Starlink Gen 3 orbital AI compute is not a live service on September 23, 2026. It is a pending FCC application covering up to 100,000 satellites, first accepted for review on September 18, 2026, with publicly claimed numbers that outrun anything SpaceX has launched or demonstrated.
Starlink Gen 3 orbital AI compute: what actually exists today
Starlink Gen 3 orbital AI compute is a pending regulatory request, not a running system. SpaceX has applied to the US Federal Communications Commission for authority to deploy up to 100,000 Gen 3 satellites, and that application entered the review process on September 18, 2026. No Gen 3 satellite with an onboard Nvidia rack has been demonstrated publicly.
The distinction matters because several numbers attached to this story arrive from different places with different evidentiary weight. A filing proves that a company asked for permission. A CEO post proves that a company said something. Neither proves that a satellite carries a working AI accelerator, that a link runs at 10 Tbps, or that 100 GW of orbital compute gets built by 2030.
The story also moved on several fronts in a single week. Elon Musk announced the payload claim, the Starlink V5 terminal went into production, a direct-to-cell post landed on September 17, 2026, and xAI released a new Grok model mid-conversation. Each belongs to a different product line, and treating them as one announcement is how the confusion starts.
SpaceX has not published a Gen 3 payload datasheet with power, thermal, compute, and link budgets that an outside engineer could check. Until that exists, treat every figure below as a claim with a named source rather than a specification.
What SpaceX filed with the FCC on September 18, 2026
The filing requests authority for up to 100,000 Gen 3 satellites across two very low Earth orbit altitude groups. Very low orbit shortens the round trip to a ground terminal, which is the mechanism behind the lower-latency and higher-capacity language in the application. The application's acceptance starts a review clock; it does not grant a licence.
Two altitude groups, rather than one, is the detail worth keeping. Satellites at different altitudes carry different latency and different orbital-debris and conjunction profiles. A constellation at that scale has to be coordinated internationally through the ITU as well as approved domestically, and a granted US licence would still leave those steps outstanding.
One caution about the satellite count. A request for authority to deploy up to 100,000 satellites is a ceiling the operator asks for, not a build plan with a schedule. Watching what SpaceX actually manufactures, launches, and files as amendments tells you more than the headline number.
What Elon Musk actually said about the payload
The onboard-compute claim traces to a post from Elon Musk, SpaceX's CEO, describing connectivity of roughly 10 Tbps per satellite in both directions with a path toward 100-plus Tbps, a payload power figure of 250 kW, and what he described as a SpaceX-designed Nvidia Vera Rubin and VL72 computer. Aaron Burnett of Mach 33 characterised it as effectively combining Starmind with Starlink in much larger satellites, with an order-of-magnitude jump in bandwidth and enough power for an NVL72 Rubin rack, and Musk replied in the affirmative.
That exchange did not reveal which Nvidia product is meant. The name reads as a compression of three separate things: Vera Rubin, the successor generation to Nvidia's Blackwell data-centre platform; the GB300 NVL72, a rack-scale system that Nvidia describes as 72 Blackwell Ultra GPUs in one liquid-cooled rack; and the NVL72 form factor itself. A single Gen 3 satellite drawing 250 kW cannot physically contain a full NVL72 liquid-cooled rack, so the comment should be read as intent about the architecture family rather than a bill of materials.
Anything about delivery timing, prioritisation, or engineering samples reaching SpaceX earlier than other customers came from a guest on the video describing how he expects the two companies to behave. His reasoning: SpaceX has committed to Nvidia as an exclusive, so getting hardware early means it gets deployed fast, which pulls forward larger orders and feeds field data back to Nvidia. That is the speaker's expectation, not an announced arrangement.
SpaceX's own Starlink page lists the consumer service and terminals that exist now. Nothing on it describes an onboard AI payload, which is why the compute claim sits at the level of a CEO statement rather than a product page.
