xAI no longer exists as a standalone company — it was absorbed into SpaceX and rebranded SpaceXAI in May 2026, so its procurement now consolidates into SpaceX’s books. And the central fact about that supply chain is one most coverage gets backwards: the binding constraint was never GPUs. It was electricity. Here is the four-layer chain, with the numbers.
👉 The framing that unlocks this topic: Colossus is not a chip project with a power problem. It is a power project with a chip budget. Every genuinely hard decision — turbines, substations, Megapacks, water — sits on the electricity side of the ledger.
- Corporate status: xAI merged into SpaceX in February 2026 at a combined $1.25 trillion and was dissolved into the SpaceXAI division in May 2026.
- Silicon: Nvidia throughout — H100 and H200 at Colossus 1, GB200 and GB300 at Colossus 2. Reported GPU counts range from roughly 200,000 to 222,000 at Colossus 1.
- Servers: Dell (PowerEdge XE9680) built about half the racks; Supermicro supplied 4U liquid-cooled systems at 64 GPUs per rack.
- Power is the real chain: 35 on-site gas turbines, grid connections via MLGW and TVA, and Tesla Megapacks — SpaceX spent $295 million on Megapacks in Q2 2026 alone.
- Vendor concentration: xAI accounts for roughly 67% of turbine supplier Solaris Energy Infrastructure’s 1,700 MW order book.
What Changed: xAI Is Now SpaceXAI
On February 2, 2026, SpaceX acquired xAI in the largest private merger in history — SpaceX valued at $1 trillion, xAI at $250 billion, the combined entity at roughly $1.25 trillion. In May 2026 Musk announced xAI would cease to exist as a separate company, with Grok and X folded into a SpaceX division branded SpaceXAI (Teslarati on the rebrand).
The stated rationale was orbital data centers. SpaceX filed with the FCC to launch up to one million satellites designed as AI compute nodes in low Earth orbit — explicitly framed as an escape from the terrestrial energy limits constraining AI development. xAI brought the software stack, Grok, X and Colossus; SpaceX brought rockets, Starlink and the capital base (analysis of the merger).
For supply-chain purposes this matters in one specific way: xAI files no periodic SEC reports, and its results now consolidate into SpaceX’s filings (xAI financials and disclosure record). Since SpaceX went public in June 2026, xAI procurement became visible for the first time — through someone else’s quarterly report. Our SpaceX supply chain breakdown covers the parent picture.
💡 Practical consequence for anyone researching this: stop looking for xAI disclosures. The numbers now appear as the AI segment inside SpaceX results — $2.56 billion of revenue in Q2 2026, up 247%, with $15.8 billion of the quarter’s $18.4 billion capex going to AI infrastructure.
The Four-Layer Supply Chain
| Layer | What it covers | Who supplies it |
|---|---|---|
| Silicon | Training and inference GPUs, networking | Nvidia — H100, H200, GB200, GB300 |
| Systems | Server racks, liquid cooling, integration | Dell (PowerEdge XE9680), Supermicro (4U liquid-cooled) |
| Power | Generation, storage, grid interconnection | Solaris Energy Infrastructure (turbines), Tesla (Megapacks), MLGW/TVA (grid) |
| Water & cooling | Closed-loop cooling, municipal water, greywater | Local utilities; on-site treatment |
Layers one and two are competitive markets where money buys throughput. Layers three and four are local, permitted, physical and political. That asymmetry explains nearly everything about how Colossus was built and where it has run into trouble.
Layer 1: Silicon
Colossus 1 launched in July 2024 with 100,000 Nvidia H100s on a single fabric, doubling to 200,000 by February 2025. Reported totals since then span roughly 200,000 to 222,000 GPUs across H100, H200 and GB200 generations, drawing around 300 megawatts at peak (the Colossus dissolution and capacity picture).
Colossus 2, at the Tulane Road site, was reported to include at least 110,000 Nvidia GB200 GPUs carrying a power load around 170 megawatts, with the site targeting gigawatt scale (Introl technical breakdown of Colossus). Nvidia CEO Jensen Huang described the original build pace as "superhuman."
⚠ Data-quality warning for this topic: GPU counts, power figures and Megapack quantities vary substantially between sources — 200,000 vs 222,000 GPUs, 168 vs 208 Megapacks, 250 MW vs 300 MW vs 1.2 GW. These reflect different dates and different scopes (Colossus 1 alone vs both sites). Any article giving you one precise number without a date is guessing.
Layer 2: Servers and Cooling Hardware
The hardware was assembled through partnerships with Dell Technologies and Super Micro Computer. Dell built roughly half the server racks using its PowerEdge XE9680 platform; Supermicro supplied the liquid-cooled systems (coverage of the Dell and Supermicro build).
