Arm AGI CPU: 136 Cores, 3nm, 2x Performance Per Rack — and 9 Companies Already Signed Up

For 35 years, Arm has been the company that designs chips but never builds them. That changed on March 24, 2026. Arm Holdings unveiled the Arm AGI CPU — its first in-house data center processor — a 136-core beast fabricated on TSMC’s 3nm process, built from the ground up for agentic AI workloads. Meta is the lead partner and co-developer. OpenAI, Cloudflare, SAP, Cerebras, F5, SK Telecom, Positron, and Rebellions have all confirmed they’re on board.

This is not an incremental product launch. It’s a complete reinvention of what Arm is as a company — from IP licensor to chip manufacturer — and a direct assault on Intel, AMD, and even Nvidia in the data center CPU space.

What Exactly Is the Arm AGI CPU?

The naming is bold (and yes, a little provocative — “AGI” here stands for Arm’s branding, not artificial general intelligence). But the hardware specs back up the ambition.

The Arm AGI CPU packs 136 Neoverse V3 cores running at up to 3.7 GHz boost (3.2 GHz base) across a dual-die design on TSMC’s N3 process. Each core gets 2 MB of L2 cache, and there’s 128 MB of shared system-level cache across the whole chip. The memory subsystem is where things get serious: 12 DDR5 channels (6 per die) running at up to 8800 MT/s, delivering 825 GB/s aggregate bandwidth — that’s roughly 6 GB/s per core with sub-100ns latency. I/O comes via 96 lanes of PCIe 6.0 and CXL 3.0 support. TDP sits at 300 watts.

Two deliberate design choices stand out. First, there’s no simultaneous multithreading — each core runs one thread, period. Arm says this is intentional for deterministic performance scaling, which matters when you’re orchestrating AI agent workloads where latency predictability is critical. Second, there are no on-chip accelerators. This is a pure CPU play, designed to handle the orchestration, scheduling, memory management, and data movement around GPU clusters — not to replace them.

Why Meta Is the Lead Customer (and Co-Developer)

Meta didn’t just sign a purchase order. They co-developed this chip with Arm, which tells you something about how seriously they take this.

Santosh Janardhan, Meta’s head of infrastructure, said the Arm AGI CPU “significantly improves our data center performance density” and will work alongside Meta’s custom MTIA silicon. The pitch is clear: as Meta scales to what they call “multi-gigawatt-scale AI deployments,” traditional x86 CPUs can’t keep up with the orchestration demands of running millions of AI agents across massive GPU clusters.

Meta plans to release board and rack designs under the Open Compute Project later in 2026, which could accelerate adoption across the entire industry.

The other eight confirmed partners cover a wide range of use cases: OpenAI for inference infrastructure, Cloudflare for edge AI, SAP for enterprise workloads, Cerebras and Rebellions for AI accelerator integration, F5 for networking, SK Telecom for telecom AI, and Positron for specialized compute.

The Numbers That Matter: Performance and Density

Arm’s headline claim is bold: more than 2x performance per rack versus x86 CPUs, which they say translates to up to $10 billion in CAPEX savings per gigawatt of AI data center capacity.

The rack-level numbers flesh this out:

Air-cooled configuration (36 kW): 30 compute blades packing 8,160 total cores
Liquid-cooled configuration (200 kW): 42 eight-node servers delivering 45,696 total cores — which Arm claims is 2.02x the density of Nvidia’s Vera ETL256

Commercial systems are already available to order from ASRock Rack, Lenovo, and Supermicro, with broader availability expected in H2 2026. The ecosystem backing is massive: AWS, Google, Microsoft, Nvidia, Samsung, SK hynix, and TSMC are all listed as ecosystem partners. Synopsys is providing full-stack design solutions for the platform.

HSBC responded to the announcement by upgrading Arm to Buy with a $205 price target, calling the AGI CPU a significant expansion of Arm’s addressable market. Arm’s stock saw intraday swings of up to 14% on announcement day, though it settled closer to flat by market close — a sign that the market is still digesting what this means long-term.

How It Stacks Up Against Intel, AMD, and Nvidia

The data center CPU market has been a two-horse race between Intel and AMD for decades, but that picture is shifting fast. AMD’s EPYC processors have clawed market share from roughly zero in 2018 to about 40% today, and Arm-based server chips (from AWS Graviton, Ampere, and others) have been steadily gaining ground in cloud workloads.

