Anthropic takes control of SpaceX's Colossus supercomputer

Anthropic just took total control of every single chip inside SpaceX's Colossus One supercomputer in Memphis, securing hundreds of thousands of NVIDIA GPUs and hundreds of megawatts of power that were originally built to train Elon Musk's XAI models. Yeah, the size of this acquisition completely rewires the power dynamics of the entire tech sector. We are going to trace exactly how a pretty catastrophic miscalculation in compute demand forced a really bitter rivalry to turn into this sprawling infrastructure alliance. Right. And, you know, how this sudden influx of hardware is unlocking artificial intelligence that literally dreams between coding sessions. So how does a company go from being publicly labeled as evil by Elon Musk to renting his entire supercomputer? And what does this bizarre alliance actually mean for the physical limits of artificial intelligence? Well, the origin of this whole crisis really comes down to a forecasting error of just historic proportions. Oh, really? Yeah, Anthropic expected their growth to multiply by like a factor of 10. But in reality, they experienced an 80-fold surge in revenue and usage. Wait, 80? 80. Yeah. And that completely exhausted their available infrastructure, just pushing their servers to the absolute breaking point. Wait, back up. How does a major tech company underestimate their own growth by that wide of a margin? I mean, forecasting a 10-fold increase is already incredibly aggressive for any enterprise operation, right? Absolutely. Absolutely. So hitting 80 times your expected growth sounds like a fundamental misunderstanding of how your own customers are actually using your product. That's exactly what it was. It was a complete misunderstanding of how users were interacting with the product. We are looking at the shift from conversational artificial intelligence to what the industry calls agentic workflows. Okay, agentic. Right. Previously, you interacted with a chatbot in a very linear, predictable way. You type a question, the model processes your text, and it generates an answer. Right. That is a single transaction. It consumes a finite, easily measurable amount of computational power. You can build out a highly accurate financial model for a data center based on, like, millions of people asking single questions. Because if I just ask for a recipe for banana bread, or, you know, I ask the AI to debug a five-line Python script, the interaction is finished the moment the text appears on my screen. Exactly. The server can immediately allocate its resources to the next person waiting in line. Right. But users stopped treating these systems like search engines. They started deploying tools like Plod Code to run continuous autonomous loops inside their software repositories. Uh-huh. You are no longer asking for a single block of code. You are telling the agent, here is an entirely new feature I want in my application. Go read my existing code base. understand how the database is structured, write the new code for the feature, run the software tests. Wow, okay. And if those tests fail, read the error logs, rewrite the code, and test it again until it passes. So the software isn't just answering a prompt. It's acting like an employee trying to solve a puzzle over and over again without stopping. And that changes the math completely. An average developer is keeping these tools running constantly. The compute consumption of an agentic loop is just astronomical compared to a simple chat request. Because it never really stops. Exactly. And understand why it uses so much power. You have to look at how these models process information through what are called tokens. Right. Tokens. A token is essentially a fragment of a word. The word hamburger might be split into three tokens by the system's tokenizer. Okay. When you ask a question, the model reads the tokens in your prompt and then generates new tokens for the answer. And every single token requires the processors to perform thousands of mathematical operations just to predict what the next fragment of a word should be. Yes. Now, apply that math to an agentic loop. Every single time the agent writes a piece of code, runs a test, and fails, it has to try again. Yeah. But to try again, it has to reread the entire history of what it just did. Oh, I see. It has to ingest your entire code base, the original instructions, the code it just wrote, the air log from the failed test, and its own internal reasoning about why the test failed. That sounds massive. It is. It has to process all of those tokens again just to understand the context of the new attempt. It consumes millions of tokens in a single afternoon. So the original infrastructure was built for people asking single questions, and suddenly millions of users are running nonstop autonomous loops that consume computing power exponentially. Yep. You have a server farm expecting people to sip water from a fountain, and instead everyone showed up with industrial fire hoses. That's a great way to put it, and that caused the product experience to degrade severely. The infrastructure literally could not handle the concurrent load. Right. Because the system is trying