Fiber, Not Chips: Musk's Colossus 2 Timetable Aims at 1.21 Million GPUs by New Year

Nearly two years after Elon Musk first promised to expand his Memphis supercomputer past a million GPUs, the end goal is finally in sight — and on September 25 he published the most detailed public timetable yet for getting there. In a post on X, reported by Bloomberg and picked up across the tech press, Musk laid out a three-batch schedule that would take Colossus 2 past 1.2 million Nvidia accelerators before the ball drops on 2027.

The numbers on the table

According to Musk’s post, Colossus 2 — the second-generation cluster near Memphis, Tennessee — currently runs 110,000 Nvidia GB200 chips and 440,000 GB300s. That fleet alone makes it one of the largest single-site AI training installations on Earth. Epoch AI’s independent data-center directory currently estimates the site at roughly 1.11 million H100-equivalents of compute, supported by about 946 MW of IT power at a capital cost of $35.8 billion.

The ramp ahead comes in unusually clean increments:

• 220,000 GB300s “fully operational next week” — the first tranche, due within days of the announcement
• Another 220,000 GB300s in November
• A third batch of 220,000 GB300s by late December — “if we get lucky,” Musk hedged

If all three batches land, Colossus 2 would finish the year at 1.21 million GB200 and GB300 processors, including 1.1 million GB300s. Add the legacy Colossus 1 site — 150,000 H100s, 50,000 H200s, and 30,000 GB200s — and SpaceXAI’s total GPU fleet in operation would reach roughly 1.44 million units.

Why 110,000? Fiber, not silicon

The strangest number in the post is the step size itself. Why does an AI cluster grow in batches of 110,000 or 220,000 chips? Musk tied it directly to networking physics rather than chip supply: the increment reflects “the number of fiber optic cables that can be plugged into a central switch.”

That detail matters because it inverts the usual narrative. For most of the past two years, the binding constraint on AI build-outs was assumed to be Nvidia’s supply of Blackwell-class GPUs — Musk himself famously lobbied Jensen Huang directly for early allocations. On this timetable, the chips are effectively in hand; what limits how fast they light up is the switch fabric, the fiber plant, and above all the power behind it. Getting GPUs is no longer the hardest part. Feeding and interconnecting them is.

The power question the post doesn’t answer

What Musk’s schedule conspicuously omits is where the electricity comes from. Reuters reported in July that 59 natural-gas turbines had been installed at the Memphis site without federal clean-air permits, drawing a lawsuit from the surrounding community. SpaceXAI has since promised to retire the temporary generation — but only over the span of a year, as its own permanent 1.2-gigawatt power plant comes online.

That plant is the quiet dependency under the whole timetable. The late-December batch is the tell: Musk already framed it as luck-dependent, and a slip there would say more about power availability and switch capacity than about Nvidia’s shipment schedule. If the 1.2-GW plant isn’t energized before winter, Colossus 2 rides through the ramp on unpermitted gas turbines — a posture that has already drawn legal action and regulatory scrutiny.

A three-year-old company at the frontier table

The scale is made more remarkable by the company’s age. SpaceXAI — xAI was folded into SpaceX in February 2026 and rebranded in July — is only about three years old, against six for Anthropic and ten for OpenAI. Yet if the December target holds, Musk clears a million-GPU bar that none of his rivals has publicly claimed for a single site. Broadcom said back in 2024 that it had three hyperscale customers gunning for million-GPU clusters by 2027, but declined to name them; Musk is the first to put a dated schedule in public.

There’s also a pragmatic shuffle visible in the fleet itself. The original Colossus 1 site mixes Hopper- and Blackwell-generation GPUs — an inefficient combination for training Grok — so SpaceXAI has rented it out to Anthropic for inference workloads instead, while Colossus 2 runs purely on Blackwell to avoid heterogeneous bottlenecks.

And Memphis is only the beginning of the ambition. Musk has said SpaceXAI will grow its data-center capacity sevenfold by 2027, has floated a target of 50 million H100-equivalent GPUs by 2030, and SpaceX is exploring an Orbital Data Center System with a million satellites — an idea Jensen Huang has politely called a “dream” for now.

What to watch

Three checkpoints will tell us whether this timetable is real:

1. Next week: do 220,000 GB300s actually come up “fully operational,” or does “operational” turn out to mean racked-but-not-networked?
2. November: the second batch lands right as winter power demand peaks in Tennessee — the first real stress test of the permanent plant’s timeline.
3. Late December: the batch Musk himself hedged. Watch the Shelby County air-permit filings and turbine-removal progress as leading indicators.

For an industry where “GPU count” has become the scoreboard for frontier ambitions, Colossus 2’s ramp is the clearest single test yet of whether the AI build-out’s bottleneck has permanently shifted from silicon to megawatts — and whether one company can outrun its own power grid to get there.