The 800G Transition

Where the world's networks are going when every hyperscaler demands four times the bandwidth per port — and why the supply chain is barely keeping up.

Standfirst: Eighteen months ago, "800G optical" was a roadmap item in an analyst slide deck. Now it is a physical constraint on AI cluster construction, a bottleneck around which entire companies have been funded, and the axis along which a new wave of networking competition is being fought. This is the story of the 800G transition — what is happening, who profits, what breaks first, and where the industry is heading toward 1.6T.

The Trigger That Shattered the Pluggable Era

AI training clusters built on NVIDIA Blackwell Ultra and later platforms require more than raw GPU compute. A single Rubin or Vera Rubin system can contain thousands of GPUs that must communicate at sub-microsecond latency across tens of thousands of links simultaneously [1]. The result: each switch port must push data faster than any commercially viable copper cable or pluggable transceiver has handled before.

Enterprises have always accepted obsolescence cycles. A network upgrade happens every seven to ten years when equipment reaches end-of-life. This cycle has been compressed dramatically. Broadcom's Tomahawk 6 — the world's first 102.4 Tbps Ethernet switch ASIC — moved from sampling to volume production in less than three quarters following its initial March 2026 ramp announcement [2]. In semiconductor time, this velocity borders on unprecedented. For a hyperscale buyer planning a multi-year datacenter program, it means the architectural decisions made today will be tested against reality in months, not years.

The infrastructure consequence is immediate and quantifiable. When Broadcom shipped the Tomahawk 6 silicon at scale, it enabled 800G per-port connectivity for the first generation of next-gen switch fabrics. The demand signal hit the optics supply chain almost overnight. Lumentum's Q4 FY2026 results confirmed the shock: revenue surged 109% year-over-year to $1.01 billion, with gross margins crossing 50% for the first time as AI optical transceivers became the primary volume driver [3]. This is not cyclical optimism. It represents a structural shift in how many optical engines an AI cluster requires — a factor of four to five times the count of a traditional cloud networking deployment.

Broadcom itself tells the most direct story. Its Q3 FY2026 semiconductor revenue reached $16.7 billion in a single quarter — a 221% year-over-year increase [4]. While the company bundles custom silicon (its core profit center) within this number, the networking component — Tomahawk switches, Ethernet PHY solutions, and the broad ecosystem surrounding them — contributes measurably. And Broadcom's full AI semiconductor trajectory projects toward $115 billion by fiscal 2027, implying that even a modest percentage share allocated to networking represents tens of billions of dollars in sustained demand for 800G-capable interconnect components [5].

This is the scale at which the 800G transition operates. We are no longer discussing incremental improvement to existing networking stacks. We are looking at a new substrate of infrastructure spending whose total value exceeds the entire enterprise-datacenter market from three years ago.

Why 800G Is Not Just "400G Times Four"

A common misconception is that moving from 400G to 800G is a matter of doubling lane count or speed. The physics do not cooperate.

At 400G densities, engineers used parallel lanes — eight lanes of 50 Gbaud PAM4 signaling — to achieve aggregate throughput. Doubling this to 800G meant either sixteen lanes (the OSFP form factor standard) or packing tighter electrical channels into QSFP-DD housings, both of which run into thermal constraints. The power consumption of an 800G transceiver module approaches 30 watts — roughly double that of a comparable 400G device. Within a single rack filled with hundreds of such modules, the cooling overhead begins to compete directly with the compute resources the network is supposed to support.

More fundamentally, electrical interconnect degradation scales non-linearly with frequency. Copper traces become unusable beyond approximately one meter at these speeds, forcing a transition strategy that involves active components, re-timed signals, and — eventually — photonic transport integrated much closer to the switching ASIC than ever before. This is where the industry's attention pivots from pluggable form factors to co-packaged optics, where the optical engine lives physically adjacent to (or even inside the same package as) the switch ASIC die.

Two critical enablers have appeared simultaneously that make the 800G transition possible right now rather than in some future decade. First, the widespread adoption of retimer chips — devices inserted between the transceiver module and the switch ASIC to regenerate degraded electrical signals over long PCB traces — has transformed reliability concerns. Companies like Credo have built their entire business models around this problem, reporting record Q1 FY2027 revenue of $479 million (+115% YoY) driven by deployment of their Screaming Eagle retimer at 100G-per-lane rates across AI deployments, with Blue Heron products pushing to 200G per lane [6]. Second, specialized PHY chip suppliers such as Astera Labs now face lead times of 16–24 weeks for early-2026 delivery, confirming capacity tightness across the entire signal chain from the switch ASIC through the transceiver to the fiber plant [7]. The supply shortage itself becomes a competitive moat: anyone already qualified into a hyperscaler's bill of materials enjoys pricing leverage that compounds with volume commitments.