The architecture detail worth knowing: HGX servers with eight GPUs each, in Supermicro liquid-cooled racks at 64 GPUs per rack, totalling around 1,500 GPU racks in the initial 100,000-GPU deployment. Rack density required liquid cooling from the outset — it was not retrofitted.
✔ Design lesson that generalizes: because liquid cooling was specified from day one, the expansion from 100,000 to 200,000 GPUs needed no fundamental redesign. Most data centers pay for that decision later. This one paid up front and bought itself scaling headroom.
Layer 3: Power — The Actual Supply Chain Story
A 300-megawatt facility standing up in months cannot wait for a utility interconnection queue. xAI’s answer was to bring its own generation, then backfill with grid power — an approach that worked technically and generated most of its legal and community problems.
Gas turbines
xAI installed 35 on-site methane gas turbines with combined capacity around 422 megawatts — more than double the permitted amount, according to environmental groups who argued this violated the Clean Air Act. After sustained pressure and a new substation coming online, an undisclosed number were removed (DataCenterDynamics on the turbines and Megapacks).
The Solaris concentration
Turbine supply runs largely through NYSE-listed Solaris Energy Infrastructure, which owns a fleet of around 600 MW of gas turbines with roughly 400 MW serving xAI. xAI represents about 67% of Solaris’s 1,700 MW order book — some 1,140 MW — with roughly 900 MW to be held in a joint venture owned 50.1% by Solaris and 49.9% by xAI (SemiAnalysis on Colossus 2 power).
💡 This is vendor concentration inverted. Normally the buyer worries about depending on a supplier. Here a public company derives two-thirds of its order book from one private customer — and then co-owns the assets with it. Solaris shareholders are effectively holding an xAI proxy position.
Tesla Megapacks
Battery storage smooths load and provides grid independence. Reported deployments range from 168 to 208 Megapacks at Colossus, and the spending is now documented: SpaceX purchased $295 million of Tesla Megapacks in the quarter ending June 30, 2026, bringing half-year spending to $329 million (CNBC on the Megapack purchases).
That is a related-party transaction between two Musk companies, now disclosed in filings because SpaceX is public. It is also one of Tesla Energy’s more significant single customers — a detail that connects this supply chain directly to Tesla’s storage business.
Grid connection
Colossus received 150 MW from Memphis Light, Gas and Water and the Tennessee Valley Authority when its first substation came online, with a second 150 MW substation following. The Tulane Road site sits adjacent to TVA’s Southaven combined-cycle plant, and the original Colossus site is across the street from TVA’s 1.1 GW Allen plant (technical site notes on Colossus).
Site selection was a power-procurement decision disguised as a real-estate decision. Both Memphis sites were chosen for proximity to existing large generation, not for land or connectivity.
| Power source | Scale | Role |
|---|---|---|
| On-site gas turbines | 35 units, ~422 MW installed at peak | Bridge power before grid connection; partially removed |
| Solaris rental turbines | ~400 MW serving xAI; 1,140 MW on order | Rapid-deploy generation, JV-owned |
| Tesla Megapacks | 168–208 units; $329M spent in H1 2026 | Load smoothing, outage backup, grid independence |
| MLGW / TVA grid | 150 MW + second 150 MW substation | Permanent baseload |
Layer 4: Water and Cooling
Liquid-cooled racks at this density consume serious water. Reporting around the Tulane Road site indicated consumption on the order of one million gallons per day. In a municipal context, water permits and aquifer impact are a slower-moving but equally hard constraint than electricity — and they attract the same community scrutiny.
👉 For anyone modelling AI data-center buildouts anywhere: the sequence of binding constraints is chips, then power, then water, then permits, then community consent. Colossus moved through the first two at record speed and is now working through the last three. That order is becoming the industry pattern.
The Pivot: When the Supply Chain Became the Product
The most consequential 2026 development is that SpaceXAI began selling Colossus capacity rather than only consuming it. Reported deals include a lease of Colossus 1 capacity to Anthropic, agreements involving Google, and a contract with Reflection AI reported at $150 million per month for GB300 access (Teslarati on the third compute deal).
The stated logic is that SpaceXAI does not currently need Colossus 1’s full capacity, having shifted internal training toward Colossus 2, so leasing monetizes capacity that would otherwise sit partially idle (analysis of the Colossus leasing strategy). Reporting describes the Anthropic arrangement as roughly a 180-day lease.
Note the strategic inversion. Colossus was built as a proprietary advantage — compute nobody else could fund. It is now partly a merchant compute business serving direct model competitors. The infrastructure went from moat to product.