Why the 100 GW orbital compute figure is a projection, not a plan
A SpaceX executive, the company's CFO, said SpaceX is on track to deploy 100 GW of orbital compute per year by the end of the decade. Elon Musk agreed with that figure. It is a revenue and capacity projection from a company officer, and it is the weakest kind of number in this story because it describes a build rate about four years out, in a unit that mixes power with business model.
The arithmetic behind the claim as presented in the video does not hold together. The speaker said current orbital compute is under 5 GW, then said it is about 2 GW by the end of 2026 and a promised 5 to 10 GW by the end of 2027, then said each 2 GW represents roughly US$100 billion per year in revenue. Those figures contradict each other rather than describing a trajectory, and no SpaceX source was cited for the revenue multiplier.
The same segment compared the 100 GW figure with roughly 40 GW of terrestrial compute projected for 2027. That comparison is category-confused: Earth-based figures cover installed data-centre capacity across many operators and technologies, while the orbital figure is a single company's stated annual deployment target. Repeating the two side by side implies a rivalry the numbers do not support.
| Figure | Value | Source | Status |
|---|---|---|---|
| Gen 3 satellites requested | Up to 100,000 | FCC application, accepted September 18, 2026 | Filing, not a build plan |
| Connectivity per satellite | 10 Tbps bidirectional, path to 100-plus Tbps | Elon Musk post | Claim |
| Payload power | 250 kW | Elon Musk post | Claim |
| Orbital compute build rate | 100 GW per year by 2030 | SpaceX CFO, endorsed by Musk | Projection |
| Current orbital compute | Under 5 GW, about 2 GW by end of 2026 | Video segment | Contradicts the 100 GW framing |
| Terrestrial compute, 2027 | About 40 GW | Video segment | Installed capacity across many operators |
The video's broader context is worth keeping too. As AI demand grows, the argument goes, the connectivity load shifts from human users to machines: humanoid robots, autonomous vehicles, and always-on agent-driven connection points. Starship's larger payload capacity is presented as the enabler for direct-to-cell and direct-to-device service as well as orbital compute.
For context on the scale involved, the US Energy Information Administration publishes national electricity generation, transmission, and capacity data. Comparing a national grid to a satellite constellation requires conversion assumptions that no one has published for this case.
How Starlink Gen 3 orbital AI compute would need to work
An orbital compute node needs to generate power, reject heat, and get data to the ground. The video's own description is a largely unchanged satellite bus with a different payload, more solar array, and a radiator that unfolds or flips out. That is the right shape of answer. Each of those three requirements is a hard constraint rather than a detail.
- Generate power. The payload figure cited is 250 kW per satellite, up from what the guest described as an earlier target nearer 100 kW. Solar array area scales with that number.
- Reject heat. Waste heat in orbit can only leave by radiation, so radiator area grows with the compute load.
- Hold orbit. A bigger radiator and array mean more mass and more drag at very low altitude, which requires more propulsion to maintain the shell.
- Get results down. Onboard inference for image or sensor processing can return small outputs over a modest radio link. Training a large model needs substantial bidirectional bandwidth and tight synchronisation across many nodes.
Power and heat scale together, and launch economics set the practical ceiling. That is why the thermal design is usually the part that decides how much compute a satellite can carry. The public framing has not said which workload is intended, and that choice drives everything else.
Starlink and Starmind: one product or two
The video's central argument is that Starlink and Starmind are converging into one satellite. The evidence for convergence is a CEO post and a short reply affirming a commentator's framing, and the guest immediately qualified it, saying SpaceX may still build two separate products because adding compute changes the cost structure.
The unresolved question is SKU count. A constellation of data-only satellites is cheaper per unit and simpler to operate. A constellation carrying accelerators, extra solar, and deployable radiators costs more, weighs more, and complicates thermal and attitude control. Nothing public resolves whether customers rent connectivity, compute, or both on the same orbital hardware.
