There’s a building in Southaven, Mississippi that draws as much electricity as a mid-sized city.
It’s xAI’s Colossus 2 — Elon Musk’s second AI supercomputer cluster — and it’s currently the largest computational machine ever assembled by a private company. Five hundred and fifty thousand GPUs. One gigawatt of power. On-site gas turbines. Tesla Megapacks for battery backup. Water-recycling cooling. The whole thing went live around mid-to-late January and has been training Grok models at a scale that didn’t exist on Earth a year ago.
And as of this week, Anthropic — the makers of Claude, xAI’s most direct competitor in the frontier-AI race — just signed a deal to rent xAI’s first supercomputer (Colossus 1) for their own training and inference workloads.
Let me say that again, because the implications are doing a lot of work in that sentence.
Anthropic is now renting compute from Elon Musk’s company.
If you use Claude through FIT, or you’re one of our nonprofit clients running Claude for grant writing and content creation, you’ve already felt the result: usage limits on Claude Code, the API, and the consumer apps quietly expanded over the last few days. That extra headroom you didn’t ask for? It’s because Anthropic just bought access to a couple hundred thousand GPUs in Mississippi.
This is one of the strangest and most telling deals in tech history, and it tells us something important about where AI is heading.
What Colossus 2 Actually Is
Numbers first, then translation:
- ~550,000 GPUs, primarily NVIDIA’s GB200 and GB300 series — the latest and most expensive AI training chips on the market
- ~1 GW of power draw — that’s a gigawatt, the output of a small nuclear plant
- Located in the Memphis-area (mostly Southaven, MS), built at record-setting speed
- Roadmap toward 1.5 GW in the next phase, with xAI’s stated goal of one million GPUs total across their cluster network
For context: Colossus 1, the previous generation, has roughly 200,000 GPUs and runs at meaningfully lower power. Colossus 2 isn’t an upgrade. It’s a category change.
It’s been described — accurately — as the world’s first gigawatt-scale AI training cluster. There has never been one before. Not Google’s. Not Microsoft’s. Not Meta’s. xAI built it from a standing start in roughly two years, which is the kind of timeline that makes infrastructure people physically uncomfortable.
The power for it doesn’t come from the grid alone. It can’t — the grid couldn’t deliver it on the timeline xAI needed. They put their own gas turbines on-site. They installed Tesla Megapacks for storage. They have water-recycling plans on hold until the rest of the system locks in. They are building a small nation’s worth of energy infrastructure to feed one building of computers.
This is what “the AI buildout” looks like up close.
The Deal: SpaceX Leasing Colossus 1 to Anthropic
With Colossus 2 fully online and absorbing xAI’s primary training workloads, Colossus 1 became, in a real sense, spare capacity. Two hundred thousand GPUs is too valuable to leave idle even for a few months. So xAI/SpaceX did the pragmatic thing: they leased it.
The customer? Anthropic. The same Anthropic that competes directly with xAI’s Grok in every benchmark, every customer pitch, every developer conversation.
A few details that have been reported:
- The deal is structured as compute rental, not a partnership or product integration. There’s no philosophy merge, no joint roadmap, no co-branded product. Just hardware time.
- Elon Musk reportedly met with the Anthropic team personally and gave them a positive read — competent, no major red flags, business-as-usual.
- This sits alongside Anthropic’s existing infrastructure deals with Amazon and Google. They’re not exclusive to anyone. They’re buying every GPU-hour they can get their hands on, from anyone willing to sell.
If the framing feels strange, it should. Imagine if Coca-Cola rented bottling capacity from Pepsi during a peak season. That’s the scale of competitive weirdness here. But it’s also the entirely rational outcome of an industry where compute is the bottleneck and capital is unlimited.
Why This Is Actually About Power, Not Chips
The thing that doesn’t get said enough about modern AI is that the binding constraint isn’t GPUs — it’s electricity.