to process billions of tokens simultaneously, paying users slammed into hard usage ceilings. The servers simply did not have the processing cores available to run the math. I mean, imagine you are a developer. You have integrated this tool completely into your daily workflow. You rely on it to hit a critical project deadline, and suddenly it just stops working. Yeah, it's a nightmare. You hit a button and get a spinning wheel, or worse, an error message telling you to come back in three hours. And it was happening constantly. Their accounts were actively throttled, particularly during peak business hours, when everyone on the East Coast and West Coast of the United States were working at the exact same time. Enterprise clients, you know, the ones paying premium corporate rates for guaranteed access, found their access restricted just when they needed it most. They were seeing timeout errors and drop sessions right in the middle of complex architectural tasks. Which creates an existential threat to the company's subscription model. If you are selling an autonomous digital co-worker, it cannot arbitrarily decide to go on strike at two in the afternoon because the server room is running hot. The reliability is the entire value proposition. If the tool is not dependable, enterprise clients will just abandon it, regardless of how intelligent the underlying model actually is. ANTHROPIC was facing a crisis where their own success was basically destroying their product experience. They needed computing power immediately, and they could not wait years to build a new data center. The only solution was to find an existing facility with the raw physical hardware already installed and ready to go. So to solve this crisis, Anthropics secured the Colossus One facility in Memphis. And this is a supercomputer built in a fraction of the usual construction time, and it is housed inside a former Electrolux appliance factory. Yeah, the physical reality of Colossus 1 is fascinating when you examine the hardware involved. It operates with hundreds of thousands of high-end NVIDIA accelerators. We are talking about H100s, H200s, and the absolute newest generation of chips designed specifically for artificial intelligence workloads. For anyone wondering why a standard computer processor can't do this job, it comes down to how the chips handle tasks. The CPU in your laptop has a few very powerful cores. It is designed to execute a few complex tasks very quickly in sequential order. Right. One after the other. But an NVIDIA GPU a graphics processing unit has thousands of smaller less powerful cores It is designed to do thousands of simple mathematical calculations simultaneously Training an AI model or running an agentic loop requires millions of simultaneous matrix multiplications. The GPUs are really the only hardware capable of that level of parallel processing. And the power density required to run hundreds of thousands of those specific accelerators in one location is just staggering. Right. An individual server rack fully loaded with these chips draws exponentially more electricity than a traditional cloud storage server rack. Hold on. Why use an abandoned appliance factory instead of building a proper data center from scratch? That's the crazy part. I mean, if you were deploying the most advanced, sensitive computing hardware on Earth, why put it in an old manufacturing plant with a leaky roof? Because of the utility grid bottleneck. Traditional data centers face interconnection queues that are becoming impossibly long. What do you mean by interconnection queues? Well, if you want to build a new hyperscale facility today, you have to ask the local utility provider for hundreds of megawatts of power. The utility cannot just flip a switch. They have to conduct massive environmental and engineering studies, physically upgrade high-voltage substations, and lay new transmission lines across miles of public and private land. Oh, wow. That queue can last well over a decade just to get the permits approved. So you simply cannot build fast enough if you rely on the standard electrical grid. You are trapped by the bureaucracy and the physical limitations of the local power company. Exactly. And if your software is crashing today, you don't have 10 years to wait for a substation. By retrofitting an industrial site, the builders bypass the grid interconnection delays almost entirely. An appliance factory already has significant industrial power infrastructure feeding into the property. But more importantly, the builders brought their own power. A what? Yeah, this site relies on dozens of massive natural gas burning turbines installed directly on the property. They essentially built their own private fossil fuel power plant just to run the servers independently of the Memphis grid. They did. And power generation is really only half the physical equation here. Right. The heat generated by hundreds of thousands of high-end accelerators operating at full capacity is immense. You cannot cool a dense deployment of these chips with traditional air conditioning. Just blowing air on them doesn't work. No. Moving cold air over the racks simply is not efficient enough to pull the thermal energy away