SALT Treaties

SALT stands for “Strategic Arms Limitation Talks,” meaning negotiations on limiting strategic weapons. It refers to talks between the USA and the Soviet Union that began in 1969. At the time, both states possessed thousands of nuclear weapons and kept building more. Both feared the other might gain a decisive advantage. So they agreed not to further increase the number of certain weapons. This resulted in two treaties: SALT I in 1972 and SALT II in 1979.

Why a cap instead of disarmament was already a success

SALT did not scrap a single nuclear weapon. The treaties only set ceilings on further expansion. That sounds modest, but it was a breakthrough. For the first time, two hostile superpowers accepted that their armaments were subject to verifiable rules.

The background was a dangerous logic. If one side arms up, the other must follow suit, or else it feels inferior. In the end, both are more expensively armed, but neither is safer. Experts call this an arms race. SALT was meant to halt exactly this spiral, without either side having to show unilateral weakness.

That is precisely why SALT is cited today in discussions about artificial intelligence. Even with very large AI systems, some argue that a race between states and corporations leads to risky shortcuts. The question then becomes: can something like SALT be built for computing power or model sizes? The parallel is popular, but contested — more on that below.

Ceilings, spy satellites, and a treaty without a signature

SALT I consisted of two parts. One froze the number of launch facilities for long-range missiles for five years. The other, the ABM Treaty, limited defense systems against such missiles. This second rule seems paradoxical: both sides voluntarily forwent protection. The reasoning behind it was that a good shield makes striking first tempting.

SALT II went further, limiting the total number of delivery systems to around 2,250 per side in 1979. Delivery systems are missiles, bombers, and submarines that carry nuclear warheads to their target. The treaty was signed but never ratified by the US Senate, so it never formally took effect. The reason was the Soviet invasion of Afghanistan. Nevertheless, both sides largely adhered to the numbers for years.

Verification was crucial. Inspectors in the other country were politically impossible. Instead, both sides allowed each other to monitor from space via satellites. This was called “national technical means.” Both also committed to not disguising their facilities. Trust thus arose not from promises, but from countability.

SALT as an argument in the AI debate

In everyday life, SALT is mostly encountered in history class, as part of Cold War détente policy. In tech news, the term appears in a different role. When researchers, politicians, or company executives call for international rules for AI, they like to cite SALT as historical proof that such a thing is possible. The International Atomic Energy Agency is also frequently cited as a model.

Concretely, this involves ideas such as reporting requirements for very large training runs or caps on the computing power used. The EU AI Act already contains such a threshold, above which a model is considered particularly consequential. That is the same idea as with SALT: a measurable number instead of a vague statement of intent.

A common mistake is stretching the analogy too far. Missiles are large, expensive, and countable by satellite. An AI model consists of numbers in a file and can be copied without anyone seeing it from the outside. What can most readily be controlled is the hardware, i.e., the specialized chips used for training. That’s why export controls on AI chips revolve around exactly this point. SALT therefore shows less a ready-made solution than the condition under which such agreements can function at all: what cannot be verified cannot be agreed upon either.