The AI boom has a public face — the GPU, the token, the model — and a quiet one. Rene Haas wants to talk about the quiet one. The CEO of Arm, speaking on the No Priors podcast, positions the microprocessor as the indispensable traffic cop of the AI era: accelerators generate tokens, he argues, but CPUs decide where those tokens go, arbitrating every transaction from data center to smartwatch. “All the AI needs some level of compute. That’s what Arm does and that compute needs to be power efficient. That’s what we’re really really good at. So those roads all lead through us,” he said.
That confidence has translated into one of the most significant strategy pivots in semiconductor history. For decades, Arm was the ultimate asset-light company — licensing CPU blueprints with a staggering 98.5% gross margin, no inventory, no returns, no scrap. Then in March 2026, it announced its first physical product, the Arm AGI CPU. The company that sold recipes started cooking.
Why Arm Abandoned a 98.5% Gross Margin
The move from IP licensor to chip seller wasn’t ambition for ambition’s sake. Haas traces it to a specific customer: Meta approached Arm seeking a general-purpose agentic CPU, and no existing supplier could deliver. The request arrived against a backdrop of lengthening chip manufacturing timelines and a customer base increasingly desperate for faster time-to-market.
What surprised Haas most was the reaction from Arm’s own licensees. Before committing, he consulted NVIDIA, Amazon, Microsoft, and Google — companies that could reasonably view Arm’s entry into physical products as a competitive threat. There was no pushback. The logic: more Arm-based software in the wild benefits everyone building on the architecture. When the AGI CPU was announced, Jensen Huang, Ronnie Boker, Amen, and James Hamilton all publicly congratulated the company.
The strategic logic is clear, but the operational leap is brutal. Physical chips require supply chain operations, memory allocation deals with Micron and SK Hynix, back-end layout engineers, and bring-up labs — none of which Arm needed as a licensor. Haas has filled leadership roles with veterans from Broadcom, Qualcomm, and NVIDIA to build these muscles quickly.
AI Inside Arm: The Genie Is Out of the Bottle
Haas is strikingly candid about how deeply AI has penetrated Arm’s own engineering workflow. Between 80 and 90 percent of Arm engineers now use AI tools daily, and he describes the dependency in terms that leave no room for hedging: “If we were to shut it off, it’s like being in the 1990s, you’ve got internet and you’re now saying, you know, only internet between the hours of two and four. After that, go to the library that we have down the hall. It’d be anarchy. The genie’s out of the bottle.”
The pattern of AI adoption in chip design follows a familiar arc. The bottleneck in any 24-to-36-month chip project isn’t the initial architecture or RTL generation — it’s verification, validation, debug, and documentation. AI excels at precisely these tasks. The immature frontier remains RTL generation and physical design implementation, where models lack training data because the information is proprietary and closely guarded.
Arm’s structural advantage, Haas contends, is its portfolio itself. “If it’s unusable and untestable, it’s actually untrainable. And if it’s untrainable, it’s not usable for AI,” he noted. Decades of documentation, test benches, and build processes give Arm a training corpus that few competitors can match. His projection: within five to ten years, straightforward designs could go from idea directly to a GDS2 file — the format sent to fabs — with minimal human intervention, though designs pushing for specific performance or efficiency targets will remain harder to automate.
The Supply Chain Gauntlet
Haas’s outlook on physical supply is sobering: three to five years of constrained conditions, driven by the transformer architecture’s insatiable appetite for compute and memory. The constraints have already migrated. Advanced packaging was the pinch point two years ago; memory is now binding certain systems. Next on the list: data center construction itself.
ConstraintStatusOutlookAdvanced packagingEased from 2024 peakStill tight for leading-edge nodesMemory (HBM, DRAM)Currently bindingRequires allocation deals with Micron, SK HynixWafer capacityTight at 3nm/2nmMulti-year expansion underwayData center buildoutEmerging bottleneckProjects behind schedule, labor-intensiveSubstratesConstrainedNew entrants must secure supply
The implication for startups is stark. Young AI chip companies with innovative designs face capital requirements and vendor relationships that many founders underestimate. Access to 3nm production lines, advanced packaging, and memory allocation will separate winners from also-rans — often before design quality even enters the equation. Haas’s advice: secure strategic partnerships early, with supply chain players, private equity, and banks alike. He notes that SoftBank’s portfolio approach, including the newly announced SoftBank Neo neocloud, can provide a home for chip startups that might otherwise struggle to secure design wins at hyperscalers.
The Data Center Backlash and the Infinite Game
Haas addresses the growing opposition to data center development with visible skepticism, attributing much of it to fear of AI-driven job loss. He cites a telling counterexample: the electricians’ labor union publicly asked not to ban data centers because the jobs are needed — highly skilled, certified work that AI creates rather than destroys.
On US semiconductor policy, Haas speaks from a distinctive vantage point — an American citizen with industry memory stretching back to the 1980s Japan Inc. memory wars, when the US launched Sematech to refortify domestic manufacturing. He argues the US needs more domestic fabs for national security and supply chain diversification, and warns against export controls that could push critical technology development offshore.
His framing of the China competition: “There’s just no downside from being the leader. There are second and third order effects that you may not like, but to be the laggard, you have the entire script dictated to you.” The race, he says, is “an infinite game” with no declared winner — but the US must remain at the forefront to drive the innovation and ecosystem jobs, from liquid cooling to energy infrastructure, that data centers create.
The CPU’s Revenge
The broader investment context validates Haas’s thesis. While Nvidia’s accelerators dominated the first wave of the AI trade, 2026 has seen a decisive rotation toward the full stack of compute. AMD has surged roughly 113% year to date, while Intel has posted triple-digit gains as investors recognize that agentic AI workloads — with their multi-step, stateful workflows — demand powerful CPUs alongside accelerators. Raymond James analyst Simon Leopold projects server CPU revenue could reach $201 billion by 2030, growing at a 44% compound annual rate.
Haas pushes back on the narrative that accelerators have displaced CPUs. Every computing system requires a microprocessor for orchestration and arbitration. As AI inference moves to edge devices — where a 50-watt GPU is physically impossible — Arm’s power-efficient instruction set architecture becomes the natural fit for on-device processing. The token factory metaphor is apt: accelerators mint the tokens, but CPUs route them to users. Remove the router, and the factory stops delivering.
This dynamic spans Arm’s entire footprint — data centers, automobiles, robots, phones, wearables. On robotics, Haas sees both humanoid and task-specific form factors thriving: humanoids for environments optimized around six-foot-tall workers, specialized robots for distribution centers and factory automation. Distribution and delivery will automate first, he predicts. Arm’s position is structurally strong: real-time sensing microprocessors at robot fingertips, perception systems, and the “brains” in current humanoids from NVIDIA and Qualcomm all run on Arm architecture.
The financing story underscores the physical reality Haas describes. Industry estimates point to roughly $3.6 trillion in AI infrastructure investment between 2026 and 2030. Nvidia has evolved from chip supplier into part-financier, offering credit support worth up to $105 billion for a single SoftBank-backed Ohio development designed to scale to 8 gigawatts. The constraint is no longer purely technical — it is physical, financial, and increasingly political. Haas’s central insight is that these are not separate problems but one interconnected bottleneck, and the companies that navigate it successfully will define the next decade of computing.