The Optical Direct-Connect Era: How Google OCS Is Rebuilding the "Highway" of AI Compute
From the OCP 2026 white paper to Lumentum's capacity lock-up race: unpacking the US$10B-scale new arena behind 13,824 optical modules and 48 OCS units
Simple Tech Trend | Flagship Premium Report | May 2026
📄 This is the preview edition. The full edition includes:
✅ Complete OCS system BOM breakdown (unit prices across seven cost items, incl. MEMS chips / FAU / alignment system / PCB / assembly fees)
✅ TPU v7 golden-ratio formula 1 : ? : ? : 1/? (the "codebook" for reading Google's order volumes)
✅ 2026-2028 quantified market sizing (TAM for four segments: DAC / optical modules / OCS / fiber)
✅ 800G / 1.6T optical module BOM comparison (EML vs SiPh; cost of laser / modulator / DSP)
✅ Chapter 6 Stock Playbook: Lumentum, Coherent, EZconn (TOP PICK), LandMark Optoelectronics, WIN Semiconductors, Browave, TFC Communication, Luxnet, FOCI, Unimicron, VPEC, GCS Holdings — full view cards on structure / catalysts / risks / tracking signals
✅ Chapter 7: Three 2026-2028 timelines (short / mid / long-term investment rhythm)
🔒 How to get the full edition 👉 Fill out the purchase form Within 24 hours of payment, the full PDF edition (in Chinese) will be sent to your inbox.
⚠️ Reader Notice & Copyright
[Not Investment Advice] The content of this report represents only the author's personal industry observations and technical analysis, intended for research reference and inspiration. It does not constitute investment, trading or legal advice of any kind. Investors should make independent judgments, evaluate carefully and bear their own investment risks.
[All Rights Reserved — No Reproduction] This is paid, exclusive in-depth research built on extensive industry surveys, third-party earnings calls, cross-checking against academic papers and supply-chain teardown work. Without the author's formal written authorization, please do not reproduce, excerpt, screenshot or publicly distribute all or any part of this content in any form.
[Support the Author & Follow the Trends] If you enjoy this kind of system-level industry analysis, visit Simple Tech Trend or follow simple_tech_trend on IG / Threads / X / FB for the latest tech insights.
Executive Takeaway
In April 2026, the Open Compute Project (OCP) released an OCS white paper co-authored by five organizations: iPronics, Lumentum, Ciena, Carnegie Mellon and Lumotive. The same month, Google TPU v7 Ironwood entered large-scale commercial shipments, and Anthropic signed a historic contract for 1 million TPUs worth more than US$52 billion in total. A month later, Lumentum's CEO revealed at the 54th JPM Global Technology Conference that the InP supply-demand gap has exceeded 30% and that the cumulative hyperscaler capex backlog has reached US$2 trillion — the optical communications industry has officially said goodbye to cyclicality.
Put these events together and 2026 is year one of the "optical direct-connect era" — optical circuit switching (OCS) is shifting from a Google in-house patent to the next-generation backbone of the entire data center industry. The core of this report:
1 — Optical communications has officially entered the "capacity lock-up war" era. The boom-bust cycle once driven by telecom carriers is dead. The hyperscalers' US$2 trillion capex backlog locks in the demand curve for the next three years; NVIDIA has already prepaid to lock up most of Lumentum's InP capacity, and other hyperscalers are lining up to sign prepayment contracts. This is the next severely supply-constrained battlefield after HBM in the 2020s.
2 — OCS is no longer Google's exclusive weapon. The April 2026 OCP white paper, backed by iPronics, Lumentum, Ciena, Coherent, Lumotive, Microsoft, NVIDIA, nEye, POLATIS and OrioleNetworks, has elevated OCS to the level of industry consensus. The formal launch of an OCP sub-project = the whole data center industry is getting ready to put OCS in its toolbox.
3 — The TPU v7 golden ratio 1 : ■ : ■ : 1/■■■ has been validated. Every 1 TPU maps to ■■■ DAC copper cables and ■■■ 1.6T optical modules, and every ■■■ TPUs "consume" one 300-port OCS. One 144-rack Ironwood Pod = 9,216 TPUs + ■■■ DACs + ■■■ 1.6T optical modules + ■■■ OCS units. 🔒 The complete golden-ratio formula and calculation logic are in Chapter 3.3 of the paid edition.
