BofA says inference shifts hardware mix

Bank of America’s latest AI hardware view is less about a single winning chip than about a wider bill of materials. The bank’s argument, echoed in recent market notes and social posts citing its research, is that inference and agent-style workloads are pushing spending toward CPUs, memory and networking alongside custom accelerators. That claim lines up with two recent datapoints: Broadcom’s triple-digit growth in AI semiconductor revenue and Intel’s June launch of Xeon-based rackscale systems aimed at inference and agentic AI. Broadcom said on September 2 that AI semiconductor revenue reached $16.7 billion in its fiscal third quarter, up 221% from a year earlier. Intel said at Computex 2026 that it was rolling out rackscale AI infrastructure built on Xeon processors and SambaNova SN-50 Reconfigurable Dataflow Units, with Foxconn providing system integration. Together, those updates show vendors selling not just chips, but full system architectures for the next phase of AI demand. ### Why would inference change the hardware mix inside a datacenter? Bank of America said in a May research note that the server CPU market could grow from about $43 billion in 2026 to $125 billion by 2030, driven by the expanding role of processors in agentic AI workloads. The bank also raised its view of AI-related infrastructure demand across compute, networking and memory, according to reports summarizing the note. (dailyaibrief.com) Inference workloads differ from training in how they move data through systems. MIT Technology Review wrote on September 4 that AI inference changes optimization from raw compute toward “coordinated infrastructure” across memory, storage and networking. That framing matches Bank of America’s view that more of the value stack sits in system design, not only in the accelerator itself. (ca.finance.yahoo.com) ### Why is Broadcom central to this argument? Broadcom gave investors the clearest recent revenue proof point. The company said AI semiconductor revenue rose 221% year over year to $16.7 billion in the quarter ended August 2, and Chief Executive Hock Tan guided to $21.7 billion in AI revenue for the fiscal fourth quarter. That growth matters because Broadcom’s AI business is tied heavily to custom silicon and networking for hyperscale customers. (technologyreview.com) Reports on the quarter said those sales were driven by demand for bespoke accelerators and related infrastructure, reinforcing the idea that hyperscalers are buying integrated platforms rather than only off-the-shelf GPUs. (dailyaibrief.com) ### What is Intel saying about Xeon’s role in inference racks? Intel used Computex in Taipei on June 2 to position Xeon as part of rack-level AI systems for inference and agentic workloads. The company said the new infrastructure combines Xeon processors with SambaNova RDUs, while Foxconn handles system integration and plans manufacturing for a CPU-dense rack architecture. (techtimes.com) Intel said those systems are designed to deliver AI inference with improved cost and power efficiency. The company’s wording is notable because it places the CPU in the control and orchestration layer of the rack, rather than treating it as a supporting part beside an accelerator. ### Which engineering jobs does that put in focus? Bank of America’s thesis implies more work at the boundaries between components. (intc.com) If inference clusters need tighter coordination among processors, memory subsystems, interconnects and custom accelerators, the hiring pull is likely to favor engineers who can work across board design, system architecture, memory behavior, thermal constraints and networking fabrics — an inference supported by the bank’s broader market forecasts and by vendors’ own rackscale product announcements. For students and early-career hardware engineers, that points less to a single “AI chip” lane than to roles in co-design, validation, packaging-adjacent memory work and datacenter networking. Broadcom’s next quarterly AI revenue target is $21.7 billion, and Intel’s rackscale inference systems remain part of its post-Computex rollout with partners including SambaNova and Foxconn. (msn.com) (ca.finance.yahoo.com)