What Is FP64? Double-Precision Computing, Explained
FP64, or double-precision floating point, is the number format that scientific computing runs on, and it has suddenly entered investor vocabulary because of Marc Chaikin’s “The End of Elon?” presentation, which builds its entire thesis on the difference between the FP16 hardware inside today’s AI data centers and the FP64 machines the Department of Energy is building at Oak Ridge. You do not need a computer science degree to follow the argument, and this article explains FP64 in plain terms, including the demo Chaikin uses to teach it.
Floating point, in one minute
Computers store numbers in a fixed amount of memory. A “floating point” format is a way of packing a wide range of values, tiny fractions to enormous quantities, into that space, at the cost of precision: some numbers get rounded. The number in the name, 16, 32, 64, is how many bits each value gets. More bits, more precision.
The trade-off is speed. FP16, “half precision,” is fast and compact, which is why the AI industry adopted it: neural networks tolerate small rounding errors fine, so training and running them on lower-precision hardware doubles or quadruples throughput. FP64, “double precision,” is the opposite trade: slower per operation, but with roughly 18 quintillion distinct values available, precise enough for physics.
The Broccoli Problem
Chaikin’s demonstration of why precision matters is the presentation’s best teaching moment, and the reason the term “the Broccoli Problem” has spread. Ask an AI model how many times the letter C appears in the word “broccoli,” and it can fail, because the underlying arithmetic happens on hardware that rounds numbers. His explanation is technically sound and worth quoting: “That lack of precision can’t be solved with better programming. You can’t solve it by feeding more data into the models. Because the floating point problem isn’t a software problem. It’s a hardware problem.”
The stakes go far beyond spelling. Simulating a fusion reactor, modeling a new drug molecule, or computing airflow over a wing means accumulating billions of tiny operations where each rounding error compounds. FP16 hardware accumulates those errors into meaninglessness; FP64 keeps the calculation honest. That is why every serious supercomputer for decades, the machines that forecast weather, certify nuclear stockpiles, and model climate, runs double precision.
The numbers in the pitch
Chaikin quantifies the gap with a raw comparison: “FP16 chips can process 65,000 unique values at once… FP64 chips can handle 18 quintillion calculations at once.” A careful reader should note that is a comparison of representable values, not an apples-to-apples performance benchmark, and his worked example (adding .00001 five times, where FP64 returns 1.00005 and FP16 “spits back the number one”) is the honest way to show it.
Where it becomes an investment thesis: the first FP64 AI machine, called Lux, is being built at Oak Ridge National Laboratory in Tennessee, “the world’s first FP64-powered AI supercomputer,” about “the area of a 7-Eleven convenience store,” drawing “as little as 24 megawatts” against the roughly two gigawatts of Musk’s Colossus campus. Lux is the first node of what Chaikin names American Atlas: a network of more than 50 DOE supercomputers, Frontier, El Capitan, Aurora, Perlmutter, and others, linked into “one giant AI breakthrough machine.”
What FP64 changes, and what it does not
The honest technical summary: FP64 machines and FP16 AI data centers solve different problems. Chaikin’s pitch is that the FP64 side, scientific AI, is where breakthroughs like fusion, drug discovery, and quantum advances will happen, and that companies supplying that buildout, his free pick is AMD, whose Instinct chips already power Frontier and El Capitan, will capture “the AI White Swan of 2027.” What FP64 does not do is render ChatGPT-style generative AI obsolete; the presentation is careful on this when you read closely, qualifying its comparisons with “regarding major breakthroughs.”
Questions readers ask about FP64
Why can’t better software fix the precision problem? Because the limit is arithmetic, not code. FP16 can represent 65,000 unique values; FP64 handles 18 quintillion. In scientific simulation, where rounding errors compound over billions of operations, no amount of model scaling recovers precision that the hardware never captured. Chaikin’s line: “Because the floating point problem isn’t a software problem. It’s a hardware problem.”
Does FP64 make ChatGPT obsolete? No, and the presentation’s own subtext, “when it comes to major breakthroughs,” concedes it. Consumer generative AI and scientific computing are different jobs: one needs fluent language at massive scale, the other needs exact numbers over long calculations. The Department of Energy’s Lux at Oak Ridge is built for the second job, and the investment thesis is that the second job produces the patentable breakthroughs, in fusion, drug discovery, and materials science.
How do you invest it? The free pick is AMD, whose Instinct accelerators power the DOE’s existing flagship machines, Frontier and El Capitan. The 17-stock American Atlas Portfolio covers the FP64 network’s contracted builders. Our full review covers the numbers and the caveats in detail.
Where to learn more
For the full thesis, start with our End of Elon? American Atlas review, then the Oak Ridge, Frontier, and Lux detail and the 17-company supercluster universe. Chaikin’s presentation, with the Broccoli demo, is at the offer page, discounted from the $499 list price with a 30-day full-refund guarantee.
NewsletterVetter is an independent publication. We receive compensation from some of the services we review through affiliate links. Nothing on this site is investment advice. Always do your own research.