Artificial Intelligence Data Centers Pivot to Agents

The biggest hardware news from Computex 2026 week points one direction: the artificial intelligence data center is being reorganized around agentic workloads, with CPUs, networking silicon, and component suppliers moving to the center of a story that used to be about GPUs alone. Here are the five developments that matter most for anyone building or operating technology and telecommunications infrastructure.

---

Intel Builds the Rack Around the Agent

Intel used Computex 2026 to make its clearest statement yet that the AI data center is no longer a GPU-only conversation. Working with Foxconn and other infrastructure partners, the company unveiled rack-scale reference designs that pack up to 128 processor sockets into a 100-kilowatt envelope. Configured with Intel's new 288-core Clearwater Forest Xeon 6+ chips, a single rack can reach 36,864 cores and up to 384 terabytes of memory. Two variants are on offer: one tuned for latency-sensitive agents, another for maximum density.

The reasoning behind the design tells you where AI infrastructure is heading. Large language models still run on GPUs, but the agent software wrapped around those models — the part that places API calls, queries databases, runs code interpreters, and drives shells — leans heavily on general-purpose CPU cores. As organizations move from single prompts to agentic AI that completes multi-step tasks, that coordination work scales fast, and it scales on cores. Intel's chief executive framed the launch as a response to customers asking the company to "think at the system level" to serve real agentic workloads at scale. Together.AI signed on as an early commercial customer.

The Clearwater Forest chips themselves are a milestone for Intel. They are the first data center processors built on the company's 18A manufacturing process, offering up to 288 efficiency cores at thermal envelopes between 330 and 450 watts. Intel claims a single Clearwater system can replace nine of its previous-generation Xeon servers, a consolidation pitch aimed at telecom, cloud, and edge operators watching both power budgets and floor space.

Intel is not alone in chasing this market. Nvidia has its own multi-core Vera CPU built for agentic reasoning, and Arm recently entered production silicon with a 136-core design aimed at the same workloads. The result is a three-way race to supply the processors that keep AI agents fed — a competition that barely existed a year ago. For operators sizing AI clusters, the lesson is direct: under-provisioned CPU capacity leaves expensive accelerators idling while agents wait on coordination work the GPUs cannot do.

Source: The Register

---

Marvell Ships the First 102.4 Tbps AI Switch

While the chip launches drew headlines, Marvell made the case that the network is now just as decisive a constraint inside the AI data center. The company announced availability of the Teralynx T100, which it describes as the industry's first 102.4 terabit-per-second switch silicon, built on an advanced 3-nanometer process and aimed squarely at large AI training and inference clusters.

The pitch centers on efficiency. A cluster of tens of thousands of accelerators only behaves like one machine if those chips can exchange data at very low latency, and the switching fabric that connects them has become a major consumer of power and a frequent source of bottlenecks. Marvell says the T100 draws up to 25 percent less power than competing parts while delivering the lowest latency in its class. Its high-radix design — supporting a 512-port scale-out — lets operators flatten their network topology, cutting the number of switching tiers and optical links a large cluster needs. Fewer tiers means fewer places for delay to accumulate, which in turn raises the utilization of the GPUs the network exists to serve.

Marvell is shipping the part in several configurations, including co-packaged optics, and supports an open software ecosystem built around the SONiC operating system and the Open Compute Project's switch abstraction interface. That open-standards posture matters to the cloud and telecom buyers who do not want to be locked into a single vendor's networking software. A company executive summarized the design goal bluntly, saying the switch was purpose-built for AI and stripped of the legacy elements that inflate power consumption in older switching platforms.

The timing is not a coincidence. As AI clusters spread across multiple buildings and even multiple sites to find available power, high-bandwidth, low-latency interconnect becomes the difference between a working system and a set of stranded islands. AI networking has quietly turned into one of the highest-stakes problems in AI infrastructure, and it is increasingly where artificial intelligence and telecommunications meet — optical transport, high-speed switching, and long-haul fiber converging on the data center. The T100 begins sampling to customers this quarter.

Source: Marvell

---

Broadcom Bets on Custom, Efficient Silicon

Not every AI chip story is about raw performance. Broadcom added South Korea's FuriosaAI to its roster of custom-silicon partners, agreeing to build third-generation accelerators using Broadcom's advanced packaging, a 2-nanometer process, and next-generation HBM4 memory. The deal is a window into how the AI chip market is broadening beyond a single dominant supplier.

