> ## Content Index
> Fetch the complete content index at: https://www.insidedeeptech.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# xAI Colossus 2: A Full Guide to the Memphis Gigawatt AI Supercomputer, GPUs, Power and Limits
- URL: https://www.insidedeeptech.com/xai-colossus-2-memphis-gigawatt-ai-supercomputer-full-guide/
- Published: 2026-10-23T06:10:47.000Z
- Updated: 2026-10-23T06:10:47.000Z
- Description: xAI Colossus 2 full guide: how SpaceX's Memphis and Southaven campus went from Colossus 1's 100,000 H100s to GB200 and GB300 clusters, how it is powered and cooled, who rents it, and where the gigawatt claims break down.
- Author: Austin Heaton
- Tags: AI, Hardware, Energy, Deep Tech

xAI bought the [Tulane Road site in Memphis in March 2025](https://www.datacenterdynamics.com/en/news/musks-xai-gets-go-ahead-for-41-natural-gas-turbines-in-mississippi-to-power-colossus-data-centers/?ref=insidedeeptech.com). It then brought Colossus 2's first cluster of [roughly 110,000 NVIDIA GB200 processors online in 91 days](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), according to SpaceX's IPO prospectus. By September 30, 2026, SpaceX was reporting [1.7 GW of nameplate compute](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), up from 0.6 GW a year earlier. Few AI builders publish their build pace in this much detail.

It's also the most argued-about data center in America. Its power comes from gas turbines across the state line, its biggest outside customer is a rival AI lab, and an independent analysis says its chips drew about half the advertised power in June.

This full guide covers what Colossus 2 is, how it grew out of Colossus 1, the GPUs, the power and cooling stack, who uses it, what SpaceX says comes next and the honest limits, all as of October 23, 2026.

Last updated: October 23, 2026.

## Key Takeaways

- **Colossus 2** is SpaceX's AI campus on Tulane Road in Memphis and Stateline Road in Southaven, Mississippi, per the [S-1](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).
- It launched with [about 110,000 GB200s in 91 days, then 110,000 GB300s in 64 days](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).
- SpaceX reported [1.7 GW of nameplate compute on September 30, 2026](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), counting installed GPUs, not actual draw.
- Power is behind the meter: MDEQ approved [41 gas turbines rated at 1.2 GW](https://www.datacenterdynamics.com/en/news/musks-xai-gets-go-ahead-for-41-natural-gas-turbines-in-mississippi-to-power-colossus-data-centers/?ref=insidedeeptech.com) for the Southaven site.
- Honest limit: one analysis found Colossus 2's chips drew [about 0.55 GW in June 2026](https://www.isambrooks.com/p/wheres-the-ai-power-boom?ref=insidedeeptech.com), roughly half the headline.

💡

Inside Deep Tech's take: Colossus 2 is real, and it's the fastest gigawatt-class AI build on public record. But "gigawatt" here is an installed-GPU number, not a measured load, and the site's power, permits and largest contracts all carry risks that a glossy build timeline hides.

## What xAI Colossus 2 actually is

Colossus 2 is the second AI training campus built by xAI, the Grok developer that SpaceX acquired in February 2026\. SpaceX's prospectus now files xAI's compute, Grok and X under one AI segment, later renamed Super Intelligence (SI), per the [S-1](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).

The S-1 is precise about the geography. "COLOSSUS" is the original site on Paul R. Lowry Road in South Memphis. "COLOSSUS II" covers the data centers on Tulane Road in Memphis and on Stateline Road in Southaven, Mississippi, which SpaceX calls part of one [coherent gigawatt-scale AI training cluster](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).

That's the key distinction. Colossus 1 is a converted factory. Colossus 2 is a multi-building campus that straddles a state border, with its own power plant.

SpaceX says Colossus 2 is training its next frontier model. In its May 2026 S-1, SpaceX named [Grok 5 as currently being trained at COLOSSUS II](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).

## Colossus timeline, 2024 to 2026

The whole program has moved in roughly 30 months. Here are the dated milestones that matter, each tied to its source.