from the silicon before the chips physically melt. You need liquid cooling. Which means pumping water directly into the server racks, circulating it over specialized cold plates attached directly to the processors, and piping the hot water away to dissipate the heat. Yes. Water is vastly superior to air when it comes to thermal conductivity. It absorbs heat much faster. Exactly. But that requires an incredible amount of water moving continuously through the system. That is why the site features a custom wastewater treatment plant. A whole treatment plant just for the servers. Yeah, they process millions of gallons of water just to handle the immense liquid cooling requirements of the facility. They created an entirely self-contained ecosystem of power generation and thermal management inside an old factory shell. And the consequence of plugging into this massive self-contained supercomputer was immediate for anthropic users. The rolling usage caps for pro subscribers basically doubled overnight. The artificial constraints simply vanished. All of the peak hour throttling that had plagued developers was entirely removed because the back-end processing power suddenly expanded to meet the demand. And for enterprise developers, the API input limits jumped by 1500%. Which is huge. Think about what that actually means for a software engineer. The context window, the short-term memory of the AI, the amount of text it can hold in its brain at one time, expanded so massively that they can process giant code bases in a single prompt. It alters the entire utility of the tool. Yeah. When you increase the input limit by that magnitude, you move from processing individual files to analyzing entire system architectures. Right. The agent can see your front-end user interface, your back-end database routing, and your security protocols all at the exact same time. It can spot an error in the database that is causing a visual glitch on the website because it has the processing power to hold the entire architecture in its active memory. So for the everyday developer, the frustrating usage limit reached screen essentially vanished overnight. Yeah. The tool changed from a novelty that might abandon you at a critical moment into a highly reliable co-worker. It transitioned from a brittle experiment into robust foundational infrastructure. But, you know, the acquisition of this specific facility reveals a very complex layer of corporate strategy, particularly when you look at the personalities involved here. Right, because Elon Musk previously attacked Anthropic publicly. Very publicly. He called them misanthropic and heavily criticized their approach to safety and alignment, suggesting they were designing models to be politically biased or restrictive. Yeah. But he personally approved this massive lease because, in his words, no one set off my evil detector during their meetings. The change in rhetoric is striking. But honestly, it is driven entirely by cold financial logic. And the structural change is happening within Musk's own organizations. You have to look at the larger corporate board. Okay. XAI, his artificial intelligence company, merged with SpaceX to form a new combined entity called SpaceX AI. And they are preparing for an upcoming public offering. And Wall Street analysts are valuing that combined entity at well over a trillion dollars. When you are taking a company public at a trillion dollar valuation, you need a flawless equity narrative to present to institutional investors. Right. Colossus One was originally built to train XAI's internal models, like Grok. But having a massive natural gas burning supercomputer that only serves your internal research division is a massive cost center. It's just burning money. It burns cash constantly just to keep the servers running and the cooling water flowing. But renting Colossus One out to a paying, well-funded enterprise customer like Anthropic changes the financial profile completely. It transforms a massive capital expenditure into a highly lucrative operational revenue stream. Exactly. It proved to investors that SpaceX AI is not just a hardware manufacturer with an artificial intelligence research wing, but a highly profitable cloud infrastructure provider. Oh, that makes sense. It secures a stable, massive revenue source for the upcoming roadshow. And it does not actually hinder Musk's own artificial intelligence development, right? His internal projects like the next generation of Grok, the computer vision models required for autonomous driving, and the neural networks powering his robotics initiatives have already migrated to the even larger Colossus 2 facility. Right. They built a second, exponentially larger supercomputer, strictly for their own internal use, which freed up the first facility to be monetized on the open market. Wow. It's a brilliant piece of balance sheet management. But, you know, the geopolitical angle surrounding this infrastructure deal is just as critical to understand. Right. This brings in the United States government dynamic. The Pentagon recently blacklisted Anthropic as a supply chain risk. That blacklisting was a major event in the tech sector The Pentagon designated them as a supply chain