AMD buys photonics startup Enosemi to supercharge rack-scale AI networking

AI clusters have grown so massive that GPUs are no longer the only bottleneck. The real headache is moving data fast enough between hundreds of servers without melting power budgets.

AMD announced the acquisition of Enosemi, a quiet Silicon Valley photonics startup. Terms of the deal were not disclosed, but AMD made its mission crystal clear: supercharging in-house co-packaged optics (CPO) development to wire up next-generation AI systems.

Wiring AI racks with light

Modern frontier models span multiple server racks packed with compute, networking gear, and cooling setups. Historically, hyperscalers had to piece together these parts manually.

Nvidia simplified this playbook with its integrated DGX systems, offering full-rack solutions ready out of the box. AMD has been sprinting to match that capability, spending $4.8 billion last year to acquire rack-scale system specialist ZT Systems. Adding Enosemi gives that strategy its optical backbone.

Sacramento-based Enosemi brings PhD-level talent with rare, proven expertise in shipping photonic integrated circuits at volume. According to Brian Amick, AMD senior vice president of technology and engineering, optical interconnects offer a compelling path forward as data movement demands accelerate at the rack level. Enosemi had already worked with AMD as an external development partner before the acquisition.

Ditching pluggable transceivers

In standard AI data centers, electrical signals from servers must convert into light to travel over fiber cables, and then switch back to electricity at the switch. Pluggable transceivers handle this translation.

The problem? Massive clusters require thousands of these transceivers. They are expensive, guzzle power, and add latency every time data makes the trip between components.

Co-packaged optics solves this by baking the optical transceiver directly into the same package as the switch or compute silicon. Eliminating external transceivers slashes hardware costs, tightens latency, and massively improves energy efficiency by letting data travel mere millimeters instead of across external boards.

Why it matters

For enterprise infrastructure buyers and cloud operators, this deal signals where AI economics are heading. Buying raw GPU flops is no longer enough; connectivity and power efficiency inside the rack will dictate who can actually scale clusters affordably. If AMD delivers working CPO systems, customers could bypass millions of dollars in external optical transceiver purchases while trimming power bills.

Across the industry, this rounds out AMD's systematic full-stack assembly against Nvidia. Over recent years, AMD has added Xilinx for adaptive silicon, Pensando for networking, Silo AI and Mipsology for software, and ZT Systems for rack manufacturing. Enosemi provides the optical interconnect puzzle piece AMD needed inside those racks.

We think acquiring a team that has already delivered photonic chips at volume is a smart way to de-risk next-gen architecture. Packaging optics directly onto silicon is notoriously tricky, with serious thermal and manufacturing hurdles. If AMD can iron out the yield and packaging challenges, it positions itself as one of the few players capable of delivering a true turnkey, photonics-enabled AI rack.