Scaling Frontier AI: Compute Growth and the Sequence of Physical and Economic Ceilings
Abstract
Artificial intelligence has advanced rapidly, yet unlike semiconductor electronics it has lacked a quantitative growth framework analogous to Moore’s Law. We compile four record-setting envelopes from a consistent database family and report each with full uncertainty quantification: frontier training compute has grown at <inline-formula> <tex-math notation="LaTeX">$4.61\times $ </tex-math></inline-formula> per year (doubling every 5.4 months), record accelerator performance at <inline-formula> <tex-math notation="LaTeX">$1.60\times $ </tex-math></inline-formula> per year, globally deployed AI compute at <inline-formula> <tex-math notation="LaTeX">$4.57\times $ </tex-math></inline-formula> per year, and frontier training cost at <inline-formula> <tex-math notation="LaTeX">$3.15\times $ </tex-math></inline-formula> per year. We connect these trends through the decomposition <inline-formula> <tex-math notation="LaTeX">$C(t) = H_{\mathrm {ref}}\cdot N(t)\cdot A(t)$ </tex-math></inline-formula>, in which <inline-formula> <tex-math notation="LaTeX">$H_{\mathrm {ref}} N(t)$ </tex-math></inline-formula> is aggregate deployed peak throughput and <inline-formula> <tex-math notation="LaTeX">$A(t)$ </tex-math></inline-formula> is the frontier-allocation residual, the product of fleet share, utilization, and run duration. The residual has been statistically flat (<inline-formula> <tex-math notation="LaTeX">$1.01\times $ </tex-math></inline-formula>/yr): frontier training compute has grown in near lockstep with deployed infrastructure, while infrastructure growth itself divides into per-chip performance gains (<inline-formula> <tex-math notation="LaTeX">$1.60\times $ </tex-math></inline-formula>/yr) and unit-count expansion (<inline-formula> <tex-math notation="LaTeX">$2.86\times $ </tex-math></inline-formula>/yr). We analyze the cascade sequence in which physical and economic ceilings bind: deployed capacity already corresponds to approximately 32 GW of power draw and is trending toward grid delivery limits within the next few years, and a single frontier training run is projected to require energy comparable to the annual output of a large power plant by the end of the decade. We show that energy infrastructure is the most consequential near-term constraint on frontier compute growth. Unlike prior scaling work that prescribes how to allocate a fixed compute budget, this framework quantifies how fast the budget itself grows, remains structurally valid as individual factors transition between growth phases, and provides a quantitative basis for forecasting frontier AI capability and identifying the transitions in energy, infrastructure, and device physics that will define the next phase of AI growth.