Chip‑stack Roles Blur
Recent reports say Arm built an internally manufactured AI chip with Meta and that Nvidia has expanded investments into startups like SiFive, indicating companies are moving beyond traditional single‑layer roles into owning larger parts of the stack. These shifts mix IP, silicon, and platform strategies as vendors and builders compete for AI‑infrastructure control. (coincentral.com) (parameter.io)
Artificial intelligence chips used to split neatly into layers. In 2026, Arm is selling its first production data-center CPU with Meta as a lead partner, while Nvidia is backing SiFive to push a rival instruction set deeper into AI servers. (arm.com) (sifive.com) Arm said on March 26, 2026 that its new Arm AGI CPU is the company’s first production silicon product, moving it beyond licensing designs into shipping an actual chip. Arm said Meta helped develop the processor and that other customers and original design manufacturers had already committed for production. (arm.com) The chip is aimed at “agentic” artificial intelligence workloads in data centers, and Arm said it delivers more than 2 times performance per rack versus x86 platforms. Arm has also been selling pre-integrated Compute Subsystems, or packaged blocks of chip design meant to cut cost and time to market for cloud and artificial intelligence hardware. (arm.com 1) (arm.com 2) SiFive, meanwhile, said on April 9, 2026 that it raised $400 million in a Series G round that valued the company at $3.65 billion. The investor list included Nvidia, alongside Atreides Management, Apollo Global Management, Point72 Turion, T. Rowe Price Investment Management, Prosperity7 Ventures, and Sutter Hill Ventures. (sifive.com 1) (sifive.com 2) SiFive sells processor intellectual property based on RISC-V, an open instruction set architecture that works like the basic grammar a chip uses to understand software. Nvidia is not only an investor there: SiFive said in January 2026 that its data-center platforms would add Nvidia NVLink Fusion so they can connect coherently to Nvidia graphics processors and other accelerators. (sifive.com 1) (sifive.com 2) That leaves fewer clean lines between “chip designer,” “platform vendor,” and “customer.” Arm now offers intellectual property, pre-validated subsystems, and finished silicon, while Nvidia is extending its artificial intelligence platform through interconnects, software, and stakes in companies building central processors around RISC-V. (arm.com) (arm.com) (sifive.com) Arm has been preparing for this shift for several product cycles. Its Neoverse Compute Subsystems package central processing cores, system fabric, and validation work into a starting point for custom cloud and artificial intelligence chips, and its Total Design program ties in foundries, design firms, and packaging partners across the supply chain. (arm.com) (arm.com) Arm and Meta also widened their relationship beyond one chip. Arm said in a recent announcement that the companies signed a multi-year partnership covering hardware and software co-design for Meta’s infrastructure and products. (arm.com) SiFive is making a similar pitch from the other direction. The company said in September 2025 that its second-generation Intelligence family added new scalar, vector, and matrix products for artificial intelligence workloads from edge devices to data centers, while Red Hat and Nvidia have both been part of its recent infrastructure push. (sifive.com) (sifive.com) The contest is no longer just over whose chip is fastest. It is over who can supply the design blocks, the links between chips, the software hooks, and, increasingly, the finished processor that sits inside an artificial intelligence server. (arm.com) (sifive.com)
