# Artificial Intelligence Data Center, Rebuilt Around Agents > A cluster of Computex 2026 chip and networking launches — Intel and Foxconn's 100kW agent-focused rack, Marvell's first 102.4 Tbps switch, Broadcom's new custom-silicon partner, and STMicroelectronics' doubled data-center forecast — shows AI data center hardware reorganizing around agentic workloads rather than model training alone. The center of gravity is moving from the GPU toward the CPUs, switches, and power systems that decide whether a fleet of agents can run, and what it costs operators to run them. Content type: article Source URL: https://www.agentpmt.com/articles/artificial-intelligence-data-center-rebuilt-around-agents Markdown URL: https://www.agentpmt.com/articles/artificial-intelligence-data-center-rebuilt-around-agents?format=agent-md Updated: 2026-06-03T13:15:53.917Z Author: Stephanie Goodman Tags: autonomous agents, Multi-Agent Workflows, AI Agents In Business, AI Powered Infrastructure, DynamicMCP, Enterprise AI Implementation, News --- # Chipmakers Spent Computex Week Rebuilding the AI Data Center Around Agents At Computex 2026 this week, Intel and Foxconn showed a server rack built to do one job: run AI agents. Their reference design packs as many as 36,864 CPU cores into a single 100-kilowatt rack, and the pitch behind it had almost nothing to do with training larger models. "Our customers are asking us to think at the system level to help them serve real agentic workloads at scale," Intel chief executive Lip-Bu Tan said. The first commercial customer is Together.AI. That framing — the rack as a machine for running agents, sized around the work agents actually create — ran through the whole week. Marvell began shipping the first 102.4 Tbps network switch. Broadcom named a new partner for custom AI silicon built around efficiency rather than peak speed. STMicroelectronics, a components supplier most buyers never think about, doubled its forecast for data-center revenue this year. Read together, the announcements describe an artificial intelligence data center whose center of gravity is shifting away from the GPU alone and toward the CPUs, switches, and power systems that decide whether a fleet of agents can run at all. ## The rack gets rebuilt for agents For most of the current artificial intelligence build-out, the story has been about graphics processing units (GPUs) — the chips that train and run large models. Agents change the mix. An AI agent rarely does one thing and stops. It plans a step, calls a tool, reads the result, calls another tool, runs a snippet of code, queries an interface, and loops. Each of those steps is coordination work, and coordination runs on conventional CPU cores, not on the tensor math inside a GPU. The model still does the thinking; a surrounding set of CPU-side "agent harnesses" handles the tool calls, API requests, code interpreters, and shells the agent leans on to get anything done. That is why Intel built its newest rack around the CPU. Working with Foxconn, it designed two blueprints: one tuned for latency-sensitive agents, the other for maximum density. The dense configuration uses Intel's new 288-core Clearwater Forest Xeon 6+, manufactured on the company's 18A process, and it is the one that reaches 36,864 cores in that 100-kilowatt rack. Intel says a single Clearwater Forest system can stand in for nine of its second-generation Xeon servers — the kind of consolidation that means something to an operator paying for floor space, cooling, and power rather than admiring a core count. Intel is not the only company deciding the CPU belongs at the center of an agent cluster. Nvidia is pairing its GPUs with a new server-CPU line of its own, and Arm has two agent-focused designs aimed at the same workloads. A three-way race over who supplies the processor in an agentic rack would have sounded unlikely a year ago, when the entire conversation was about GPU supply. For the people sizing these clusters, the practical lesson is that CPU headroom is now a first-order budget decision. Under-provision it, and expensive GPUs sit idle waiting on the coordination around them to feed work through. Plan the cores poorly and the most costly part of the rack runs below capacity. That efficiency argument is also where the software running on top of this hardware starts to matter. Faster cores make each agent cheaper to run; whether that headroom turns into lower cost depends on how the orchestration manages it. AgentPMT, an integration platform that lets agents call thousands of managed tools, uses an approach it calls [Dynamic MCP](https://www.agentpmt.com/dynamic-mcp) to deliver a tool's definition on demand, only at the moment the agent needs it, instead of loading every available tool into the model's context up front. As the number of tool calls per agent climbs — exactly the curve this hardware is being built to absorb — keeping that context lean is part of what keeps the bill from climbing with it. ## The bottleneck moves to the network and the supply chain Once you can pack tens of thousands of cores and a wall of GPUs into a rack, the constraint moves to whether all that silicon can talk to itself fast enough. A large training or inference job is split across thousands of accelerators that have to exchange data constantly; if the network between them stalls, the chips wait. That is the gap Marvell is aiming at with Teralynx T100, which it announced this week as the industry's first 102.4 Tbps switch chip. Built on a 3-nanometer process, the T100 is designed to flatten the network — using higher-capacity switches to cut the number of switching tiers and optical links a signal has to cross between two chips. Fewer hops mean lower latency and fewer expensive optical transceivers. Marvell says the part draws up to 25 percent less power than competing silicon, and across the largest artificial intelligence data centers, where racks already pull well over 100 kilowatts, a quarter off the networking power budget is headroom that can go back to compute or cooling. Efficiency