- **2024:** xAI brings the first Colossus cluster online in South Memphis in [122 days, in a repurposed factory](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).
- **October 28, 2024:** NVIDIA confirms Colossus runs on Spectrum-X Ethernet and that xAI is [doubling it to 200,000 Hopper GPUs](https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus?ref=insidedeeptech.com).
- **February 2025:** xAI says Colossus reached [full scale at 200,000 GPUs, 276 days from groundbreaking](https://x.ai/colossus?ref=insidedeeptech.com).
- **July 2025:** Shelby County permits [15 turbines at Colossus 1](https://www.cnbc.com/2025/07/03/musks-xai-gets-permit-for-turbines-to-power-supercomputer-in-memphis.html?ref=insidedeeptech.com), and xAI buys the [Southaven power plant site](https://www.datacenterdynamics.com/en/news/musks-xai-gets-go-ahead-for-41-natural-gas-turbines-in-mississippi-to-power-colossus-data-centers/?ref=insidedeeptech.com).
- **January 2026:** Colossus 2 comes online; Musk calls it the [first gigawatt training cluster in the world](https://www.isambrooks.com/p/wheres-the-ai-power-boom?ref=insidedeeptech.com).
- **February 2026:** SpaceX acquires xAI, folding Colossus into its [AI segment](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).
- **March 2026:** Mississippi approves the [41-turbine, 1.2 GW Southaven plant](https://www.datacenterdynamics.com/en/news/musks-xai-gets-go-ahead-for-41-natural-gas-turbines-in-mississippi-to-power-colossus-data-centers/?ref=insidedeeptech.com).
- **May 2026:** Anthropic takes [all of Colossus 1's capacity](https://www.anthropic.com/news/higher-limits-spacex?ref=insidedeeptech.com); SpaceX files its [S-1 on May 20](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).
- **August 14, 2026:** SpaceX [closes its Cursor acquisition](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).
- **September 30, 2026:** SpaceX reports [1.7 GW of nameplate compute](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).

## Colossus 1: the 122-day build that set the template

You can't understand Colossus 2 without the first site. xAI took over a former Electrolux factory in South Memphis in 2024 and, per the S-1, brought online [about 100,000 H100 processors and about 130 MW of compute power in 122 days](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).

NVIDIA's own release says it took [19 days from the first rack rolling onto the floor until training began](https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus?ref=insidedeeptech.com). xAI's site says the cluster was then [doubled to 200,000 GPUs in 92 days](https://x.ai/colossus?ref=insidedeeptech.com).

The hardware was conventional for 2024, built at unconventional speed. ServeTheHome's tour, republished by Supermicro, describes liquid-cooled racks of [eight 4U servers with eight H100s each, or 64 GPUs per rack](https://www.supermicro.com/CaseStudies/Success%5FStory%5FxAI%5FColossus%5FCluster.pdf?ref=insidedeeptech.com), grouped into 512-GPU blocks.

The network was Ethernet, not InfiniBand. NVIDIA says Colossus used Spectrum-X and held [95% data throughput, against about 60% for standard Ethernet at that scale](https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus?ref=insidedeeptech.com).

Power was the first fight. Memphis Light, Gas and Water said in May 2025 that xAI was [receiving 150 MW at Paul Lowry Road and had requested another 150 MW](https://www.mlgw.com/images/content/files/pdf/new/5-5-25%20xAI%20Update.pdf?ref=insidedeeptech.com). The gap was filled with gas turbines, and the Shelby County Health Department issued a permit for [15 turbines on July 2, 2025](https://www.cnbc.com/2025/07/03/musks-xai-gets-permit-for-turbines-to-power-supercomputer-in-memphis.html?ref=insidedeeptech.com).

Today Colossus 1 belongs to someone else's workload. In May 2026, Anthropic agreed to use [all of Colossus 1's compute, more than 300 MW and over 220,000 NVIDIA GPUs](https://www.anthropic.com/news/higher-limits-spacex?ref=insidedeeptech.com).

## Colossus 2 by the numbers: clusters, GPUs and gigawatts

The cleanest timeline comes from SpaceX's own filings. Everything below is either a company disclosure or is labeled as an estimate.

| Milestone                | What was disclosed                                    | Source                                                                                                                                         |
| ------------------------ | ----------------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
| First Colossus 2 cluster | About 110,000 GB200s, about 210 MW, online in 91 days | [SpaceX S-1](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com)         |
| Second cluster           | 110,000 GB300s and 220 MW, online in 64 days          | [SpaceX S-1](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com)         |
| Next phase (planned)     | At least 220,000 more GB300s and over 400 MW          | [SpaceX S-1](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com)         |
| March 31, 2026           | 1.0 GW nameplate compute draw (Colossus 1 + 2)        | [SpaceX S-1](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com)         |
| June 30, 2026            | 1.4 GW nameplate compute                              | [Q3 2026 release](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com)    |
| September 30, 2026       | 1.7 GW nameplate compute                              | [Q3 2026 release](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com)    |
| End of Q4 2026 (target)  | 2.5 to 3 GW nameplate compute                         | [Q3 2026 call](https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/?ref=insidedeeptech.com) |

Source: [SpaceX S-1, May 2026](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), [SpaceX Q3 2026 earnings release](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com) and the [Q3 2026 earnings call transcript](https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/?ref=insidedeeptech.com).