4 — Silicon photonics OCS will replace MEMS in 2028-2030. MEMS switches in milliseconds, silicon photonics MZI in nanoseconds — once AI training needs to "dynamically reconfigure network topology between iterations" (the academic papers ACTINA and MixNet already propose solutions), milliseconds are not fast enough. MEMS is the present; silicon photonics is where the future will be decided.
Verdict 5 — Taiwan's real foothold is three things: CW lasers + FAU + InP, not DSP. Taiwanese vendors' scarcity in upstream components, stacked on the non-China geopolitical tailwind, makes LandMark Optoelectronics, WIN Semiconductors, Browave, FOCI and EZconn the five most direct beneficiaries of this wave. 🔒 Entry catalysts, risks and tracking signals for these five stocks are in the paid edition's Chapter 6 Stock Playbook.
Three Must-See Numbers
Number | What It Means |
■■ OCS units / 9,216 TPUs | The optical backbone scale of one Pod — OCS usage ratio formula 1/■■■ 🔒 |
$■■k BOM vs $■■■-■■■k selling price | OCS system gross margin exceeds ■■%, which explains why Google keeps ■■,■■■ units in-house 🔒 |
$2T / 30% / 8× | Hyperscaler capex backlog / InP supply-demand gap / Lumentum InP capacity already 8× its 2023 level |
📖 Full Table of Contents
Chapter | Content | Preview Status |
Executive Takeaway | Five verdicts + three must-see numbers | ✅ Preview |
Chapter 1 | From EPS to OCS — why is it different this time? | ✅ Preview |
1.1 Hitting the Bottleneck | When network power starts eating the data center power budget | ✅ |
1.2 OCP 2026 White Paper | OCS goes from Google's exclusive weapon to industry consensus | ✅ |
1.3 Five Flaws of Clos | Why traditional electronic switching is reaching its end | ✅ |
Chapter 2 | OCS teardown — a "mirror maze" for light | ✅ Preview |
2.1 Glass Window vs Processor | Think of OCS as a "glass window" | ✅ |
2.2 Six OCS Technology Routes | MEMS / LC / Robotic / Piezo / SiPh / Metasurface | ✅ |
2.3 Five OCS Battlegrounds | Spine replacement, scale-up, scale-out, redundancy, slicing | ✅ |
2.4 ACTINA + MixNet | The academic frontier reveals the next-gen answer | ✅ |
Chapter 3 | Google's calculus — a 144-rack supercluster | Partial preview |
3.1 TPU v7 Ironwood Cube | The 4×4×4 cube design philosophy | ✅ Preview |
3.2 Cube Position Sets Module Count | Engineering memory formula | 🔒 Full edition |
3.3 ★ Golden Ratio 1:■:■:1/■■■ | The codebook for reading Google's order volumes | 🔒 Full edition |
3.4 Anthropic's 1M-TPU Deal | Igniting the 2027-2028 demand curve | Partial preview |
3.5 TPU Shipment Forecasts | Cross-checking multiple institutions' numbers | 🔒 Full edition |
3.6 2026-2028 Market Size | DAC / optical module / OCS / fiber TAM | 🔒 Full edition |
Chapter 4 | BOM teardown — who is making money from optics? | 🔒 Full edition |
4.1 OCS System $■■k BOM vs $■■■k Price | Breaking down the 80% gross margin | 🔒 |
4.2 Companion Modules Cost ■× the OCS | The bigger-ticket hidden battlefield | 🔒 |
4.3 WDM + Circulator | The strategic meaning of hardware decoupling | ✅ Preview |
4.4 Coherent-Lite as the "Perfect Teammate" | The key partner in the 1.6T era | ✅ Preview |
4.5 800G / 1.6T BOM Comparison | EML vs SiPh cost breakdown | 🔒 |
Chapter 5 | The OCS ecosystem — who is reshaping this "web of light"? | Partial preview |
Chapter 6 | Stock Playbook (2026-2028) | 🔒 Full edition |
6.1 Lumentum (LITE) | Halfway to the North Star | 🔒 |
6.2 Coherent (COHR) | A shovel seller with two weapons | 🔒 |
6.3 ★ EZconn (6442) | TOP PICK, the purest Google-chain proxy | 🔒 |
6.4 LandMark Optoelectronics (3081) | The scarcity of non-China InP | 🔒 |
6.5 WIN Semiconductors (3105) | The hidden beneficiary of Lumentum's CW outsourcing | 🔒 |
6.6 Other Watchlist Names | Browave, TFC, Luxnet, FOCI, Unimicron, VPEC, GCS Holdings | 🔒 |
6.7 Tracking Indicators | Four things to watch next quarter | 🔒 |
Chapter 7 | Three 2026-2028 timelines | 🔒 Full edition |
Conclusion | When optics becomes the highway of new compute | ✅ Preview |
🔓 Preview Content Begins
Chapter 1: From EPS to OCS — Why Is It Different This Time?
