Networking Watch · November 2, 2026
Marvell's $115 Billion Lumentum Deal, AI Data Center Connectivity and the Optical Networking Race — Why the AI Infrastructure Battle Is Moving Beyond Chips
|
THE CODEW · Partner Intelligence
Ship better with data and AI skills
Hands-on Python, SQL, and AI learning from DataCamp, for technical teams and founders who want to build with data.
> python · sql · data analysis · ai
|
Run DataCamp → |
The AI Infrastructure Race Is Moving Beyond Chips to Connectivity
The next phase of AI infrastructure competition is not only about who builds the fastest processor. It is about how efficiently those processors communicate. Marvell Technology's $115 billion all-stock acquisition of Lumentum, announced November 2, is the clearest signal yet that optical connectivity has become a strategic layer of the AI stack rather than a commodity component. The deal combines Marvell's custom silicon and electrical interconnect with Lumentum's lasers, photonics, and optical circuit switches — creating what both companies call a category leader in AI connectivity.
The transaction is also a competitive move against Broadcom, which has built a broad AI data center hardware portfolio and remains Marvell's larger rival in custom accelerators and networking. And it lands in a market where the numbers are extraordinary: TrendForce estimates AI networking and interconnect revenue will grow from $37.9 billion in 2024 to $279.8 billion in 2027, a 95% compound annual growth rate. Optical pluggables alone are projected to reach $84.1 billion in 2027, with 1.6T modules contributing about 56% of that revenue.
This edition examines the Marvell–Lumentum deal in detail, then connects it to the broader networking landscape: Arista's Ethernet challenge to Nvidia's proprietary NVLink, the Cisco–Nvidia USFP optical form factor, and HPE's raised networking outlook. The through-line: as AI clusters scale to tens of thousands of accelerators, the bottleneck shifts from compute to data movement, and the companies that control the optical and Ethernet layers will capture a growing share of the AI infrastructure economy.
Dating note: The Marvell–Lumentum transaction is a November 2 announcement. The Arista, Cisco–Nvidia, and HPE developments are background from October and September, included to explain the competitive context. Each item states its original date.
Lead Story
Marvell to Acquire Lumentum in $115 Billion All-Stock Deal, Strengthening Its AI Data Center Networking Position
Announced November 2, 2026.
What happened: Marvell Technology agreed to acquire optical components maker Lumentum Holdings in an all-stock transaction valued at about $115 billion including debt, with an equity value of $102.54 billion. Lumentum shareholders will receive a fixed exchange ratio of 4.5 Marvell shares for each Lumentum share, implying $1,132 per share — a 23.9% premium to Lumentum's unaffected closing price on October 28. After closing, Marvell shareholders will own about 67% of the combined company and Lumentum shareholders about 33% on a fully diluted basis. Marvell CEO Matt Murphy will remain chairman and CEO, and three Lumentum directors, including CEO Michael Hurlston, will join Marvell's board. The companies expect more than $1 billion in annual run-rate cost synergies and said the deal should be accretive to non-GAAP EPS in the first year after closing, expected in the first half of calendar 2028. Marvell said the combination expands its total addressable market to $550 billion.
Why it matters: Murphy framed the deal as a response to a fundamental architectural shift: “AI infrastructure is undergoing a fundamental architectural shift toward optical connectivity, creating an opportunity that comes along once in a generation.” That is not marketing language. As AI clusters grow from hundreds to tens of thousands of accelerators, copper interconnects run into reach and signal-integrity limits, and optical links become the only way to move data at the required bandwidth and distance. Marvell already buys lasers from Lumentum; owning the source of those lasers removes a supply dependency and gives Marvell a photonics portfolio that complements its custom AI silicon and DSPs. It also directly strengthens Marvell against Broadcom, which has built a broader menu of AI data center hardware and is the dominant custom-chip partner for Google.
The market's reaction: Marvell shares fell sharply on the dilution implied by an all-stock deal of this size, with reports of declines ranging from about 8% to 13% at various points, before recovering to a smaller decline on the day. Lumentum shares rose. The divergence reflects the classic acquirer-discount problem: Marvell is issuing a large amount of stock to buy a company roughly half its size, and investors are pricing the near-term dilution against the long-term strategic benefit. Marvell also reaffirmed its preliminary third-quarter revenue guidance in line with its prior outlook, and Lumentum separately reported $1.26 billion in first-quarter revenue, slightly above the $1.25 billion consensus, led by pump lasers and optical circuit switches.
