The instruction set architecture is the contract between software and silicon. For decades that contract was licensed, and ARM's version became the default for everything from phones to accelerators. RISC-V offers the contract for free. The debate is not about whether the instructions are better — they are broadly similar — but about what an open standard costs you in places a licence fee used to cover.
ARM sells a verified, toolchain-complete design with a licence; RISC-V gives you the specification for free and leaves you to pay for the verification, the toolchain maturity and the ecosystem that a licence used to bundle.
Part One: What is being compared
ARM licenses IP cores and the instruction set. A customer pays an upfront licence and a per-chip royalty, and in return gets a core that has been silicon-proven, a compiler that has been debugged against it for years, and a support organisation. The design is closed: you cannot add your own instructions to the ARM ISA, though you can add custom accelerators around the core.
RISC-V is an open standard. Anyone can implement the base instruction set without a licence, and the modular design lets a chipmaker add custom instructions for a specific workload — a matrix-multiply opcode, a crypto extension, a domain-specific accelerator — without asking anyone's permission. The specification is maintained by a foundation, and implementations range from hobbyist cores to production silicon.
The comparison is between a bundled product and a specification. One arrives finished; the other arrives as a promise that you can finish it yourself.
Part Two: The arithmetic
The licence fee is the easy number. ARM core licences can run from hundreds of thousands to millions of dollars, with per-chip royalties often in the range of one to two percent of die cost — for a high-volume accelerator that royalty is a line item worth eliminating. For a high-volume accelerator shipping millions of units, that royalty is a real line item, and removing it is the headline reason to consider RISC-V.
The hidden cost is verification. A production CPU core needs exhaustive verification — billions of cycles of simulation, formal proofs for critical properties, silicon test chips. An established ARM core has already paid that cost across thousands of customers. A team building its own RISC-V core either buys a verified implementation from a vendor, which reintroduces a licence fee, or pays the verification cost itself, which can exceed the saved royalty by an order of magnitude for a first design.
Custom instructions are where RISC-V genuinely wins in AI accelerators. Adding a bespoke matrix opcode can cut instruction count and energy for a specific kernel by a meaningful factor, something impossible with a licenced ARM core. But the compiler must be taught the new instruction, the toolchain must emit it, and the software stack must be maintained — work that benefits the whole ecosystem only if the extension is standardised. A proprietary custom instruction saves silicon but forks the toolchain.
The ecosystem numbers favour ARM today. Android, most mobile toolchains, and a vast body of verified libraries target ARM first. RISC-V has strong Linux and GCC/LLVM support for the base ISA, but optimised libraries for advanced features, and the long tail of vendor-specific extensions, remain fragmented.
Part Three: Where the model breaks
The "free ISA" argument breaks on implementation quality. Not all RISC-V cores are equal, and the specification permits behaviours that make portability harder than the marketing suggests. A core that implements the base ISA but not the optional vector or bit-manipulation extensions will run generic code but miss the performance that made the architecture attractive. Two "RISC-V" chips can be as different as two unrelated architectures once their extensions diverge.
The vector extension is a case in point. It is powerful and still evolving; toolchain support has lagged the specification, and early implementations made different choices about register width and configuration. Code tuned for one vector implementation may not port cleanly to another, which is precisely the fragmentation the open model was supposed to avoid.
Security verification is another soft spot. ARM's trusted execution environments have years of adversarial review. RISC-V implementations of equivalent isolation — the trusted execution platform specifications — are newer and less battle-tested. For an accelerator handling sensitive data, that maturity gap matters more than the licence fee.
Part Four: The later cost
The strategic cost is lock-in of a different kind. With ARM, you are locked to a vendor's roadmap and licence terms. With RISC-V, you are locked to your own implementation decisions: once you ship custom instructions, every piece of software that depends on them is your maintenance burden forever. The freedom to customise is also the freedom to own a fork.
The talent cost is real. ARM expertise is abundant; engineers who have verified a high-performance out-of-order RISC-V core are rarer and more expensive. A team choosing RISC-V for a cutting-edge design is betting it can hire or grow that expertise, and that bet is easy to lose under schedule pressure.
There is also a supply-chain dimension. An open ISA lowers the barrier for new entrants, which is why it has attracted significant investment in China and in startups worldwide. That is good for competition, but it means the ecosystem is splitting along geopolitical lines, with regional extensions and toolchains that may not converge. A global product may have to support several dialects.
Part Five: Outside the team
For the product user, the difference is almost invisible — it shows up as battery life, price and which AI features run locally. A chip that used its instruction-set freedom to add an efficient matrix engine may offer on-device inference that a licenced core could not match at the same power. A chip that used the freedom badly may simply be slower.
For the software team, the choice decides portability. An ARM target is well understood: one binary family, mature simulators, predictable behaviour. A RISC-V target may require choosing an extension profile, verifying which implementation it will run on, and maintaining separate tuned paths for different vendors.
For a company building chips, the choice is strategic. RISC-V removes a royalty and grants design freedom, which compounds over millions of units. But it transfers the burden of ecosystem maturity from the IP vendor to the chip company, and that burden is paid in engineering years, not licence fees.
A fair comparison fixes the performance target, the volume, the required extensions and the available verification budget, then measures total cost per shipping chip including tools, verification and software maintenance. Comparing a royalty-free specification against a licenced, verified core is comparing unequal packages.
Market sizing check
If open instruction sets continue to gain share in embedded and accelerator silicon, RISC-V-based chips and the surrounding IP, verification and tooling market is plausibly $10B to $25B a year by 2030. The swing factor is whether the ecosystem converges on a stable set of extensions — especially the vector spec — or fragments into vendor dialects that erode the portability advantage.
These figures are a synthesis of published market ranges and a stated model, not a forecast of any specific vendor.
The comparison that matters
The comparison that matters is between bundled maturity and design freedom. ARM wins when time to market, toolchain depth and verified security dominate, and the royalty is acceptable. RISC-V wins when volume is high enough to amortise verification, custom instructions deliver a real workload advantage, and licence independence is strategically valuable.
The ranking assumes you can actually exploit the freedom. If your design does not need custom instructions and your volume is modest, the open ISA delivers freedom you will not use and a cost you will pay.
RISC-V does not make chips free — it moves the cost from a licence invoice into verification, toolchains and ecosystem maintenance, and that trade only wins when the freedom to customise is worth more than the maturity you gave up.
SOL: HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS
