Starmind: SpaceX and Nvidia's Plan for AI Computers in Orbit.

SpaceX has a product page for Starmind, a plan to put AI computers on satellites and send the results back through Starlink. Nvidia’s site says the AI1 compute payload uses a Vera Rubin NVL72. Elon Musk has posted two dates: a space-optimized rack aimed at orbit in the fourth quarter of next year, and volume production of a “Star Mind” computer next year. No Starmind satellite is described, on the pages reviewed here, as already in orbit.

This page is the receipt. It sits on the AI Wars hub, next to Colossus.

What SpaceX’s page publishes

Fetched October 18, 2026, spacex.com/spacexai/starmind describes an “AI1” satellite: localized compute, power and cooling in space, results beamed to the Starlink constellation by laser. The specification block on that fetch:

The same page says the satellites “use proven technology and are more streamlined than the Starlink satellites we already manufacture at scale,” that the architecture is “AI chip vendor agnostic,” and that cooling power overhead drops “by an order of magnitude” because heat radiates into vacuum. Label: Company claims on the page. Vendor-agnostic is a design sentence. The cooling comparison is a claim, not a measured test this desk has seen.

The Gigasat Factory in Bastrop “is planned to enable rapid production and deployment of thousands of AI satellites starting as soon as late 2027.” Label: Company plan. “As soon as” is the page’s hedge. Thousands is not a manifest.

Chips: “SpaceX plans to produce advanced AI chips at scale in its new fab facility, a joint project with Tesla.” That plan is Terafab. The Starmind page does not, in the text fetched here, name Vera Rubin.

What Nvidia’s page says

Fetched the same day, Nvidia’s space-computing page says: “SpaceX’s Starmind AI1 satellite compute payload is powered by NVIDIA Vera Rubin NVL72.” A separate block says the NVIDIA Space-1 Vera Rubin module “delivers up to 25x more AI compute per GPU for space-based inferencing and orbital data centers.”

Label: Nvidia product-page copy. The NVL72 sentence is Nvidia’s description of the payload. It is not a photograph of a satellite, and it is not a launch. The “up to 25x” line is Nvidia’s claim for that module, not a Starmind benchmark with a published method. The Register (Simon Sharwood, August 25, 2026) quoted an Nvidia release that the satellites “will be based on” the optimized rack, future tense, and noted that neither company had said what “space-optimized” changes inside the rack.

The Register also described a full terrestrial NVL72 as 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs, and BlueField-4 DPUs, about 4,000 pounds, drawing 120 kW. Label: The Register’s account of Nvidia’s terrestrial rack. Do not paste those counts onto the satellite. The Register said it had asked whether “space-optimized” changes the parts list, and it printed SpaceX’s then-published satellite as a 210 kW solar array. That August figure is not the “up to 250 kW peak” line on the SpaceX page this desk fetched in October. The page has moved, or the two sentences describe different power numbers. They are not one spec.

What Mr. Musk posted in August

On August 24, 2026, at 17:22 UTC, Mr. Musk posted (status 2091939113008238838), quote-posting Nvidia:

SpaceX, in partnership with Nvidia, has designed a space-optimized Vera Rubin NVL72 system for launch to orbit in Q4 next year, with significant scale in 2028

Four minutes later (status 2091940127866257601):

Our design is significantly simpler, lower cost, denser and lighter than a traditional rack

Label: Primary for what he wrote. “Q4 next year,” written on August 24, 2026, is the fourth quarter of 2027. “Designed” is a design claim. “For launch” is a target. “Significant scale in 2028” is his phrase, not a number of satellites. The second post is a comparison with a terrestrial rack. It does not list a part that was removed.

Nvidia’s quoted post the same day says SpaceX “is deploying NVIDIA Vera” for agentic AI, “From gigawatt AI factories to orbit.” Label: Primary for Nvidia’s post. “Is deploying” is the company’s present tense. It is not a statement that a Starmind satellite has reached orbit.

Mint’s August 25 story (Swati Gandhi), carrying the same sentence, says the company had earlier pointed at 2028 and that this post pulled a first launch toward the fourth quarter of 2027. Label: Reported timeline history. The post itself is the dated target.

The October post says “Star Mind”

On October 15, 2026, at 17:48 UTC, Mr. Musk posted (status 2110789964430786745):

Old friends are the best friends. Jensen rocks.

The Star Mind computer being co-engineered by SpaceX and Nvidia is far better than almost anyone expects and will reach volume production next year.

Working with the incredibly talented Nvidia engineering team is pure joy!

Label: Primary for those words. He wrote “Star Mind” as two words. The product page says Starmind. “Far better than almost anyone expects” is a boast. “Volume production next year” is a target for 2027. It is a different sentence from “launch to orbit in Q4 next year.”

The post quote-posts @firesidealpha, whose caption presents Jensen Huang on the BG2 Pod, including “I’m delighted that I love working with Elon” and “One particular new market that we’re going to go explore together is putting data centers out in space.” Label: The caption is what that account posted, and it is on the quote-post this desk opened. This desk did not open the BG2 episode. The caption is not a Huang transcript from Nvidia.

Benzinga (Ananya Gairola, October 16, 2026) quotes the “Jensen rocks” line and the “far better” line, and says the system would reach volume production in 2027. Benzinga’s page says the article “was partially produced with the help of AI tools and was reviewed and published by Benzinga editors.” The post above is the primary. Benzinga is the writeup, with that disclosure.

One million satellites is a requested maximum

On January 30, 2026, Space Exploration Holdings applied for a new non-geostationary system. The FCC Space Bureau’s public notice DA 26-113, released February 4, 2026, accepted the application for filing and sought comment. File number SAT-LOA-20260108-00016. The notice’s words: “a new NGSO satellite system of up to one million satellites to operate as the ‘SpaceX Orbital Data Center system.’” Altitudes 500 km to 2,000 km. The notice says the application calls the system a first step toward a Kardashev II civilization. That phrase is in the application, as the FCC quotes it.

Label: Confirmed as an application the FCC accepted for filing. Acceptance for filing is not a license, not a launch grant, and not a build plan for a million satellites. “Up to one million” is the requested ceiling. Comment dates on the notice ran into March 2026. This page does not have a later grant.

Benzinga’s October 16 story calls a million satellites a “proposed maximum constellation rather than a firm deployment commitment.” That characterization matches the FCC notice. The notice is the document. Benzinga, again, disclosed partial AI production.

The orbital-compute argument the company made when it said it had acquired xAI is on xAI joins SpaceX. Those gigawatt figures are the post’s arithmetic. They are not Starmind satellites on orbit. Getting a 30-meter satellite up in any number is a Starship problem. The Register’s point, that the launch vehicle was unspecified in August and that Mr. Musk’s public promises often slip, is press commentary. Label: The Register’s view, not a finding that the date will slip.

Published spec versus target

On the map

Terrestrial compute for Grok is Colossus and xAI and Grok. The laser path home is Starlink. The rocket is Starship. The chip plant named in the Starmind copy is Terafab. The corporate post that argued for compute in orbit is xAI joins SpaceX. Starmind is the satellite design sitting under that argument. A design page is not a constellation.

Sources

Checked October 18, 2026.