Evidence status: Company-confirmed transaction announced November 2, with terms, premium, ownership split and synergy targets disclosed by both companies and reported by Reuters, Bloomberg, The Wall Street Journal and CNBC. Deal completion is expected in the first half of 2028 and remains subject to regulatory approval. Connect to: Company Intelligence · Tech M&A Watch · Semiconductor Intelligence
Supporting Analysis
Arista Expands Ethernet's Challenge to Nvidia's Proprietary AI Networking Architecture
Background: October 7, 2026.
What happened: Arista Networks rolled out new Ethernet reference architectures and a group of integration partnerships — including AMD, Arm, Broadcom, Meta, Microsoft, and Qualcomm — aimed at making Ethernet-based scale-up and scale-out designs viable for AI infrastructure. The company is positioning standard Ethernet, through the Open Compute Project's ESUN initiative and the Ultra Ethernet Consortium, as an open alternative to Nvidia's proprietary NVLink and InfiniBand stacks. Arista's 7060EX7 switches, fiber patch panels, liquid-cooling manifolds, rack management and power shelves are all built on Ethernet rather than proprietary interconnect.
Why it matters: Scale-up networking — the interconnect between accelerators inside a rack — has historically been Nvidia's fortress. Arista's argument is that as the accelerator market diversifies beyond Nvidia to Microsoft Maia, Meta MTIA, AMD Helios and Google TPU, customers will prefer an open Ethernet fabric that works across vendors over a proprietary interconnect tied to one supplier. If that thesis is right, Ethernet becomes the default interconnect for heterogeneous AI clusters, and Nvidia's NVLink advantage narrows as its chip share declines. The ESUN backer list is notable: it includes Nvidia itself, alongside AMD, Arista, Broadcom, Cisco, HPE, Marvell, Meta, Microsoft, OpenAI and Oracle — a sign that even the incumbent sees value in supporting an open standard while defending its proprietary position.
Evidence status: Company-confirmed product and partnership announcements, reported by Network World and Fierce Network. Adoption claims are company-described. Connect to: Enterprise AI Intelligence · Hardware Watch
Cisco and Nvidia Develop USFP Optical Form Factor for 400G-Per-Lane AI Connectivity
Background: Announced October 12–13, 2026.
What happened: Cisco and Nvidia announced plans for Ultra-Dense Small Form Factor Pluggable (USFP), a new optical module form factor designed to support electrical signaling at 400 Gb/s per lane and aggregate capacity of up to 12.8 Tb/s. The specification was announced at the Open Compute Project Global Summit and is positioned as the first module designed for 400G-per-lane electrical signals, roughly doubling the per-lane bandwidth of current 200G optics while increasing density and reducing the network's share of power consumption.
Why it matters: Every generation of AI networking is constrained by how much data can move through a pluggable module at a given power and footprint. Doubling per-lane bandwidth to 400G lets AI builders scale compute without proportionally expanding the network's physical footprint or power draw — which matters because data center power is increasingly the binding constraint on AI expansion. USFP is also a standards play: by co-developing it with Nvidia, Cisco ensures that the optics ecosystem aligns with the accelerator roadmap. For Marvell and Lumentum, higher per-lane bandwidth increases the value of the optical components they supply, which is part of why the Marvell–Lumentum combination is strategically coherent.
Evidence status: Company-confirmed announcement, reported by Cisco, TelecomTV and Futurum. The specification is planned; production timelines are not yet disclosed. Connect to: AI Infrastructure Watch · Semiconductor Watch
HPE Raises Networking Growth Outlook and Lands a $1.2 Billion AI Infrastructure Order From Vultr
Background: September 30, 2026.
What happened: Hewlett Packard Enterprise raised its long-term networking revenue growth forecast and announced a $1.2 billion order from cloud provider Vultr to supply AMD Helios AI racks equipped with AI chips and HPE networking switches and software for Vultr's US data centers. HPE now expects networking segment revenue to grow at a compound annual growth rate in the high teens from fiscal 2026 to fiscal 2029, up from a previous forecast of 5% to 7%. Its data center networking revenue is targeted to grow at a low-to-high 50% CAGR through fiscal 2029. The company also raised its Juniper integration cost-savings target to $800 million in annual run-rate by the end of fiscal 2028.