✔ Watch this line item specifically. If merchant compute revenue keeps growing, the supply chain reframes entirely: turbines and Megapacks stop being cost centers supporting Grok and become revenue-generating assets. That changes how the capex should be judged.
The Risks That Are Not About Suppliers
- Regulatory and legal. Clean Air Act challenges over the turbines, an NAACP lawsuit connected to the Southaven gas plant, and sustained community complaints about emissions, odour and noise.
- Community consent. Memphis residents reporting health effects and pollution have fuelled a broader national backlash against data-center developers — a factor that now shapes site selection industry-wide.
- Concentration in both directions. Nvidia on silicon, Solaris on turbines, Tesla on storage. Each is a single dominant relationship, and two of the three are related parties or joint ventures.
- Latency and topology. Reporting indicates Grok teams hit latency problems connecting Memphis to other facilities — a reminder that distributed compute has physics constraints money cannot remove.
The Next Supply Chain: Orbit
The orbital data-center plan is not a side project — it is the stated reason the merger happened. If SpaceX launches compute nodes as satellites, the supply chain changes shape completely: no turbines, no water, no municipal permits, but launch cadence, radiation-hardened electronics and thermal rejection in vacuum become the constraints instead.
Every problem in the Memphis chain is a terrestrial problem. That is the bet: trade permitting and community consent for launch capacity and space-qualified hardware, in a company that already owns the rockets.
Checklist: How to Read Any Colossus or xAI Number
- Check the date and the scope. Figures for Colossus 1 alone, Colossus 2 alone, and both sites combined are routinely mixed together.
- Separate installed from operational. Turbine capacity installed is not turbine capacity permitted or running.
- Look for the power figure, not the GPU figure. Megawatts tell you more about actual capability than chip counts do.
- Treat related-party deals as disclosed, not neutral. Megapack purchases and the Solaris JV involve Musk-linked entities on both sides.
- Find the number in SpaceX filings now. Anything sourced to "xAI disclosed" after May 2026 is second-hand by definition.
FAQ
Does xAI still exist?
Not as a standalone company. SpaceX acquired xAI in February 2026 and dissolved it into a division called SpaceXAI in May 2026. Grok and X continue under that branding.
Who supplies xAI’s GPUs?
Nvidia, across H100, H200, GB200 and GB300 generations. Dell built roughly half the server racks using PowerEdge XE9680 systems, with Supermicro supplying 4U liquid-cooled racks at 64 GPUs each.
How is Colossus powered?
A hybrid of on-site gas turbines, rental turbines from Solaris Energy Infrastructure, Tesla Megapack battery storage, and grid power from MLGW and TVA via two 150 MW substations.
How much does xAI spend on Tesla Megapacks?
SpaceX purchased $295 million of Tesla Megapacks in Q2 2026 for the Colossus sites, bringing first-half 2026 spending to $329 million.
What is the real bottleneck for Colossus?
Power, water and permits — not chips. GPUs and servers are competitive markets; generation capacity, water rights and community consent are local, slow and political.
Is Colossus only used by SpaceXAI?
No. Capacity is being leased to external customers, with reported agreements involving Anthropic, Google and Reflection AI, the latter reported at $150 million per month for GB300 access.
The Bottom Line
xAI’s supply chain is a masterclass in solving the wrong-looking problem first. Buying 200,000 Nvidia GPUs is a financing exercise. Getting 300 megawatts to a Memphis site in months is an engineering, legal and political one — and that is where the real work went.
The 2026 restructuring changed who reports the numbers, not what they are. Nvidia, Dell, Supermicro, Solaris and Tesla remain the chain; the difference is that it now sits inside a public company’s filings, where the related-party transactions and the capex are visible.
The number to watch through the rest of 2026 is merchant compute revenue. If leasing Colossus capacity keeps scaling, the turbines and Megapacks stop being the cost of Grok and start being a business. Track the parent picture in our SpaceX supply chain analysis and the storage side in our Tesla supply chain breakdown.
💡 We track the whole Musk hardware ecosystem — Tesla, SpaceX, SpaceXAI, Neuralink, The Boring Company — with confirmed and reported figures kept in separate columns. Join the community to get each breakdown as it lands.
Sources: CNBC, Teslarati, DataCenterDynamics, SemiAnalysis, Introl, Tom’s Hardware, The AI Rankings, glennklockwood.com. GPU counts, Megapack quantities and power figures vary by source and date and are reported as ranges where sources conflict. Last fact-check: August 5, 2026.
NEVER MISS AN OPTIMUS UPDATE
We track every Tesla Optimus development — specs, deployment milestones, pricing and competitive moves — updated as news breaks.