1.1 Hitting the Bottleneck: When Network Power Starts Eating the Data Center Power Budget
In an era of explosive AI compute growth, we have run into two huge bottlenecks: the power wall and the IO wall.
Today's data center architecture faces a severe "scale tax." As switches are stacked layer upon layer, cost and energy consumption grow geometrically. The OCP white paper gives hard numbers: if a large data center uses 64 16-slot chassis spine switches at 30 kW each, the spine layer alone draws 1.9 MW at full load. At the projected 100 Pbps scale, optical network equipment power would reach 48.7 MW — beyond the provisioning limits of most power facilities.
The more practical bottleneck is energy efficiency. How much energy does it take to move 1 bit? This metric is pJ/b (picojoules per bit).
Traditional pluggable optics + electronic switch architecture: a typical 2-tier fat-tree topology consumes about 83 pJ/b
CPO at the endpoints + OCS in the network: compressed to 30-31 pJ/b
That is nearly a 3× gap. This is not just about saving power — it is the only way to let compute clusters scale from 1,000 chips to 9,216 chips and beyond ten thousand. Without this "slimming revolution," under a finite power budget, hyperscalers' GPU/TPU additions will never outpace their power additions.
1.2 The OCP 2026 White Paper: OCS Goes from Google's Exclusive Weapon to Industry Consensus
Over the past three years, the most-cited yet most-misunderstood document in optical communications is Google's "Jupiter Evolving" paper presented at SIGCOMM 2022. That paper revealed one thing: Google had already replaced the electronic packet switches (EPS) in its data center spine layer with OCS, cutting network power by 41% and CAPEX by 30% while delivering a 5× throughput improvement.
At the time, it looked like "Google using yet another in-house technology nobody else can replicate." But in April 2026, this OCP white paper changed everything — its authors come from five players with very different positions: iPronics, Lumentum, Ciena, Carnegie Mellon and Lumotive.
The most important signal in this white paper is not its technical content, but the fact that it exists at all:
Founding members of the OCP-OCS sub-project (launched July 2025)
Role | Company |
Co-leads | iPronics, Lumentum |
Founding members | Coherent, Google, Lumotive, Microsoft, nEye, NVIDIA, OrioleNetworks, POLATIS (HUBER+SUHNER) |
Putting these 10 companies on the same page amounts to a declaration: Microsoft, Meta, Amazon and NVIDIA are all evaluating or piloting OCS architectures internally. OCS has been upgraded from "Google's exclusive weapon" to "the next-generation data center backbone formally endorsed by the OCP community."
China is moving in parallel:
On 2025/10/12, China Mobile Cloud and the Institute of Computing Technology, Chinese Academy of Sciences released the "Cloud Intelligent Computing Optical Interconnect Development Report," explicitly planning, in every network plane, to use OCS to replace the existing Super Spine
On 2025/10/14, China's MIIT launched the metro-area "Millisecond Computing Access" special action, promoting OCS deployment in computing centers
The whole industry — including Taiwan's supply chain — should start putting OCS in its toolbox instead of treating it as a lab curiosity. From wait-and-see to readying capacity, Taiwanese vendors probably have only an 18-24 month window left.