risk specifically because Anthropic refused to remove safety restrictions for military operations Wait really Yeah. The military wanted to integrate the models into certain logistical and analytical applications, and Anthropic's leadership decided those applications violated their internal usage policies regarding potential harm and autonomous decision-making in combat scenarios. So they walked away from massive, highly lucrative defense contracts to maintain their internal safety standards. And as a result, the Department of Defense effectively banned them from certain federal integrations. They did. And counterintuitively, it created a highly unique and valuable market position for them. How so? Well, if you look at the other major cloud providers, the companies that typically rent out the computing power needed to run these massive models, they are deeply entangled with defense contracts. Oh, wow. They provide the foundational cloud infrastructure for global military logistics, intelligence gathering, and drone operations. But Anthropic is now relying exclusively on SpaceX hardware for their primary computing needs. And this provides them with a form of sovereign compute. Because they are running on an independent infrastructure stack, completely separated from the traditional massive cloud providers who are deeply embedded with the Pentagon, they offer a very distinct alternative to the corporate world. This appeals heavily to enterprise clients who want to completely distance their workflows from any military applications. Exactly. Imagine you are the chief compliance officer of a massive European financial institution, a global health care conglomerate, or a multinational pharmaceutical corporation operating in regions with strict data sovereignty and privacy laws. Those executives might be extremely hesitant to run their highly proprietary regulated corporate data through a cloud infrastructure system that is intimately linked with United States military operations simply due to the risk of espionage, regulatory blowback or public relations disasters. Right. Anthropic's isolated non-military compute environment opens up entirely new corporate markets for them globally. So having this limitless hardware finally secured gave Anthropic the physical freedom to explore software features they simply could not afford to run when they were strapped for compute. Yes, and this is where it gets really interesting. They launched a feature called Dreaming, which allows AI agents to review their own past sessions and rewrite their memory stores while the system is idle. This represents a fundamental shift in how artificial intelligence retains and processes information. Dreaming operates as an asynchronous background job. When the agent is not actively working on a coding task or answering a prompt for a user, the servers do not just sit quiet. Right. The agent initiates this reflection process on its own. Wait, I need clarification on how this is actually different from an AI just remembering what you told it. If I tell a standard chatbot, my name is John, it remembers my name for the rest of the conversation. Sure. How is dreaming functionally different from standard memory? We have to distinguish between in-session memory and cross-session consolidation. A standard artificial intelligence is stateless across sessions. Meaning what? It starts every single interaction as a completely blank slate, relying only on its original frozen training data. It forgets its repeated mistakes every time you start a new task. So if I open a new chat window and ask the agent to format a SQL database and it makes a syntax error, I correct the error and it finishes the job. Yep. But tomorrow, when I open a brand new window and ask it to format a similar database, it will make the exact same syntax error all over again because the new window has no idea the previous window ever existed. Exactly. Because it has no cumulative operational experience. Yeah. But a dreaming AI behaves entirely differently. It curates its own persistent memory file. As you interact with the agent across weeks and months, the system writes continuous notes to a persistent file about your preferences, your coding environment, and the mistakes it has made. But if you just let an AI write infinite sticky notes to itself, eventually the desk is completely covered and the system slows down because it has to read every single sticky note before it takes an action. That's the danger, yeah. The agent might learn a highly specific workaround for a software bug on Monday, write a note about it, but by Friday, that bug is patched by the vendor and the workaround is completely obsolete. And the persistent memory file becomes bloated and highly contradictory. The agent ends up wasting massive amounts of your computing power reading outdated instructions and conflicting rules, which severely degrades its performance and reasoning abilities. Right. That degradation is a phenomenon called context rot. Dreaming is the cure for context rot. So how does the asynchronous job actually clean up the sticky notes? During the idle time, the agent actively evaluates all of its stored notes against the transcripts of its most recent