Musk Says AI Data Centers Are Limited by Physical Infrastructure
Elon Musk stated that launching large AI data centers requires not only GPUs and power supply but also the simultaneous construction of power generation facilities, transformers, liquid cooling circuits, large chillers, and complex networks for training clusters.
He previously judged that the approximately 15GW of AI computing power produced in 2027 may not be operational that year. According to him, the constraints are not about whether chips can be manufactured but whether the transformation, distribution, wiring, heat dissipation, and network facilities from the power grid or self-built power sources to server racks can be delivered on time.
The power access chain includes power generation or grid power sources, high-voltage transmission, project-specific substations, step-down transformers, switching equipment, uninterruptible power supplies, and rack distribution systems. If any link cannot be completed on time, the already purchased and installed GPU servers will also be unable to complete training or inference tasks.
Musk specifically mentioned liquid cooling and large chillers because the power and heat dissipation density of high-density AI training racks have increased, making traditional air cooling ineffective. Liquid cooling circuits, pump groups, heat exchangers, chillers, pipelines, and water treatment systems need to be constructed in sync with the design of the machine room, power load, and GPU rack deployment, and cannot be simply retrofitted after the servers arrive.
Training clusters also rely on low-latency, high-bandwidth networks. Large model training requires a large number of GPUs to continuously synchronize parameters and exchange data, and the network interconnection, optical modules, switches, wiring topology, and cluster scheduling determine GPU utilization; if the network cannot match the scale of computing power, expensive chips may not achieve the expected training efficiency due to communication bottlenecks, even if powered on.
Large power transformers are considered one of the most scarce links, with reports in the public market indicating that the delivery cycle for some high-voltage equipment has been extended to 48 to 60 months. Queueing for grid access, transmission expansion, permitting processes, switching equipment, and gas turbine supply will also delay the actual operation time of data centers.
In market mechanisms, buyers include OpenAI, xAI, Google, Meta, Microsoft, cloud service providers, and companies building AI clusters; sellers are no longer just Nvidia and server manufacturers but also power developers, gas turbine manufacturers, transformer and switching equipment suppliers, liquid cooling manufacturers, machine room engineering companies, and optical communication and network equipment vendors. Funds are flowing from GPU procurement further into grid capacity, self-built power sources, transformers, cooling systems, transmission and distribution, and network interconnection; suppliers with deliverable power, grid connection permits, or key equipment capacity benefit, while project parties with only chips but lacking power, cooling, or network face the risk of asset idleness.
Source: Public Information
ABAB AI Insight
AI infrastructure was previously primarily constrained by GPU supply. Nvidia controlled the main supply of high-end training chips during the H100 and Blackwell accelerator cycles, with cloud vendors competing for chip orders, HBM memory, advanced packaging, and server cabinets. Musk's proposed 15GW gap indicates that the bottleneck is shifting from semiconductor manufacturing to the physical engineering "post-chip": the arrival of cabinets does not equal the launch of computing power; the truly sellable product is GPU hours that are powered, cooled, networked, and operating stably.
The capital path has thus been extended. In the past, AI companies purchased GPUs from cloud vendors, who procured equipment from Nvidia, ODMs, and data centers; now cloud vendors must also secure grid access, transformers, substations, transmission capacity, cooling equipment, construction teams, and potential natural gas or nuclear power resources in advance. Project capital expenditures have expanded from server procurement to multi-year energy and engineering commitments, and the equipment value realization cycle has been lengthened by grid connection permits and infrastructure construction. GPU procurement lacking power delivery capability can lead to "accounting power, zero operational income" stranded assets.
Historically, during the internet bubble, fiber optic networks were laid first, followed by traffic demand, leading to some communication assets being long idle; the cloud computing era saw AWS, Azure, and Google Cloud improve server utilization through shared data centers. The difference in the AI cycle is that the power density of training clusters is far higher than that of traditional servers, and the deployment window is directly related to model competition. Established data center operators with ready power, land, permits, and cooling capabilities may convert capital expenditures into billable computing power leasing income sooner than companies that merely announce GPU orders.
This represents a restructuring of the supply chain. The value chain of AI is shifting from "model companies purchasing chips" to "model companies competing for a complete set of operational industrial systems": fuel or grid power, transformers, transmission and distribution, liquid cooling, construction, networks, servers, and operations and maintenance must all be in place simultaneously. Any single link can gain phase pricing power; when transformer delivery times extend to several years, its scarcity may more directly determine project production than a batch of interchangeable servers. The next layer of AI competition is not just model capability but the speed at which capital, energy, and engineering are organized into continuous production capacity.
ABAB News · Cognitive Laws
- Chips determine capability, power delivery determines income
- Computing power is not servers; operation is the asset
- In the end, software competition is about industrial delivery.