1.3 Five Flaws of the Clos Architecture: Why Traditional Electronic Switching Is Reaching Its End
To understand why OCS is unavoidable, first look clearly at how many bottlenecks electronic packet switching (EPS) has hit in the AI training era:
1. The "scale tax" of hierarchical expansion — To connect more servers you must stack switches layer by layer (Leaf → Spine → Core); every extra layer makes cost and power soar geometrically.
2. Moore's Law is slowing — Network bottlenecks used to be solved by chip upgrades. Now CMOS scaling has hit its limits: Tomahawk 6 (102.4T) already relies on TSMC 3nm, and the next generation is two to three years out.
3. The "downclocking and idling" of incremental upgrades (Stranded Capacity) — A newly purchased 800G leaf switch is forced to run at reduced speed as long as the upper spine is still 400G. This is why Jensen Huang keeps stressing co-design at NVIDIA — chips, NICs and switches must advance together; upgrading a single point does not help.
4. Over-provisioning for the worst case — To handle 1% of extreme traffic, traditional Clos wastes cabling 99% of the time.
5. Frequent O-E-O conversion — Every time data passes through an EPS node it must be translated "optical → electrical → electrical → optical." OEO already eats more than 30% of a switch's power.
OCS's solution is brute force: don't convert to electrical at all. Light enters an input port and is steered directly to an output port; the entire light path stays in the optical domain.
Chapter 2: OCS Teardown — A "Mirror Maze" for Light
2.1 Think of OCS as a "Glass Window," Not a "Processor"
The OCP white paper has a particularly good analogy: OCS should be understood as "a transparent pane of glass" — it simply guides light through, without reading it, without processing it, without even knowing what the light carries.
This is a completely different species from EPS. EPS is like a post office: every packet that arrives must be opened, its address read, a forwarding decision made, and then resealed. OCS is like a railway switch: flip the points and the whole train changes direction — what cargo is in the cars has nothing to do with it.
This species difference brings four direct consequences:
✅ Benefit 1: Protocol-agnostic + rate-agnostic — Because it doesn't read content, the same OCS can carry 100G today, 400G tomorrow and 1.6T the day after, with no hardware replacement at all.
✅ Benefit 2: Ultra-low power, ultra-low latency — No OEO conversion and no electronic processing mean extremely low power, extremely low latency and near-zero jitter.
⚠️ Cost 1: Circuit switching — Not suited to highly dynamic, bursty, unpredictable traffic.
⚠️ Cost 2: Blind to content — QoS, traffic inspection and telemetry are all impossible.
2.2 Six OCS Technology Routes
The OCP white paper groups the physical technologies currently able to implement OCS into six categories:
Technology Route | Port Count (Radix) | Insertion Loss | Switching Time | Best Fit | Representative Vendors |
3D MEMS | Medium (300+) | Medium (1.5-2.7 dB) | Milliseconds | Mainstream in volume production | Lumentum R300, Google Palomar |
Liquid Crystal | Medium (64-512) | Medium (2.3 dB) | Microseconds-milliseconds | More stable switching, no mechanical failure | Coherent DLX |
Robotic (mechanical) | Large | Very low (0.5 dB) | Seconds-minutes | Slowly changing, ultra-large scale | Telescent |
Piezo (piezoelectric) | Medium | Medium (1.5-2.5 dB) | 10-75 ms | Precise alignment, ultra-low power | Polatis |
Silicon Photonics MZI | Low-medium | High (3.7 dB) | Nanoseconds-microseconds | Fastest switching, manufacturable at scale | iPronics, nEye.AI |
Metasurface | Large | Medium | Microseconds | Compact, polarization-independent | Lumotive (dark horse) |
Three key industry judgments:
Judgment 1: MEMS is the present. Google's TPU v4/v7 Superpods use MEMS-based OCS, and it will remain mainstream for the next 3-5 years.
Judgment 2: Silicon photonics OCS is where the future will be decided. Its switching time is nanoseconds to microseconds — 1,000× faster than MEMS. When AI training needs to "dynamically reconfigure network topology between iterations," millisecond-class MEMS is too slow; only silicon photonics can keep up.
