Musk Claims AI Will Complete All Digital Work by End of Next Year, Model Competition Focuses on Algorithms, Scale Competition on Power
Elon Musk predicted during a video speech at the G20 finance ministers and central bank governors meeting in North Carolina that AI software will become extremely powerful within 12 to 18 months, capable of completing all digital work that does not require human modification of physical objects by the end of next year at the latest.
He stated that AI programming capabilities will reach "Stockfish level" at some point next year: just as top human chess players cannot compete with Stockfish in regular chess games, human programmers will struggle to compete with AI in terms of software writing speed and capability. This is Musk's personal prediction and not a product roadmap or validated capability metric provided by OpenAI, xAI, or other model companies.
Musk estimates that AI could increase the global economic scale by 20% to 30%, equivalent to an additional $20 trillion to $30 trillion in output each year. This figure assumes that AI capabilities will widely penetrate software, engineering, finance, media, administration, R&D, and other digital work, and that society can simultaneously complete business process transformations, infrastructure construction, regulatory adaptations, and labor reallocations.
Regarding humanoid robots, Musk predicts that there will be over 1 billion humanoid robots globally within a decade, each with a productivity at least five times that of humans, resulting in a total productivity exceeding that of all humans combined. He summarized the practicality of robots as the product of AI software, AI chips, and electromechanical flexibility, believing that robots will eventually participate in manufacturing other robots, triggering recursive production expansion.
He warned that the near-term bottleneck for AI development is power rather than chips. Musk stated that by 2027, the available power corresponding to AI chips may face a shortfall of at least 15GW; chip production capacity is expected to grow at an annual rate of about 40% to 50%, while the annual growth rate of available power outside of China is only 10% to 20%, potentially leading to stranded computing power where "servers are available but cannot be powered on" in data centers.
Musk mentioned that SpaceX is addressing some of the power supply issues by building its own power generation facilities and has leased computing power to several AI companies. Public reports indicate that Anthropic has agreed to pay SpaceX $1.25 billion per month until May 2029 for computing resources from the Colossus and Colossus II training data center clusters; additionally, it is reported that Anthropic plans to lease AI computing power from Nscale valued at approximately $45 billion.
Musk called on regulators to adopt a "default legal rather than default illegal" approach to new technologies like AI. The logic is that if licensing, grid connection, construction approvals, energy equipment delivery, and regulatory restrictions cannot keep pace with chip deployment speeds, the constraints on the AI industry will shift from model development to power, transformers, land, cooling, and data center construction cycles.
In market mechanisms, buyers include xAI, OpenAI, Anthropic, Google, Meta, Microsoft, and enterprise customers that need to train and deploy large models; sellers range from chip manufacturers like Nvidia to power developers, gas turbine and transformer manufacturers, data center operators, liquid cooling suppliers, network equipment providers, and land asset holders with grid connection permits. Capital is shifting from GPU procurement to self-built power generation, long-term computing power leasing, grid access, and cooling infrastructure; suppliers with delivered power and operational clusters benefit, while projects that only hold chip orders but cannot secure energy and infrastructure face idle risks.
Source: Public Information
ABAB AI Insight
Musk's "Stockfish level" metaphor continues his long-standing radical assessment of AI capability curves. Stockfish represents a closed system where machines have long surpassed top human players under fixed rules and clear win/loss outcomes; however, software engineering includes ambiguous requirements, legacy systems, cross-team communication, accountability, security audits, and operational maintenance in production environments. GitHub Copilot, Cursor, Claude Code, and OpenAI coding models have integrated code generation into the development process, but there remains a gap between "being able to write code" and "independently taking full responsibility for software delivery," which still involves requirement definition, testing, deployment, incident response, and business decisions. Musk's prediction effectively bets that these non-code aspects will also be rapidly toolized by proxy systems.
The capital pathway has already shown that AI competition is shifting from model financing to infrastructure locking. Anthropic's commitment to SpaceX for $1.25 billion per month until May 2029 for computing power indicates that model companies are no longer just purchasing cloud services on demand but are beginning to reserve scarce training capacity with multi-year, multi-billion dollar contracts; its leasing arrangement with Nscale for approximately $45 billion further reflects this trend. Capital initially flows to GPUs, servers, and data centers, and must subsequently lock in power, land, transformers, liquid cooling, and networks to convert chip procurement into operational model training and inference revenue.
Historically, railroads, power grids, internet backbones, and cloud computing have all experienced a process of "first building general infrastructure, then scaling applications." The difference in the AI cycle is that demand and hardware procurement are exploding simultaneously, while power construction, transmission approvals, and large transformer manufacturing cycles are measured in years. Amazon Web Services gained scale advantages early on by centrally building data centers and renting computing resources on demand; if SpaceX can provide powered training clusters through self-built power generation and data centers, its competitive advantage will not only come from the number of GPUs but also from the certainty of production bypassing public grid queues.
This represents a restructuring of the industrial chain. In the past, software companies primarily purchased cloud resources, and cloud vendors mainly procured servers; AI extends the chain to "model—chip—cabinet—network—cooling—transformer—grid or power plant—fuel and permits." Musk's proposed 15GW shortfall means that model capabilities are no longer solely determined by algorithms and parameters, but by who can coordinate physical engineering and energy assets. When chip growth outpaces available power, power access rights will gain higher marginal pricing power than chip orders; if policies maintain a "default illegal" approval logic, AI capital expenditures will first be stalled by construction permits and energy delivery rather than research laboratories.
ABAB News · Cognitive Laws
Model competition focuses on algorithms, while scale competition focuses on power.
Chips can be bought, but powered capacity cannot.
Automation will first replace software, then compete for the real world.