In Terminator 3: Rise of the Machines, released in 2003, the military switches on Skynet and someone proudly reports that it is running at 60 teraflops per second.
Things go badly from there.
It is Halloween, so we did what Flopper does for fun. We took the number seriously.
Every part of it has aged in a different and slightly horrifying way.
Per second, per second
Start with the units.
A teraflop is not a quantity of something. TFLOPS stands for trillions of floating-point operations per second. The rate is already inside the word.
So "60 teraflops per second" is 60 trillion operations per second, per second. That is not a speed. It is an acceleration.
Read literally, Skynet was getting 60 TFLOPS faster every second it stayed switched on.
Which, to be fair to the screenwriters, is roughly the plot.
In 2003, 60 TFLOPS was genuinely frightening
It is easy to laugh at the number now. It was not a silly number at the time.
When the film came out, the fastest computer on the TOP500 list was Japan's Earth Simulator. It filled a purpose-built building in Yokohama and measured 35.86 TFLOPS running Linpack, the standard supercomputer test.
Skynet, at 60 TFLOPS, was about 1.7 times faster than the most powerful computer on the planet.
For a film about a military AI that outgrows its makers, that is a well-chosen number. Big enough to be a leap. Small enough to sound plausible.
Then consumer hardware kept going.
Rookie numbers
The RTX 4090 (Ada Lovelace, 2022) is a gaming GPU. It is rated at 82.6 TFLOPS FP32, draws 450 W and sits in a tower under a desk.
That is 1.4 Skynets. In a machine people mostly use to play video games.
Its successor, the RTX 5090 (Blackwell, 2025), reaches 104.8 TFLOPS FP32. Call it 1.7 Skynets, or one Earth Simulator with room to spare for a browser.
You do not even need to buy one. RTX 4090s appear on our GPU rental price tracker for well under a dollar an hour.
Judgment Day, now available on demand.
The machine that ended the world in 2003 is a line item on a cloud invoice in 2026.
The twist: which flops?
Every good horror film has a twist. Here is ours.
Skynet's 60 TFLOPS came with no precision. Our own style guide forbids that, and for good reason.
In 2003, a supercomputer figure meant FP64, 64-bit double precision. That is what Linpack measures and what the Earth Simulator was built for.
Modern GPUs are designed for AI, and AI mostly runs at much lower precision: BF16, FP8 and now FP4. Those are the numbers on the box. FP64 has been left behind, and on some chips it has been cut almost to nothing.
So let's count again, in the precision Skynet's engineers would have meant.
The RTX 4090 manages 1.29 TFLOPS FP64. You would need about 47 of them to build one Skynet.
The H100 SXM5 (Hopper, 2022) still takes double precision seriously. With its FP64 Tensor Cores it reaches 67 TFLOPS, so one H100 is about one Skynet.
Then there is the GB300 NVL72 (Blackwell Ultra, 2025), a whole rack of 72 GPUs drawing about 142 kW. Blackwell Ultra trades FP64 for low-precision AI throughput, and each GPU is rated at 1.39 TFLOPS FP64.
The full rack comes to roughly 0.1 PFLOPS FP64.
That is 1.7 Skynets. For a rack.
Switch the same rack to FP4 and it is rated at 1,080 PFLOPS. That is 18,000 Skynets.
Same hardware, same power bill, and a difference of more than ten thousand times depending on which number you are shown.
Skynet never told us its precision. In 2026, that is the scariest part of the story.
The Skynet scale
For anyone who wants to measure their own hardware in Skynets, here is the conversion. One Skynet equals 60 TFLOPS. All figures are dense peak ratings from the Flopper database, except the Earth Simulator, which is its measured Linpack result.
| Hardware | Year | Peak compute | Skynets |
|---|---|---|---|
| Skynet | 2003 | 60 TFLOPS, precision unknown | 1 |
| Earth Simulator (FP64) | 2002 | 35.86 TFLOPS FP64 | 0.6 |
| RTX 4090 (FP32) | 2022 | 82.6 TFLOPS FP32 | 1.4 |
| RTX 4090 (FP64) | 2022 | 1.29 TFLOPS FP64 | 0.02 |
| RTX 5090 (FP32) | 2025 | 104.8 TFLOPS FP32 | 1.7 |
| H100 SXM5 (FP64) | 2022 | 67 TFLOPS FP64 Tensor Core | 1.1 |
| H100 SXM5 (FP8) | 2022 | 1,979 TFLOPS FP8 | 33 |
| DGX Spark (FP4) | 2025 | 500 TFLOPS FP4 | 8 |
| B200 (FP4) | 2024 | 9,000 TFLOPS FP4 | 150 |
| GB300 NVL72 rack (FP64) | 2025 | 0.1 PFLOPS FP64 | 1.7 |
| GB300 NVL72 rack (FP4) | 2025 | 1,080 PFLOPS FP4 | 18,000 |
Read down the last column and the lesson is plain. A performance figure without a precision is not a specification. It is a plot device.
What Skynet would worry about today
A 2026 Skynet would not be short of compute. It would be short of other things.
It would want memory, because a large model has to fit somewhere, and a 4090's 24 GB will not hold much of a world-ending intelligence.
It would want bandwidth, because inference speed is usually limited by how fast weights move, not how fast they multiply.
It would want power. A single GB300 NVL72 rack draws about 142 kW, and grid connections are now one of the hardest things in AI infrastructure to get.
And it would want a good price per GPU-hour, like everyone else.
The machines did not need to rise. They needed a purchase order and a capacity reservation.
Happy Halloween
In 2003, 60 TFLOPS was the stuff of nightmares. In 2026 it is a gaming GPU, a fraction of a single B200, or a rounding error in an AI rack.
The number did not become less impressive. The industry just moved it, precision by precision, from FP64 down to FP4.
So this Halloween, if someone tells you a system runs at a frightening number of teraflops, ask the question the film never did.
Which precision?
Then look it up. Every GPU on Flopper lists its peak figures at every precision it supports, side by side, so nobody gets scared by the wrong number.