experiences. It scans dozens of past sessions. Oh, wow. It recognizes the pattern of its own failures. It independently decides to prune the obsolete workarounds, merge duplicate instructions, and synthesize a clean, highly optimized memory store. That's incredible. And it executes its entire cleanup operation without a human ever explicitly telling it what to remember or what to forget. Anthropic also introduced a parallel feature alongside Dreaming, where up to 20 specialized agents work simultaneously on a single project. And all 20 of those agents are feeding their individual experiences back into this shared Dreaming memory. This concept is called multi-agent orchestration. You have a lead agent acting as a primary project manager, and it autonomously delegates tasks to highly specialized sub-agents. How does that work in practice? Well, you might have one agent tasked exclusively with searching air logs, a second agent tasked entirely with reading external API documentation, and a third agent writing the actual Python code. All at the same time. They operate in parallel, simultaneously, which drastically reduces the physical time it takes to complete a highly complex software architecture project. The pricing structure for this orchestration is highly unique. Users pay a few cents per active session hour, plus the standard token cost for the input and output. You are effectively paying for the infrastructure time, the active session hour, to keep the autonomous loop open and running on the server, in addition to the raw intelligence, which is the tokens. Right. It reflects the new physical reality that hosting an active looping agent requires dedicated, sustained server resources, rather than just a momentary blip of processing power. This alters how Soshware is built from the ground up. Agents no longer suffer from context rot, where their instructions become bloated and confused. No, they don't. They transform into self-improving entities that accumulate genuine operational experience over countless tasks. A development team working with a dreaming agent for six months will have a fundamentally different tool than a team that just started using the exact same model today. Because it's learned their habits. The agent has actively adapted to their specific coding standards, learned their preferred architectural patterns, and deeply internalized all the strange quirks of their legacy systems. It evolves the artificial intelligence from a generic digital assistant into a highly specialized customized colleague But even with Colossus 1 the physical demands of these dreaming looping agents are still growing exponentially The electrical grid on Earth simply cannot sustain this trajectory. It really can't. This leads to the ultimate escape valve. The infrastructure agreement includes a provision for Anthropic and SpaceX to jointly explore building gigawatt-scale data centers in Earth's orbit. The appeal of placing supercomputers in space is driven by the absolute exhaustion of terrestrial resources. The electrical grid on Earth is entirely tapped out. We discussed the decade-long interconnection cues just to build a substation. Furthermore, local communities are actively and aggressively protesting the environmental impact of massive data centers. Right. People do not want dozens of natural gas turbines running constantly next to their suburban neighborhoods. And they absolutely do not want millions of gallons of their local municipal water supply evaporating into the atmosphere just to cool server racks for a tech company. Exactly. Orbit offers an environment entirely free from those terrestrial constraints. There are zero zoning laws in low Earth orbit. There are no local community boards denying building permits or water usage rights. Right. And most importantly, Orbit offers continuous, unimpeded solar energy. A solar panel positioned in space can generate significantly more power than the exact same panel on Earth because there is no atmosphere, no clouds to filter the light, and depending on the specific orbital path, there is no night cycle. You have access to a permanent gigawatt-level power source. The energy generation is flawless, but the concept hits a massive physics wall the moment you try to manage the heat generated by the servers. It really does. This is a fundamental flaw in the space is cold argument that most people naturally assume. People assume that because the void of space is freezing, cooling a massive data center in orbit would be incredibly easy. You just open a vent and let the cold in. That's what everyone thinks. But the vacuum of space is actually a near perfect insulator because there is absolutely no air to carry the heat away. On Earth, we rely heavily on convection and conduction to cool our electronics. You blow cold air over a hot surface, the air physically absorbs the heat energy, and the fan moves that hot air away from the device. Or, you pump cold water over a hot surface, the water absorbs the heat, and you pipe it away to evaporate in a cooling tower. Convection requires a physical medium-like air or water to transfer the thermal energy from the hot object to the surrounding environment. Think of space not as a