Elon Musk: Aiming for AI Top Spot in Six Months, Benchmarking Against Anthropic's Six Years and OpenAI's Ten Years
Elon Musk responded to concerns that "SpaceX AI is already behind," stating that the company's artificial intelligence has only been in development for three years, compared to Anthropic's six years and OpenAI's ten years. He believes that if the second derivative remains strong, they could reach the top spot in about six months. He also expressed cautious optimism about delivering a model on the scale of Fable/GPT-6 within two to three months.
The second point reduces the competition to a lower dimension: once a type of task far exceeds the required intelligence, adding more intelligence becomes meaningless, "there's no need to cram Newton-level intelligence into a toaster." The third point identifies hardware as the bottleneck: rapidly scaling up computing power is extremely difficult, and SpaceX claims to have proven its capabilities in this area and will continue to accelerate.
On the same day, he also mentioned that computing power in space will ultimately "round up to 100% of all computing power," echoing Google's plan to use Falcon 9 to send TPUs into space for testing orbital data centers. On the ground side, the financial report indicates that by the end of the year, computing power will exceed 2 gigawatts, up from 400 megawatts a year ago, 1 gigawatt in the first quarter, and 1.4 gigawatts this quarter. They aim to accumulate nearly 10 gigawatts online by the end of 2027, with power and cooling targets of 20 gigawatts, and actual figures potentially close to 15 gigawatts. The chip strategy has been set to follow only the NVIDIA Vera Rubin architecture.
After completing the largest IPO in history in June 2026, SpaceX will acquire Anysphere, the company behind the programming tool Cursor, for approximately $60 billion in stock, and will use the Memphis Colossus cluster for both its own model training and external rentals. Anthropic had previously signed a monthly cloud service agreement to use its supercomputing resources. Musk likened rocket engineering to data centers: taking a small portion of the engineering capability used to send satellites and applying it to ground facilities is like "the Yankees playing Little League baseball."
In terms of market mechanisms, what is sold is the speed of going online and gigawatt-level power, while what is bought is the narrative of model rankings and tenant contracts. Buyers of computing power include Anthropic, their own Grok/Starmind training, and future orbital payloads, while sellers are NVIDIA racks and SpaceX's self-built substations. Beneficiaries are infrastructure builders who can bundle launch, cooling, and power delivery cycles, while those under pressure are model labs that only have software and cannot rent power. Funds flow from the capital expenditures of public companies into GPUs, substations, and Starship launch pads, and then exchange for valuations through cloud leases and model releases.
Source: Public Information
ABAB AI Insight
Musk integrates the xAI narrative into the publicly listed SpaceX, using the "three years versus ten years" argument to explain the lag as a slope advantage. OpenAI started in 2015, and Anthropic split from it in 2021, with SpaceX's large-scale AI investments corresponding to the Colossus and Grok cycles, starting around 2023. The six-month target for the top spot is a bet on the second derivative: it’s not that the parameters are already larger, but that the speed of deploying gigawatts can still accelerate. The toaster metaphor gives him an escape route—if the model doesn't reach the top spot, he can still argue that the task doesn't require Newton.
The capital path subsidizes data centers with launch capabilities. Starship and Starlink write "extreme hardware construction" as a transferable skill, with the Memphis data center, space-based computing payloads, and TeraFab chip experiments all emerging from the same engineering organization. Cursor is integrated into this organization to lock the IDE workflow into proprietary models and computing power. Renting out Colossus is equivalent to using competitors' training budgets to pay for their own substations.
Similar structures can be seen with Amazon using AWS to support retail, Google using TPUs to support search and cloud, and Microsoft binding OpenAI with Azure. SpaceX is in an infrastructure expansion phase: first turning 2 gigawatts into 10 gigawatts, then discussing the next chapter with orbital cooling, while model rankings serve as milestones for the capital market, not the only product.
The structural judgment belongs to capital concentration before technological replacement. The constraints of the intelligence race have shifted from algorithm papers to transformer power and cooling tonnage. The mechanism is: whoever can schedule electricity, chips, racks, and launch windows into the same Gantt chart will own the next training window; toasters don’t need Newton, but training clusters need substations.
ABAB News · Cognitive Laws
The second derivative is a better price determinant for rankings than the year of origin.
Toasters don’t need Newton, but training clusters need substations.
Hardware is difficult, particularly in terms of power and cooling, not just slogans.