In February 2026, Elon Musk posted on X that Anthropic was “misanthropic” and its $380 billion valuation was “insane.” Three months later, on May 6, his company SpaceX signed over the entire Colossus 1 supercomputer to run Claude. The facility holds more than 220,000 NVIDIA GPUs and delivers over 300 megawatts of compute power. That reversal tells the real story of the AI industry in 2026: compute demand is so extreme that it overrides personal grudges, competitive rivalries, and three months of public insults.
What SpaceX Handed Over
Colossus 1 sits in Memphis, Tennessee. SpaceX inherited it when the company acquired xAI, Musk’s AI lab, earlier in 2026. The facility ranks among the largest AI supercomputers ever assembled: 220,000 NVIDIA GPUs generating more than 300 megawatts of compute capacity. Anthropic gets the full data center, not a partition or a time-share, available within one month of signing.
For context, a typical hyperscaler data center operates on 50 to 100 megawatts. Colossus 1 alone delivers three to six times that. Anthropic announced the deal at the Code with Claude developer conference in San Francisco, alongside immediate changes to Claude’s usage limits for every paid subscriber.
This new capacity stacks on top of an already massive compute footprint. Anthropic had previously locked in five gigawatts through Google and Broadcom partnerships, plus another five gigawatts via Amazon Web Services through a $100 billion spending commitment over ten years. Even ten gigawatts was not enough to keep up with demand.
80x Growth That Broke the Plan
CEO Dario Amodei described Q1 2026 in blunt terms: “just crazy” and “too hard to handle.” Anthropic’s revenue and usage grew 80x on an annualized basis during the quarter, against an internal projection of 10x. Annualized revenue hit $30 billion, tripling year over year. Over 1,000 enterprise customers now spend more than $1 million annually on Claude.
The surge came primarily from Claude Code. Software engineering teams adopted it for agentic development workflows, and major enterprise clients including Uber and Netflix integrated it into production pipelines. Amodei pointed out that “software engineers are the ones who are fastest to adopt new technology,” a pattern that tends to pull the rest of the enterprise behind it within 12 to 18 months.
The demand overwhelmed Anthropic’s inference infrastructure. Throughout April 2026, paid users hit rate limits after just a few coding sessions. Developer forums filled with complaints. Anthropic was posting record revenue growth while simultaneously degrading the experience for the customers generating that revenue. When demand outstrips your internal forecast by 8x, software optimization cannot close the gap. You need physical hardware, and you need it before your fastest-growing user base loses patience.
Why Musk Said Yes
The deal reads as paradoxical until you look at the balance sheet from SpaceX’s side. xAI built two supercomputers: Colossus 1 and the larger Colossus 2. After SpaceX absorbed xAI, it inherited both facilities. But xAI’s flagship product, Grok, never grew into the capacity that Colossus 1 provided. By one estimate, xAI utilized only 11% of its total chip allocation. Roughly 200,000 GPUs were sitting idle in Memphis, consuming power and producing zero revenue.
With SpaceX’s IPO expected in June 2026, converting that idle asset into a high-margin revenue stream makes straightforward financial sense. Analysts estimated the deal could generate $3 to $4 billion in annual revenue for SpaceX. Turning a potential write-down on an underutilized facility into a multi-billion-dollar income line right before going public is exactly the kind of financial engineering that IPO investors reward.
Musk’s public comments reflected the pragmatism. He said he was “impressed” with Anthropic’s team and that Claude will “probably” be good for humanity. Then he added the most Musk caveat imaginable: “We reserve the right to reclaim the compute if their AI engages in actions that harm humanity.” xAI will continue operating on Colossus 2, its newer and larger facility.
The Enemy of My Enemy Gets 220,000 GPUs
The strategic subtext runs deeper than a landlord and tenant arrangement. Musk is simultaneously suing Sam Altman and OpenAI over their transition from a nonprofit to a for-profit entity. Anthropic was founded by Dario and Daniela Amodei, who left OpenAI in 2021 over disagreements about safety practices and commercialization strategy. One analyst summarized the dynamic without subtlety: “Elon’s enemy is Sam. Dario’s enemy is Sam. Enemy of my enemy is a compute partner.”
Anthropic’s growth directly pressures OpenAI’s market position. Claude Code competes with GPT-5.5 for developer adoption, and Anthropic’s revenue trajectory now challenges OpenAI’s lead in enterprise AI. By providing the compute that lets Anthropic scale its developer tools, Musk funds the most effective challenger to his primary rival without writing an investment check. He gets paid to do it.
This deal also introduces a new category in the compute arms race. The traditional model was hyperscaler partnerships: Anthropic with Amazon and Google, OpenAI with Microsoft. OpenAI’s $20 billion Cerebras deal for non-NVIDIA inference chips already stretched the definition of “compute partner.” An AI lab renting full capacity from a space company’s subsidiary was not a partnership model anyone predicted even twelve months ago. The industry’s compute hunger is producing increasingly creative supply arrangements, and the strangeness is accelerating.
What Changed for Claude Users
The user impact landed immediately on May 6. Five-hour rate limits for Claude Code doubled across Pro, Max, Team, and Enterprise plans. Peak-hour restrictions, which had throttled Pro and Max users during high-demand periods, were eliminated entirely. API rate limits for Claude Opus models also increased considerably.
For developers who had been hitting capacity walls after a few coding sessions, the change was material. Claude Code subscribers who previously exhausted their allocation by mid-afternoon could now work through full development cycles without interruption. In enterprise environments where engineering teams depend on Claude Code for daily output, the reliability improvement matters as much as the raw throughput increase.
The rate limit complaints had become Anthropic’s most visible product problem in April 2026. The company was growing revenue faster than any AI company in history while simultaneously frustrating the developers generating that revenue. The SpaceX deal bought capacity, and the doubled limits bought back trust.
Orbital Compute: The Clause Nobody Expected
Buried in the announcement was a sentence that read like science fiction: Anthropic expressed interest in developing “multiple gigawatts of orbital AI compute capacity” in partnership with SpaceX. In January 2026, SpaceX had filed with the FCC to deploy a million-satellite orbital AI data center megaconstellation.
The physics, in principle, favor the concept. Data centers generate enormous heat and consume enormous power on Earth. In orbit, solar energy is abundant and continuous. Radiative cooling in the vacuum of space is vastly more efficient than any terrestrial system. No land acquisition, no water rights, no local permitting fights over power consumption. The engineering challenges remain significant: latency to ground stations, launch costs per GPU, radiation shielding for sensitive chips, and the basic problem of maintaining hardware that no technician can physically access.
The filing exists. The partnership language is in the deal. When two companies whose combined market expectations exceed $1.5 trillion both put orbital compute into a signed agreement, the question shifts from “is this serious” to “what is the timeline.”
The Constraint Rewriting Every AI Alliance
Anyone who has spent decades running enterprise IT infrastructure recognizes the pattern playing out across the AI industry right now. When a capacity crunch hits, you stop caring about vendor politics and start caring about who can deliver hardware this quarter. The entire AI industry reached that stage in 2026.
Anthropic’s total compute commitments now span Amazon, Google, Broadcom, and SpaceX. OpenAI’s supply chain stretches across Microsoft, Cerebras, Oracle, and its own $20 billion chip investments. Every frontier lab is chasing the same scarce resource through increasingly creative channels. The deals are getting larger, the counterparties are getting stranger, and the urgency is not slowing down.
When a founder calls your company “misanthropic” in February and rents you 220,000 GPUs in May, the compute shortage is not a talking point. It is the force rewriting the AI industry’s alliances in real time.