Co-Packaged Optics: From Lab to Commercial Reality

If 800G is the present, co-packaged optics (CPO) is the near-term future — and the boundary between demonstration and production is collapsing with unusual speed in September 2026.

Cisco made what analysts described as a significant public statement recently: CPO is ready for mass deployment amid surging AI data-center demand [8]. The company noted that pluggable modules, near-package optics, and co-packaged optics will coexist, but emphasized that the architecture shift could reshape how hyperscale infrastructure is designed globally. This is notable coming from Cisco — a vendor whose installed base of proprietary pluggable-optics ecosystems gave it every incentive to delay CPO adoption.

Ciena took a parallel path. The company announced Vesta — described as the industry's first high-density, low-power, open-ecosystem pluggable CPO solution — with sample availability starting in Q2 FY2026 [9], [10]. The Vesta platform offers 206.4 Tbps switching density using pluggable CPO connectors that interface directly with Broadcom and Marvell switch ASICs. Ciena's own Q3 FY2026 revenue jumped 37% year-over-year to $1.67 billion, corroborating the demand environment for optical systems [11].

Broadcom — the incumbent switch ASIC leader — introduced its third-generation CPO technology operating at 200 Gbps per lane, a step-up from the 100 Gbps per lane used in its first-generation Tomahawk 4 Humboldt chipset launched in 2021 [12]. The third-gen implementation targets the 400G per lane regime that underpins 1.6T-scale architectures, effectively embedding the CPO capability directly into the switch ASIC package so that the optical engine sits millimeters rather than centimeters from the silicon core. Power savings approach 3.5× compared to traditional pluggable architectures, according to independent analyses [13]-How-It-Works,2026-Platforms,Benefits,and-Challenges.jsp).

NVIDIA's entry into the optical interconnect space through its NVLink Fusion program further accelerates adoption. Lightmatter, the photonic-interconnect startup founded in 2017 and valued at $4.4 billion, joined NVIDIA's NVLink Fusion ecosystem in June 2026, committing to supply co-packaged and near-package optics products designed specifically for NVIDIA's optical SerDes technology — reducing fiber and connector requirements by 50% compared to conventional electrical interconnect [14], [15]. At OFC 2026, Lightmatter showcased hardware spanning near-package optics, co-packaged optics, and photonic interposers, positioning silicon photonics as a foundational layer for next-generation AI systems [16].

The ecosystem consolidation visible in mid-2026 — multiple competing open-source alliances formed in parallel (Open CPX offering socketed optical engines, the Open Compute Initiative defining optical interconnect protocols, and the XPO specification for ultra-dense pluggable optics) — indicates that the industry recognizes fragmentation risk but also urgency [17]. No single consortium wins; instead, buyers evaluate interoperability while vendors race to qualify their designs first.

New Entrants Raise Hundreds of Millions

Perhaps the most revealing signal about the 800G transition's maturity comes from the venture-capital activity surrounding companies building 800G-capable optical infrastructure. Lumilens, a San Jose startup founded in 2024 by former Aruba Networks CTO Ankur Singla, emerged from stealth in August 2026 having raised more than $700 million in a Series C round that brought its total funding above $900 million at a $5.5 billion valuation [18], [19]. Critically, Lumilens reported that it had already begun shipping optical transceivers into the production data centers of a large cloud provider — meaning this valuation attaches to actual supply chain participation, not merely prototypes and partnership announcements [20].

Lumilens' technical strategy is instructive: it builds for both near-package / co-package optics on the scale-up side (connecting GPUs within a single chassis) and pluggable transceivers at 800G, 1.6T and beyond on the scale-out side (connecting racks and pods) [21]. Most competitors specialize in one architecture; Lumilens' dual approach suggests that hyperscalers expect both paradigms to coexist during the transition, each serving different topology requirements within a single cluster.

Applied Optoelectronics (AAOI), another relatively unknown manufacturer, has similarly demonstrated hyperscale qualification for 800G production volumes, commanding premium valuations as the sole verified supplier in certain capacity segments [22]. The market increasingly treats a hyperscale-qualified 800G transceiver line not as an optionality but as essential infrastructure — and prices accordingly.

These new entrants operate alongside established players whose earnings numbers tell the same story. Credo's continued triple-digit growth across seven consecutive quarters — $479 million in Q1 FY2027 alone — underscores how broadly-based the demand is: retimers, active ethernet cables (HiWire), optical DSP, and optical engines are all growing simultaneously because 800G systems require all of them, not just the lasers themselves [23].