Judgment 3: Metasurface is the dark horse. Lumotive uses ±80° wide-angle beam steering and wavelength-scale sub-pixels to build a compact, high-density, polarization-independent free-space OCS.
2.3 Five OCS Battlegrounds in the AI Data Center
# | Application | What It Replaces | Real-World Players |
1 | Spine layer replacement | Spine EPS in a Clos architecture | Google Jupiter |
2 | Scale-Up (Intra-POD) | NVSwitch / TPU ICI | TPU v4 / v7 Ironwood |
3 | Scale-Out (Inter-POD) | EPS cross-POD connections | Microsoft ProjecToR |
4 | Redundancy resource pooling | Idle spare racks | Huawei RDC, Google ICI failure rerouting |
5 | Physical data center slicing | Multi-tenant logical isolation | Google TPU slicing |
2.4 Academic Frontier: ACTINA + MixNet Reveal the Next-Generation Answer
Two academic papers published in 2025 point to the same thing: integrating OCS directly into GPU transceivers.
Paper 1: ACTINA (Columbia + NVIDIA, SC 2025) — The first to bring OCS down to the "GPU port level." Each GPU integrates silicon photonics transceivers directly, dynamically reallocating optical-path bandwidth within a single training iteration. Measured iteration time is 1.84× faster and energy consumption 1.72× lower.
Paper 2: MixNet (HKUST, SIGCOMM 2025) — Designed specifically for MoE training. On the Mistral 8×7B MoE model, training was 1.6× faster than on a static fat-tree.
STT View: MEMS is the money-making tool for 2026-2027, but don't misjudge the timeline — after 2028, only silicon photonics OCS can truly keep up with the dynamic-topology demands of AI training.
Chapter 3: Google's Calculus — A 144-Rack Supercluster
3.1 TPU v7 Ironwood: The Cube Design Philosophy
In Google's design, the individual server has faded into the background; the basic unit of compute is now the rack. The soul of TPU v7 Ironwood is a cube built from 64 TPUs arranged as a 4×4×4 3D torus. This is not just a cool name — it is a precise calculation of space efficiency and choice of medium.
Physical structure:
Each Ironwood rack consists of 16 TPU trays
Each tray packages 4 TPUs (dual-chip package architecture)
16 × 4 = 64 TPUs per rack
Each TPU has 4 OSFP cages (for ICI interconnect) + 1 CDFP PCIe cage
Connection logic: each TPU connects to adjacent nodes in all three dimensions X, Y and Z, so every chip has 6 neighbors.
Media split:
Inside the rack (within the cube): mostly PCB traces + copper cables (DAC)
Between racks (outside the cube): optical modules + OCS are required
3.2 Cube Position Determines Optical Module Count: An Engineering Memory Formula 🔒
This is the most elegant part of Google's Ironwood architecture. How many optical modules a TPU needs depends entirely on its position in the 4×4×4 cube. Exclusive teardown in the full edition: ✅ How many DACs + optical modules each of the four positions — corner / edge / surface / interior — needs ✅ Why 96 optical modules per 64-TPU rack is a physical necessity ✅ STT's exclusive "engineering memory formula" — one sentence to calculate how many optical links any TPU in the cube needs ✅ Full mathematical verification: 8×?+24×?+24×?+8×?=160 interfaces = 80 OSFP copper cables 🔒 [Buy the full edition to view]
