Arm Goes Down to Make CPUs: The Biggest Bottleneck in Chip Design Is Verification, Not Design
The real time sink in the chip cycle is verification, debugging, and documentation, not RTL; Arm already has 80%-90% of engineers using AI daily, and has broken its 98.5% gross margin licensing model to make CPUs itself.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Arm's move to make CPUs itself is customer demand
Arm originally only sold IP licenses with a 98.5% gross margin, never touching inventory or wafers. The step forward to compute subsystems already had customers grabbing the pre-integrated 'subsystems' of CPU, GPU, and memory as modules. The further step came because 'customers wanting general-purpose CPUs for the large-model era couldn't find a supplier' and came directly to Arm. Before the announcement, Arm asked nearly all its licensees; opposition was minimal: a larger software ecosystem benefits every customer, so Nvidia, Amazon, Microsoft, and Google—companies making Arm server chips—all publicly supported it.
— Rene HaasBanning AI tools is like sending engineers back to the library
The total chip design cycle is 24-36 months; the most time-consuming parts are not RTL generation or architecture mapping but verification, validation, debugging, and documentation. AI is especially suited to these. Internally, about 80%-90% of Arm engineers already use AI daily; turning off the tools would be like in the 1990s letting people use the internet only two hours a day and sending them back to the library downstairs for the rest. Currently AI's weakest areas are RTL generation and physical design, because models rely on public corpora while the industry's most critical details are private data—exactly the gap companies like Arm can fill.
— Rene HaasFrom idea to tapeout takes five years, not two or three
When asked whether AI, by eating the most time-consuming verification step, could shrink 24-36 months to 6-12, Rene Haas gave a measured answer: 2-3 years is unrealistic; beyond five years, for designs with less complex constraints, going from idea all the way to GDSII (the tapeout file handed to the foundry) is possible. If you feed the tool 'make a chip 10% faster, 20% cheaper, and 30% more power-efficient than a reference,' don't expect a one-click result. His judgment is that in 5-10 years, the industry's design methodology will change significantly.
— Rene HaasAI supply and demand are far apart; the bottleneck is at construction sites, not wafers
Discussing the AI bubble, Rene Haas distinguished between stock-level valuation froth and the industry chain itself; he stated clearly that supply compared with demand 'is still far off.' Transformer inference and training consume both compute and memory, so demand won't stop. The real hard constraint will shift to data center construction: many projects are being delayed and need more manpower, and there is growing resistance to new data centers in various places. Over the next 3-5 years, the constraint may not be wafer capacity or memory but the persistently tight physical construction.
— Rene HaasSoftBank's cloud move gives chip companies a customer
For chip/AI entrepreneurs with heavy assets and strong capex: capital availability itself is a gate, and strategic partnerships should reach the supply chain, PE, and bank levels early. Behind Arm is SoftBank, not just as a shareholder; SoftBank just announced it will build a 'neo cloud'-type service, allowing young chip companies to avoid having to knock on Microsoft's or Google's door for design wins, and instead form a direct customer relationship with the Arm/SoftBank ecosystem—an extra path to market.
— Rene HaasHumanoids won't unify; specialized robots will coexist long-term
Rene Haas believes Robot 1.0 is purpose-built: both mechanics and software are optimized for a single action, and changing a production line means almost starting over. New-generation robots that can redefine tasks through training and have a general-purpose mechanical body will first eat repetitive labor in construction, infrastructure, and security. Form factors won't converge on a single direction: factories, warehouses, and tools are all designed to human dimensions, so humanoids have an advantage; other tasks are better suited to specialized forms. Costs must come down first, and delivery and warehousing will be the earliest automated scenarios.
— Rene HaasChip restrictions on China are an unwinnable infinite game
He is direct on US manufacturing: America should build more fabs domestically, whether for national security or supply chain diversification. On export controls, he speaks of an 'infinite game'—restricting chips to China cannot produce a victory with an end, and may instead drive key technologies out of the US. He also denies that data centers are 'big warehouses with no jobs': the ecosystem chain of energy, liquid cooling, and more can create jobs in America.
— Rene HaasIn a token world, every path goes through the CPU
Using the token factory metaphor for accelerators, Rene Haas thinks the real system question is: who orchestrates and arbitrates where tokens go and delivers them to users? He believes every path goes through the CPU: beyond training, there is inference scheduling, system orchestration, and context and tool invocation. At the edge, on devices that can't fit a 50W GPU, Arm's power advantage is more pronounced. The CPU hasn't disappeared; instead, it creates sustained demand alongside accelerators.
— 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
| Arm's traditional licensing gross margin | 98.5% | 6:02 |
| Share of Arm engineers using AI tools daily | 80%-90% | 8:07 |
| Total chip design cycle | 24-36 months | 8:07 |
Glossary
- IP licensing
- A business model where a chip company sells design blueprints rather than physical chips.
- compute subsystem
- A pre-integrated semi-finished design from Arm that combines CPU, memory, and other components, leaving customers less room for modification.
- RTL
- Register-transfer level: a code describing a chip's logic hierarchy, sitting between architecture and physical layout.
- GDSII
- The physical layout file format that chip designs are finally handed to the foundry for manufacturing.
How to listen
AI infra investors, chip entrepreneurs, cloud decision-makers, and hardtech practitioners who want to understand why Arm is moving from IP to selling CPUs.
Listeners familiar with Arm's licensing ecosystem can skip 0:49-2:37; first-timers can jump straight into the chip-making story at 2:37.