Elon Musk said SpaceX could begin launching the first units of Starmind, its proposed network of solar-powered orbital data centres, in the third quarter of 2027, speaking at the 33rd Annual Baron Investment Conference. Musk framed the date as an estimate rather than a committed launch schedule, according to gsmdome, which reported the remarks on 9 November 2026.
Days earlier, Ron Baron, chief executive of the US investment firm Baron Capital, told CNBC on 6 November that SpaceX is expected to start orbital launches in 2027, with significant deployments projected for 2028, as reported by Crypto Briefing.
Everything that follows comes from three published accounts of public remarks and company specifications. None of the outlets quotes a SpaceX document committing to a launch date.
Key facts
- The programme is referred to as Starmind, or AI1 — both labels describe the same SpaceX orbital AI compute initiative, per Crypto Briefing.
- The satellites are intended to hold AI compute in orbit, powered by near-continuous solar energy and connected to the Starlink network through laser links, gsmdome reported.
- SpaceX applied to the US Federal Communications Commission in January 2026 for approval of up to 1 million orbital compute satellites; Crypto Briefing cautioned that the request signals intent, not permission.
- No regulatory approvals tied to the plan had been confirmed as of November 2026, and no customer contracts had been confirmed either, the same report said.
Why it matters
Musk’s stated goal is to support AI infrastructure on the scale of hundreds of gigawatts per year, which he argues may be difficult to build using only terrestrial facilities, gsmdome reported. Baron pointed to power access, land use and community resistance as the limits that space could sidestep, Crypto Briefing said. The outlet noted that Baron Capital’s chief executive has been a prominent public backer of Musk’s companies, so his comments carry weight but also reflect a believer’s view.
None of the three reports puts a figure on what it would cost to lift the racks. SpaceX’s heavy-lift programme continues in parallel: Starship, whose next test flight is targeted for mid-November.
Nor does any of the reports describe a change to Starlink satellite internet service — not to coverage, not to any Starlink price. The connection they draw is technical: laser links between Starmind satellites and the Starlink constellation. SpaceX’s AI work is already visible elsewhere in the business — its Grok assistant is already answering about 20,000 Starlink support calls a day, and SpaceX’s AI unit overtook Starlink in its latest quarter.
Where the sources disagree
Timing
The reports do not line up on when the first launch is meant to happen, and the figure attributed to Musk has moved:
- 24 August 2026: Musk posted to X confirming a fourth-quarter 2027 Starmind orbital launch, with what he called “significant scale” to follow in 2028, Tech Times reported.
- 6 November 2026: Baron’s CNBC remarks gave only years — 2027 and 2028 — according to Crypto Briefing.
- 9 November 2026: gsmdome put Musk’s Baron conference estimate in the third quarter of 2027.
Power
gsmdome said SpaceX has presented an AI1 satellite capable of up to 250 kW of peak computing power. Tech Times cited SpaceX’s published AI1 specifications as a 150-kilowatt satellite with a 70-metre wingspan, rated at 150 kW peak and 120 kW average continuous compute — power ratings the outlet said roughly match one Nvidia GB300 rack’s terrestrial draw. Crypto Briefing reported first-generation peak compute of 120-150 kW per satellite, with initial capability around 120 kW, and solar arrays said to achieve approximately 250 W/m² efficiency.
The Nvidia hardware behind it
The Hot Chips announcement
At Hot Chips 2026 in Palo Alto on 25 August, Nvidia announced SpaceXAI’s deployment of its Vera CPU to power the next generation of Grok’s agentic AI infrastructure, Tech Times reported, making SpaceXAI the second hyperscaler after Meta to commit to the chip outside full Vera Rubin rack configurations. The first Starmind generation is set to use a space-optimised version of Nvidia’s Vera Rubin NVL72 system.
The NVL72 rack
That rack integrates 72 Rubin GPUs and 36 Vera CPUs in a single enclosure connected by NVLink 6’s 260-terabyte-per-second all-to-all fabric, with no internal copper cables, according to Tech Times. The cable-free midplane design cuts assembly time from 100 minutes to five compared with the prior generation; the outlet said those properties become mission-critical in orbit, where cables cannot be rerouted after launch and mechanical connections can fatigue during a Starship ascent. Musk described the space-optimised design as “significantly simpler, lower cost, denser and lighter than a traditional rack.”
The sceptics
Tech Times reported that the standard critique of orbital AI data centres — articulated by SoftBank’s Masayoshi Son and echoed by OpenAI’s Sam Altman — focuses on all-reduce latency: distributed AI training needs GPU-to-GPU synchronisation that takes microseconds on terrestrial InfiniBand networks but cannot go below roughly 25 milliseconds at orbital distances, even over laser inter-satellite links. The research firm SemiAnalysis concluded in June 2026 that the maths does not close for large-scale distributed training in orbit, the outlet said.
Where the programme stands
For now the work appears focused on prototype development, with SpaceX expanding manufacturing operations in Bastrop, Texas, Crypto Briefing reported. gsmdome said SpaceX is building a factory intended for high-volume production of the systems.
Crypto Briefing also recorded Musk’s January 2026 economic claims: that space-based solar AI data centres would become the lowest-cost option within two to three years, and that monthly deployments of orbital computing could exceed terrestrial deployments by the end of 2028. The outlet noted market scepticism about whether orbital computing can compete economically with terrestrial alternatives in the near term. The key unresolved point, gsmdome wrote, remains whether the first launches can meet Musk’s provisional 2027 target.
Frequently Asked Questions
Has any of this computing hardware actually been tested in space yet?
Musk said Nvidia GPUs and Google TPUs have already been tested in orbit without major problems, according to gsmdome’s report of his Baron conference remarks. The outlet did not say when those tests took place or on which spacecraft.
How would an orbital data centre be cooled?
Ron Baron highlighted passive cooling as an advantage because radiators can shed heat without the water-hungry chillers that ground facilities often need, Crypto Briefing reported. Tech Times reported that SpaceX’s published AI1 specifications show solar arrays and liquid radiators providing roughly 110 square metres of heat-rejection surface.
What are the milestones to watch next?
Crypto Briefing lists FCC action on the one-million-satellite application, the first prototype launches in 2027, and any signed customer contracts. It also points to the terrestrial side: SpaceX is aiming for 5-10 GW of ground-based data centre capacity by 2027, which the outlet says will show whether the company can execute on infrastructure it can already build.
Could software written for Grok on Earth run on a Starmind satellite?
Tech Times said the shared Vera architecture means software and tools developed for Grok’s ground-based agents could in principle run on Starmind satellites without a complete rewrite. Nvidia framed the approach as a single unified architecture rather than two separate product lines.
![Using Your Own Router with Starlink: Bypass Mode [2026]](https://starlink-news.com/wp-content/uploads/2026/09/using-your-own-router-with-starlink-bypass-mode-20-f6c09f95-120x86.png)
