There's a war raging inside the AI buildout that has nothing to do with GPUs — and everything to do with who owns the toll booth. On one side is ARM, the quiet British company whose chip designs sit inside virtually every smartphone on Earth, collecting a royalty on each one. On the other is RISC-V, an open chip architecture anyone can use for free. The fight between them is quietly deciding who controls the brains of the AI era.
Here's the plain-English version. A chip's architecture is its native language — the basic vocabulary of instructions it understands. ARM owns the most popular language in computing and rents it out. RISC-V is a rival language that costs nothing to speak. Think of ARM as the toll road and RISC-V as the public highway: no tolls, but you build the road yourself.
That used to be a niche engineering argument. Now it's central to the AI buildout — the biggest tech companies on Earth have all become chipmakers.
Look at the hyperscalers. GOOGL has shipped six generations of Tensor Processing Units and built Axion, its own data-center CPU. AMZN has Graviton CPUs and Trainium accelerators. MSFT has Cobalt and Maia. META has MTIA. They all found the same math: when you're spending billions on silicon, paying a royalty on every chip starts to feel like a tax you can engineer away.
ARM is the incumbent for good reason: founded in 1990, it designs blueprints, not factories, and sells access in two tiers — a core license that rents a ready-made Cortex design, or a pricier architectural license for customers like Apple who design their own cores around ARM's vocabulary. More than 300 billion Arm-based chips have shipped, and 35 of the TOP500 supercomputers run on Arm, including the world's fastest as of June 2026.
It's also a wild corporate story. SoftBank bought ARM in 2016, Nvidia's $40 billion takeover collapsed in February 2022 under regulatory pressure, and ARM IPO'd on Nasdaq in September 2023 at a roughly $54.5 billion valuation, with SoftBank keeping about 90%.
RISC-V is the challenger, born in a UC Berkeley lab in May 2010. Its spec is open and royalty-free — the Linux of chips — and it's governed by RISC-V International, a nonprofit that launched in 2015 with 36 founding members including Google, Qualcomm, and IBM, and now counts 4,500+ members, including Nvidia, Intel, and Samsung. The catch: the language is free, but the chip is not. You still have to design, validate, and software-enable it yourself, or pay a vendor like SiFive to do the heavy lifting.
And yet RISC-V is already everywhere you don't notice. NVDA shipped a billion RISC-V cores in 2024 — embedded microcontrollers running inside its own GPUs. Meta's MTIA is RISC-V-based. Europe's processor initiative is building RISC-V silicon. And China has made it a national strategy, with Alibaba, StarFive, and Espressif selling RISC-V chips. Server-class parts like Ventana's 192-core Veyron V2 began shipping in 2025.
But the crown jewel — the data-center CPU — remains Arm and x86 territory. Graviton, Axion, Cobalt: all Arm. RISC-V servers are still early, and analyst consensus says that holds. RISC-V grows from the edges — microcontrollers, accelerator control planes, IoT — upward.
Here's why this war matters. In February 2026, OpenAI and Amazon announced a strategic partnership: OpenAI committed to consume 2 gigawatts of Trainium capacity, with Amazon investing $50 billion. Trainium isn't an ARM-vs-RISC-V story on its face — it's a custom accelerator — but it shows exactly where the value is going: away from licensed architectures, toward custom silicon and the arms dealers who build it. AVGO co-designs Google's TPUs, Meta's MTIA, Microsoft's Maia, and OpenAI's inference chip — the toll booth of the custom-silicon era.
For a hyperscaler already spending billions on custom design, skipping the architecture royalty is pure margin. ARM's revenue runs around $4 billion a year, the vast majority of it royalties on those 300+ billion chips — a blended average well under a dollar per embedded chip, a few dollars for a phone-class SoC. Small tolls, colossal volume. RISC-V doesn't just threaten that income; it threatens the whole idea that chip vocabulary should be rented at all.
So who wins? Probably both, for now. ARM keeps the data-center CPU crown and the smartphone forever. RISC-V eats the edges and grows upward, one billion-core GPU shipment at a time. But for investors, the architecture war is the sideshow. The main event is the custom-silicon era — where the value is captured by whoever builds the chip, not whoever owns the language.
Disclosure: This article is for informational purposes only and does not constitute financial advice. The Signal holds no position in any company mentioned. Positions may change.
The toll-booth model that made ARM one of the most quietly profitable businesses in computing is meeting a free alternative at exactly the moment every hyperscaler became a chipmaker. ARM keeps the data-center CPU crown for now, while RISC-V grows from the edges upward — microcontrollers, accelerator control planes, and China's entire chip strategy. For investors, the ISA war is the sideshow: the value is migrating to custom accelerators like TPU and Trainium, and to the arms dealers like Broadcom who build them.




