Two papers this summer point the same way. One is a community guide to running models on a gaming laptop. The other is from the director of Japan's national supercomputing centre. Both say the same thing: the maths behind AI is moving to 8 bits, and that makes every AI read of your website cheaper.
What FP8 Is
A model is a very large pile of numbers. Traditionally each one was stored in 32 bits. FP8 stores each in 8. The Hugging Face guide uses the right analogy: it is like compressing a photo. You lose a little detail, and the file loads much faster.
The arithmetic is simple. A 7 billion parameter model needs about 28 GB of memory in 32-bit. In FP8 it needs about 7 GB. That is the difference between a data centre card and the graphics card in a laptop. The guide claims modern FP8 models keep 90 to 95% of the original's performance, and that the speedup is often two to four times.
Where it runs today
NVIDIA H100 and the RTX 40-series (Ada) have FP8 tensor cores, and so does AMD's MI300. Older cards such as the RTX 30-series and the A100 do not, whatever a community post says. There the gain is memory only, not speed.
What it does to a laptop
A 12 GB laptop card holds a 3 billion parameter model in full precision, with no room left for the conversation. In FP8 it holds an 8 billion parameter model with room to spare. That is the size at which a local model becomes a usable assistant.
What it does to a server
The guide's example: a single cloud instance that served one user at a time serves three or four in FP8. The cost per query drops by the same ratio.
The Chips Have Already Chosen
The second paper is the interesting one. Satoshi Matsuoka runs the RIKEN Center for Computational Science, home of the Fugaku supercomputer. In May he posted a paper titled FP8 is All You Need, and has revised it three times since. It is about scientific computing, not language models, and that is exactly why it matters here.
Science has always wanted 64-bit numbers. Matsuoka's point is that NVIDIA's newest chips have all but stopped providing them. On the B300 the native 64-bit rate has fallen to about 1.3 trillion operations per second, a fraction of what the previous generation offered. The silicon went to 8-bit tensor cores instead, because that is what AI inference buys.
His answer is to stop fighting it. Using a number-theory trick called the Ozaki scheme, a 64-bit multiplication can be rebuilt from a stack of 8-bit ones. The paper's figures: the emulated 64-bit ceiling rises from about 1.3 to about 135 trillion operations per second on the B300, and to about 473 on the coming Rubin chip. NVIDIA has since added tensor-core emulation of 64-bit maths to its own libraries.
Read that as a website owner and the message is plain. The most expensive chips on earth have been rebuilt around cheap 8-bit inference, and even physics is being rewritten to fit. The supply of AI reads is going up for years. The price of each one is going down.
Sources
FP8 Quantization: Your GPU's New Best Friend, Thomas Andrew Barrie, Hugging Face community article, 21 June 2026. FP8 is All You Need (Part 1): Debunking Hardware FP64 as the HPC Holy Grail, Satoshi Matsuoka, arXiv 2606.06510, submitted 28 May 2026, revised 8 September 2026. Chip figures are the paper's own projections. AI Read Cost benchmark, phase 1, llmstxt.studio, September 2026.
What Changes for Your Website
Three things, and they compound.
More of your readers will be machines. Every drop in the cost of a token makes another agent worth running. We wrote about the agentic web last week. FP8 is part of why it arrives on schedule.
More of those readers will be small. The model reading your page will often be an 8 billion parameter model on someone's laptop, not a frontier model in a data centre. These models are quantized, so they have less accuracy to spare, and they are run with short context windows to fit in memory. A few thousand tokens is common. They cannot afford a page that buries its content in navigation, scripts and boilerplate.
Compute got cheaper. Attention did not. FP8 shrinks the maths. It does not enlarge the window a model reads through. That budget is the scarce thing, and it is where our own measurements bite.
In our read-cost benchmark we sent agents to answer questions about 20 sites, by crawling the site and by reading its llms.txt. The result split cleanly on one variable: file size. Under about 10,000 tokens the file made the site twice as cheap to read. Over that line it made the site nine times more expensive, because the agent had to load the whole file before it could use any of it. A local model with a small window does not get nine times more expensive. It fails to read the file at all.
What to Do
Treat your llms.txt as an index, not a dump. Keep it under 10,000 tokens, which is roughly 40 KB. Give your key pages a clean Markdown copy, which version 2 of the spec now expects, so a small model reads the content and nothing else. Then check the whole site the way a machine would. Run the AI Readiness Audit. It checks whether AI systems can reach, find and read your site, and it flags a file that is too big to be useful.
Questions
What is FP8?
A way of storing numbers in 8 bits instead of the 32 bits AI models were trained in. A model held in FP8 takes a quarter of the memory and moves a quarter of the data, so it runs faster and fits on smaller hardware. Newer GPUs run FP8 maths directly in their tensor cores.
Does FP8 change how AI reads my website?
Not the mechanics. It changes who does the reading. Cheaper inference means more agents making more requests, and more of them running on small local models with tight context windows. Those readers are the least forgiving of a large or messy page.
How big should an llms.txt be?
Under about 10,000 tokens, which is roughly 40 KB of text. In our read-cost benchmark, files under that line made a site about twice as cheap to read as crawling it. Files over it made the site about nine times more expensive, because the file no longer fits the reader's budget.