Networking Watch: AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI Details Progress on Optical Compute Interconnect Standard
Six AI Giants Just Agreed on How Light Should Move Inside a GPU Rack
The Electrical Domain Just Ran Out of Room
At last week's Hot Interconnects Symposium, engineers from AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI detailed progress on the Optical Compute Interconnect (OCI) standard — an open, silicon-centric wavelength-division-multiplexing spec the six companies jointly launched in March specifically to move data optically inside a single GPU rack, not just between racks. Meta's principal optical infrastructure engineer put the problem plainly: as scale-up compute domains grow, clusters are "exceeding the bounds of what can be attached in an electrical domain of a single rack." Copper has run out of reach.
OCI is one of six interconnect Multi-Source Agreements the industry signed in March 2026 — alongside Open Co-packaging, eXtra-dense Pluggable Optics, Linear Pluggable Optics, Expanded Beam Optical, and multi-core fiber — all aimed at the same underlying shift: optics moving from the edges of the data center into the rack itself. Analysts covered in today's Semiconductor Watch tie the shift to hard numbers — the global AI data-center optical interconnect market is projected to reach $144.4 billion by 2030, up from $13.7 billion in 2024, a 48.1% compound annual growth rate. Arista's own supply-chain posture, detailed below, shows a vendor racing to secure component supply well ahead of that curve.
01 — The Networking Lead
Nvidia's Biggest Rivals and Customers Just Standardized Optical Scale-Up Together
The Optical Compute Interconnect MSA is a notable document for who signed it as much as what it specifies: AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI — companies that compete fiercely for AI accelerator and cloud market share — co-authored an open standard for the one layer none of them wants to solve alone. OCI targets scale-up networking specifically, the tightest and most latency-sensitive tier connecting GPUs within a single compute domain, where Meta's engineers say electrical interconnects are now the binding constraint on how large that domain can grow before signal loss and power draw make copper impractical.
The standard is one of six MSAs signed in March spanning the full optical transition: Open Co-packaging (Open CPX), eXtra-dense Pluggable Optics (XPO), Linear Pluggable Optics (LPO), Expanded Beam Optical (EBO), and four-core multi-core fiber (SDM4 MCF) round out the set. Together they represent an unusual level of pre-competitive coordination among AI's biggest buyers and sellers of compute — all racing toward the same conclusion: within a few years, effectively every high-bandwidth interconnect inside an AI data center, from GPU-to-GPU links to inter-facility fiber, will be optical.
02 — AI Networking Watch
Arista Networks' latest quarter is the clearest read yet on how tight AI networking supply has become. Revenue crossed $3 billion for the first time — $3.036 billion, up 37.7% year over year — and CEO Jayshree Ullal described demand as "significant" even as the company works through a "tight" supply chain. The more consequential number sits below the revenue line: Arista's multi-year purchase commitments to component suppliers nearly tripled year over year, from $3.6 billion to $9.7 billion, as the company locks in memory, integrated circuits, and optical components years in advance rather than risk getting caught short.
Arista's Etherlink AI fabric customer base has grown from four or five hyperscaler deployments in 2024 to more than 100 cumulative customers today, and management is explicitly pushing further into non-Nvidia accelerator networking — AMD GPUs, Google TPUs, custom hyperscaler silicon — where network architecture isn't locked to a single vendor's proprietary interconnect the way InfiniBand ties customers to Nvidia. That diversification strategy is precisely the dynamic the OCI standard above is built to reinforce: an open interconnect layer that works across accelerator vendors rather than favoring whichever company owns the switch silicon.
03 — Optical Networking
The market data backs up the standards activity. A China Insights Consultancy report, drawing on LightCounting data, projects the global AI data-center optical interconnect market growing from $13.7 billion in 2024 to $144.4 billion by 2030. Silicon photonics is expected to drive the shift, climbing from 16.6% of that revenue in 2020 to 63.7% by 2030, with its share of the market compounding at 68.5% annually — more than double the growth rate of every other optical technology combined. The crossover point, where silicon photonics passes half the total optical interconnect market, is projected around 2027.
Broadcom's Tomahawk 6-Davisson — the industry's first 102.4 Tbps Ethernet switch with integrated co-packaged optics — has already moved from announcement to production deployment inside Meta's AI clusters, an important reliability proof point given how much the industry's optical transition depends on CPO matching or beating pluggable transceivers on uptime, not just bandwidth. Marvell, meanwhile, is leaning on its Celestial AI acquisition's Photonic Fabric technology to position itself as a credible optical-interconnect challenger to Nvidia, particularly with hyperscalers like Amazon and Meta building their own custom accelerators and needing an optical layer that isn't tied to Nvidia's stack.