1.6T on the Horizon — and the Architecture Wars Ahead

The 800G transition does not end with 800G. Every major switch ASIC vendor is already designing for the next speed class.

Broadcom's Tomahawk 7 — taped out in August 2026 as the industry's first 200 Tbps Ethernet switch — explicitly targets "scale-in" networking, moving the fabric architecture away from two-tier Clos topologies into flattened single-stage switching that reduces hop count and latency for large-scale AI training [24]. The Tomahawk 7 naturally supports 1.6T per port when paired with appropriate optics. Volume production timing remains uncertain given the complexity jump from 102.4 Tbps (Tomahawk 6) to 200 Tbps (Tomahawk 7); however, following the <3 qsample-to-production precedent set by Tomahawk 6, proto samples may arrive in 2027 with volume likely in 2028 [25].

Lightmatter's Passage CPO chiplet achieved a record 1.6 Tbps per fiber in March 2026, demonstrating photonic interconnect at speeds that exceed current electrical lane limits and making it directly applicable to the 1.6T optical transceiver timeline [26]. This is not a simulation or lab demonstration — the product was sampled, meaning hyperscale customers now hold reference designs capable of supporting 1.6T per-port speeds.

For the optical transceiver industry, the 1.6T roadmap introduces several compounding engineering challenges. Each 800G→1.6T generational step doubles the data rate per channel, requiring either higher-order modulation (PAM6 versus the current PAM4), wider wavelength spacing in WDM implementations, or more lanes per package. The power budget grows proportionally — approaching 60+ watts per transceiver at 1.6T — pushing pluggable-form-factor solutions toward their thermodynamic limit and accelerating migration to co-packaged implementations where heat dissipation is managed at the package level rather than through individual module fans.

Market projections for the 1.6T co-packaged optics segment itself suggest substantial commercial opportunity: estimates place the global 1.6T CPO switch market at $1.8 billion in 2025, expanding to $12.4 billion by 2034 at a 23.9% CAGR [27]. More conservatively, the broader data-center interconnect market — encompassing all interconnect technologies including electrical, optical, and hybrid — is projected to reach $38.2 billion by 2032 at a 16.8% CAGR, with AI training-cluster interconnects (800G/1.6T coherent optics) representing the single fastest-growing product category at 42% CAGR [28].

What Breaks First

In every major infrastructure transition, one constraint proves harder to solve than the rest. For the 800G transition, three candidates compete:

1. Optical-engine manufacturing capacity. Unlike electronic semiconductors, where fab capacity expands through capital expenditure over 12–18 months, optical assembly involves delicate manual processes — fiber coupling, hermetic sealing, laser tuning — that scale slower than purely automated semiconductor fabs. Lumentum's achievement of >50% gross margins for the first time signals that the industry is reaching the inflection point where volume scaling finally provides margin leverage, but this milestone came after years of constrained output that forced buyers to sign long-lead-time contracts [29]. Lead times beyond 40 weeks for 800G and 1.6T transceivers (versus 8–14 weeks for enterprise 100G parts) indicate the market has not fully cleared.

2. Retimer/PHY supply. Astera Labs' 16–24 week lead times and Credo's record revenue figures confirm that signal integrity components are as constraining as the transceivers themselves [30], [31]. A well-designed 800G system cannot function without re-timed electrical paths; the retimer is not optional. Multiple suppliers (Credo, Astera Labs) facing capacity constraints means buyers cannot simply diversify to an alternate source without risking redesign cycles.

3. Thermal management at the pod level. As optical transceivers consume 30+ watts each and hyperscale pods pack thousands of them, the cooling architecture transitions from a secondary concern to a first-class design variable. Vertiv's own FY2026 guidance — raised to $14 billion sales with a record $14.06 billion backlog — reflects the cascading demand for liquid cooling and advanced thermal solutions that 800G densification necessitates [32]. This connects back to the power constraint discussed in separate analysis: you cannot ship 800G optics if the datacenter cannot handle the BTU load they generate.

Forecast Scenarios

Based on the evidence surveyed above, here are three plausible pathways through 2026–2028:

Base Case: 800G Dominance Through 2027, 1.6T Entering Production

800G optical transceivers remain the dominant speed class through 2027, with ASPs stabilizing after initial spikes as supply chain capacity catches up. Tomahawk 6-based switch fabrics dominate new deployments. CPO solutions begin commercial pilot programs at 2–3 hyperscalers but do not achieve volume production until late 2027 or 2028. 1.6T optical transceivers enter proto-sample phase in 2027 aligned with Tomahawk 7 evaluation cycles. Total TAM for AI-relevant optical interconnect reaches approximately $30–35 billion annually by 2027. Retimer and PHY lead times compress to 12–16 weeks as capacity expands. Probability: 50–60%.

