Explainer

Compute Thresholds, Explained: The 10²⁶ Question

The most consequential number in AI policy is a count of floating-point operations. Understand the 10²⁶ question and you can read every serious frontier-AI bill — including exactly where its drafters made design choices.

The number that governs the frontier

Somewhere around 10²⁶ FLOPs — one hundred million trillion trillion operations — sits the line that modern AI governance has converged on. Training runs below it: notice. Training runs above it: reporting obligations, and increasingly, licensing conversations. The figure traces back to the 2023-era U.S. reporting framework for large compute clusters and was carried forward into the EU’s “systemic risk” tier for general-purpose AI models. It was never a magic number. It is a tripwire: a bright, checkable line that says “if you’re spending this much, the rules change.”

Why that scale? Because at current frontier prices, a 10²⁶ training run costs hundreds of millions of dollars in hardware time alone. Below roughly that line, you can’t plausibly build a frontier system; above it, you’re unmistakably in frontier territory. The threshold isn’t where anyone has located danger. It’s where everyone can agree the stakes have changed.

Why compute is the dial most bills use

Regulators can’t measure “intelligence,” “alignment,” or “intent.” They can measure a datacenter. Training compute has three properties that make it the most enforceable regulatory dial available:

• It’s observable. Chips draw power and appear on import manifests, utility interconnect queues, and cloud billing. You cannot hide a gigawatt.
• It’s scarce and lumpy. Capability at the frontier comes from enormous, centralized training runs — not from a laptop overnight. The number of entities able to cross a 10²⁶ threshold is small, and the capital requirements grow each generation.
• It’s upstream. Regulating the training run touches risks before deployment, when the irreversible decisions (data, scale, evaluation) are still being made.

Every alternative dial is worse along some dimension. “Harm-based” rules act only after victims exist. “Model-based” rules require capabilities the regulators don’t have. Compute is where the leverage physically lives: you can’t train a superintelligence in a basement.

What a threshold actually does

Critics describe thresholds as “banning innovation at an arbitrary line.” That’s a category error. A threshold attaches obligations — disclosure, evaluation, safety cases — to scale. It’s closer to how other high-consequence systems are treated: an airliner with more than a certain number of seats isn’t banned, but it carries different certification duties than a Cessna. The threshold is the point where the operator is asked to demonstrate, in advance, that they know what they’re doing.

The four failure modes — and how bills differ on them

Thresholds have known weaknesses. When you read a bill, check which of these the drafters addressed — or left open:

1. Efficiency drift

Algorithmic progress means each year, less compute buys more capability. A fixed FLOP threshold ratchets: today’s frontier tripwire becomes, in five years, a mid-tier training run. Bills that index the threshold to hardware cost or capability benchmarks, or that include a review clause, address this. Bills that hard-code the number let the obligation quietly decay — sometimes by design, sometimes by neglect.

2. Offshoring

Compute obeys physics, not jurisdiction. A threshold that binds only U.S. developers relocates the frontier to whichever jurisdiction will take it — which is why hardware export controls and international compacts aren’t adjuncts to compute thresholds; they’re load-bearing walls. The drafting question to check: does the bill’s obligation bind training for U.S. deployment or merely training by U.S. companies? The first closes the gap; the second leaves it open.

3. Below-threshold amplification

Fine-tuning runs below the line can still produce dangerous systems if the base model was dangerous — the way the safest chemistry lab rules mean little if the stockroom already contains the wrong reagents. The fix is attaching obligations to the base model’s capability tier, not just the incremental run. Bills that ignore this handle the easy case only.

4. Lock-in

The threshold that made sense in 2024 was set by 2024 hardware. If legislation treats it as permanent, the practical effect is a gap: obligations apply to systems that are no longer frontier and never apply to the ones that are. A periodic review mandate — unglamorous, essential — is what prevents it.

What the design space looks like

Synthesis: a threshold regime designed for longevity (1) indexes the trigger so it tracks the frontier, (2) scales obligations with capability evaluations, not FLOPs alone — compute gets attention, evaluations determine requirements, (3) requires a safety case at the trigger, filed before the training run rather than defended after an incident, and (4) is internationally compatible, so compliance in one allied jurisdiction transfers to the others where the frontier actually trains.

Every bill in the tracker that touches thresholds is logged against exactly this checklist — which of the four failure modes the text addresses, in which section. The 10²⁶ question isn’t really “how many FLOPs is too many?” It’s “which design choices did the drafters make?” The number is where the reading starts — not where it ends.