3.3 ★ STT Golden Ratio Formula: 1 : ■ : ■ : 1/■■■ 🔒
Pulling together the whole rack's numbers, STT proposes an exclusive "golden ratio formula" — the "codebook" for reading Google's order volumes:
Component | Ratio (vs TPU) |
TPU chip | 1 |
OSFP copper cable (DAC) | ■ 🔒 |
Optical module | ■ 🔒 |
OCS switch | 1 : ■■■ 🔒 |
Pod-level verification (9,216 TPUs = 144 racks)
Optical module demand: ■■■■■ units (all 1.6T optical modules)
DAC demand: ■■■■■ cables
OCS demand: ■■ units (Lumentum R300 300×300 spec)
Investor shorthand example: Google announces a data center with 100,000 TPUs —
Optical module demand: ■■■,■■■ units
DAC demand: ■■■,■■■ cables
OCS demand: about ■■■ units
🔒 [Buy the full edition to view the complete formula and calculations]
3.4 Anthropic's 1-Million-TPU Deal: Igniting the 2027-2028 Demand Curve
In November 2025, Google announced a historic contract with Anthropic for 1 million TPUs, rolled out in two phases:
Phase | Scale | Model | Value |
Phase 1 | 400,000 TPU v7 Ironwood | Direct supply (by Broadcom) | ~US$10 billion (rack-level) |
Phase 2 | 600,000 TPUs | Leased via GCP | ~US$42 billion RPO |
Converting this contract with the STT golden ratio formula gives the specific demand for OCS / 1.6T optical modules / OSFP DAC copper cables, and its share of Google's 2026 internal demand — 🔒 [Buy the full edition to view the detailed calculation]
Google's New "Credit Guarantee" Model for Breaking the Power Bottleneck
Market research reveals an important industry innovation: Google did not sign a long-term lease directly with Fluidstack; instead it stepped in via a "credit guarantee" — if Fluidstack cannot pay its data center rent, Google steps in to settle it as the holder of an "off-balance-sheet IOU." This model solves the neocloud industry's most painful "maturity mismatch" problem and has become the de facto new financing template in this space.
3.5 TPU Shipment Forecasts: Cross-Checking Multiple Institutions 🔒
Because Google does not disclose TPU shipments, the market has to estimate them through supply-chain surveys. The three major research houses do not fully agree. This report compiles: ✅ Asian investment bank (bullish camp) 2026/2027/2028 shipment estimates ✅ US investment bank (mainstream upward-revision camp) 2026/2027/2028 shipment estimates ✅ Local investment advisory (neutral camp) 2026/2027/2028 shipment estimates ✅ STT's blended baseline estimate (accounting for InP supply bottlenecks and power constraints) 🔒 [Buy the full edition to view the complete comparison table]
3.6 2026-2028 Market Size 🔒
Using the STT golden ratio formula + blended TPU shipment estimates, here is the market size of each supply-chain segment over 2026-2028:
Year | TPU Shipments | DAC | Optical Modules | OCS | Fiber | Total Value |
2026 | ■■■ ×10k units | $■■■M | $■.■■B | $■.■■B | $■.■■B | $■.■■B 🔒 |
2027 | ■■■ ×10k units | $■■■M | $■.■■B | $■.■■B | $■.■■B | $■.■■B 🔒 |
2028 | ■■■ ×10k units | $■■■M | $■.■■B | $■.■■B | $■.■■B | $■■.■■B 🔒 |
What this table means: a single TPU v7 SKU will generate more than $20B in supply-chain value within three years — larger than the entire 800G optical module market in 2024.