Two numbers need context. First, SpaceX defines nameplate compute draw as [installed GPUs multiplied by their all-in power draw, which does not represent actual power consumption or utilization](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com). It also excludes cooling and facility overhead.

Second, the GPU count. Elon Musk's breakdown, quoted in an October 2026 analysis, put Colossus 2 at [110,000 GB200s and 440,000 GB300s](https://www.isambrooks.com/p/wheres-the-ai-power-boom?ref=insidedeeptech.com) as of mid-2026\. That's a social post, not a filing, so treat it as directional.

Independent trackers land in the same zone. Epoch AI estimates Colossus 2 at [about 1.11 million H100-equivalents on 946 MW of IT power, at a capital cost near $35.8 billion](https://epoch.ai/data/ai-data-centers/directory/colossus-2?ref=insidedeeptech.com), built from satellite imagery and cooling models.

⚠️

Nameplate is not load. SpaceX's 1.7 GW figure counts the GPUs it has installed times their rated draw. It tells you capacity, not how hard those chips are running or how much electricity the campus actually uses.

## Colossus 1 vs Colossus 2 side by side

| Attribute           | Colossus 1 (Paul R. Lowry Rd)                                                                                                                                             | Colossus 2 (Tulane Rd + Southaven)                                                                                                                                                                            |
| ------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Building            | Converted Electrolux factory                                                                                                                                              | Multi-building campus across the TN/MS line                                                                                                                                                                   |
| First cluster       | [\~100,000 H100, \~130 MW, 122 days](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com)            | [\~110,000 GB200, \~210 MW, 91 days](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com)                                                |
| Main GPUs           | H100, H200, some GB200                                                                                                                                                    | GB200, then GB300                                                                                                                                                                                             |
| Grid power          | [150 MW, plus 150 MW requested](https://www.mlgw.com/images/content/files/pdf/new/5-5-25%20xAI%20Update.pdf?ref=insidedeeptech.com)                                       | [Primarily a dedicated gas plant](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com)                                                   |
| Turbine permit      | [15 turbines, Shelby County, July 2025](https://www.cnbc.com/2025/07/03/musks-xai-gets-permit-for-turbines-to-power-supercomputer-in-memphis.html?ref=insidedeeptech.com) | [41 turbines, 1.2 GW, MDEQ, March 2026](https://www.datacenterdynamics.com/en/news/musks-xai-gets-go-ahead-for-41-natural-gas-turbines-in-mississippi-to-power-colossus-data-centers/?ref=insidedeeptech.com) |
| Cooling (estimated) | [Evaporative cooling towers](https://epoch.ai/data/ai-data-centers/directory/colossus-1?ref=insidedeeptech.com)                                                           | [Air-cooled chillers and dry condensers](https://epoch.ai/data/ai-data-centers/directory/colossus-2?ref=insidedeeptech.com)                                                                                   |
| Main user today     | [Anthropic, all capacity](https://www.anthropic.com/news/higher-limits-spacex?ref=insidedeeptech.com)                                                                     | Grok training, Anthropic, Cursor                                                                                                                                                                              |

Source: [SpaceX S-1](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), [MLGW](https://www.mlgw.com/images/content/files/pdf/new/5-5-25%20xAI%20Update.pdf?ref=insidedeeptech.com), [CNBC](https://www.cnbc.com/2025/07/03/musks-xai-gets-permit-for-turbines-to-power-supercomputer-in-memphis.html?ref=insidedeeptech.com), [DCD](https://www.datacenterdynamics.com/en/news/musks-xai-gets-go-ahead-for-41-natural-gas-turbines-in-mississippi-to-power-colossus-data-centers/?ref=insidedeeptech.com), [Anthropic](https://www.anthropic.com/news/higher-limits-spacex?ref=insidedeeptech.com) and Epoch AI's [Colossus 1](https://epoch.ai/data/ai-data-centers/directory/colossus-1?ref=insidedeeptech.com) and [Colossus 2](https://epoch.ai/data/ai-data-centers/directory/colossus-2?ref=insidedeeptech.com) pages.

