AI & Compute
How Compute Is Counted And Why It Is Controlled
Restrictions on advanced chips depend on measuring computing capability, and the metrics chosen determine which hardware falls inside the rules and which does not.
Governments have moved to restrict access to advanced computing hardware, which requires defining what counts as advanced. The definitions are technical, and they shape both policy and product design.
Capability is measured, not described
Rules cannot reference product names, since those change constantly. They instead reference measurable properties such as arithmetic throughput and interconnect bandwidth.
Those figures are published by manufacturers and can be verified, which makes them administrable in a way that qualitative descriptions would not be.
The consequence is that a threshold expressed in numbers determines which products may be sold where, with no discretion at the point of sale.
Thresholds create design incentives
When a limit exists, products appear just beneath it, configured to deliver as much capability as the rules permit.
Manufacturers can reduce one measured property while preserving others, producing hardware that is compliant yet substantially useful for the workloads of concern.
Regulators respond by adding or adjusting metrics, which produces an iterative cycle rather than a stable settlement.
Interconnect matters as much as raw speed
Large models are trained across many chips, so the rate at which those chips exchange data limits what can be built regardless of individual performance.
This is why communication bandwidth appears alongside arithmetic throughput in restrictions, and why limiting it is an effective lever.
It also means that a large number of modest chips is not simply equivalent to a smaller number of powerful ones, since the coordination overhead grows.
Access can be rented rather than owned
Controls on physical shipment do not by themselves prevent remote use of hardware located elsewhere, since computation can be purchased as a service.
Addressing that requires rules about who may be provided with capacity, which involves service providers rather than manufacturers and is harder to verify.
The distinction between owning hardware and using it has therefore become a substantive policy question rather than a commercial detail.
Measurement lags the technology
Metrics that captured capability well at one point can become poor proxies as architectures change and specialised units for particular operations appear.
A chip optimised for one numerical format may score modestly on a general measure while performing extremely well on the workloads that matter.
Keeping definitions current is an ongoing administrative task, and the gap between technical reality and written thresholds is where most of the practical difficulty lives.