The guest's read is that this is schedule pull-in and de-risking rather than a finished product decision: SpaceX has the Gen 3 satellite in design and an AI platform for future space-based data centres, so bolting compute onto an existing bus is cheaper than starting from nothing. The positions below separate what was claimed from what is verifiable.
Starlink V5 terminals and V2 direct-to-cell are shipping or announced, and different from Gen 3
Two other announcements in the video are easy to confuse with the orbital-compute story. Elon Musk said the Starlink V5 terminal is in production, describing a smaller and lighter unit with greater power efficiency and speeds up to 375 Mbps for home internet. A terminal is ground equipment, and its production says nothing about satellites.
The V5 terminal is manufactured at SpaceX's Bastrop, Texas factory, where resin arrives at one end of the line and finished antennas leave the other. The guest noted that printed circuit board assembly of this kind has largely moved out of the US, and that the economics only work with high volume and a small number of SKUs. Change the mix constantly and the automation advantage disappears.
A post dated September 17, 2026 described a next-generation direct-to-cell service with an Australian partner carrying 5G-class speeds from space, with the V2 satellite network expanding on V1 at 100 times the data density and 100 times more bandwidth. V2 is a separate constellation generation for direct-to-device service; the 100x figure is a claimed improvement, not a measurement. Starship's larger payload bay is the stated enabler for that generation.
Keep the three generations apart. The V5 terminal is customer premises equipment, V2 is a direct-to-cell satellite generation, and Gen 3 is the high-capacity satellite generation attached to the 100,000-satellite filing. SpaceX's Starlink page is the primary source for what the consumer service offers today, and Starlink's own direct-to-cell page is the primary source for that service.
The model and vendor claims in the same video
The video also covered a newly released Grok model from xAI, alongside Claude from Anthropic models from OpenAI, with pricing and benchmark comparisons. xAI released Grok 4.7 during the recording, positioned as its most powerful model for coding and knowledge, designed to work longer on difficult tasks and handle more context. Musk walked back the claim that it would beat everything else, saying that would have to wait for Grok 5.
According to the slide shown, Grok 4.7 scored higher than rivals on legal and electrical engineering, sat behind Fable 5.1 and GPT on software engineering, and trailed on multi-hour office work and software terminal work. Grok already prices below GPT from OpenAI and Claude from Anthropic the input-token cost was described as much cheaper than Fable. Those comparisons came from a slide, not from a published evaluation report.
The guest's argument is that cost per unit of intelligence is what matters, not benchmark rank alone. Frontier models from OpenAI and Anthropic still produce more intelligence, but they cost much more, and cheaper alternatives keep closing the gap. At some point a crossover becomes rational for buyers, which is the competitive pressure the frontier labs face.
Model releases move faster than articles. Exact version numbers, prices, context limits, and benchmark standings from a September 2026 video should be checked against the vendors' current release notes, pricing pages, and system cards before anyone repeats them. A cheaper per-token price is not the same as a lower cost per completed task, and neither is independent of the specific workload.
Nothing here changes the orbital-compute picture. A model's benchmark scores do not establish that a satellite can host it.
What to watch to track this story
Four documents settle most of these questions: an FCC grant or amendment order, a published Gen 3 payload specification, a launch manifest showing Gen 3 satellites flying, and independently measured terminal throughput. Until those appear, orbital AI compute remains a claim under review.
A practical reading habit helps here. Ask which document a number came from, and whether that document is a filing, a press statement, a slide, or a measurement. The orbital-compute story mixes all four, and the strongest evidence in it is a regulatory acceptance rather than a working satellite.
Also watch what is not in the filing. No published power budget, no thermal analysis, no radiation-tolerance data for accelerators, and no cost per kilogram to orbit for a compute payload have surfaced. Absence of those documents is itself information: it means nobody outside SpaceX can check the 250 kW figure against a real design.
FAQ
Is Starlink Gen 3 approved yet?
No. SpaceX filed for authority to deploy up to 100,000 Gen 3 satellites, and the FCC accepted that application for review on September 18, 2026. Acceptance starts a process rather than granting a licence, and international spectrum coordination would still be needed.
