Enter the gigafactory of compute. From empty warehouses to the world’s first gigawatt-scale AI training clusters — the full story of Colossus 1 & 2.
A structured cinematic deep-dive designed for engineers, researchers, founders, and the simply curious.
Big ideas, simple words. No jargon left unexplained.
Colossus is like a giant video-game console… but instead of one chip, it has hundreds of thousands of super-smart chips working together. They help teach computers to think, talk, and create — like training a robot friend named Grok.
Colossus is the name of xAI’s (now SpaceXAI’s) giant computer buildings in Memphis, USA. Think of it as a special factory whose only job is to make artificial intelligence smarter.
Every important word from the workshop, translated into 9-year-old language:
That’s more brain-chips than there are people in many big cities — all working together as one team.
IT Power is the electricity that goes straight into the computers and chips — not the cafeteria lights or the office air conditioning. It’s the “food” the brain factory eats.
Later in this workshop you’ll find a spinning 3D model of the Memphis campus. It’s not a photo — it’s a fun interactive picture you can control with your mouse or finger.
Because bigger and faster “brain factories” help AIs learn faster and become more helpful. Colossus is one of the places where the next generation of smart computer helpers is being trained — so they can answer questions, invent new ideas, help scientists, and maybe one day help us understand the universe a little better.
How an empty Electrolux plant became the largest coherent AI training system on Earth in just four months.
In early 2024, conventional wisdom said a 100,000-GPU training cluster would take 18–24 months. xAI (now SpaceXAI) decided that timeline was unacceptable.
They acquired a decommissioned Electrolux appliance factory at 3231 Paul R. Lowry Road in South Memphis — 785,000 sq ft on 81 acres — and treated the build like a rocket launch: remove every non-essential step, parallelize everything possible, and pay above market to attract the best talent and contractors.
Jensen Huang called the achievement “superhuman.” The industry still has not matched the velocity.
Three strategic advantages converged:
The founding site that proved the model: speed, scale, and coherent training at unprecedented size.
Colossus 1 is 150k H100, 50k H200 and 30k GB200 — a multi-generation cluster. It started pure Hopper (H100), added H200 memory-enhanced parts, and later incorporated early Blackwell GB200 systems. This mixed fleet is still treated as a single coherent training domain — a non-trivial software and networking achievement.
| Chip | Count | Generation | Role |
|---|---|---|---|
| NVIDIA H100 | 150,000 | Hopper | Original workhorse |
| NVIDIA H200 | 50,000 | Hopper (HBM3e) | Memory-intensive workloads |
| NVIDIA GB200 | 30,000 | Blackwell | Next-gen density & interconnect |
In a landmark deal, Anthropic agreed to rent essentially the entire compute capacity of Colossus 1. Reported figures put the contract at roughly $1.25 billion per month through May 2029 — one of the largest known AI infrastructure leases in history.
This transformed Colossus 1 from a proprietary training asset into a high-value multi-tenant platform while freeing SpaceXAI’s own frontier training to move fully onto Colossus 2.
MACROHARD. The first gigawatt-scale AI training cluster on Earth. Blackwell Ultra at planetary scale.
Colossus 2 occupies a ~1,008,000 sq ft converted warehouse at 5420 Tulane Road in Whitehaven, Memphis, on ~79 acres. The rooftop lettering “MACROHARD” has become an iconic landmark visible from satellite imagery.
Musk declared it operational on 17 January 2026 as the first gigawatt training cluster in the world. Colossus 2 is 110k GB200 and 440k GB300 as of late September 2026, with three further waves of 220k GB300 each planned through the end of the year.
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 2026.
If we get lucky, yet another 220k GB300 by late December 2026.
— Elon Musk on X, 25 September 2026
Elon Musk publicly detailed the remaining 2026 build-out:
| Timing | Addition | Cumulative Colossus 2 |
|---|---|---|
| Sep 2026 (current) | 110k GB200 + 440k GB300 | ~550,000 |
| Next week | +220k GB300 | ~770,000 |
| November 2026 | +220k GB300 | ~990,000 |
| Late Dec 2026* | +220k GB300 | ~1,210,000 |
* “If we get lucky” — Musk
The 110k / 220k step sizes are driven by networking constraints: the number of fiber-optic cables that can be plugged into a central switch.
Colossus 2 is powered primarily by a behind-the-meter gas-turbine plant across the state line in Southaven, Mississippi. Temporary mobile turbines enabled rapid scale-up while a permanent 1.2 GW gas plant is under construction. Regulatory agreements require removal of temporary units by mid-2027.
Rotate, zoom, and explore a stylized representation of the dual-site Memphis campus. Drag to orbit.
Blue = Colossus 1 · Purple = Colossus 2 · Glowing nodes represent GPU racks
The invisible systems that make a million-GPU coherent cluster possible.
At this scale, the network is the cluster. Colossus uses high-bandwidth, low-latency fabrics (NVIDIA NVLink / InfiniBand-class) designed so that any GPU can efficiently communicate with any other during distributed training. The 110k-chip “slice” size is explicitly limited by how many fiber connections a central switch can accept.
Modern Blackwell systems (especially NVL72 racks) are liquid-cooled. The campus deploys massive cooling capacity measured in hundreds of megawatts of heat rejection. Tesla Megapacks provide short-term buffering, while the gas turbines supply the continuous baseload.
Running multi-generation hardware (H100, H200, GB200, GB300) as a single training domain requires sophisticated scheduling, mixed-precision handling, and fault-tolerant distributed frameworks. This is one of the least visible but most impressive engineering achievements of the Colossus project.
| Metric | Colossus 1 | Colossus 2 (current) | Combined Target |
|---|---|---|---|
| Physical GPUs | ~230k | ~550k+ | 1M+ |
| H100-equivalents | ~276k | ~1.11M | ~1.8M+ |
| IT Power | ~340 MW | ~946 MW | ~2 GW |
| Primary Users | Anthropic (lease) | SpaceXAI, Cursor, Anthropic, Google (upcoming) | Multi-tenant + internal |
What a gigawatt of coherent compute unlocks — and where the campus is headed next.
Colossus 2 is the primary home for SpaceXAI’s Grok training runs. The ability to scale to hundreds of thousands of next-generation GPUs in a single coherent domain is a decisive advantage for training the largest and most capable models.
Beyond internal use, the campus already hosts Anthropic (Colossus 1), and has signed capacity agreements with Google (~110k GPUs from Oct 2026) and others. This hybrid model — own the iron, lease the surplus — turns infrastructure into a high-margin revenue stream while still prioritizing frontier research.
Additional buildings (Minihard, Macrohardrr / Colossus 3) are rising in the Memphis–Southaven corridor. The long-term vision is a multi-gigawatt campus capable of supporting continuous frontier training and large-scale inference simultaneously.
Put your knowledge to the test. Play, match, race, and quiz — earn points and master the Colossus story.
Click a fact, then click its matching value. All pairs correct = +50 pts.
Spam the button to “build” Colossus 1 from 0 → 100k GPUs. Beat 8 seconds for a bonus!
Drag events into chronological order. Correct sequence = +40 pts.
Answer as fast as you can. Each correct = +10 pts.
Question 1 / 6
Click an answer — feedback is instant. +15 pts per correct answer.
Colossus is more than a collection of GPUs. It is a demonstration that industrial-scale AI infrastructure can be built at rocket-company velocity. The race is no longer about who has the best algorithm alone — it is about who can deliver coherent compute at the frontier first and keep scaling.