Table of Contents
AI infrastructure demand is broadening from accelerators into full-stack systems, while custom silicon is becoming a meaningful counterweight to merchant GPUs
First things first, Dell Technologies and Broadcom this week delivered monster earnings reports. But beyond the financials, commentary from management provided a useful view of the AI infrastructure market from two different points in the value chain. Dell sits close to the deployment layer, integrating compute, networking, storage, power and cooling into systems that customers can put into production. Broadcom sits further upstream, where custom accelerators, networking silicon and advanced packaging increasingly determine the economics of the largest AI clusters.
The shared message that I took away was that AI infrastructure has moved beyond a simple GPU procurement cycle. Demand remains extraordinary, but the opportunity is widening around the accelerator while supply constraints are becoming more distributed across the stack.
Dell’s fiscal second quarter numbers make the scale of demand clear. The company booked $60.9 billion of AI server orders in the quarter, recognized $16.4 billion of AI server revenue and exited with $95 billion in backlog. Its pipeline continued to grow sequentially and remained “multiples” of backlog despite Dell converting $131.7 billion of AI demand into orders during the preceding 12 months. Dell subsequently raised its full-year AI server revenue forecast to $74 billion.
But the more interesting figures highlighted where that demand is going. Dell’s AI Factory customer count has surpassed 6,500, and adoption is accelerating with 3,300 of those customers added during the past three quarters, compared with eight quarters to reach the first 3,200. Enterprise revenue, repeat buyers and the enterprise pipeline all increased, with those customers also tending to buy more networking and storage as part of a more complete solution.
That supports the thesis behind the AI Factory model. As Dell COO Jeff Clarke put it, “AI infrastructure requires much more than assembling and delivering components.” Some customer engagements require more than 50 individual designs as Dell optimizes systems around workload performance, power, cooling and the physical data center environment.
In other words, the value pool is spreading outward from the GPU. Dell reported traditional server and networking revenue up 122% and storage revenue up 26%; company leaders attributed part of both trends directly to agentic workloads. AI agents create CPU demand for orchestration and runtime functions while also generating logs, traces, files and KV-cache requirements that increase demand for storage and data management.
This is an important evolution of the enterprise AI infrastructure story. Dell AI Factory started as a way of simplifying access to accelerated computing; now, it looks like a mechanism for modernizing the entire data center around AI. Dell’s latest portfolio expansion reflects that shift, spanning Rubin-based PowerEdge systems, networking, storage, data platforms and CPU infrastructure for agentic AI.
Broadcom’s results show a complementary transition happening at the highest end of the market. The company reported $16.7 billion of AI semiconductor revenue in fiscal Q3, up 221% year-over-year, and expects that figure to reach $21.7 billion in Q4. More tellingly, custom XPU shipments increased more than 3.5x and represented 73% of AI revenue during the quarter, while AI networking revenue increased more than 2.5x. The company named Google, Anthropic, OpenAI and Meta among its six XPU customers and detailed multiple generations of custom accelerators already shipping or in development.
Importantly, this does not suggest merchant GPUs are being displaced broadly. It does suggest that custom silicon remains concentrated among a small number of frontier model developers that have the workload scale and engineering resources to justify it. But within that cohort, custom acceleration is rapidly becoming a major architectural choice.
The economics explain why. Broadcom CEO Hock Tan explained that an XPU co-designed for a specific model can deliver better performance on that workload and run at “half the cost of a GPU” or less. That is Broadcom’s claim rather than an independent benchmark, but the deployment trajectory is significant regardless. Anthropic is expected to move from one gigawatt of Broadcom-enabled TPU capacity in 2026 to another five gigawatts in 2027, while OpenAI and Meta are also moving through successive generations of custom silicon.
Broadcom consequently expects fiscal 2026 AI semiconductor revenue of approximately $58 billion, followed by roughly $115 billion in fiscal 2027 and $230 billion in fiscal 2028. Even more striking than those numbers was Tan’s qualification: “Our demand actually exceeds this outlook, and we will work to improve supply.”
This is where Dell and Broadcom converge. For Dell, the constraint isn’t accelerator availability. Clarke summarized the memory access situation as “DRAM, DRAM, DRAM followed by NAND, NAND, NAND,” before listing shortages in CPUs, disk drives, mature-node power components, substrates, glass, optics, cooling distribution units and power racks. “The AI supply chain is working red lines all out,” he said.
Broadcom described essentially the same problem from the semiconductor side. Its outlook incorporates availability of leading-edge wafers, advanced substrates and HBM, but also whether customers can actually bring land, power and data center shells online in time. Tan called AI deployment a “multidimensional problem” in which any one of those inputs can become the bottleneck. Broadcom is responding by bringing substrate capacity online in Singapore and more than tripling capacity across EML, CW and VCSEL optical components, with management saying laser demand is already far exceeding industry supply.
The implication for the broader AI infrastructure supply chain is that scarcity is becoming systemic rather than component-specific. The first phase of the AI buildout was characterized primarily by accelerator scarcity. The next phase is constrained by the rate at which the entire ecosystem can manufacture, power, connect, cool, finance and deploy increasingly dense compute.
That also changes where competitive advantage resides. Broadcom’s moat goes beyond designing accelerators to simultaneously securing wafers, HBM, packaging, substrates and networking and helping customers plan multi-gigawatt deployments years ahead. Dell’s advantage goes beyond selling servers to integrating those scarce components into functioning AI factories and getting customers from an order to what Clarke repeatedly described as the “first token.”
Dell and Broadcom provided insight into two parallel growth engines. Frontier model builders are concentrating enormous spend into customized, gigawatt-scale infrastructure optimized for inference economics. At the same time, enterprise AI is distributing demand across thousands of customers buying integrated compute, CPU, networking, storage and data infrastructure.
The outlook for AI demand remains exceptionally strong. And the key to continuing this exceptionally strong outlook is mustering a supply chain that can turn demand into deployed, powered and productive capacity.
