Armstrong: Compute, chips and data limit AI progress

Brian Armstrong, CEO of Coinbase, called access to large-scale compute the biggest bottleneck for artificial intelligence development in recent public comments. He identified limits in GPU supply, datacenter capacity and high-quality training data as the key constraints slowing progress.

Armstrong described two linked hardware issues. The physical supply of accelerators such as GPUs and other AI chips is constrained by semiconductor manufacturing timelines and memory supply chains. At the same time, power, cooling and networking requirements for hosting these chips at scale are growing more expensive and harder to deploy.

Those combined pressures raise the cost and delay the timeline for training and running leading-edge models, Armstrong explained. He pointed out that the largest AI models require clusters of specialized accelerators, and shortages of those chips and of datacenter space concentrate capability among organizations with deep pockets or direct cloud access.

Data availability is another constraint he highlighted. Training high-performing models requires large volumes of clean, labeled and diverse data. Assembling such datasets is costly and time-consuming, and privacy rules plus commercial ownership of sources limit what developers can use, particularly in specialized domains.

Armstrong also raised talent scarcity as a limiting factor. A small pool of engineers and researchers can design, tune and deploy state-of-the-art systems, and competition for that talent increases costs for startups and smaller companies.

To ease the bottlenecks, he recommended increasing investment in semiconductor capacity and expanding datacenter buildout. He also suggested improving data-sharing practices that protect privacy while making more training material available and advancing tooling and model efficiency to reduce overall compute needs.

Armstrong noted that cloud providers and chipmakers have been adding capacity, but demand from AI developers has at times outpaced supply, contributing to higher prices for GPU instances and specialized hardware.

Industry work on model efficiency, new accelerator designs and reuse of pretrained models can reduce the impact of hardware scarcity, he said. Improved software and model architectures can stretch limited compute further, and shared tooling and benchmarks can cut duplicated effort across teams.

He framed the constraints in economic terms, saying current limits favor well-funded companies that can secure long-term hardware deals and operate large datacenters. Smaller entrants may need to focus on niche datasets, optimized models or partnerships to compete. He did not propose regulatory changes and emphasized the role of private investment and market incentives in expanding capacity.

Background material notes that training state-of-the-art AI models consumes large amounts of compute and energy. Semiconductor manufacturers operate multi-year production cycles, making it hard to meet sudden surges in demand. Cloud operators must balance capital expenditures on datacenters with fluctuating customer demand, and data protection laws can restrict how companies share or pool training data.

Armstrong’s remarks came as companies across industries race to develop and deploy large language models and other advanced systems that require substantial compute.