Bull Case: CPO Adoption Accelerates, 800G Supply Surges

Multiple hyperscalers commit to CPO-integrated switch fabrics in 2026–2027 based on the validation provided by Cisco and Ciena's public announcements. Lightmatter's NVLink Fusion integration enables rapid prototyping of photonic interconnect at NVIDIA GPU scale, creating a self-reinforcing cycle where GPU architects design around optical I/O rather than electrical. 800G transceiver supply capacity scales faster than expected as Lumentum and other manufacturers add shifts and new assembly lines. 1.6T samples arrive ahead of schedule, enabling earlier migration planning. Probability: 20–25%.

Bear Case: Supply Chain Friction Delays Deployments

Geopolitical disruptions (export controls on silicon photonics or laser components) constrain optical engine manufacturing capacity. Retimer shortages persist past 2026 due to wafer-allocation competition with memory and logic foundries. Hyperscalers delay new cluster construction pending resolution of the CPO vs. pluggable debate, leaving the industry in a holding pattern. Some buyers explore alternative architectures (electrical interposers, advanced copper) to defer the optical transition. Total optical spend growth moderates below consensus expectations. Probability: 15–20%.

Buyer Guidance

What should procurement and architecture teams responsible for datacenter infrastructure do now?

Immediate actions (Q3–Q4 2026):

1. Lock in 800G transceiver supply contracts now. With Lumentum, Credo, and Astera Labs all signaling supply constraints, securing committed volume through multi-year agreements at negotiated pricing is essential. Delaying risks paying spot-market premiums during periods of acute shortage, particularly given Lumentum's guidance of $1.225–$1.275B for Q1 FY2027 (midpoint ~$1.25B, up ~130% YoY) indicating demand continues accelerating rather than normalizing [33].

2. Evaluate CPO readiness with your switch vendor. If working with Broadcom, Marvell, or Cisco platforms, initiate conversations about CPO-enabled designs and understand the qualification timeline. Early engagement influences roadmap prioritization and gives buyers visibility into whether CPO-ready switch ASICs will align with their cluster-deployment cadence.

3. Design for retimer diversity. Do not architect systems that depend exclusively on Credo or Astera Labs retimers without qualifying an alternative. Dual-sourcing strategy mitigates single-supplier capacity risk, though as Astera Labs' IP licensing model demonstrates, some suppliers create barriers by combining chip sales with licensed SerDes intellectual property that locks customers into specific implementation paths [34].

Medium-term planning (2027–2028):

4. Budget for the 1.6T migration window. Systems deployed with 800G optics in 2026–2027 will face natural lifecycle pressure to migrate to 1.6T-compatible architectures by 2028–2029. Factor replacement or augmentation costs into capital plans beginning in 2027, when prototype availability creates procurement options.

5. Monitor the open-ecosystem alliances. The concurrent formation of Open CPX, OCI, and XPO specifications (March 2026) represents the industry's recognition that proprietary optics lock-in carries unacceptable risk at this scale. Buyers should track which alliance achieves interoperability certification first — that designation determines vendor options and price negotiation leverage going forward.

The additional source registry for this edition is maintained in the evidence ledger at data/platform-data.json; methodology and as-of conventions: methodology.

Key Metrics at a Glance

| Metric | Value | Source |

|--------|-------|--------|

| Tomahawk 6 volume production start | March 2026 (<3q from sampling) | Broadcom IR press release |

| Tomahawk 7 tape-out date | August 2026 | Q3 FY2026 earnings call |

| Credo Q1 FY2027 revenue | $479M (+115% YoY) | Yahoo Finance / Converge Digest |

| Lumentum Q4 FY2026 revenue | $1.01B (+109% YoY), GM >50% | BestStocks / Blockonomi |

| Astera Labs retimer lead time | 16–24 weeks (early 2026) | Intuition Labs supply chain analysis |

| Broadcom Q3 FY2026 AI semiconductor rev | $16.7B (+221% YoY) | LightHouse Canton |

| Broadcom AI revenue target | $115B by 2027 | WhatJobs / investor presentations |

| Ciena Q3 FY2026 revenue | $1.67B (+37% YoY) | Converge Digest |

| Lumilens total funding | >$900M at $5.5B valuation | Wall Street Journal / LiveMint |

| Lightmatter valuation | $4.4B (2024) | Various financial reports |

| 1.6T CPO switch market size (2025 est.) | $1.8B | DataIntel report |

| 1.6T CPO switch market (2034 proj.) | $12.4B (23.9% CAGR) | DataIntel report |

| AI DCI interconnect CAGR (800G/1.6T) | 42% | Iraqi Weekly / market data |

Sources

[1] NVIDIA Developer Blog: "Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era," Aug 2025.

