Elon Musk says his Colossus 2 data center in Memphis will have more than 1.2 million Nvidia GB200 and GB300 chips online by the end of 2026, capping a year in which the site adds roughly 660,000 accelerators and pushes Musk's total AI compute fleet toward 1.44 million GPUs in operation.
The figures, laid out by Musk in a series of posts on X and reported by Benzinga, Wccftech, and Tom's Hardware, describe a breakneck installation schedule: 220,000 additional GB300 GPUs expected online by the end of this week, another 220,000 by the end of November, and a final 220,000 batch before the calendar flips to 2027. For more context on this story, see our ongoing AI news.
The Numbers Behind the Build-Out
Colossus 2 currently runs on about 110,000 GB200 chips and 440,000 GB300 chips, according to Musk's posts. The three planned waves of 220,000 GB300s each add up to exactly the 660,000 GPUs Tom's Hardware cites as arriving this year — bringing the Colossus 2 site alone to more than 1.2 million Blackwell-generation packages.
The older Colossus 1 facility, which started construction in 2024 and was originally built by xAI, contributes another 230,000 chips: about 200,000 H100 processors from Nvidia's previous Hopper generation plus 30,000 GB200s. Combined, the two Memphis sites are nearing 1.44 million GPUs in operation by the end of 2026.
For perspective, a single GB200 or GB300 package pairs Nvidia's Grace CPU with Blackwell GPUs and is aimed at the kind of large-scale training and inference workloads that frontier labs buy by the hundreds of thousands. Colossus 2's year-end count would rival the total accelerator fleets that most national governments have procured.
Power Is the Real Bottleneck
Chips are only half of the equation. Tom's Hardware reports the firm is building a 1.2-gigawatt power plant to bring the expanded systems fully online — an infrastructure project on the scale of a mid-sized city's electrical demand, dedicated to a single AI campus.
Musk himself framed the constraint plainly in his posts, writing that bringing "massive compute online rapidly is incredibly difficult." The comment is a rare acknowledgment from the project's chief proponent that the binding constraint on AI ambitions in 2026 is not chips, which can be bought, but transformers, substations, and turbines, which take years to commission.
Musk has previously pursued unconventional routes to speed up that timeline, including on-site turbine generation and gas turbines mounted on mobile frames — approaches that drew regulatory scrutiny in Memphis even as they kept Colossus 1's construction ahead of schedule.
Why the GB300 Gets the Nod
The new waves are all GB300s, Nvidia's Blackwell Ultra variant. The GB300 shares its Grace CPU with the GB200 but carries a newer, faster GPU die and more memory bandwidth, and it has become the volume workhorse for 2026's frontier training runs.
Musk's comments indicate Colossus 2 will rely primarily on the GB300 going forward, with the GB200 contingent serving as the installed base. That ratio will shift again when Nvidia's next-generation Rubin platform — the subject of Musk's earlier exclusive GPU deal with Nvidia for space-based deployment — begins shipping in volume.
The Race Against Anthropic and OpenAI
The install schedule is not just an infrastructure flex; it is the opening move in Musk's public campaign to close the model gap with the frontier leaders. In the same posts, Musk claimed his AI efforts — which he says are only three years old, against six for Anthropic and ten for OpenAI — could match Anthropic's Fable-class models and OpenAI's GPT-6 within two to three months, and that if the effort's "second derivative remains strong," the company would "reach pole position in about 6 months."
"We will keep accelerating," Musk wrote.
Those are bold claims from a lab whose most recent public model releases have trailed the frontier on most agentic benchmarks, and rivals' cadence gives little reason for pause: Anthropic and OpenAI both shipped major coding and agentic model updates this quarter. But the compute trajectory lends the argument a material basis. Training compute has remained the strongest predictor of frontier capability, and no lab is adding training capacity faster than Musk's operation right now.
What It Means for the AI Hardware Market
The expansion reinforces two trends that have defined the 2026 build-out. First, the concentration of frontier training capacity into a handful of gigascale campuses — Colossus 2 joins Stargate-linked sites, Amazon's Anthropic campuses, and Google's TPU fleets as single-site installations measured in gigawatts. Second, Nvidia's grip on that spending: essentially every accelerator in Musk's fleet is an Nvidia part, sustaining the demand assumptions that have made the chipmaker the central asset of the AI economy.
SemiAnalysis, the research firm that previously projected Musk's compute operation could exceed 10 gigawatts of capacity by 2027 with Microsoft as its largest computing customer, has argued that even imperfect infrastructure is being absorbed by extreme GPU demand. The Colossus 2 schedule — three 220,000-GPU waves in under three months — is the most visible test yet of whether that thesis holds.
Whether the new waves arrive on Musk's timeline will say as much about the global GPU supply chain as about Memphis: every GB300 in the plan is a chip Nvidia did not sell to a rival, and every megawatt is power the grid did not have to spare.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
