xAI says Colossus 2 could hit 1.21 million Nvidia chips by year-end -- more than double what's running today

xAI's Colossus 2 could reach 1.21 million Nvidia GPUs by the end of the year -- more than double the roughly 550,000 it's running today -- according to the most detailed deployment timeline Elon Musk has given for the Memphis-area cluster. Replying to a question on X on Sept. 25, Musk broke down both sites: Colossus 1 holds 150,000 H100s, 50,000 H200s and 30,000 GB200s; Colossus 2 currently runs 110,000 GB200s and 440,000 GB300s. Three more waves of 220,000 GB300s each are due next week, in November, and -- his words -- "if we get lucky," by late December.

GB300 is Nvidia's current top-tier Blackwell-generation training chip, the successor to the GB200s that made up Colossus 2's first buildout; both are sold as part of full rack-scale systems rather than standalone processors, which is part of why a chip count alone understates how much else -- networking, cooling, power distribution -- has to arrive and get commissioned alongside each wave. xAI's own history with this site is one of the reasons Musk's numbers get taken seriously rather than dismissed as a founder's usual optimism: Colossus 1 reportedly went from an empty building to a running training cluster in 19 days, against an industry-standard permitting-to-power timeline measured in years. That reputation for speed is doing real work in how this announcement lands -- the same claim from an operator with a slower, less-verified build history would reasonably draw more skepticism than this one has.

By early this year, independent trackers already put Colossus 2 at roughly 555,000 GPUs, 2 gigawatts of capacity and about $18 billion in cumulative hardware spend -- figures that line up closely with the 550,000-chip current count Musk gave in September, once GB200/GB300 mix and a few months of incremental deployment are accounted for. Musk's new numbers, in other words, aren't a change of plan so much as the next scheduled step in a buildout that has consistently moved faster than the industry's normal permitting-to-power timeline.

Laid end to end, two of those three future waves carry no qualifier at all in Musk's own account -- only the last one does. That distinction matters more than the eye-catching 1.21 million total, because it separates what Musk is stating as scheduled from what he's explicitly flagging as contingent.

“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.” — Elon Musk, on X, Sept. 25, 2026

Even the conservative reading is a large number: the two waves Musk stated without any qualifier would take Colossus 2 alone to roughly 990,000 chips by November, before the hedged December wave is even counted. Combined with Colossus 1's 230,000, Musk's own figures put the Memphis site as a whole within reach of 1.44 million GPUs if every wave lands on schedule -- a threshold xAI has been publicly chasing since it first talked about a million-GPU Memphis footprint earlier this year, and one it would now clear with room to spare even under the more conservative reading.

The comparison undercuts a clean "who's bigger" headline as much as it supports one. Colossus 2's figures are Musk's own raw chip tally; Stargate Abilene's, per Epoch AI's independent tracker, are a normalized 509,000 H100-equivalents today, projected toward 1.02 million by year-end, on 421 megawatts of IT power scaling to about 843 megawatts and a reported $15.9 billion in capital cost growing toward $31.9 billion. Nvidia's GB300 is also meaningfully more capable per chip than the H100 generation Epoch's normalization is anchored to, which means a straight chip-for-chip count actually understates Colossus 2's relative compute versus a raw comparison of the two totals. xAI's site currently runs on roughly a fifth of Stargate Abilene's projected year-end power draw relative to its chip count, though, which is the more interesting number: compute without commissioned power behind it doesn't train anything.

For a reader tracking xAI's position against OpenAI, Google and Anthropic in the compute race, the more durable fact isn't the 1.21-million-chip target -- it's that a single site in Memphis is now routinely discussed in the same breath as a national-scale, multi-site project like Stargate, which spans seven planned locations rather than one. xAI has consistently chosen to concentrate its buildout at one campus instead of spreading it across sites the way OpenAI, Microsoft and Google have -- a bet that a single, faster-to-permit location beats the redundancy and grid access multiple sites would offer. That concentration is also what makes Colossus 2 unusually exposed to a single point of failure: one delayed substation, one contested permit, or one supply disruption slows the whole cluster rather than one site among several.

Whether Musk's schedule holds past the next self-reported milestone is exactly what the apply items above are watching for -- and it's worth remembering that this isn't the first time a Colossus timeline has been stated in public before it was independently confirmed. The pattern so far has been that the numbers arrive roughly on schedule, if not always exactly as first described; that track record is the actual reason this projection is worth covering rather than dismissing, not proof that this specific 1.21-million figure will land precisely as stated.