The same hardware can claim 1 exaflop or 16 exaflops depending on how you count, without either figure being wrong. That is why comparing power claims between countries or between companies is almost always misleading, and why you need to look at the small print.
We reported this morning on a plan targeting 9,800 exaflops. Here is what that figure means.
The basic unit
A flop is one operation on floating-point numbers, performed in one second. It is the elementary unit of work for a processor doing scientific computing or AI.
Prefixes stack in multiples of a thousand: teraflop for a thousand billion, petaflop for a million billion, exaflop for a billion billion operations per second.
The scale is hard to picture. A decent laptop reaches a few teraflops. An exaflop is roughly a million times that.
The trap: precision
Here is what invalidates most comparisons.
A number can be represented with more or less precision: on 64 bits, 32, 16, 8, or even fewer. The lower the precision, the more operations the processor performs per second, because each operation is simpler.
The gap is not trivial. Between high-precision computing, used for scientific simulations, and very low-precision computing used to serve AI models, the same hardware can show a difference of a factor of ten or more.
This is exactly what we explained about quantization: reducing precision costs a bit of quality and buys a huge amount of speed. That is acceptable for a language model, and not for a climate simulation.
When an announcement talks about capacity intended for AI, it is usually counting at reduced precision, the one that gives the highest figures. When a ranking of scientific supercomputers publishes its results, it counts at high precision. Both are legitimate and they are not measuring the same thing. Comparing a national target expressed in intelligent computing to the performance of a scientific supercomputer is like comparing fuel consumption in litres per 100 km to range in kilometres.
The three other caveats
Theoretical power is not achieved power. A claimed figure often corresponds to the theoretical maximum. In real conditions, you typically get half or less, depending on the task and the efficiency of the software.
Memory often matters more. As we explained about dense and mixture-of-experts architectures: to serve a model, the constraint is frequently the amount of available memory and the speed at which you access it, not raw power.
A national total is not usable capacity. Adding up the power of thousands of machines spread across a country gives an impressive figure. That does not mean you can put them all to work on a single problem: that requires very high-bandwidth links between them, which is another matter entirely.
How to read an announcement
Three questions are enough.
At what precision? If it is not stated, assume the one most favourable to the person making the claim.
Theoretical or measured? Independent rankings measure by running a reference program. Announcements add up spec sheets.
Distributed or clustered? A national total and a single machine do not enable the same uses.
What to take away
These figures do have a real use: they indicate orders of magnitude and trajectories. A capacity that grows by 177% in a year says something, whatever unit is used.
What they do not allow is ranking. When you read that one country is overtaking another in computing power, there is a good chance the two figures come from sources that are not measuring the same thing. It is the same problem we described for model benchmarks: a number without its protocol is a number without meaning.