04 — Cloud & Enterprise Networking
Data-center operator Switch Inc. has filed confidentially for a U.S. IPO, working with Bank of America, Citigroup, Goldman Sachs, JPMorgan, and Morgan Stanley on an offering that could launch as soon as November — another sign that public-market investors want direct exposure to the physical infrastructure layer underneath AI, not just the model and chip companies sitting on top of it. On the enterprise side, HPE continues pushing its Juniper-integrated, GPU-utilization-focused networking strategy toward a cumulative "Networks for AI" order target of at least $2 billion by the end of fiscal 2026 — evidence that the scale-up and optical trends dominating hyperscaler headlines this week are trickling down into mainstream enterprise AI deployments on a lag, not staying confined to the handful of trillion-dollar cloud platforms building frontier models.
05 — Wireless & Edge
Away from the data center, researchers have developed a rice-grain-sized chip aimed at next-generation millimeter-wave connectivity, part of a wider push to shrink the silicon needed for dense 5G-and-beyond wireless deployments. Separately, chip designers are exploring giving Arm-based processors dedicated support for virtual radio access networks (vRAN), a move that would let telecom operators run more of their radio infrastructure on general-purpose, software-defined hardware rather than proprietary baseband silicon. Neither story rivals the scale-up optical news above, but both point to the same underlying pattern: connectivity silicon is getting more specialized and more purpose-built at every layer of the stack, from rack-scale GPU interconnects down to the radio edge.
06 — Network Economics
Arista's decision to nearly triple its forward purchase commitments is itself an economic signal worth reading closely: when a vendor with $13.34 billion in cash on hand chooses to lock in $9.7 billion of multi-year component orders rather than buy on the spot market, it is pricing in sustained scarcity, not a temporary blip. Management specifically cited securing memory supply — DDR4, DDR5, and NAND — into 2027, alongside qualifying new suppliers to reduce single-vendor risk, a hedge against exactly the kind of concentrated supply chain that made last week's FCC transceiver-restriction news (covered in a prior edition of this series) so consequential in the first place.
That policy backdrop still shapes who benefits economically from the optical build-out described above. Chinese vendors currently account for nearly two-thirds of global optical transceiver unit shipments, led by Innolight's 27% revenue share, with Coherent second at 17%, Eoptolink third, and Lumentum around 6%. As silicon photonics revenue compounds toward that $144.4 billion 2030 figure, even modest share shifts — whether from policy restriction, reliability track record, or the new open MSA standards reducing switching costs between suppliers — translate into billions of dollars moving between a small set of companies.
07 — Capital & Competition
The competitive landscape in optical interconnects is unusual in that its biggest players are simultaneously collaborators (through the OCI and related MSAs) and rivals racing to out-invest one another. Nvidia has committed roughly $4 billion combined to lock up supply from Coherent and Lumentum; Marvell paid up to $5.5 billion for Celestial AI to build its own competing Photonic Fabric; and private optical-interconnect startups continue drawing outsized rounds — Ayar Labs closed a $500 million Series E earlier this year at a $3.75 billion valuation, backed by Nvidia, AMD, and the Qatar Investment Authority among others, betting its TeraPHY chiplets can compete directly with Marvell's acquired technology. Every one of those companies also sits inside the same standards bodies working to keep the underlying interconnect layer open enough that no single vendor can fully lock in the market the way Nvidia did with InfiniBand. Whether that tension resolves toward genuine multi-vendor interoperability or toward a handful of vertically integrated winners is the central competitive question this series will keep tracking through year-end.
Three Networking Signals
- Optics has moved inside the rack, not just between racks. The OCI standard's entire premise is that electrical interconnects can no longer serve even scale-up, intra-rack GPU connectivity at current bandwidth demands — a threshold the industry didn't expect to cross this early in the AI buildout.
- AI's biggest rivals are choosing to standardize the interconnect layer together rather than compete on it. AMD, Broadcom, Meta, Microsoft, Nvidia, and OpenAI co-authoring OCI suggests the industry sees more collective value in an open, interoperable optical layer than any one company sees in trying to own it outright — for now.
- Component scarcity is now a multi-year planning problem, not a quarterly one. Arista nearly tripling its forward purchase commitments to $9.7 billion, and specifically locking in memory supply through 2027, shows vendors treating AI networking component scarcity as a structural feature of the market rather than a temporary bottleneck to wait out.
The CODEW Take
The Optical Compute Interconnect standard is a quiet admission from the industry's most powerful companies that no single vendor — not even Nvidia — wants to solve rack-scale optical interconnect alone, because the physics problem is now bigger than any one company's roadmap. That is a meaningfully different posture than the GPU market, where competitive advantage comes from proprietary differentiation, not open standards. Arista's purchase-commitment numbers show why: when a networking vendor is willing to lock in $9.7 billion of multi-year component orders and specifically cite memory and optics scarcity extending into 2027, that is a company planning for a structural shortage, not a cyclical one. Layer in a $144 billion projected optical interconnect market by 2030, a Chinese supply chain that still controls roughly two-thirds of global transceiver shipments, and six of AI's largest companies choosing collaboration over competition on the interconnect standard itself, and the picture that emerges is of an industry where the network has become too important, and too physically constrained, for any single company to fully own — even as billions of dollars in capital keep flowing toward exactly that goal.
Reviewed by Erwin Castro
on
Monday, August 31, 2026
Rating: