Arena
Physica

A model that does in milliseconds what used to take all night

An RF engineer hits run. Once, that meant a coffee, a meeting, maybe a walk around the block while a simulator chewed through Maxwell's equations. Now the answer is already on the screen.

That is the room Arena Physica is trying to build. The New York company has spent years on an unglamorous bet: that the next wave of technology will be limited less by software and more by physics - specifically by electromagnetism, the force humming inside every phone antenna, every satellite link, every radar, and every chip running the AI boom. Its foundation model, Heaviside, claims to predict electromagnetic behavior from a design's geometry in roughly 13 milliseconds. The commercial solvers it competes with can take hours.

Arena Physica calls the goal electromagnetic superintelligence. It is a big phrase. The interesting part is that there are customers - AMD, Anduril, Sivers Semiconductor - willing to put it to work on hardware they actually ship.

The numbers Arena Physica likes to lead with. The one engineers actually feel is the first one.

Electromagnetism runs everything and almost nobody can master it

Here is the awkward truth the industry rarely says out loud. The physics that decides whether a 5G chip works, whether a fighter jet's radar sees clearly, or whether a data center can move heat and signal without melting is the same physics most engineers find deeply unintuitive. Electromagnetism does not behave the way mechanical parts do. You cannot eyeball it.

Arena Physica's thesis is that modern devices have crossed a line - what they call the electromagnetic inflection. Performance, cost, and failure modes are now governed by how well a team understands fields and waves, not how cleverly they bolt parts together. And the tools for that work are, to put it gently, a museum: fragmented toolchains across physics domains, verification that demands hours of simulation and physical testing, and a talent pool of RF experts so rare that hiring one feels like a minor miracle.

Most companies respond to this by hiring more specialists and waiting longer. Arena Physica decided the bottleneck was not effort. It was that no AI had ever truly learned the physics. So they set out to build one - which, naturally, was the easy part to say and the hard part to do.

Two physicists, one inconvenient conviction

Arena Physica was started in 2019 by Pratap Ranade and Engin Ural. Ranade studied physics at Stanford and Columbia, did a stint at McKinsey, then co-founded Kimono, a startup acquired by Palantir. He is, in other words, a physicist who has shipped software and sold it - a useful combination when your pitch is that physics is the next software.

The bet was contrarian in 2019 and only slightly less so now: that general-purpose AI would never be enough for hardware, because the world's most capable language models do not know what an S-parameter is, let alone how to design for one. Arena Physica's leadership has since grown into a bench of Columbia professors, ex-AWS engineers, and distinguished RF specialists - the kind of people who argue about Maxwell over lunch and mean it.

A founding team that treats "grounded in applied physics" as a hiring filter, not a tagline.

From AI hardware engineer to electromagnetic superintelligence

Six years of heads-down physics, then a very loud spring. The rebrand to "Physica" was not subtle, and it was not meant to be.

Atlas, and the model named after a forgotten genius

Arena Physica's platform is Atlas: an AI hardware engineer that works across the stack, starting with electrical and RF engineering. Underneath it sits Heaviside, a physics-native foundation model trained on millions of designs and more than twenty years of proprietary simulation data. The name is a quiet flex - Oliver Heaviside is the engineer who rewrote Maxwell's equations into the compact form everyone uses today and got comparatively little credit for it.

The headline capability: Heaviside-0 predicts S-parameters straight from geometry in about 13 milliseconds, roughly 0.3ms per board when batched on a GPU. That is the 800,000x claim in practical terms. And in the company's own tests, frontier general-purpose models reportedly failed the same RF tasks while burning through 50,000-plus tokens per attempt - a reminder that knowing language is not the same as knowing physics.

Four products, one obsession: closing the loop between a design on screen and an electron in the real world.

The receipts: customers, speed, and a few Fortune 500 logos

Speed claims are cheap. What gives Arena Physica's a little more weight is who is running it. Atlas has been deployed in production at Fortune 500 companies including AMD and Bausch & Lomb, with partners spanning aerospace, automotive, medical devices, and defense - among them Anduril and Sivers Semiconductor. According to the company, customers have reported around a 35% reduction in engineering hours, multi-month-faster time to market, and a better-than-3% improvement in product quality.

Reported by Arena Physica. Treat percentages as the company's own scorecard - but the customer list is harder to argue with.

Make electromagnetic engineering 10x cheaper and 10x faster

Strip away the superintelligence framing and the mission is concrete: accelerate the design, development, and deployment of the devices that will sense, communicate, compute, and actuate in an autonomous future. Arena Physica wants to make electromagnetic hardware engineering roughly 10x cheaper and 10x faster - enough to make ordinary devices more capable and to make exotic, once-impossible devices merely difficult.

The culture borrows openly from Theodore Roosevelt's "man in the arena." Credit, the company insists, belongs to those actually in the fight - which in practice means embedding alongside customers on their hardest problems rather than tossing software over a wall. Backers include Founders Fund, Initialized, Goldcrest, Shield Capital, and, tellingly for a company courting defense, General David Petraeus.

If hardware is the new bottleneck, this is the new pickaxe

Every projection about AI, autonomy, and connected everything quietly assumes the hardware will keep up. More antennas, denser chips, smarter sensors, all built faster than before. That assumption rests on a small number of people who understand fields and waves - and there are not enough of them, and there never will be at the pace the roadmaps demand.

That is the gap Arena Physica is wagering on. If a model can carry real electromagnetic intuition, the rare expert stops being a bottleneck and becomes a force multiplier. The skeptic's fair question is whether physics-native AI generalizes beyond curated demos to the messy edge cases of real fabrication. Arena Physica's answer is its strange, expensive data loop: it builds, simulates, fabricates, and measures its own designs, then feeds reality back into the model. That is hard to fake and harder to copy.

So picture that engineer again. She hits run. The answer is already there - and so is the next design, and the one after that.

The all-nighter did not get shorter. It disappeared. And the question stopped being "will the simulation finish in time" and became "what would we build if it always already had." That is a smaller promise than "superintelligence," and a more useful one.