Arm Launches First Silicon CPU, Targets Data Center Agentic AI Workloads

Arm spent 35 years in the business of not making things. On March 24, 2026, at its Arm Everywhere event in San Francisco, that ended. The company announced the AGI CPU, its first production silicon — a 136-core, TSMC 3nm chip co-developed with Meta, and the most consequential strategic pivot in Arm's history. Citi analysts called it just that: the most significant shift in the company's history. They're not overstating it.

The company that built its empire on licensing instruction-set architecture to every chipmaker on earth is now competing directly with those chipmakers. Arm has always collected royalties whether its licensees won or lost. Now it wants to win the socket itself — and it has recruited the single largest infrastructure bet in computing history to make the opening credible.

Meta is spending $135 billion in capital expenditure on AI infrastructure in 2026 alone, according to CNBC. That makes it the anchor customer for Arm's first silicon. Santosh Janardhan, Meta's vice president of infrastructure, said in Arm's announcement that the company worked alongside Arm to develop the AGI CPU to "significantly improve our data center performance density and support a multi-generation roadmap for our evolving AI systems." A $135 billion customer betting on your first product is not a soft endorsement. It's a strategic alignment that doesn't come along often.

The AGI CPU is not positioned against Nvidia's GPUs for training workloads. It's targeting the orchestration layer — the tool processing, function calling, and coordination work that agentic AI systems spend most of their time doing. Research from Georgia Tech and Intel has shown that CPU-side tool processing accounts for up to 90.6 percent of total latency in representative agentic workloads, according to Futurum analyst Brendan Burke. Anyscale has demonstrated an 8x reduction in GPU requirements by disaggregating CPU- and GPU-intensive pipeline stages. The industry is reorganizing around the insight that the CPU matters again — not as a host, but as an orchestrator. Arm wants to own that job.

The hardware is not modest. Each chip carries 136 Arm Neoverse V3 cores running at 3.7 GHz on TSMC's 3nm process, in a dual-chiplet design with 12-channel DDR5 memory at up to DDR5-8800, ServeTheHome reported. Arm's target is 4–6 GB/s per-core memory bandwidth, which it calls the sweet spot for agentic workflow performance. The company is positioning against AMD's EPYC processors — notably not Intel's Xeon line — in its competitive comparisons. Commercial systems are available to order from ASRock Rack, Lenovo, and Supermicro now, with broader production expected in the second half of 2026.

The financial case is where the optimism gets expensive. Arm projects $15 billion in annual AGI CPU revenue by fiscal 2031, against a $4 billion revenue base in fiscal 2025, according to CNBC. That's a nearly fourfold revenue expansion in five years. Arm estimates it can earn $500 in gross profit per chip — roughly 50% margin on a roughly $1,000 chip — versus $50 from IP royalties, a 10x improvement on per-unit economics, as Arm CFO Jason Child outlined in the company's investor guidance. But $15 billion in AGI CPU revenue implies moving roughly 30 million units annually at that margin — a volume that would require roughly 100 large hyperscalers each deploying 300,000 chips. That math is not impossible. It is a specific future that has to be built.

Analysts are split accordingly. Moor Insights analyst Matt Kimball told Data Center Knowledge that Arm has credibility in its architecture but zero credibility as a product company. "Arm has credibility in its architecture. It doesn't have credibility as a product company," he said. Tirias Research founder Jim McGregor was blunter: "There is no room in our industry anymore for an IP-only company. You have to have other revenue streams, other product lines and other services." Both things are true at the same time. The architectural credibility is real. The product company credibility does not exist yet — Arm has never managed a supply chain, a customer support organization, or a hardware failure return flow. That is a different business, and Arm is now in it.

Futurum's Burke offered a more structural read: Arm occupies a position no other company can replicate, compensated whether it wins the socket directly or its licensees do. That dual-monetization math is the reason Arm's stock rose 16 percent the day of the announcement, per CNBC, and why Barclays raised its price target to $200 and Evercore ISI raised its target to $227, both with Buy ratings. The market is buying the optionality, not the silicon.

The pressure lands on the x86 incumbents. Counterpoint co-founder Neil Shah told EE Times that the AGI CPU puts "pressure on the x86 camp to protect its market share and position. It remains to be seen how Intel and AMD react." Neither has responded publicly yet. They have time — broader production doesn't ship until the second half of 2026. But Meta's roadmap, if it holds, gives Arm a reference deployment at a scale that is difficult to dismiss.

Meta's own silicon program gives context for how seriously the company is treating custom silicon. The company is deploying hundreds of thousands of its own MTIA inference chips across its infrastructure, with four new generations in development and a new chip every six months, according to Meta's announcement. This is not a company hedging its bets. It is building out a multi-source silicon strategy where Arm's AGI CPU handles the orchestration layer and MTIA handles inference acceleration. That architecture — disaggregated, purpose-fit, heterogeneous — is the actual model being bet on.

The AGI CPU is real hardware. The anchor customer is real demand. The architectural thesis — that CPU-side orchestration latency is the next bottleneck to solve in agentic AI — is supported by credible research. Whether Arm can execute as a product company, scale the supply chain, and hit the revenue targets it has projected is an open question that won't be answered in a press release. Synopsys used its full design stack to tape out the chip on TSMC's 3nm node, which is a genuine engineering achievement. But TSMC's 3nm capacity is constrained, allocated across Apple, Nvidia, AMD, and now Arm. Volume at 30 million units annually is not a given. It is a ambition.

The second half of 2026 is when this becomes verifiable. Benchmark data against AMD's EPYC and Intel's Xeon lines will tell the real story. For now, the headline is simple: the company that invented the business model of not making things just made something. And one of the largest infrastructure bets in corporate history is betting it works.