If you want the rack-level detail on the chips inside, Inside Deep Tech's [GB200 NVL72 full guide](https://www.insidedeeptech.com/gb200-nvl72-nvidia-ai-rack-full-guide/) and the [Vera Rubin NVL72 vs GB300 vs GB200 comparison](https://www.insidedeeptech.com/vera-rubin-nvl72-vs-gb300-vs-gb200-full-guide/) cover the hardware Colossus 2 is built from.

## How Colossus 2 is powered: turbines, Megapacks and Southaven

Here's the problem Colossus 2 solved by force. The local utility said in July 2025 that it was not supplying power to Colossus 2, as cited in [this analysis of the site's power records](https://www.isambrooks.com/p/wheres-the-ai-power-boom?ref=insidedeeptech.com). So xAI built its own.

The answer sits just across the border. xAI bought a former Duke Energy power plant site in Southaven, Mississippi, and in March 2026 the Mississippi Department of Environmental Quality granted a permit for [41 natural gas turbines generating 1.2 GW](https://www.datacenterdynamics.com/en/news/musks-xai-gets-go-ahead-for-41-natural-gas-turbines-in-mississippi-to-power-colossus-data-centers/?ref=insidedeeptech.com) to serve Colossus 2 and a planned Colossus 3.

SpaceX describes the result plainly. The S-1 says Colossus 2 [is capable of operating entirely by a self-built behind-the-meter gigawatt-scale natural gas power plant](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), integrated with what it calls the world's largest Megapack deployment.

The Tesla batteries matter more than they look. Large training jobs can ramp power up and down quickly, and the S-1 says Megapacks [deliver critical redundancy and help stabilize operations during peak demand](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).

Going forward, SpaceX says Colossus 2 will be [primarily powered by a dedicated natural gas power plant, supplemented over time by grid capacity it funds through local utilities](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com). On the Q3 call, Musk said SpaceX now casts turbine blades and vanes in-house and claimed that lets it bring power online [12 to 18 months faster than competitors](https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/?ref=insidedeeptech.com).

This is the largest real-world test of the approach Inside Deep Tech covers in its [behind-the-meter generation full guide](https://www.insidedeeptech.com/behind-the-meter-generation-ai-data-centers-full-guide/): skip the interconnection queue, burn gas on site and deal with the permits and community fallout later.

## Cooling and networking inside Colossus 2

GB200 and GB300 racks need liquid cooling at the chip. SpaceX's S-1 says its data centers use [advanced direct-to-chip cooling to support higher rack densities](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), on a brownfield retrofit strategy that reuses existing industrial sites.

The heat rejection outside the building is where Colossus 2 differs from Colossus 1\. Epoch AI's satellite review counts [217 air-cooled chillers and 144 air-cooled condensers, about 412 MW of cooling](https://epoch.ai/data/ai-data-centers/directory/colossus-2?ref=insidedeeptech.com) at Colossus 2, while Colossus 1 relies on [33 evaporative cooling towers](https://epoch.ai/data/ai-data-centers/directory/colossus-1?ref=insidedeeptech.com).

Dry cooling trades water for electricity: air-cooled systems avoid evaporating water but spend more fan and compressor power, especially on hot Memphis summer days. For the full cooling stack, see Inside Deep Tech's [data center liquid cooling full guide](https://www.insidedeeptech.com/data-center-liquid-cooling-ai-full-guide/).

SpaceX hasn't published Colossus 2's network design. Colossus 1 used NVIDIA Spectrum-X Ethernet with [SN5600 switches at up to 800 Gb/s per port](https://nvidianews.nvidia.com/news/spectrum-x-ethernet-networking-xai-colossus?ref=insidedeeptech.com), and xAI's site lists [2.8 Tb/s of network bandwidth per server](https://x.ai/colossus?ref=insidedeeptech.com) for the original cluster. Whether Colossus 2 repeats that choice is undisclosed.

## Who uses Colossus now: Grok, Anthropic and Cursor

Colossus started as a private training cluster for Grok. In 2026 it became a cloud business, and that changes how you should read it.

