What happened
Epoch AI’s changelog records the initial release of its AI chip imports explorer on October 27, 2026. The explorer is part of Epoch AI’s data offerings and is accompanied by documentation, methodology and downloads. Its stated purpose is to estimate how much AI compute arrives in each country, using national customs records together with Epoch AI’s AI chip sales data.
The estimates are quarterly and cover the period since 2022. Epoch AI says the dataset includes destination-country estimates in three forms: customs trade dollars, chip units and H100-equivalents. These measures answer related but different questions. Trade value describes the value recorded in customs data; unit counts describe estimated quantities of chips; and H100-equivalents express the estimated arrivals in a common chip-equivalence measure. The changelog does not define the conversion details in its release entry, so the equivalence should be read as a standardized estimate rather than a direct count of H100 hardware.
The documentation distinguishes two record sources by destination coverage. For sixteen destinations, the estimates are built from national customs records. For all other destinations, Epoch AI uses exporters’ records. The changelog does not name the sixteen destinations in its release entry, nor does it specify the countries represented by each exporter record. It describes the resulting figures as estimates, not as a complete census of chip shipments.
The release is an explorer and a data resource, rather than a new chip, model or cloud service. Epoch AI says the data draws on its AI chip sales data as well as customs records. The changelog links readers to methodology and downloads, but its entry does not report a price, subscription tier or access restriction. It also does not give a single headline estimate for global arrivals or a country ranking in the changelog itself. The key operational distinction is therefore between a newly accessible way to inspect country-level estimates and a verified account of how much functioning compute each country can use.
What the archive adds
New Runtime’s earlier coverage, Where compute ends up, country by country (November 11, 2026), described Epoch AI’s interactive database estimating where AI chips had arrived since 2022, using customs records and domestic production data. This release is related to that earlier coverage rather than a wholly new topic: the changelog documents the explorer’s initial release and its stated measurement dimensions. New Runtime also covered Epoch AI’s Inference Performance Index on November 10, 2026, a separate effort to compare modeled concurrent inference capacity rather than raw FLOP/s alone.
Trend signal
This release does not yet fit a pattern New Runtime tracks. It contributes to infrastructure measurement, but the listed patterns concern shifts in agent systems, product boundaries, organizational practice and related operational changes; the changelog does not establish one of those shifts. Country-level chip-arrival estimates may become useful evidence for future analysis, but the release alone does not show a change in agent economics, deployment practices or the distribution of usable compute.
Connections to other events
New Runtime’s earlier coverage of the Inference Performance Index described a different measurement problem: comparing chips by modeled concurrent inference capacity rather than relying on raw FLOP/s alone. The imports explorer measures a different part of the infrastructure picture—estimated arrivals by destination and quarter. Read together, the two efforts distinguish where hardware is estimated to go from what a chip may be able to serve under a modeled inference workload. Neither measure alone establishes the compute available to a particular team or country.
Why it matters
For teams choosing where to train, serve or host AI workloads, the geography of hardware supply can matter alongside model capability and cloud pricing. A country-level series offers a starting point for asking whether chip arrivals are concentrated or changing over time, and for comparing a trade-based estimate with other infrastructure indicators. That is useful context for strategy, investment research and policy analysis.
The figures should not be treated as a proxy for immediately available compute. Customs entries and exporter records track estimated arrivals; they do not, by themselves, show whether chips were installed, commissioned, retained in the destination country or made available to external customers. Nor does a count of chips capture the full differences among hardware configurations or utilization. Builders can use the explorer to frame questions and identify places for further investigation, but decisions about capacity still require evidence about deployed systems, access, networking, power and service availability.
Limits and uncertainty
Epoch AI describes the figures as estimates and uses different underlying records for sixteen destinations and all others. The changelog does not quantify uncertainty or explain the conversion from chip units to H100-equivalents. Arrival estimates do not establish installation, ownership, operational status or utilization, and customs-trade dollars are not equivalent to compute capacity. The release entry also does not provide a headline total against which to assess the scale of the dataset. These limits matter when using the explorer to compare countries or infer actual compute access; the methodology documentation is the relevant place to inspect how the estimates are constructed.
Reader takeaway
Epoch AI has made country-level, quarterly estimates of AI-chip arrivals inspectable across three measures. Treat the explorer as a map of estimated hardware flows—not a census of deployed compute or a guarantee of accessible capacity—and pair it with evidence about installation and actual service availability before making infrastructure decisions.