Why it matters: HPE is the clearest example of a legacy enterprise networking vendor repositioning around AI infrastructure. Its networking unit, expanded through the Juniper acquisition, is now the fastest-growing part of the business, and the Vultr order shows that AI cloud providers — not just hyperscalers — are becoming a meaningful source of demand. HPE's framing is worth noting: Rami Rahim, who heads the networking unit, said: “AI is reshaping the technology stack, making the network more strategic and driving significant new demand from enterprises and service providers.” That is the same thesis driving Marvell, Arista and Cisco, and it explains why networking has become the most contested layer of the AI infrastructure stack.
Evidence status: Company-confirmed guidance and order announcement, reported by Reuters and Bloomberg. Connect to: Enterprise AI Intelligence · The Term Sheet
Market Context: The Numbers Behind the Optical Race
TrendForce's November 2 forecast puts AI networking and interconnect revenue at $279.8 billion in 2027, up from $37.9 billion in 2024 — a 95% CAGR. The composition is revealing: switches remain the largest segment at about $105 billion (roughly 38%), pluggable optical modules reach $84.1 billion, and NICs and DPUs reach $57.4 billion. Together, those three categories account for about 88% of the market.
Scale-up is the fastest-growing interconnect segment. Its share of AI networking value is projected to rise from 11% in 2024 to 24% in 2027, as GPU platforms move from server-board architectures to rack-scale designs with dedicated switch trays. Copper cable remains dominant inside the GPU rack in 2027, but Google's TPU already uses optical interconnect with optical circuit switching to scale clusters — a sign of where the architecture is heading.
The optical module mix is shifting to 1.6T. Scale-out bandwidth upgrades move the pluggable optics revenue center from 800G to 1.6T in 2027, with 1.6T expected to contribute about 56% of pluggable revenue. Even as 1.6T prices decline with volume, they remain more than twice the price of 800G, so the mix shift itself lifts revenue. TrendForce projects pluggable optics revenue growth of 76% year over year in 2027, outpacing unit growth.
New optical architectures redistribute value. Co-packaged optics moves the optical engine onto the switch package, shifting some optical value from pluggable modules into the switch itself. Optical circuit switching supports large-cluster interconnect through optical path switching rather than electrical packet switching. Both change which suppliers capture the value — which is precisely why Marvell wants to own lasers, photonics and switching silicon under one roof.
Why It Matters
Optical connectivity is becoming strategically important because AI clusters have outgrown copper. A single AI rack now connects dozens of accelerators that must exchange data continuously during training. At 200G-per-lane and beyond, copper's reach and signal integrity degrade, and optics become the only viable medium for the distances involved. That is why Lumentum's lasers and optical circuit switches matter to Marvell, and why Cisco and Nvidia are co-developing 400G-per-lane optics.
Marvell can strengthen its position against Broadcom, but the deal is not a knockout. Marvell gains laser supply, photonics technology and a broader connectivity portfolio. Broadcom retains a wider AI data center hardware menu, the Google custom-chip relationship and a strong carrier Ethernet position. The competitive question is whether customers prefer an integrated optical-plus-silicon supplier or a broader one-stop AI hardware vendor. Marvell's bet is that optical depth wins.
The optical landscape will consolidate around a few integrated suppliers. Lasers, DSPs, optical engines and switching silicon are increasingly designed together, and the companies that can co-optimize across those layers will have an advantage. Lumentum competing against Coherent and China's Zhongji Innolight was already a difficult position for a components maker; inside Marvell, it becomes part of a larger platform. Expect further consolidation among optical component suppliers.
Ethernet will become a stronger alternative to proprietary AI networking, but not immediately. Arista's ESUN and UEC efforts, plus Cisco's participation, show that the industry wants an open fabric. But Nvidia's NVLink remains the highest-performance scale-up interconnect, and Nvidia itself backs ESUN — a hedge that lets it support openness while defending its proprietary advantage. The transition will be gradual, and it will accelerate as non-Nvidia accelerators take share.
For hyperscalers and AI infrastructure providers, the implication is supplier concentration risk. If optical components consolidate into a handful of integrated suppliers, buyers will have fewer alternatives for a component that is becoming a bottleneck. That argues for multi-sourcing strategies and for supporting open standards that keep the module ecosystem competitive.
