Musk's GPU Blitz: 660,000 Nvidia GB300s by December

AI INFRASTRUCTURE

Elon Musk says 220,000 more Nvidia GB300 GPUs go live at his Colossus complex "next week," with two more 220,000-chip waves by December — a buildout that would push his Memphis-area fleet toward 1.44 million AI accelerators and an estimated $58 billion in capital cost.

Published September 28, 2026 · By Daniel Rao

Elon Musk speaking at an event in National Harbor, Maryland, in February 2025. His September 25 post on X laid out the Colossus GPU counts and a three-wave GB300 rollout running through December. Photo: Gage Skidmore, via Wikimedia Commons (CC BY-SA 3.0)

Elon Musk just published the clearest snapshot anyone has ever given of the largest private AI supercomputer on Earth — and the numbers are difficult to process. In a September 25 post on X, Musk disclosed the exact GPU counts inside his two Colossus clusters near Memphis, then laid out a deployment schedule that would add as many as 660,000 Nvidia GB300 chips before the year ends. If all three waves land, the combined fleet approaches 1.44 million accelerators. Independent modelers at Epoch AI put the full buildout's capital cost at roughly $58 billion.

This is not a press release. It was a reply to a user asking about cluster size. But in the AI infrastructure race — where the scarcest commodity is no longer the chip but the power to run it — Musk's casual Friday post was the most consequential hardware announcement of the month.

Colossus GPU count: the numbers, straight from Musk's post

Musk's post, replying to the X user @minchoi, broke the fleet down by site and chip generation. The wording, quoted verbatim by multiple outlets:

"Colossus 1 is 150k H100, 50k H200 and 30k GB200. Colossus 2 is 110k GB200 and 440k GB300. Another 220k GB300 will be fully operational next week and another 220k in November. If we get lucky, yet another 220k GB300 by late December."

Colossus 1, the original Memphis cluster built in a converted Electrolux appliance plant, runs 230,000 chips across two Nvidia generations: the older Hopper-based H100 and H200 cards plus 30,000 of the newer Blackwell GB200s. Colossus 2, the newer and far larger site, runs 550,000 Blackwell-generation chips — 110,000 GB200s and 440,000 GB300s. The GB300 is an advanced variant of the GB200, carrying the same Grace CPU but a newer GPU die that Nvidia calls Blackwell Ultra, according to Wccftech's technical coverage.

Note the dates carefully. The post went up Friday, September 25 — so "next week" means this week, the week of September 28. The November wave is a stated schedule. The December wave is explicitly conditional: "if we get lucky." Musk's own qualifier is the most important word in the post.

There is a revealing engineering detail buried in the counts. The repeated 110,000-chip increments are not arbitrary: they reflect, Musk said, "the number of fiber optic cables that can be plugged into a central switch." The cluster's unit of expansion is literally the switch — networking, not silicon, sets the rhythm of the rollout.

1.44 million chips: the math and the money

Add Musk's numbers up and the scale becomes concrete. Today's fleet: 230,000 at Colossus 1 plus 550,000 at Colossus 2 — 780,000 accelerators already running. The three planned waves add 660,000 GB300s. If every wave lands by year-end: approximately 1.44 million GPUs across the Memphis complex, with Colossus 2 alone holding roughly 1.21 million Blackwell chips.

One caution on the arithmetic. Independent tracking does not fully agree with Musk's counts. Epoch AI, which monitors the campus from satellite and drone imagery, estimated on September 24 — the day before Musk's post — that Colossus 2 held 440,000 Nvidia GPUs: 110,000 B200s and 330,000 B300s, drawing 946 megawatts of IT power at a capital cost of $35.8 billion. The total matches Musk's 550,000; the mix differs, with Epoch attributing 110,000 fewer chips to the newest GB300 generation. As RuntimeWire noted, Epoch's figures are outside estimates built from imagery and disclosures, "not a chip count supplied by Musk or SpaceXAI" — but neither is Musk's post independently audited. The honest reading: the total is in the same ballpark; the exact generational split is contested.

On Epoch's modeling, the full buildout — all three waves installed and operational — lands at approximately $58 billion in capital cost, with the site projected to reach about 1.825 million H100-equivalent units of compute, per Crypto Briefing's reporting of the Epoch projections. For context, Tesla investor Sawyer Merritt estimated the chips across both sites represent more than $100 billion in GPUs and over two gigawatts of compute — a back-of-the-envelope figure, but directionally consistent with Epoch's. Either way, this is the most expensive single-site compute buildout in history, and it is not close.

"Pole position in about 6 months"

The GPU post was part of a broader set of comments in which Musk made his most aggressive capability claim yet. "Our AI efforts are only 3 years old, versus six and 10 years old for Anthropic and OpenAI," he wrote on X, in remarks verified by Barron's. "If our second derivative remains strong, SpaceX will reach pole position in about 6 months." He added: "Hardware is hard. Bringing massive compute online rapidly is incredibly difficult. SpaceX has demonstrated exceptional ability in this regard and will only get better."