[2] Broadcom Investor Relations: "Broadcom Now Shipping World's First 102.4 Tbps Switch in Production," Mar 12, 2026.

[3] BestStocks/Blockonomi: "Lumentum Stock Rockets 8% on Record-Breaking Q4 Results," Aug 28, 2026.

[4] LightHouse Canton: "Broadcom Q3 FY2026 Earnings: AI Revenue Guide to $230bn by 2027," Sep 3, 2026.

[5] WhatJobs aggregation of investor presentations and analyst estimates on Broadcom's $115B AI revenue projection, Sep 2026.

[6] Yahoo Finance/Credo Technology Group: Q1 FY2027 Earnings Report, Sep 1, 2026.

[7] Intuition Labs: "NVIDIA GB200 Supply Chain: The Global Ecosystem Explained," Apr 18, 2026.

[8] Digitimes: "Cisco: CPO is ready for mass deployment amid surging AI demand," Sep 1, 2026.

[9] MoatMap deep dive: "CIEN Ciena Corporation — Buy," citing Ciena Vesta 206.4T CPO sample announcement.

[10] Motley Fool: Ciena Q1 FY2026 Earnings Call Transcript, Mar 5, 2026.

[11] Converge Digest: "Ciena Q3 2026: AI & Cloud Drive Optical Networking Growth," Aug/Sep 2026.

[12] Finviz/Broadcom PR: "Broadcom Unveils Third-Gen CPO with 200G/lane Speed."

[13]-How-It-Works,2026-Platforms,Benefits,and-Challenges.jsp) IC Components: Comprehensive CPO guide covering 2026 platforms, benefits, and architectural challenges.

[14] DataCenter News: "Lightmatter Joins NVIDIA NVLink Fusion AI Ecosystem," Jun 2, 2026.

[15] Morningstar/Business Wire: Lightmatter-NVLink Fusion partnership details, Jun 2, 2026.

[16] Converge Digest: "Lightmatter Unveils Liquid-Cooled Laser NIC for AI Scale-Up Optics," May 21, 2026.

[17] LinkedIn/Wei Dong: "2026 may become the turning point" — analysis of Open CPX, OCI, XPO alliance announcements.

[18] RuntimeWire: "Lumilens raises more than $700 million to connect AI data centers."

[19] LiveMint: "Startup Raises $700 Million to Replace Data-Center Wires With Light," citing $5.5B valuation.

[20] Quasa: "Lumilens Raises $700M at a $5.5B Valuation," noting shipment to large cloud provider confirmed by WSJ.

[21] People Are Geek: Analysis of Lumilens dual-strategy (near-package/co-package for scale-up, pluggable 800G/1.6T for scale-out).

[22] 24/7 Wall St.: "This Little-Known AI Company Is Quietly Building for the Data Center Boom," Aug 31, 2026.

[23] Converge Digest: "Credo Revenue Jumps 115% as Optics Become Major Growth Driver," Sep 1, 2026.

[24] Yahoo Finance: Broadcom Inc (AVGO) Q3 FY2026 Earnings Call Transcript, Sep 2, 2026.

[25] Broadcom IR: Tomahawk 6 volume production milestone.

[26] Lightmatter PR: "Record 1.6 Tbps Per Fiber," Mar 11, 2026.

[27] DataIntel: 1.6T CPO Switch Market Research Report 2034.

[28] Iraqi Weekly: "$38.2 Billion by 2032" — datacenter interconnect market analysis, 42% CAGR for AI interconnects.

[29] BestStocks: Lumentum Q4 FY2026 results, gross margin crossover to >50%.

[30] Intuition Labs: Astera Labs retimer supply constraints and 16–24 week lead times.

[31] Yahoo Finance: Credo Q1 FY2027 record $479M revenue.

[32] Motley Fool: Vertiv earnings analysis — FY2026 guidance raised to $14B with $14.06B record backlog.

[33] BestStocks: Lumentum Q1 FY2027 guidance $1.225–$1.275B midpoint (~$1.25B).

[34] Intuition Labs: Astera Labs dual business model (IP licensing + retimer chip sales).

The additional source registry for this edition is maintained in the evidence ledger at data/platform-data.json; methodology and as-of conventions: methodology.