What to Watch
- Marvell–Lumentum regulatory review: The deal is expected to close in the first half of 2028. Watch for antitrust scrutiny, given the scale and the optical market concentration it creates.
- Marvell share reaction: Whether the dilution-driven sell-off persists or reverses as investors digest the strategic logic.
- Broadcom's response: Whether Broadcom accelerates its own optical and photonics roadmap or announces a competing acquisition.
- Lumentum's NPO/CPO wins: Lumentum disclosed a second major NPO/CPO win backed by more than $5 billion in purchase commitments — watch whether that converts into revenue.
- Arista ESUN adoption: Whether hyperscalers deploy Ethernet-based scale-up fabrics at scale, or continue to default to NVLink for Nvidia clusters.
- USFP specification progress: Whether the Cisco–Nvidia 400G-per-lane form factor is ratified and when products ship.
- HPE networking execution: Whether the high-teens CAGR guidance is met and whether the Vultr order expands.
- TrendForce forecast accuracy: Whether 1.6T optics ramp on schedule and whether pluggable revenue growth approaches 76% in 2027.
- Coherent and Zhongji Innolight: Whether these Lumentum competitors gain share as customers seek multi-sourcing.
- Co-packaged optics commercialization: Whether CPO moves from sampling to volume, and how it redistributes value between module makers and switch vendors.
Marvell's $115 billion bet on Lumentum is the clearest statement yet that the AI infrastructure trade is moving from compute to connectivity. For two years, the market rewarded whoever made the fastest accelerator. Now the constraint is moving data between accelerators, and the companies that control lasers, optics, and switching silicon are capturing a growing share of the value. Murphy's “once in a generation” framing is aggressive, but the architectural logic is sound.
The market's negative reaction to the deal is about dilution, not strategy. Marvell is issuing roughly a third of the combined company to Lumentum shareholders, and investors are pricing in the near-term EPS impact. That is the standard acquirer penalty for a large all-stock deal. If the optical thesis is right, the dilution will look cheap in retrospect. If the AI build-out slows, Marvell will have paid a very high price for laser capacity it could have sourced contractually.
Broadcom remains the more diversified competitor, and that matters. Marvell is buying depth in optics. Broadcom already has breadth across custom accelerators, Ethernet switching, optical DSPs and carrier networking. The competitive question is whether customers want one deep optical partner or one broad platform vendor. In AI infrastructure, where integration reduces risk, breadth usually wins — which is why Marvell's deal is a necessary move rather than a sufficient one.
Ethernet's challenge to NVLink is real, but it is a slow-burn story. Arista, Cisco, Broadcom, AMD, Meta and Microsoft are all pushing open Ethernet fabrics, and Nvidia's own participation in ESUN acknowledges the direction. But NVLink remains the performance leader for Nvidia clusters, and most AI workloads still run on Nvidia accelerators. Ethernet wins share as the accelerator market diversifies — not before.
The biggest risk in the optical trade is overcapacity. TrendForce's $279.8 billion 2027 forecast is a 95% CAGR from 2024, and forecasts of that magnitude invite capacity additions. If AI capex slows while optical capacity ramps, the component market could swing from shortage to glut quickly. Marvell is betting that demand outruns supply for years. The memory cycle suggests that bet does not always pay off.
Sources and Evidence Notes
Marvell–Lumentum transaction (November 2) — Reuters via Moneycontrol and The Globe and Mail; Bloomberg via Yahoo Finance; The Wall Street Journal; CNBC; Yahoo Finance. Arista Ethernet scale-up (October 7) — Network World; Fierce Network. Cisco–Nvidia USFP (October 12–13) — Cisco; TelecomTV; Futurum Group; Data Centre Magazine. HPE networking outlook and Vultr order (September 30) — Reuters via Investing.com; Bloomberg; Constellation Research; CRN; Fortune. AI networking market forecast (November 2) — TrendForce via iConnect007, CNA and 金融界.
Date distinction: The Marvell–Lumentum deal is a November 2 announcement. The Arista, Cisco–Nvidia and HPE items are background developments from October and September, included to explain the competitive context. They are not presented as November 2 breaking news.
Evidence labeling: The Marvell–Lumentum terms and HPE guidance are company-confirmed. TrendForce's forecast is an analyst projection. The USFP specification is planned, not yet in production. This briefing does not provide investment advice.
Reviewed by Erwin Castro
on
Monday, November 02, 2026
Rating:
No comments: