Arm Launches First Own Chip in 35 Years, Targets $15B by 2031

Arm Holdings' launch of the AGI CPU on March 24, 2026 marks the structural inflection point where the world's dominant chip architecture licensor becomes a direct silicon vendor — a transition that reshapes competitive dynamics across AI infrastructure, challenges the x86 incumbent duopoly, and forces Arm's own licensees to reckon with a partner that is now also a rival. Arm shipped its first production processor in 35 years, co-developed with Meta, built on TSMC's 3nm process, and priced at approximately 50% gross margins — a signal, as Citi analysts wrote the following day, of 'the most significant shift in the company's history.' Arm stock rose 16% on March 25, the day after the announcement, pricing in what the company's own CEO Rene Haas framed explicitly: $15 billion in AGI CPU revenue by 2031, within a $25 billion total annual revenue target and $9 earnings per share.

The market Arm is entering is substantial and structurally underpenetrated by any single architecture optimised for agentic AI workloads. Arm characterises the addressable opportunity as a $1 trillion AI CPU market — a figure that, while unverified independently, is consistent with the trajectory of hyperscaler capital expenditure: Meta alone spent over $37 billion on capex in 2025, and has guided up to $135 billion for 2026 as it builds multiple gigawatts of AI data center capacity. The five dominant players in AI data center silicon today are Nvidia (GPU accelerators), AMD (EPYC CPUs and Instinct GPUs), Intel (Xeon CPUs and Gaudi accelerators), Amazon (Graviton and Trainium custom silicon), and a growing cohort of custom ASIC vendors including Broadcom and Marvell. Arm's AGI CPU enters this field not as a general-purpose processor but as a purpose-built orchestration and inference engine for agentic AI — the CPU-side workload that coordinates accelerators, manages data movement, and runs the inference loops that GPU-centric architectures handle inefficiently. Nvidia itself recently told CNBC that CPUs are 'becoming the bottleneck' as agentic AI reshapes compute requirements, effectively validating the market Arm is entering.

The AGI CPU's technical specifications establish its competitive positioning precisely. The chip packs up to 136 Neoverse V3 cores running at 3.2 GHz all-core and 3.7 GHz boost across two dies, within a 300-watt thermal design power. It supports 12 channels of DDR5 memory at up to 8,800 MT/s, delivering more than 800 GB/s of aggregate memory bandwidth — or 6 GB/s per core — with a target sub-100ns latency. A standard air-cooled 36kW rack holds 30 blades for 8,160 total cores; Arm has also partnered with Supermicro on a liquid-cooled 200kW configuration housing 336 chips and more than 45,000 cores. Arm claims the chip delivers more than 2x performance per rack versus x86 CPUs, which the company translates into up to $10 billion in capital expenditure savings per gigawatt of AI data center capacity — a figure that, if accurate at Meta's projected 5-gigawatt Hyperion facility in Louisiana, implies tens of billions of dollars in avoided spend. Commercial systems are available now from ASRockRack, Lenovo, and Supermicro, with volume shipments expected by end of 2026 and material revenue impact from 2028 onward. (Technical note for engineers: the Neoverse V3 core is Arm's highest-performance server core lineage, successor to the V2 used in AWS Graviton4 and Google Axion — giving the AGI CPU a validated performance baseline even before independent benchmarks at scale.)

Three structural forces converged to make this moment possible in early 2026 rather than five years ago. First, agentic AI has fundamentally changed the CPU-to-accelerator workload ratio: as AI systems move from batch inference to continuous, multi-step agent orchestration, the CPU becomes the scheduling and coordination bottleneck — a role for which x86's legacy instruction overhead is increasingly penalising. Second, TSMC's 3nm node reached sufficient yield maturity to make a 136-core, two-die configuration commercially viable at the volume Arm requires; the same node underpins Apple's M4 and Nvidia's Blackwell variants, meaning process risk is substantially de-risked by prior high-volume production. Third, the capital environment at hyperscalers has reached a scale — Meta at $135 billion in 2026 capex, Microsoft at a projected $80 billion, Google at over $75 billion — where even a 10% improvement in performance per dollar of rack spend represents procurement decisions worth billions annually. The precedent most directly analogous is Amazon's Graviton program: AWS began developing its own Arm-based server CPUs in 2018, and by 2024 Graviton instances represented a material and growing share of EC2 capacity — demonstrating that hyperscaler-scale custom silicon can achieve rapid adoption curves when performance-per-dollar advantages are real. Arm's AGI CPU is the logical extension of that dynamic, with Arm itself capturing the margin rather than ceding it to a hyperscaler.

