What Is FP64 in Computing? A Plain Guide
“What is FP64?” has moved from computer science textbooks into investor searches, and the reason is Marc Chaikin’s “The End of Elon?” presentation, which builds its entire American Atlas thesis on the difference between two number formats: FP16, the half-precision format inside today’s AI data centers, and FP64, the double-precision format that scientific computing has always required. You do not need any technical background to follow the argument, and this guide keeps it plain.
The one-minute version
Computers store numbers in a fixed amount of memory. The “floating point” format packs a huge range of values, microscopic fractions to astronomical quantities, into that space, and the price of the packing is rounding: some numbers cannot be represented exactly and get approximated. The number in the name is how many bits each value gets. FP16 gives each number 16 bits, enough for about 65,000 distinct values. FP64 gives each number 64 bits, enough for roughly 18 quintillion.
Why the AI industry chose small: speed and cost. Neural networks are tolerant of small errors, so running them on 16-bit hardware doubles or quadruples throughput per dollar. It was the right choice for chatbots. Why science cannot: precision compounds. A simulation adds and multiplies billions of times, and each tiny rounding error accumulates until the answer is noise. Weather models, nuclear stockpile certification, drug molecule dynamics, fusion reactor design: all of it runs on FP64, and has for decades, because a 65,000-value ruler is too coarse for physics.
The Broccoli Problem
Chaikin’s demonstration is the friendliest entry point to the idea. Ask ChatGPT how many C’s are in “broccoli.” It can fail, not because the model is dumb, but because the arithmetic underneath runs on rounding hardware. His worked example: add .00001 five times. FP64 returns 1.00005. FP16 returns 1. The gap is invisible in a chat and decisive in a simulation. His summary, and the presentation’s central teaching: “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.”
Why FP64 suddenly matters to investors
The Department of Energy is building AI-era machines in FP64. The first, Lux, is being assembled at Oak Ridge National Laboratory: “the world’s first FP64-powered AI supercomputer,” about “the area of a 7-Eleven convenience store,” drawing “as little as 24 megawatts,” yet positioned to do scientific work that Musk’s two-gigawatt Colossus campus cannot honestly attempt. Lux seeds a network, “more than 50” DOE supercomputers including Frontier, El Capitan, Aurora, and Perlmutter plus nine new micro clusters, that Chaikin calls American Atlas. His claim for the era: “a sudden 360-fold acceleration in major AI-powered breakthroughs,” and a $248 trillion opportunity aggregated from AI, fusion, cancer research, and quantum computing forecasts.
The picks follow the hardware. The free one is AMD, whose Instinct accelerators already power the DOE’s Frontier and El Capitan, “the first company on my American Atlas ‘buy’ list.” The paid report covers “17 companies that have signed deals… They’re all working directly on building American Atlas.”
The honest technical footnote
A careful reader should hold one caveat alongside the enthusiasm. The presentation’s comparison, “FP16 chips can process 65,000 unique values at once… FP64 chips can handle 18 quintillion calculations at once… That makes them nearly 283 trillion times more powerful,” compares the count of representable values, not apples-to-apples performance. FP16 and FP64 machines are built for different jobs, generative AI and scientific simulation respectively, and FP64 does not render ChatGPT obsolete. The presentation itself qualifies its claims with “regarding major breakthroughs,” and that qualifier is the technically accurate reading. The investable idea is the buildout, not the obsolescence narrative.
Where to learn more
Start with our End of Elon? American Atlas review, then the Oak Ridge, Frontier, and Lux detail and the 17-company supercluster universe. The full presentation is at the Chaikin Analytics offer page, discounted from $499 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.