NVIDIA can ship more chips. Capital is available — there’s a tidal wave of money flowing into AI infrastructure right now, more than any sector has absorbed in modern history. What’s actually scarce is the ability to physically deliver one gigawatt of clean power to a single building, reliably, by next Tuesday.
The reason xAI is winning the build-speed race isn’t that they’re smarter. It’s that they’re willing to put gas turbines on-site, eat the regulatory friction, and build their own grid. Most other companies are playing by traditional utility rules. xAI is essentially building their own utility.
And the reason Anthropic just signed up to rent from xAI is that even with deals from Amazon and Google, they can’t get enough power-backed compute to keep up with demand. The bottleneck has moved from “do we have a model” to “can we run the model at the scale our customers need.” The world’s most sophisticated AI lab just admitted, in business terms, that they need to borrow another lab’s spare capacity to keep their products running.
That’s the story.
What This Means for Small Business
Here’s the part that matters if you run a small business in Waseca, Montgomery, Le Sueur, or anywhere else where the AI infrastructure conversation feels distant.
One. The capability ceiling for AI tools you use every day — Claude, ChatGPT, Gemini, the rest — is going to keep going up for the foreseeable future. The companies powering them have access to compute scales we couldn’t imagine eighteen months ago. The free or cheap version of these tools will keep getting more capable, faster, with longer memory and bigger context windows. If you’ve felt your AI tools “leveled up” recently, you’re not imagining it. The infrastructure under them is changing in real time.
Two. The companies building this infrastructure are no longer treating each other as zero-sum competitors. They are treating each other as suppliers and customers depending on the day. This means the AI industry is consolidating into a small number of compute providers (xAI, Microsoft, Google, Amazon, Meta) selling capacity to a small number of model labs (Anthropic, OpenAI, xAI, Google DeepMind, Meta). Some companies are on both sides of that equation. The boundaries are blurring.
Three. The advantage of being early is widening, not narrowing. Every month that small businesses wait to integrate AI is a month their competitors who already integrated are training their own internal processes on a more capable substrate. The cost of waiting compounds.
Four — and this is the FIT-specific one. The reason we’re aggressive about getting our clients onto Claude, M365 Copilot, and the various AI tools we recommend isn’t because we love new shiny tech. It’s because we can see this curve from where we sit. The infrastructure is being built right now, today, in Mississippi, in Iowa, in Northern Virginia. The capabilities are going to keep arriving on a timeline that gets shorter, not longer. Adopting early means your business is built on top of a substrate that is appreciating, not depreciating. The companies that figure this out in 2026 will look very different from the ones that figure it out in 2028.
The Bigger Picture
The Anthropic-SpaceX deal is the first major public sign that the AI industry has crossed a threshold: the labs that compete against each other in the press release also need to cooperate on infrastructure to function. The compute crunch is so severe that ideological differences, brand loyalty, even direct competitive rivalry — all of it bends to the gravitational pull of “we need GPUs running by next quarter or we lose ground.”
There will be more of these deals. Probably soon.
And the Colossus 2 cluster — that gigawatt machine running quietly in Mississippi right now — is going to be looked back on in a few years as a turning point. Not because of what it computes. Because of what its existence proves: that private capital can build infrastructure at this scale, this fast, when the incentive structure aligns. The same playbook will be run again. Bigger buildings. More turbines. More chips.
The machines aren’t coming. They’re already here, drawing a gigawatt of power and running twenty-four hours a day.
The only question for the rest of us is what we do with the capabilities they’re shipping out into the world.
Need help figuring out where your business sits on the AI adoption curve, or what tools actually deserve a place in your day-to-day operations? That’s what we do. Schedule a conversation.
— Matt Stoltz, Flower Insider Technologies
Sources: Reporting on Colossus 2 capacity, power draw, and operational status as of May 2026 from xAI’s public Colossus page and contemporaneous coverage. Anthropic-SpaceX compute rental announcement public as of May 6, 2026. Intel synthesis assisted by Ringmaster Blu (Grok / SuperGrok) reporting pipeline.