Arena Physica launches Heaviside, the first foundation model for electromagnetism, trained on a self-built data factory

Summary

Arena Physica has launched Heaviside, what the company describes as the first foundation model for electromagnetism. The pitch is that RF and antenna design — a field gated by a small pool of specialists and simulation tools built in the 1980s — can be unlocked the same way language was: by pointing a large model at the underlying physics and scaling it up.

How it works

A user types a natural-language prompt — say, a 10 GHz bandpass filter for a satellite application. An LLM interprets and structures the request, filling in missing specifications where needed, then hands it off to Heaviside, the EM-specific foundation model, which generates the physical design. The output Arena Physica brought to the conversation is a fabricated RF circuit with what the founder, Pratap, describes as an "alien geometry" — a structure that looks nothing like a human-designed component but functions as a working phased array filter. The model runs 800,000 times faster than a commercial EM solver.

Where they are on the curve

Arena Physica is candid about maturity. Pratap places it at a "GPT-1 moment" rather than a ChatGPT moment. The training set currently contains tens of millions of data points, all synthetically generated in a proprietary data factory because labeled EM data doesn't exist at scale in the wild. The model can produce components equivalent to what a human expert would design, but is not yet superhuman. The near-term value proposition is cost and speed, not capability beyond human ceiling — yet.

The scaling-law bet is the core thesis. Just as language models stopped solving translation, summarization, and spam detection separately and generalized once pointed at language as a substrate, Arena Physica believes a single large EM model can eventually power any industry bottlenecked by electromagnetism, without needing domain-specific fine-tuning for each application.

Market map

The addressable surface is wide. Pratap breaks it into satellite communications (phased arrays for SATCOM, including user terminals and satellites — a Starlink satellite reportedly needs to track 64 simultaneous ground locations via phased array), radar (counter-drone, soldier-portable, ship-based systems, with Raytheon as the named incumbent), chip-to-chip data center interconnects, and all-weather imaging using radar as a substitute for LiDAR in robotics. Pratap argues that published market forecasts for phased array and RF components were built by propagating historical numbers forward and don't account for the expected boom in space infrastructure and AI data centers.

Early customers include AMD and Anduril.

Business model

Arena Physica positions itself closer to ARM or Palantir than to a chip manufacturer — licensing IP and deploying forward-deployed engineers (FDEs) who are both software and RF engineers. The current product stack includes Atlas, an agentic tool for debugging hardware design workflows, alongside the newly launched Heaviside. Engagement cycles are long: Pratap describes typical projects running a year or more, with the goal of becoming a persistent partner across a customer's full portfolio of electromagnetically constrained products. Pratap prior company was acquired by Palantir, which informs both the FDE model and the IP-licensing framing.

The open question Pratap flags explicitly is whether the ARM-style IP licensing model will evolve as Heaviside becomes an IP-generating machine rather than simply a tool. If the model can generate novel antenna geometries on demand, the value of traditional licensed IP libraries looks different — and Arena Physica hasn't settled on how to price that.