This follows our May 3 coverage of Epoch AI’s compute concentration data. This item covers the complementary angle: capability pace.
The numbers come from Epoch AI’s Epoch Capabilities Index, a measure that tracks frontier model performance over time on a normalized scale. According to Epoch AI’s May 2026 update, the pace of frontier AI capability improvement has approximately doubled from pre-2024 levels. The reported figures, roughly 8 points per year before 2024, approximately 15.5 points per year as of the latest measurement cycle, have not been confirmed against the source page directly due to a source verification gap in the current package. They should be treated as Epoch-attributed figures pending direct source confirmation, not as independently verified hard numbers.
What the figures mean in practice: the rate at which frontier AI systems are getting better is now materially faster than it was two years ago, and it has stayed that way. This isn’t a single data point. It’s been consistent across the April 22, April 27, April 29, and May 3 measurement cycles tracked by the hub. The acceleration has held.
The consumer hardware implication is the less-covered part of the story. Historical data from Epoch AI’s tracking suggests frontier AI capabilities, the kind that require significant compute at release, have tended to reach consumer-accessible hardware within roughly 12 months of initial deployment. That’s a pattern, not a guarantee, and Epoch characterizes it as an observation rather than a projection. But it matters for planning: capabilities that are enterprise-only today are likely consumer-accessible within a year.
One practical consideration the acceleration data doesn’t surface: the ECI measures what models can do, not what running them reliably at production scale requires. Capability pace tells you how fast the frontier is moving. It doesn’t tell you how fast the operational infrastructure, observability tooling, fine-tuning pipelines, security hardening, is keeping up. For most enterprise teams, that infrastructure gap is the binding constraint, not the capability ceiling.
The May 2026 ECI update appears in the same data release as the compute concentration figures published on May 3. That coverage focused on compute growth, 44x annual increase, with a significant share concentrated among a small number of hyperscalers. The capability-pace angle and the compute-concentration angle are two ways of reading the same dataset. Together they describe a frontier that is accelerating in capability while centralizing in the infrastructure required to produce it. Both dynamics compress the timelines that compliance and planning functions are working with.
The April 29 brief on ECI scores covers the specific model rankings that give these trajectory numbers context. For teams tracking both capability pace and individual model benchmarks, that brief and this one should be read together.