Colossus 2 adds 660,000 GPUs before its power plant is finished

Elon Musk put a build schedule on X on 25 September: another 220,000 Nvidia GB300 GPUs operational "next week," another 220,000 in November, and a final 220,000 by late December "if we get lucky." Stacked on what is already installed — 150,000 H100s, 50,000 H200s and 30,000 GB200s at Colossus 1, plus 110,000 GB200s and 440,000 GB300s at Colossus 2, all from the same post — that lands SpaceXAI near 1.44 million accelerators in operation by the end of the year, about 1.1 million of them Blackwell Ultra.

The arithmetic is Tom's Hardware's sum of Musk's figures, not a disclosed build plan, and every input traces to one person describing his own site. Musk has missed these deadlines before and built the hedge into the sentence: the third tranche arrives only with luck.

Half of that fleet is not a training fleet

The headline number is a count of installed chips. The number that decides anything is how many can train a frontier model, and those differ by a lot here.

Colossus 1 is a Hopper-plus-Blackwell mix, and by Musk's own account that combination is inefficient for training Grok — so the cluster was rented out to Anthropic for inference instead. The usable training fleet is Colossus 2: 550,000 GPUs today, 1.21 million if all three tranches land on schedule. That is still the largest concentration of training accelerators any single lab can point at one run, but the honest comparison against OpenAI's and Anthropic's distributed footprints is 1.21 million, not 1.44 million, and the rented-out half is a revenue line rather than a capability claim.

Whether install speed converts into model quality is a separate question, and the answer so far is no. In a China Media Group interview published this weekend, Musk said Grok 4.7 is "not as good as, say, Opus 5.5 that just got released," and offered compute scale as the mechanism for closing the gap rather than a result — the admission we covered in Musk: Grok 4.7 trails Opus 5.5, expects to catch up next year. A fleet this size is a necessary condition for his "pole position in about six months" claim, not evidence for it.

The new GPUs inherit 58 temporary turbines

The power side has a harder record than the GPU side, because regulators write it down.

SpaceXAI's own Memphis updates page says the site's permanent source is a 1.2-gigawatt plant running 41 turbines, authorized under a Clean Air Act permit granted in March 2026, replacing 69 temporary trailer-mounted gas turbines. Those turbines are the subject of a Clean Air Act suit brought by the Southern Environmental Law Center and Earthjustice on behalf of the NAACP, alleging they ran without required permits or pollution controls; the case is undecided, and the Justice Department has moved to intervene on the company's side, arguing the technology has been used in US military operations and that the litigation threatens national and energy security.

Under an agreed order with the Mississippi Department of Environmental Quality, removal of all 69 runs to July 2027. On 31 August the company said 11 had been powered down. Powered down is not removed — the obligation in the order is removal from the site — which leaves 58 temporary units running while the GPUs arrive, roughly ten months of overlap by the company's own schedule.

Now put the two halves together. An Nvidia GB300 NVL72 rack — 72 Blackwell Ultra GPUs — draws about 135 to 142 kilowatts, depending on whose configuration you read. Extrapolated across 1.1 million GB300s that is on the order of two gigawatts of rack power before cooling, against a permitted plant rated at 1.2 gigawatts, with the local grid making up the difference. Musk has never published the site's total draw, and the company describes the plant as coming online rather than complete. Nobody outside the company can tell you which megawatts those new GPUs will run on, and that is the gap worth watching.

The enforcement contrast is instructive. In New Jersey, a data center was fined $1.07 million over 62 unpermitted gas generators — the case we covered in New Jersey fines an AI data center $1.07M over 62 gas generators. In Southaven, an equipment category in the same class produced an agreed order in one state, a pending federal suit in another, and a federal intervention on the operator's side. The variable is not emissions; it is who shows up for the defendant.

Speed is the product

Every major lab now says power, not chips, is what limits AI buildout. Musk's advantage is not that he has more megawatts — it is that he treats permitting as a retrofit instead of a precondition. Colossus went from site work to 200,000 GPUs in about seven months, per the company's account; interconnection queues and air permits run in years. That is the entire mechanism, and it is not transferable to a lab that shares a grid and a regulator with the neighborhoods it sits in, which is nearly all of them — the dynamic behind the quarter in which local opposition blocked $68 billion of US data centers.

Read the GPU schedule as the second half of the same bet. Musk is not proving the permitting model is legal; he is installing hardware fast enough that the hardware becomes the reason to legalize it.

What to watch

Whether the November and December tranches land, and whether anyone independent ever counts them. Whether Mississippi environmental regulators publish a compliance record on the July 2027 removal milestones, since the company's announcements are not the order. And the only number that would settle the actual argument: a Grok release measured against Opus 5.5 on something other than a launch-stage comparison.

If a lab can install a gigawatt of compute faster than a regulator can permit it, whose clock should the rules run on? Tell us in the comments.

Sources: Elon Musk on X · Tom's Hardware — SpaceXAI to add another 660,000 AI GPUs this year · SpaceXAI — Memphis updates · The 901 Report — SpaceXAI says it will remove 69 turbines · Action News 5 — SpaceXAI shuts down 11 turbines at Southaven facility · Data Center Dynamics — GB300 NVL72 rack power