- **Grok:** SpaceX says [Grok 5 is being trained at COLOSSUS II](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), and the next expansion phase is designed for that model.
- **Anthropic:** in May 2026, Anthropic signed cloud services agreements covering Colossus 1 and 2, agreeing to pay [$1.25 billion per month through May 2029](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).
- **Cursor:** SpaceX signed a compute deal with Cursor in April 2026, then [completed the acquisition on August 14, 2026 at an implied $60 billion equity value](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).

The money shows up fast. SpaceX's SI segment reported [$7.6 billion of revenue in Q3 2026, up 195% sequentially](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), driven mainly by new cloud services agreements and the Cursor deal.

So does the spending. SI capex was [$16.4 billion in Q3 2026 alone](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), and SpaceX guided to [$35 billion to $40 billion of total Q4 capex](https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/?ref=insidedeeptech.com) for the Memphis and Southaven build.

Compare that with the other gigawatt project in the news. OpenAI's flagship campus runs on a very different partner and financing model, covered in Inside Deep Tech's [OpenAI Stargate full guide](https://www.insidedeeptech.com/openai-stargate-abilene-sites-power-full-guide/).

## What comes next: 3 GW, Vera Rubin and Colossus 3

SpaceX's guidance is aggressive even by AI standards. The CFO said on the Q3 call that the company expected [nearly 2.2 GW of nameplate compute by the end of October and 2.5 to 3 GW by year-end](https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/?ref=insidedeeptech.com).

Musk went further on the same call. He said he'd be surprised by [less than 13 GW of compute by the end of 2027](https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/?ref=insidedeeptech.com), and he described a planned training cluster of [440,000 Vera Rubin systems by the end of Q2 2027](https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/?ref=insidedeeptech.com).

On the ground, the next steps are already filed. DCD reported a [$659 million building on a parcel next to Colossus 2](https://www.datacenterdynamics.com/en/news/musks-xai-gets-go-ahead-for-41-natural-gas-turbines-in-mississippi-to-power-colossus-data-centers/?ref=insidedeeptech.com) and a third data center in Southaven, and Epoch AI projects the site reaching [about 1,531 MW of IT power in Q1 2027](https://epoch.ai/data/ai-data-centers/directory/colossus-2?ref=insidedeeptech.com).

Treat the 2027 numbers as targets, not plans of record. Musk himself said on the call that he doesn't expect everything to go right.

## Honest limits: where the Colossus 2 story breaks down

Colossus 2 is a genuine engineering feat. It also has five weak spots you should weigh before quoting its numbers.

### 1\. Nameplate gigawatts are not measured load

An independent analysis of xAI's turbine records and pipeline gas nominations estimated Colossus 2's chips drew [about 0.55 GW of IT power in June 2026, about half the roughly 1.1 GW implied by filings](https://www.isambrooks.com/p/wheres-the-ai-power-boom?ref=insidedeeptech.com). It also found turbines able to make 1.45 GW running well below capacity.

That doesn't prove the GPUs were idle. It does mean "first gigawatt training cluster" describes what's installed, not what's running.

### 2\. The power plant is under legal challenge

In April 2026, the NAACP sued xAI, alleging the [mobile gas turbines powering COLOSSUS II violate the Clean Air Act](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), and sought a preliminary injunction in May. At Colossus 1, the SELC said xAI had operated [dozens of unpermitted turbines](https://www.selc.org/press-release/groups-appeal-permit-for-xais-south-memphis-data-center-decisions-around-unpermitted-methane-gas-turbines/?ref=insidedeeptech.com), and Ars Technica reported [24 turbines on site against a 15-turbine permit](https://arstechnica.com/tech-policy/2025/07/xai-gets-an-air-permit-to-power-its-supercomputer-but-pollution-fears-remain/?ref=insidedeeptech.com).

### 3\. Customer concentration and short exits

Anthropic's contract is huge, but the S-1 says the agreements can be [terminated by either party upon 90 days' notice](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com). A single tenant renting a whole campus on that clause is revenue with a fuse attached.

### 4\. Estimates fill the disclosure gaps

SpaceX doesn't break out GPUs, IT power or cost by site. Epoch's figures are model-based estimates from satellite imagery, and the GPU count comes from social posts. Both could move.

### 5\. Gas lock-in

Simple-cycle turbines are fast to deploy but burn more fuel: one analysis notes they take [about 46% more gas per kilowatt-hour than a combined-cycle plant](https://www.isambrooks.com/p/wheres-the-ai-power-boom?ref=insidedeeptech.com). Musk said on the Q3 call that SpaceX also [wants solar and wind](https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/?ref=insidedeeptech.com), but it hasn't disclosed a timeline.

