-limit changes, regional availability, data rules, and documented fallback plans.

Developers should also avoid designing workflows that assume one model will always remain accessible. Model abstraction layers, exportable records, and tested alternatives can reduce switching costs.

None of this means Musk plans to break his promise. There is no verified evidence that he intends to interfere with Anthropic’s service.

The risk comes from structural incentives and operational complexity, not proof of bad faith. That distinction keeps the analysis grounded while recognizing why users raised the concern.

Three Signals Will Show Whether Musk’s Assurance Holds

The next test is not another supportive post. It is whether Anthropic’s capacity, model access, and supplier independence remain stable under pressure.

The first signal is Claude’s actual availability. Anthropic connected the SpaceX agreement to higher usage limits, making product performance the earliest measurable outcome.

Users should watch whether those higher limits persist during major model launches and demand spikes. Stable access would support Anthropic’s claim that Colossus 1 materially expanded its capacity.

Repeated reductions would weaken that claim, even if the underlying cause remained unclear. Capacity promises only matter when customers can complete their work reliably.

Latency also deserves attention. Higher request limits do not fully solve the problem if responses slow sharply during busy periods or agent sessions fail unpredictably.

The second signal is Anthropic’s next allocation of frontier workloads. The company has already spread capacity across several providers and chip architectures.

Future announcements will show whether Colossus 1 becomes a central dependency or one component in a balanced portfolio. A wider distribution would strengthen the case that supplier leverage remains limited.

The placement of Mythos successors will be especially revealing. Training and serving the next model generation will require large, coordinated resources.

If Anthropic assigns substantial new workloads to SpaceX, that decision would signal confidence in the relationship. If it shifts growth elsewhere, supplier risk may have influenced planning.

The third signal is how xAI and SpaceX behave during direct competitive pressure. A major Grok release could create simultaneous demand for chips and infrastructure.

That moment would test whether outside customers receive predictable treatment when Musk’s own AI company needs resources. Transparent capacity reporting would make the outcome easier to judge.

It would also clarify SpaceX’s long-term identity. The company can operate as a strategic extension of xAI or as an infrastructure supplier serving multiple frontier laboratories.

Those roles can coexist, but only if governance and contracts keep competitive decisions from undermining customer service.

Model competition will provide another useful reference point. OpenAI, Google, and other developers will continue releasing systems aimed at coding, research, and long-running agents.

Independent evaluations should reveal whether Mythos and Fable maintain the lead that Musk described. Anthropic’s position will look different if rivals match its capabilities with fewer access constraints.

Safety performance matters as much as benchmark leadership. Anthropic must show that Fable’s classifiers block harmful requests without making legitimate professional work impractical.

Researchers should watch reported jailbreaks, false-positive rates, and updates to trusted-access programs. Those indicators will test the company’s central product compromise.

A trustworthy outcome would combine three results: stable Claude capacity, continued infrastructure diversification, and competitive neutrality from SpaceX during periods of scarcity.

A weaker outcome would feature recurring restrictions, concentrated dependence, or allocation decisions favoring xAI without transparent explanations.

For developers and enterprise buyers, Elon Musk praising Anthropic and promising not to cut off its compute is useful context, not a reliability guarantee.

Treat the statement as the beginning of an observable test. Track Claude’s limits, Anthropic’s next infrastructure commitments, and SpaceX’s conduct when Grok and Claude compete for scarce capacity.

The reversal already tells us that Anthropic changed one prominent rival’s assessment. The coming months will show whether the unusual alliance can also change how AI competitors share the infrastructure beneath their models.