The competitive implications sort cleanly into three tiers. Intel faces the most acute pressure: its Xeon roadmap is already under margin compression from AMD's EPYC, and a third credible x86-alternative CPU architecture — with claimed 2x rack density at 50% gross margins — narrows Intel's window to reclaim AI infrastructure share before its next major node transition. AMD's Lisa Su struck a measured tone, stating 'competition is healthy for the industry,' which is the response of a company that is not currently the primary target but recognises the threat vector. The second-order disruption falls on Arm's own licensees: Apple, Nvidia, Amazon, and Google all build chips using Arm's architecture under licensing agreements. Arm now competes directly with Amazon's Graviton for data center CPU workloads and with Nvidia for the inference orchestration layer — a dynamic that will stress the commercial terms of those licensing relationships at the next renewal cycle. The value chain shift is explicit: Arm previously captured royalties of roughly 1–2% of chip selling price; at 50% gross margins on $15 billion in projected AGI CPU revenue by 2031, Arm captures orders of magnitude more value per unit of compute deployed. Cloudflare, OpenAI, Cerebras, SAP, and SK Telecom are among the launch partners, which signals adoption intent across CDN, frontier AI, accelerator, enterprise, and telco verticals simultaneously.

Our read: Arm's move into production silicon is not a hedge — it is a calculated bet that agentic AI will sustain CPU demand at a scale large enough to justify the channel conflict with its licensees and the execution risk of becoming a chip vendor after 35 years as a pure IP business. The hypothesis is testable on a clear timeline. If Meta confirms AGI CPU deployment inside Hyperion by end of 2026, and if Arm reports initial silicon revenue in its fiscal Q3 or Q4 2026 earnings, the model is validating on schedule. The disconfirming scenario is slower-than-expected hyperscaler adoption — particularly if AWS, Google, or Microsoft decline to qualify the AGI CPU in favour of their own custom silicon, leaving Arm dependent on Meta as a near-sole anchor customer through 2027. The 50% gross margin claim also warrants scrutiny: fabless chip economics at TSMC 3nm carry significant per-wafer costs, and at initial volumes below those of Apple or Nvidia, Arm's blended margin may compress before it expands. CFO Jason Child's public 50% gross margin disclosure suggests confidence, but the figure needs three to four quarters of reported results to be verified independently.

Four specific indicators will resolve the strategic picture by end of 2027. First, Meta's Hyperion deployment confirmation: watch for any Meta infrastructure disclosure — earnings call, data center announcement, or infrastructure blog post — that names the AGI CPU as a production component inside the Louisiana facility; a confirmed deployment at gigawatt scale would validate the density claims and anchor the 2028 revenue materialisation timeline. Second, Intel and AMD analyst day disclosures in Q2 2026: both companies are expected to respond with updated EPYC and Xeon roadmap data directly addressing rack-level performance density — any acceleration of their respective next-generation node timelines would signal that the 2x density claim has been taken seriously internally. Third, TSMC Arizona qualification: Arm's Mohamed Awad stated the company would 'love to manufacture here,' and any announcement of domestic 3nm production for the AGI CPU would simultaneously qualify it for U.S. federal procurement programs, opening a channel currently closed to Taiwan-fabricated silicon. Fourth, Arm licensee contract renewals: Apple, Nvidia, and Amazon all have architecture licensing agreements with Arm that are periodically renegotiated — any public disclosure of revised terms, renegotiation timelines, or licensee-developed alternative architectures in the 12 months following this launch would be the clearest leading indicator of whether channel conflict is materialising faster than Arm's revenue ramp can offset it.