🧭

Inside Deep Tech's take: Colossus 2 proves that power, not chips, is now the gating item for frontier AI, and that a builder willing to own its own generation can move several times faster than the grid. The cost is regulatory and community risk that hyperscalers have chosen not to take. Expect copycats, and expect the same lawsuits to follow them.

## What Colossus 2 means for you

- **If you run AI infrastructure:** the 91-day build is a benchmark for brownfield sites with on-site gas and batteries, not for grid-tied greenfield campuses.
- **If you buy compute:** large blocks of GB200 and GB300 capacity are now rentable outside the big three clouds, but read the notice periods.
- **If you invest:** separate nameplate compute from utilization, and watch SI revenue per gigawatt as the real signal.
- **If you work on policy or permits:** Southaven is the template other gas-first AI campuses will cite, for better or worse.

[Behind-the-Meter Generation for AI Data Centers: A Full GuideGas, fuel cells, batteries and SMRs for AI campuses that can't wait for the grid, with the permitting and air-quality limits.![](https://www.insidedeeptech.com/favicon.ico)Inside Deep Tech](https://www.insidedeeptech.com/behind-the-meter-generation-ai-data-centers-full-guide/)

[Read the OpenAI Stargate Full Guide](https://www.insidedeeptech.com/openai-stargate-abilene-sites-power-full-guide/)

---

## FAQ

#### What is xAI Colossus 2?

Colossus 2 is SpaceX's second AI training campus, spanning Tulane Road in Memphis and Stateline Road in Southaven, Mississippi. It launched with [about 110,000 NVIDIA GB200s](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com), and SpaceX's May 2026 filing said it was training Grok 5.

#### How many GPUs does Colossus 2 have?

SpaceX hasn't filed a site-level count. Musk's mid-2026 breakdown, quoted in [this analysis](https://www.isambrooks.com/p/wheres-the-ai-power-boom?ref=insidedeeptech.com), was 110,000 GB200s and 440,000 GB300s, and Epoch AI estimates [about 1.11 million H100-equivalents](https://epoch.ai/data/ai-data-centers/directory/colossus-2?ref=insidedeeptech.com).

#### Is Colossus 2 really a gigawatt?

On installed capacity, yes: SpaceX reported [1.7 GW of nameplate compute across its sites on September 30, 2026](https://www.sec.gov/Archives/edgar/data/1181412/000162828026066482/spaceexplorationtechnologi.htm?ref=insidedeeptech.com). On measured load, one analysis put June 2026 chip draw at [about 0.55 GW](https://www.isambrooks.com/p/wheres-the-ai-power-boom?ref=insidedeeptech.com).

#### How is Colossus 2 powered?

Mostly by natural gas turbines in Southaven, Mississippi, where regulators approved [41 turbines rated at 1.2 GW](https://www.datacenterdynamics.com/en/news/musks-xai-gets-go-ahead-for-41-natural-gas-turbines-in-mississippi-to-power-colossus-data-centers/?ref=insidedeeptech.com). Tesla Megapacks buffer load swings and grid supply is meant to grow over time.

#### Does Anthropic use Colossus?

Yes. Anthropic agreed to use [all of Colossus 1's capacity, over 220,000 GPUs](https://www.anthropic.com/news/higher-limits-spacex?ref=insidedeeptech.com), and SpaceX's S-1 says its contracts span Colossus 1 and 2 at [$1.25 billion per month through May 2029](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com).

#### What's the difference between Colossus 1 and Colossus 2?

Colossus 1 is a converted factory with [about 100,000 H100s at launch](https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm?ref=insidedeeptech.com). Colossus 2 is a newer multi-building campus built around GB200 and GB300 racks with its own gas power plant.

## What to watch next

Three dates will tell you whether Colossus 2 keeps its lead. First, SpaceX's Q4 2026 report: did nameplate compute actually hit [2.5 to 3 GW](https://www.fool.com/earnings/call-transcripts/2026/10/20/spacex-spcx-q3-2026-earnings-call-transcript/?ref=insidedeeptech.com)? Second, the NAACP case over the Southaven turbines. Third, the first Vera Rubin racks in Memphis in 2027.

If you're planning your own AI campus, start with the power plan, not the GPU order. Colossus 2 shows that's where the schedule is won or lost.