Will each Gen 3 satellite really carry an Nvidia supercomputer?
Elon Musk described a SpaceX-designed Nvidia Vera Rubin and VL72 computer in a post. No physical specification has been published, and a satellite drawing 250 kW cannot contain a full liquid-cooled NVL72 rack as sold for data centres.
What does 10 Tbps per satellite mean?
It is claimed bidirectional capacity, not guaranteed throughput for any user. Aggregate satellite capacity is shared across terminals in view, so per-user speed depends on how many terminals compete for that capacity at the same time.
Is 100 GW of orbital compute per year realistic by 2030?
It is a projection from SpaceX's CFO, which Elon Musk endorsed. SpaceX has not published a procurement, launch, thermal, or power plan showing that build rate, and the supporting arithmetic in the video contradicts itself across successive years.
Does the Starlink V5 terminal have anything to do with orbital AI?
No. The V5 terminal is customer premises equipment for home internet, announced as in production with speeds up to 375 Mbps. It is ground hardware and says nothing about onboard compute.
Could Nvidia racks be launched into orbit at all?
The physics allow it in principle. Power generation, radiative heat rejection, radiation tolerance for accelerators, and the cost per kilogram to orbit set the practical limits, and none of those budgets has been published for this design.
Why did the video say Starlink and Starmind are the same thing?
That was an interpretation of Musk's post and his brief reply agreeing with a commentator. The guest on the video immediately said SpaceX may still build two separate products for cost reasons, and that data-only V3 satellites might remain cheaper to operate.
Is the V2 direct-to-cell network the same as Gen 3?
No. V2 is a direct-to-cell satellite generation, and the September 17, 2026 post describes 100 times the data density over V1. Gen 3 is the high-capacity satellite generation attached to the 100,000-satellite filing. The V5 terminal is ground equipment.
Do the Grok, Claude OpenAI comparisons in the video matter for Starlink?
They matter for the model market, not for orbital hardware. Those comparisons came from a slide, and versions and prices change quickly enough that they need rechecking against each vendor's current documentation.
Why did the article say orbital compute is under 5 GW when the target is 100 GW?
Because those figures describe different things and come from different sentences in the same segment. The under-5 GW figure describes what exists now, the 100 GW target describes an annual build rate by 2030, and the same segment also said 2 GW by end of 2026. The numbers are inconsistent, which is why they are treated as claims rather than a trajectory.
From satellite filings to written analysis
Regulatory applications, design commentary, and investor framing rarely arrive in a form anyone wants to read, and yet the details that decide a story usually sit inside that material. The gap between a 19-minute discussion and a usable written analysis is mostly structure and source-checking.
If your own YouTube videos carry that kind of knowledge, whether it is an explanation, an interview, or an opinion built on primary documents, Skalablog turns the video into a written article: paste a YouTube URL, get a transcript, generate the article, and edit it into something you would sign your name to.
Note on authorship and scope
This article is an independent analysis written by Gustavo Dev Doido and is based on the public filings, company statements, and vendor documentation linked above. It is not affiliated with SpaceX, Nvidia, xAI, Anthropic OpenAI, and it is not investment advice. More work from the author is available at crazystack.com.br.
Every forward-looking figure attributed to a company executive or a video guest is labelled as such. Where the underlying document could not be located, the claim is described rather than repeated as fact, and no benchmark in this article was reproduced by the author.
If the numbers in this story are hard to pin down, the same problem applies to the analysis sitting in your own video library: the value is real, but it lives in a format nobody can quote, search, or cite.
Fork this article
Start a new branch from the same video, shaped your way. You keep the credit; the original keeps the attribution.
A fork in another language is filed as a translation of this article, so the two pages point at each other. You can unlink it later from the editor.
0/240
You are creating
- Format
- For
- Language
- Source
- Your angle
You will be asked to sign in before it is generated.
Buy credits