Elon Musk: SpaceX AI Will Surpass OpenAI in Six Months
Elon Musk recently stated that SpaceX's AI development has only been underway for about three years, while Anthropic has been around for about six years and OpenAI for about ten years. If the current acceleration continues, SpaceX could achieve a leading position in about six months. He also mentioned that once intelligence far exceeds what is required for certain tasks, further investment becomes less meaningful. Hardware deployment remains challenging, and large-scale computing power expansion is extremely difficult, but SpaceX has proven adept at building infrastructure and will continue to improve.
SpaceX has integrated the former xAI, making AI its third business segment. In the second quarter, total revenue was approximately $7.8 billion, with AI contributing about $2.56 billion to $2.6 billion, a year-on-year increase of about 247%. The connectivity business (including Starlink) generated about $4.29 billion, and aerospace operations brought in about $962 million. Musk stated in an all-hands meeting in August that AI revenue would "probably surpass the rest of the company" in September, and would significantly exceed it in the fourth quarter, claiming that AI could account for about 99% of the company's value in four to five years. He set a target of 10 gigawatts of computing power by the end of 2027, with a valuation of $30 to $50 per watt, corresponding to annual revenues in the range of $300 billion to $500 billion.
The Memphis and Southaven cluster has a nameplate capacity of about 1.4 gigawatts. Anthropic has leased capacity from Colossus, which was publicly stated to be about $1.25 billion per month. Musk clarified in May that the initial lease is for 180 days, after which either party can terminate with 90 days' notice; short-term leases are a requirement from SpaceX to reclaim capacity during tight computing power situations. The agreement with Google is about $920 million per month for approximately 110,000 NVIDIA GPUs, with plans to start leasing in the fall. Training will remain on the ground, while daily inference is envisioned to move to an orbital data center. The Grok route was mentioned to enter internal testing at SpaceX and Tesla at version 4.5, with plans for a tighter version release schedule; Starlink and Starship engineers have been described by Musk as sources of acceleration.
Mechanically, the purchase is for computing power that can tie power plants, clusters, and launches to the same balance sheet, while selling relies on the speed of lab expansion that depends solely on cloud vendors. The event-driven aspect is that the leader has compressed the timeline for "catching up with models" to six months and emphasized the priority of self-use with the ability to reclaim leases. Funds from tenants like Anthropic and Google enter the cluster's depreciation and the next batch of cabinets; benefiting SpaceX, which can expand on its own electricity and land, while putting pressure on tenants that rely on the same cluster for training—short-term lease terms turn their capacity into retrievable inventory. If model rankings are truly rewritten within six months, it hinges on the speed of hardware delivery continuing to outpace competitors' algorithm accumulation.
Source: Public Information
ABAB AI Insight
Musk frames the years of OpenAI and Anthropic as a disadvantage: they started earlier, while SpaceX compensates with three years of launches and power station capabilities. A six-month lead is a rallying call for shareholders and engineers, not an indication that rankings have already changed. The real revenue already coming in is from computing power rentals: the AI segment generated $2.6 billion in the second quarter, mainly from cluster contracts, not Grok subscriptions. A monthly rental of $1.25 billion, if annualized, would surpass the entire company revenue level for 2025, but the 180-day plus 90-day exit clause indicates this revenue is not locked in as an annuity until 2029. After Google enters the market in the fall, the AI segment will have a better chance of exceeding Starlink in a single month.
The capital path is "first sell computing power, then train its own models." Colossus rents out externally for cash and timelines, while retaining the right to reclaim internally, effectively turning competitors into transitional tenants. If coding tools like Cursor are integrated, it would supplement Grok's product offerings on the developer side, rather than proving that the foundational model has surpassed Opus. The 10 gigawatts and $30 to $50 per watt are directly multiplying power construction into the revenue model, while orbital inference writes Starship's capacity into future marginal costs. Engineers drawn from Starlink and Starship to AI are migrating the rocket company's delivery culture to cluster timelines.
This is analogous to how Amazon uses AWS to support retail, Microsoft uses Azure to back OpenAI, and Google utilizes TPUs for both internal use and external leasing: all infrastructure companies discover that training is their best customer. The difference is that SpaceX simultaneously owns launch, satellites, and ground power stations, allowing it to integrate "where the power comes from and where the cards go" into the same timeline. The industry phase is shifting from comparing parameters and evaluations to who can first plug in gigawatts. The six-month window bets on delivery, not papers.
Structural judgment belongs to the overlap of technological substitution and capital concentration. Pricing power shifts from the release rhythm of model labs to integrated companies that can deliver gigawatts monthly. The mechanism is that the bottleneck of cutting-edge training has shifted from algorithms to power, land, and supply chains; whoever can write leases for 180 days can switch between self-use and rental. If a six-month lead occurs, it will manifest in cabinets first, and then appear on the rankings later.
ABAB News · Cognitive Laws
- Laboratories that start earlier may not outpace companies that plug in power first.
- Short-term lease terms turn tenants' computing power into landlords' retrievable inventory.
- The timeline for model leadership is often written behind infrastructure timelines.