Arm Is Building Its Own CPU: Chip Design's Biggest Bottleneck Is Verification, Not Design
The real time sink in a chip cycle is verification, debug and documentation, not RTL; 80%-90% of Arm's engineers already use AI every day, and the company is breaking its own 98.5%-gross-margin licensing model to go build a CPU itself.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Arm is building its own CPU because customers came asking for it
Arm used to sell only IP licenses: 98.5% gross margin, no inventory, no wafers. One step forward from that, its compute subsystem business already has customers grabbing pre-assembled ‘subsystems’ of CPU, GPU and memory as a module. One step beyond that, the push came from customers who wanted a general-purpose CPU for the large-model era and could not find a supplier, so they came to Arm directly to place the order. Before the announcement Arm asked nearly every one of its license customers; objections were few, because a bigger software ecosystem benefits every customer — which is why Nvidia, Amazon, Microsoft and Google, all of them companies building Arm server chips, publicly backed it.
— Rene HaasBanning AI tools just sends your engineers back to the library
A chip design cycle runs 24-36 months, and the biggest time sink is not RTL generation or architecture mapping but verification, validation, debug and documentation. AI is particularly well suited to exactly those stages, and roughly 80%-90% of Arm's engineers already use it every day. Turning the tools off is the equivalent of letting people use the internet only two hours a day in the 1990s and sending them back down to the library for the rest. Where AI is weakest right now is RTL generation and physical design, because the models depend on public corpora while the industry's most critical details are private data — which is precisely the gap a company like Arm can fill.
— Rene HaasIdea straight to layout is a five-year prospect, not two or three
The host asks whether, once AI eats verification — the most time-consuming step — the 24-36 months could shrink to 6-12. Rene Haas gives a measured answer: two to three years is unrealistic; beyond five years, for designs whose constraints are not too complex, going from an idea all the way to GDSII (the layout file handed to the foundry) is possible. If what you feed into the tool is ‘build me a chip 10% faster, 20% cheaper and 30% lower power than this reference’, do not expect one click to finish it. His judgment is that in 5-10 years the whole industry's design methodology will look noticeably different.
— Rene HaasAI supply still trails demand; the choke point is construction, not wafers
On the AI bubble question, Rene Haas argues you have to separate valuation froth at the stock level from the industry chain itself; he says explicitly that supply and demand are ‘nowhere close’. Transformer inference and training eat both compute and memory, so demand is not going to stop. The real hard constraint moves toward data-center construction: plenty of projects are being delayed and need more people, and resistance to new data-center builds is showing up in various places. Over the next 3-5 years the binding constraint may not be wafer capacity or memory but a physical build-out stage that stays tight throughout.
— Rene HaasSoftBank going into cloud makes it a customer for chip companies
For founders in asset-heavy, high-capex chip and AI businesses: access to capital is itself a gate, and strategic partnerships need to reach the supply chain, PE and bank layers early. Behind Arm is SoftBank, and not merely as a shareholder — SoftBank has just announced it will build a ‘neo cloud’-type cloud service, which means a young chip company does not have to compete for a design win only at Microsoft's or Google's door, but can form a direct customer relationship inside the Arm/SoftBank ecosystem: one more path to market.
— Rene HaasHumanoids will not win outright; purpose-built robots will coexist for years
Rene Haas describes robotics 1.0 as purpose-built: the mechanics and the software are both optimized for one single motion, so switching production lines means tearing it down and starting again. The new generation — robots whose task can be redefined through training and whose mechanical body is genuinely general — will first take on repetitive labor in construction, infrastructure and security. Form factor will not converge on a single direction: factories, warehouses and tools are all designed to human dimensions, which favors humanoids; another set of tasks is better served by a specialized form. Cost has to come down first, and delivery and warehousing will be the earliest scenarios to be automated.
— Rene HaasRestricting chips to China is an infinite game no one wins
He is direct on the American manufacturing question: the US should build more fabs on home soil, for national security and for supply-chain diversification alike. On export controls he talks about an ‘infinite game’ — restricting chips to China does not manufacture a victory with an endgame, and may instead push critical technology out of the United States. He also rejects the notion that data centers are ‘big warehouses with no jobs’: energy, liquid cooling and the rest of that ecosystem chain are all employment that can be created in the US.
— Rene HaasIn a token world, every road still runs through the CPU
If you use the token factory metaphor for accelerators, Rene Haas argues the real system question is who orchestrates and arbitrates where those tokens go, and gets them to the user. He thinks every road passes through the CPU: beyond training there is inference scheduling, system orchestration, context and tool calls. Out at the edge, on devices that cannot house a 50W GPU, Arm's power advantage is even more pronounced. The CPU has not disappeared; instead it is generating sustained demand right alongside the accelerator.
— Rene HaasIn their own words · checked verbatim
There's no computing problem that's ever been invented that doesn't utilized and can't utilize the microprocessor. It is the heart of everything. All roads lead through it around it past it. Something has to do the orchestration, arbitration decision around where those tokens go. That's what CPU is doing.
Rene Haas0:00
So now we're in that soup ourselves from the standpoint of.We're also having to figure out how to buy substrates and buy wafers and buy memory, et cetera, et cetera.
Rene Haas2:37
but I think in five to 10 years... We're going to see some amazing differences relative to how chips are designed.
Rene Haas10:35
I call BS on that because if you think about whether it's around energy.Liquid cooling.All of the things that make the data center better.That's all those are all jobs that can be created and done here.
Rene Haas28:20
Figures
| Gross margin of Arm's traditional licensing business | 98.5% | 6:02 |
| Share of Arm engineers using AI tools every day | 80%-90% | 8:07 |
| Total chip design cycle | 24-36 months | 8:07 |
Glossary
- IP licensing
- The business model in which a chip company sells design blueprints rather than physical chips.
- compute subsystem
- A half-finished design in which Arm has pre-assembled components such as CPU and memory, leaving customers less room to modify.
- RTL
- A form of code describing a chip's logic layer, sitting between the architecture and the physical layout.
- GDSII
- The physical layout file that a finished chip design hands to the foundry for manufacturing.
How to listen
AI infra investors, chip founders, cloud decision-makers, and hard-tech people who want to understand why Arm is moving from licensing IP to selling CPUs.
Listeners who already know Arm's licensing ecosystem can skip 0:49-2:37; first-time listeners can go straight into the chip-building story at 2:37.