The Arm AGI CPU enters this landscape with a different angle. Instead of competing head-to-head on general-purpose compute like Intel Xeon or AMD EPYC, it’s targeting a specific workload category: AI agent infrastructure. The thesis is that as AI deployments scale, the CPU’s job isn’t to run the models — that’s what GPUs and accelerators do — but to orchestrate everything around them. Code execution, scheduling, memory management, reinforcement learning frameworks, tool calls — this is what eats CPU cycles in an agentic AI stack.

Compared to Nvidia’s Vera CPU (88 custom Arm cores, also TSMC-fabbed, paired with Rubin GPUs), the Arm AGI CPU goes wider with 136 cores and positions itself as vendor-neutral — it works with any accelerator, not just Nvidia’s. Nvidia’s Vera claims up to 1.5x higher agentic sandbox performance versus x86 under full-socket load, but independent benchmarks aren’t available yet for either chip.

Against AMD EPYC 9965 (192 cores, Zen 5), Arm trades raw core count for architectural efficiency at the rack level. AMD has the edge in pure multithreaded throughput, but Arm’s argument is about total cost of ownership and power efficiency across an entire rack — not single-socket benchmarks.

Intel’s position is the most precarious. Their Xeon lineup has been losing share to AMD for years, and the Arm AGI CPU represents yet another front in that war. Intel’s Granite Rapids (up to 128 cores) competes on core count, but Arm’s 3nm process advantage and power efficiency claims put real pressure on Intel’s roadmap.

What This Means for the AI Hardware Market

This launch matters beyond the specs. When a company that has spent 35 years as a pure IP licensor — powering 99% of the world’s smartphones without ever making a chip — decides to enter the merchant silicon market, it signals a fundamental shift in how the AI infrastructure stack is being built.

Arm is betting that the agentic AI era needs a different kind of CPU, and that the company best positioned to build it is the one whose architecture already powers the most energy-efficient processors on the planet. Whether that bet pays off depends on execution: can they deliver at scale, maintain competitive pricing, and build the software ecosystem needed to make enterprises comfortable switching from x86?

The fact that Meta co-developed the chip and that the OCP community will get open rack designs suggests Arm is playing the ecosystem game, not just the hardware game. That’s a smart move in a market where software compatibility and vendor lock-in have historically been Intel’s strongest moats.

Frequently Asked Questions

How much does the Arm AGI CPU cost?
Arm has not disclosed per-chip pricing. Commercial systems are available to order from ASRock Rack, Lenovo, and Supermicro, but pricing likely varies by configuration and volume commitments. Arm’s performance-per-watt claims suggest the total cost of ownership story is their primary selling point rather than upfront chip price.

When will the Arm AGI CPU be widely available?
Systems are orderable now from select OEMs (ASRock Rack, Lenovo, Supermicro). Broader availability, including Meta’s at-scale deployment and OCP-based rack designs, is expected in the second half of 2026.

Does the Arm AGI CPU compete with Nvidia GPUs?
No. The Arm AGI CPU is designed to work alongside GPUs and accelerators, not replace them. Its role is to handle orchestration, scheduling, and data movement in AI infrastructure — the CPU-side work that surrounds GPU-accelerated training and inference. It actually competes with Intel Xeon, AMD EPYC, and Nvidia’s Vera CPU.

What workloads is the Arm AGI CPU optimized for?
Arm is targeting agentic AI workloads specifically — AI agents that execute code, make tool calls, manage memory, and coordinate across large clusters. This includes AI inference orchestration, reinforcement learning frameworks, and general data center compute tasks where power efficiency matters.

How does it compare to AWS Graviton and other Arm server chips?
AWS Graviton processors are custom Arm-based chips designed exclusively for Amazon’s cloud. The Arm AGI CPU is different — it’s Arm’s own merchant silicon, available to any buyer through OEM partners. It uses Arm’s latest Neoverse V3 cores at 3nm, while Graviton 4 uses Neoverse V2 cores. The two could coexist, with Graviton serving AWS customers and the AGI CPU serving everyone else.

Top AI Product

Leave a comment Cancel reply