The model roadmap behind the boast is equally ambitious. Musk said he is "cautiously optimistic" that SpaceXAI will release an AI model on par with Anthropic's Fable or OpenAI's GPT-6 within two to three months, according to Analytics India Magazine and Asianet News. He has sketched a rapid Grok release cadence — Grok 4.8, a 2.5-trillion-parameter model, expected to finish initial training imminently; Grok 4.9 positioned in the class of OpenAI's Astra and Anthropic's Fable; Grok 5, in his words, "better than anything." Markets noticed: shares of SpaceX (SPCX) climbed 0.4% in premarket trading on Friday after the posts, though the stock remains on track to snap a four-week winning streak.

There is an irony here worth naming. Earlier this month, Musk publicly endorsed Anthropic CEO Dario Amodei's call for the industry to slow frontier AI development, posting "Dario is right" — even as his own company races to install more than a million accelerators. Slowing model development and pouring data-center concrete can technically coexist. But the man predicting a safety slowdown and the man ordering 660,000 GPUs by Christmas are the same person, and the tension between those two Musks is one of the defining contradictions of the AI race.

The power problem is the real story

Chips, it turns out, are the easy part. The binding constraint on Colossus is electricity — and the numbers are as extreme as the GPU counts. Colossus 2 currently relies in part on temporary gas turbines parked outside the facility, a stopgap that has drawn regulatory scrutiny over emissions, Crypto Briefing reported. The permanent answer is a 1.2-gigawatt power plant in Southaven, Mississippi, expected to reach full operational capacity by mid-2027.

The broader Memphis complex — Colossus 1, Colossus 2, and future buildings internally referred to as "Minihard" — is targeting aggregate power capacity of roughly 2 gigawatts. Barron's, citing company figures, reported that SpaceX ended the second quarter with about 1.4 gigawatts of AI computing capacity, up 250% year over year, with end-2026 capacity targeted at about 2 gigawatts and ambitions of up to 10 gigawatts by the end of 2027. Morgan Stanley analyst Stephen Byrd wrote recently that SpaceX "has proven capabilities in scaling [electricity-generating] solutions faster and cheaper than its competitors" — a judgment the 122-day construction of the original Colossus supports.

This is the dimension of the race that matters most for ordinary Americans. Data centers at gigawatt scale consume electricity on a level utilities were never designed for, and the costs are already spilling into household bills. The infrastructure Musk is building to win the AI race is the same infrastructure reshaping the American power grid — and the political fight over who pays for it has barely begun. Musk himself made the economic stakes explicit earlier this month, predicting on September 18 that AI would roughly double US GDP growth next year. The GPU blitz is the physical bet behind that forecast.

Who else is running on Colossus

The Memphis complex is not just Musk's training ground — it is becoming a commercial compute utility, if single-outlet reporting holds up. A Crypto Briefing report citing SpaceX's IPO disclosures says a May 2026 agreement gives Anthropic access to compute capacity across Colossus and Colossus 2 at $1.25 billion per month through May 2029, and describes a Google cloud-services agreement under which payments rise to $920 million a month starting in October, with Google reportedly utilizing around 110,000 GPUs on the campus. Neither figure has been corroborated by a second outlet, so treat these as reported, not confirmed.

Consider what that means. Two of Musk's fiercest competitors in the model race — Anthropic and Google — are paying his company billions to run on his hardware. The AI race is simultaneously a war and a landlord-tenant relationship, and Musk owns the building. That commercial leverage is part of why the December wave matters: every additional 220,000 GB300s is not just training capacity for Grok, but rentable compute in the tightest hardware market in history.

What could go wrong

Separate the claims from the caveats. The 220,000 chips "next week" and the November wave are Musk's stated schedule; the December wave is an explicit hope. None of it is independently verified, and as RuntimeWire observed, a facility can have equipment on order or in transit without that capacity being online and available for training. Power, networking, and cooling must arrive together — a 660,000-GPU deployment is a systems-integration project at a scale no one has attempted, and the Southaven plant that ultimately powers it will not be fully operational until mid-2027.

Then there is the question the hardware cannot answer: what the compute is for. Musk's "pole position" claim assumes that more chips convert into better models on a six-month timetable — that the second derivative of capability stays strong as the second derivative of spending goes vertical. The history of AI scaling suggests compute matters enormously. It also suggests that data, algorithms, and talent are not purchased by the megawatt. Musk is making the largest bet in the industry's history that the first factor dominates. By December, with 1.44 million chips or something short of it, we will start to learn whether he is right.

Sources: Elon Musk's September 25, 2026 post on X (via reporting by Barron's, Wccftech, Seeking Alpha via TradingView, Benzinga via StockTwits/Asianet News); Epoch AI projections via Crypto Briefing and RuntimeWire; Analytics India Magazine; Interesting Engineering; Morgan Stanley (via Barron's); TechPulse USA's own reporting.

Related: Elon Musk Says AI Will Double US GDP Growth to 4% Next Year — Economists Are Not Convinced.