In early May, Anthropic struck a deal with SpaceX to use the full computing capacity of Colossus 1, its data center in Memphis, Tennessee. The AI group has also unveiled plans that seem to blur the line between reality and science fiction.
The capacity offered through SpaceXAI amounts to about 300 megawatts of computing power. It should accelerate the development of large Claude Opus language models and shorten waiting times for subscribers.
Anthropic, whose Pentagon work came under scrutiny earlier this year over military AI safeguards, unveiled a new feature in its models, “dreaming”, during a developers’ day in San Francisco, California.
How Claude Learns to Dream
Boris Cherny, who leads the Claude Code project, said the shift meant users would no longer have to prompt the coding tool step by step. Instead, Claude would increasingly direct Claude Code itself.
“The default isn’t, ‘I’m going to prompt Claude Code.’ The default is now, ‘I will have Claude prompt Claude Code’”, he said.
A guide to Claude Code’s Auto Dream feature describes it as a tool that consolidates memory files, prunes stale notes and merges insights. The process is “like REM sleep for your AI agent”, the guide says.
In general, the phase known as rapid eye movement (REM) sleep is associated with vivid dreams and plays an important role in memory consolidation. During REM sleep, the brain reviews recently learned information and helps decide what should be stored and what can be discarded.
“The naming is deliberate, and the metaphor is surprisingly accurate,” the guide notes. According to the guide, Auto Dream reviews what Auto Memory has collected, strengthens what remains relevant, removes outdated material and reorganizes the rest into clean topic files.
The guide compares Auto Memory to Claude’s daytime brain, in which it takes notes as it works: debugging patterns, build commands, architecture decisions and user preferences. Without a consolidation step, the notes accumulate like unconsolidated short-term memories. Contradictions persist, outdated entries remain and the signal-to-noise ratio deteriorates with every session.
The official material, however, makes no claim that the feature shows any form of self-awareness or consciousness. It describes a technical process for reorganizing stored information, not human-like dreaming.
Back to Space
On the other side of the equation is SpaceXAI, the Musk-linked venture that has signed an agreement with Anthropic to provide access to Colossus 1, one of the world’s largest AI supercomputers. Anthropic plans to use the additional computing power to improve capacity for Claude Pro and Claude Max subscribers.
As part of the agreement, Anthropic has also expressed interest in partnering to develop multiple gigawatts of orbital AI computing capacity.
“The compute required to train and operate the next generation of these systems is outpacing what terrestrial power, land and cooling can deliver on the timelines that matter”, xAI said.
SpaceX is also set to be joined by another big tech company. Alphabet, the owner of Google, has expressed interest in orbital data centers. For the company run by Sundar Pichai, the move is also pragmatic, since Google already owns about 6% of SpaceX.
Google has also turned to other companies that launch cargo rockets. According to Wall Street Journal sources, Jeff Bezos’ Blue Origin is also under consideration. Google is still negotiating these deals, the anonymous officials acknowledged.
Last year, Google announced its own plans to launch prototype satellites by 2027 as part of an ambitious initiative called Project Suncatcher. It is working with Planet Labs to develop satellites that collect energy from the sun, the Journal recalled.
“We will send small racks of devices, put them in satellites, test them and then start expanding their use”, Alphabet CEO Sundar Pichai told Fox News in a November interview.
“I have no doubt that ten years from now we’ll see this as a pretty common way of building data centers”, he declared.
Will Robots Dream in China?
AI models could theoretically “learn to dream” from human beings by analyzing brain activity in real time. Musk’s Neuralink is one possible route into that field, but China has also made brain-computer interfaces (BCI) a national priority.
One sign of Beijing’s ambitions is Charles Lieber, the 67-year-old nanotechnology pioneer and former Harvard professor who moved to Shenzhen in April 2025, according to Reuters. Lieber now works with the state-backed Shenzhen Medical Academy of Research and Translation (SMART), where he has rebuilt his brain-computer research project, known as i-BRAIN.
Lieber’s move is politically sensitive. In 2023, he was sentenced to six months of house arrest over undisclosed payments from Beijing while he was at Harvard University. He had previously joined China’s Thousand Talents Program, which the US Department of Justice has described as a vehicle for intellectual property theft.
The Chinese i-BRAIN project also includes a primate research facility with 2,000 cages, while Lieber’s team is seeking researchers for electrophysiological studies on macaques. The animals are widely used in China’s scientific sector, including in studies on reversing aging.
The Chinese military has also begun examining possible military uses for BCI technology, including the creation of a kind of “super soldier” whose brain chip could improve mental alertness and reduce reaction times in combat.