FuriosaAI's current-generation RNGD accelerator competes on efficiency rather than raw horsepower. It delivers roughly 512 teraflops of FP8 compute at about 180 watts. An Nvidia B200 offers far more throughput while drawing close to 1,000 watts. Furiosa's wager is that efficiency wins a meaningful slice of the market: a chip modest enough to run in standard, air-cooled data centers is easier and cheaper to deploy when power and cooling are the scarcest resources in the building. Electronics maker LG is already using the chips for AI inference.

For Broadcom, the partnership underscores a strategic shift that has become hard to miss. Custom accelerator work — designing bespoke AI chips for large customers rather than selling off-the-shelf parts — now accounts for roughly 65 percent of the company's quarterly revenue. Broadcom has moved from a behind-the-scenes supplier to a visible participant in the AI ecosystem, with publicly acknowledged work for hyperscale customers. Adding a partner like Furiosa extends that model to a new class of buyer.

The broader signal for the technology and telecommunications sector is that the AI chips powering data centers are diversifying. Hyperscalers are designing their own accelerators to reduce dependence on any one vendor, custom-silicon specialists are filling efficiency niches, and the assumption that every AI workload must run on the most powerful — and most power-hungry — GPU is loosening. For operators, more silicon options mean more ways to match a chip to a workload's actual latency, efficiency, and budget requirements, rather than over-buying performance they will not use.

Source: The Register

---

STMicroelectronics Doubles Its Data Center Forecast

A quieter announcement from STMicroelectronics offered a useful gauge of how far the AI infrastructure boom now reaches. The European chipmaker raised its 2026 data center revenue target to roughly $1 billion — double its earlier projection — and said it expects that figure to double again in 2027. The company attributed the upgrade to continued strong, AI infrastructure-led demand and recent progress ramping up its manufacturing capacity.

STMicroelectronics does not build the marquee GPUs that dominate AI coverage. It supplies the less visible components — power management, analog, and related silicon — that every rack of accelerators depends on to run reliably. That is exactly why the forecast matters. When a broad-line component supplier doubles its data center expectations, it confirms that the money flowing into AI is fanning out well beyond a handful of GPU vendors and into the wider supply chain that makes high-density computing possible.

The revision lands amid a string of data points pointing the same way. As racks climb from tens of kilowatts toward hundreds, the power-delivery and thermal-management components inside them become both more numerous and more demanding, and suppliers that can meet those specifications at volume are seeing demand they did not forecast a year ago. For a company that serves more than 200,000 customers across many industries, a billion-dollar data center business is a meaningful share of attention — and a sign of where it expects growth to come from next.

For readers tracking the technology and telecommunications sector, the takeaway is that AI data center spending is becoming a rising tide for the entire semiconductor and components industry, well beyond the chips that grab attention. The suppliers positioned in power and analog stand to benefit from the same buildout driving the headline GPU and CPU launches, and their order books are an early indicator of how durable the demand really is.

Source: STMicroelectronics

---

The Buildout Shifts From Chips to Systems

Tying the week together, a hardware roundup from Data Center Knowledge captured the through-line: the AI buildout is moving from chasing chips to engineering the systems that keep racks running at scale. The framing is a useful corrective to a year of coverage fixated almost entirely on GPU supply.

The evidence is spread across the industry. AMD posted booming growth in its data center business, a reminder that the accelerator market has more than one serious player. Google and Blackstone formed a multibillion-dollar venture focused on custom AI chip infrastructure. GPU rental prices, after a long climb, have begun to show signs of compression as supply normalizes in parts of the market — a notable turn after years of scarcity. And operators are rethinking the unglamorous systems around the silicon: battery storage is starting to replace diesel as backup power, and hollow-core fiber is under evaluation for lower latency and energy use inside facilities.

What connects these threads is a maturing view of what an AI data center actually requires. Power delivery, cooling, networking, and components are increasingly co-designed rather than bolted together after the fact, because at megawatt-scale rack densities, a weakness in any one of them caps the performance of the whole. The center of gravity in AI infrastructure is moving up and out from the individual chip toward the system that surrounds it.

For technology and telecommunications operators, that shift reframes the buying decision. Sourcing the fastest accelerator is no longer enough; the harder questions are whether the network fabric, power envelope, and cooling can keep that accelerator busy, and whether the agentic AI workloads running on top can be operated within a predictable budget. The companies that treat AI infrastructure as a system — and that keep tight control over what their agents cost to run — are the ones positioned to turn this buildout into something durable rather than a sequence of expensive experiments.

Source: Data Center Knowledge

---

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