How fast will AI computing improve?

A historical trend is not a 2030 promise

Epoch AI measured roughly 49% average annual improvement in chip performance per dollar since 2023. That spending-weighted trend moved in steps as new generations entered purchasing mixes rather than rising smoothly. It also differs from the cheapest available single chip. Fivefold growth between the selected quarters is a testable extrapolation, not a fact already observed.

The 2028 semiconductor roadmap

TSMC has described A14 as targeting volume production in 2028 and highlighted improvements in performance, density and energy efficiency relative to N2. This is a manufacturer roadmap, not a measured benchmark outcome. Timing, yield, packaging, product availability and accelerator software can all change real value delivered to buyers.

Hardware means systems, not only transistors

AI computing performance depends on compute silicon, high-bandwidth memory, interconnects, networking, cooling, power delivery and software. NVIDIA's Vera Rubin platform illustrates the move towards integrated racks rather than stand-alone GPU comparisons. A smaller manufacturing node does not guarantee an equal gain in end-to-end training or inference efficiency.

What exactly is performance per dollar?

A valid measurement must freeze arithmetic precision, workload relevance, inflation-adjusted prices, shipment weighting and the hardware class included. If the underlying Epoch AI series is retired, Cortist would require a prospectively specified equivalent before this draft is eligible to be scored. Vendor peak-throughput numbers are not interchangeable with an independent system-wide benchmark.

What would push the forecast up or down?

Evidence supporting a higher estimate could include independently documented jumps in shipment-weighted accelerator capability, falling relative purchase costs and better memory or packaging efficiency. Supply bottlenecks, expensive advanced packaging, constrained chip yields, export restrictions and power delivery limits could reduce or delay improvement. Separate price-performance from absolute computing spending.

Why AI compute matters beyond chip companies

Faster and less expensive compute can make more model training and inference economically feasible. However, greater efficiency can also expand demand and total electricity consumption rather than automatically reduce it. Our Energy, AGI and AI Agents topics examine different consequences; none is resolved simply by the next semiconductor generation.

How to read Cortist's 55%

This is the first subjective editorial F-005 estimate for the selected 2030 event, not a calibrated probability. Move the timeline to see the disclosed scenario assumptions, not a live benchmark. The 2026 v0.1 snapshot is a historical version; any new estimate requires a new date, primary-source reasoning and a separate file.

Resolution questions remain open

Before an official forecast lock, independent reviewers must decide whether to freeze the original Epoch series or define an equivalent replacement, how to handle changes in precision conventions and what external benchmark constitutes reproducibility. Until then the percentage is clearly labeled a draft rather than a scored forecast.

How evidence will change this forecast

Cortist monitors original datasets, official measurements and independent evaluations rather than treating manufacturer, researcher or policymaker claims as settled outcomes. Any material reassessment must disclose the previous value, the new evidence, counterarguments and the reason for changing the subjective assessment. The older snapshot will remain at its original versioned URL.

Before this question becomes eligible for formal scorekeeping, independent reviewers must agree on the boundary of the event, adjudication rules, data sources and any necessary fallback measurement. If an essential definition changes, a fresh forecast question will be published instead of rewriting an old one.