freezer, but as a giant vacuum-sealed thermos. If you put boiling hot coffee inside a high-quality thermos, it stays hot for hours because the vacuum layer prevents the heat from escaping. Perfect analogy. There is no air inside the vacuum to pull the heat away from the liquid. A data center in orbit is trapped inside a cosmic thermos. If you put a hot computer chip in the vacuum of space and run a calculation, it will simply bake itself until the silicon physically melts because the thermal energy has nowhere to go. The only physical mechanism available to remove heat in a pure vacuum is through thermal radiation. You have to literally radiate the heat away as infrared light. Which is really hard. Very hard. This entire process is governed by the Stefan-Boltzmann law, a principle of physics which details exactly how energy radiates from a black body in direct relation to its temperature and its surface area. And radiating heat as infrared light is incredibly inefficient compared to terrestrial liquid cooling. To dissipate just a small fraction of the heat generated by these advanced AI chips, a single satellite requires a radiator surface area roughly equal to four professional tennis courts. You cannot just attach a small metallic fin to the side of the satellite and expect the servers to survive. The radiators have to be massive to provide enough physical surface area for the infrared light to escape into the void. So the hotter the chip, the better it radiates. The hotter the chip is allowed to get, the more efficiently it radiates heat away. But these highly sensitive silicon chips have strict thermal limits. If they reach a certain temperature, the logic gates fail and the chip is destroyed. So you have to keep them somewhat cool. So you have to maintain a relatively low operating temperature, which mathematically dictates that the radiators have to be exponentially larger. Scaling this mathematical reality to the gigawatt-level data centers, they're actually proposing means deploying square kilometers of cooling fins into orbit. We are talking about delicate metallic structures the size of small cities floating in space, existing solely to act as heat sinks for the processors. The mechanical engineering complexity required to launch, unfold, deploy, and maintain square kilometers of fragile cooling fins in an environment filled with micrometeoroids and dangerous orbital debris is staggering. Just one little piece of debris could ruin it. A single piece of space junk traveling at orbital velocity could sever a coolant line. You have to pump liquid Kuhlman through miles of delicate piping exposed to the harsh radiation and vacuum of space, constantly circulating the fluid back to the server core. This physical constraint severely limits what orbital computing can actually achieve in practice. It forces a complete shift in strategy. It really does. Space-based data centers will likely be restricted to specific inference tasks rather than replacing the massive, tightly coupled terrestrial clusters needed for core AI training. Training a frontier artificial intelligence model from scratch requires hundreds of thousands of chips communicating with each other instantaneously, sharing data across fiber optic cables with zero latency. And you can't do that in space. The latency requirements and the thermal management necessary for a months-long training run are virtually impossible to achieve in an orbital environment with our current materials science and launch capabilities. But inference, which means using an already fully trained model to analyze new data, is much more feasible. If an observation satellite is taking high-resolution, multispectral photos of the Earth's surface, instead of beaming terabytes of raw image data down to a terrestrial ground station, which consumes massive amounts of bandwidth, the Orbital Data Center can analyze the images directly in space. Exactly. The model can identify the relevant information, like a specific crop yield or a fleet of ships, and only beam down a few megabytes of text summarizing the findings. That solves a massive bandwidth bottleneck for satellite communications. The Orbital Data Center effectively becomes an edge computing node, processing the heavy data directly where it is generated. It is a highly specialized, valuable application for military and commercial intelligence. But it absolutely does not solve the fundamental terrestrial power crisis for the companies trying to train the next massive generation of language models. The insatiable demand for autonomous coding tools has pushed terrestrial power grids to their absolute limits. This reality is forcing fierce competitors to share hardware and driving engineers to look to the vacuum of space just to keep the servers running. If our software requires supercomputers the size of appliance factories just to function today, how much of the Earth's surface will we need to dedicate to cooling fans tomorrow? If you're not subscribed yet, take a second and hit follow on whatever app you're using. It helps us keep making this. We appreciate you being here. Also, check out our YouTube channel for more business and tech updates. There's a link in the description.