is also the bet behind a quieter announcement. Broadcom added FuriosaAI to its roster of custom-silicon partners — companies it helps design chips tuned for one workload rather than the broad market a flagship GPU chases. Furiosa's current accelerator, RNGD, is built around power draw: it runs at roughly 180 watts, a fraction of the near-kilowatt a top-end GPU can demand, and it fits in standard air-cooled server rooms instead of requiring liquid cooling. It gives up raw throughput to do it. For an operator who cannot [get more power into a building](https://www.agentpmt.com/articles/energy-ai-meets-flexible-compute-nvidia-and-emerald-ai-build-data-centers-that-bend-to-grid-demand) or more cooling into a room — an increasingly common ceiling — a chip that does useful work inside those limits can beat a faster one that cannot be deployed at all. Custom accelerators like these now make up the majority of Broadcom's chip revenue, a sign of how far the market has moved past one-size-fits-all GPUs. The silicon underneath agents is splintering across Nvidia, Intel, Arm, and a growing list of custom parts from Broadcom's partners and the hyperscalers' own labs. For anyone building above it, that fragmentation is an argument for staying portable: AgentPMT, for one, [runs the same workflow](https://www.agentpmt.com/marketplace/agentpmt-workflow-creator) across different models and providers, so a change in the underlying chip or runtime does not force a rebuild of the agent. ## From buying chips to engineering systems Add these announcements up and a pattern shows in how the money is being spent. Data Center Knowledge, surveying the month's hardware news, described the shift as moving from chasing chips to engineering the systems that keep racks running at scale — power delivery, cooling, networking, and components designed together rather than bought piecemeal. Its roundup ranged across battery storage replacing diesel backup generators, hollow-core fiber under evaluation for lower-latency links, and a multibillion-dollar venture between Google and Blackstone to build custom AI chip capacity. The common thread is that the spending is fanning out well past the GPU. STMicroelectronics made the same point from the supply side. The company, which makes the power-management and analog components that sit alongside the marquee processors, doubled its 2026 data-center revenue target to roughly one billion dollars and signaled it expects to double again the year after, citing demand it traces directly to AI infrastructure. When a parts supplier most agent builders have never heard of suddenly sees a billion-dollar line of business, it is a fair measure of how much of the build-out happens below the level of the chips that get the headlines. The same forces reach past the hyperscale data center. [Telecom carriers](https://www.agentpmt.com/industries/technology-telecommunications) standing up edge sites for 5G and connected devices face the identical tradeoffs in miniature — CPU headroom, network capacity, and power, only in a cabinet at the base of a tower rather than a hall full of racks. Intel is pitching Clearwater Forest at exactly those telco, cloud, and edge deployments, and as agents move closer to users to cut latency, the cost of running them travels out to the edge with them. The case for AI for telecommunications ends up looking like the case for the core: pick the processor and the network with the power bill in mind. For AI and telecommunications operators, it is one continuous problem from the central office to the cell site, and the components that make this artificial intelligence technology run are the same ones in short supply everywhere. For the enterprises that will actually deploy agents on all this hardware, the systems-engineering mindset has a software counterpart. As compute gets scarcer and more expensive, the cost of running an agent is set less by the price of a chip than by how disciplined the agent is about spending it. A single agent can fan out into hundreds of billable tool and API calls in a way no human operator ever would, and without limits that can become a runaway bill as easily as a finished task. This is where [spend controls](https://www.agentpmt.com/articles/budget-ai-agents-like-cloud-not-like-headcount) stop being an afterthought. AgentPMT, again as one example, caps each agent with its own budget, restricts which tools and vendors it can spend against, meters usage in credits that refund automatically when an action fails, and records every call in an audit trail with human approval on hand for anything sensitive. Across a rack full of AI and machine learning workloads, that kind of governance is what keeps [an agent fleet](https://www.agentpmt.com/articles/from-assistant-to-workforce-operating-an-agent-fleet) from becoming the least predictable line in the budget. None of this means the GPU is being dethroned. It still does the heavy computation at the core of every model. What changed this week is where the rest of the attention went. The cores that coordinate the work, the switches that move it, the power systems that feed it, and the components most buyers never name are all being redesigned, at once, around the assumption that the workload is a fleet of agents. The operators who come out ahead will be the ones who engineer the whole system instead of shopping for the fastest chip — and who know, down to the individual agent, what that system costs them to run. * * * ## Sources - Intel and pals cram 36,864 CPU cores into a 100kW rack while chasing the agentic AI dragon — The Register - Intel launches 288-core Clearwater Forest Xeon 6+ on 18A — The Register - Marvell Announces Availability of Industry's First 102.4 Tbps Switch (Teralynx T100) — Marvell - Broadcom lands FuriosaAI as latest custom AI chip partner — The Register - STMicroelectronics raises its revenue ambition for Data Centers — STMicroelectronics - Data Center Hardware Highlights: June 2026 — Data Center Knowledge - Data Centers Accelerate in 2026 — StartupHub.ai