🔒 [Buy the full edition to view the complete quantitative estimates]
Chapter 4: BOM Teardown — Who Is Making Money from Optics? 🔒
4.1 The OCS System: $■■k BOM vs $■■■-■■■k Selling Price 🔒
If OCS is a precision "mirror maze," what makes this maze truly valuable is a gross margin above ■■%. The full edition's exclusive OCS system BOM teardown:
Component | Unit Price | Qty/Unit | Subtotal | Supplier |
MEMS micromirror array chip | $■,■■■ | ■ sets | $■,■■■ | 賽微電子 (exclusive prototyping at Silex, Sweden) |
2D FAU fiber array | $■■■ | ■ pcs | $■,■■■ | Corning, TFC Communication |
Alignment system (complete) | $■,■■■ | ■ sets | $■,■■■ | Lumentum/Coherent |
Filter/isolator/lens kit | $■■■ | ■ sets | $■■■ | 騰景科技 |
PCB (38-layer high-speed) | $■,■■■ | ■ boards | $■,■■■ | Unimicron |
Chassis/power/control ASIC | – | – | $■,■■■ | System integrators |
System assembly fee | – | – | $■,■■■ | Celestica, Lumentum, Coherent |
BOM total | ~$■■,■■■ | |||
System price (300×300 port) | $■■■,■■■-■■■,■■■ | |||
Gross margin (est.) | ~■■% |
🔒 [Buy the full edition to view the complete BOM]
Why Does Google Build 12,000 Units In-House and Buy 3,000? 🔒
Google's 2026 OCS demand is about 15,000 units of the 300-port spec, but its procurement strategy is shrewd: the 4:1 in-house/external split is itself evidence of price divergence. The gap between Lumentum/Coherent's external selling price and Google's internal Celestica manufacturing cost lets Google save $■■-■■k per unit × 12,000 units = a direct saving of $■-■■ ×100M. 🔒 [Buy the full edition to view the complete price-divergence analysis]
4.2 Companion Optical Modules: ■× the Cost of the OCS Itself 🔒
Every OCS port needs an 800G/1.6T optical module. Scaled to the Pod level: one Pod's 48 OCS units × $■■■,■■■ = $■.■■M in OCS systems 🔒 13,824 optical modules × $■■■ = $■.■■M in optical modules 🔒 Total companion optical module cost is ■.■× that of the OCS systems What does this mean? Optical communications investing cannot focus on OCS alone — look back at optical modules, which are the bigger-ticket piece. 🔒 [Buy the full edition to view the complete reverse calculation]
4.3 WDM + Circulator: The "Technology Premium" of OCS Modules
In Chapter 3 we said one optical module maps to exactly one OCS optical port. But insiders know this is not physically natural — optical transmission needs at least two channels, one transmit and one receive. Google achieves this "one-to-one" miracle thanks to two high-value components: WDM (wavelength division multiplexing) and the circulator (optical circulator).
Component | Role |
WDM (wavelength division multiplexing) | One fiber carries 4-8 wavelengths = the same fiber jumps from 200G → 1.6T instantly |
Circulator (optical circulator) | Forces light into one-way circulation so Tx + Rx pass each other cleanly in the same main fiber |
Strategic meaning: breaking the curse of "incremental upgrades" — "hardware decoupling"
Scenario | Traditional Architecture | OCS + WDM + Circulator |
800G → 1.6T upgrade | Spine switches must be replaced | Only the optical modules at both ends change; the OCS stays in place |
1.6T → 3.2T upgrade | The whole set of network equipment must be replaced | OCS is rate-transparent; only modules change |
This is why OCS is a "sell once, use for years" business. The OCS itself can serve 1-2 generations of compute — but companion optical modules must upgrade with each speed step, so the optical module market keeps getting more investment.
4.4 Toward the 1.6T Era: How Coherent-Lite Becomes OCS's "Perfect Teammate"
As AI compute demand rushes from 800G to 1.6T and even 3.2T, traditional IM-DD (intensity modulation / direct detection) is running into severe physical bottlenecks. Coherent-Lite (lightweight coherent technology) has emerged in response.
1. Breaking the "dispersion barrier" and improving the "link budget" — Coherent-Lite provides an extra 10 dB of sensitivity advantage, letting signals easily push through the insertion loss of OCS MEMS micromirrors. OCS + Coherent-Lite are a natural pair.
2. Semiconductor economics: subtract in optics, add in silicon — Shift the burden from "optical components that are hard to scale" to "DSP chips that follow Moore's Law." As processes move to 3nm, the higher the data rate, the more dominant Coherent-Lite's cost-performance becomes.
3. Proven in practice: the OFC 2026 Marvell + Lumentum demo — Marvell Aquila 1.6T Coherent-Lite DSP + Lumentum R300 OCS cut switching latency by up to 98% and saved 65% of energy.
Coherent-Lite does not replace OCS — it is the "optical link amplifier" that lets OCS hold up into the 1.6T era. Marvell, Nokia, Ciena and Cisco (Acacia) are all on this track.
4.5 800G / 1.6T BOM Comparison: EML vs SiPh 🔒
The full edition's exclusive teardown of the complete 800G and 1.6T optical module BOM, item by item for the EML vs SiPh approach: Laser (unit prices by EML spec / by CW laser spec) Modulator PD Driver + TIA DSP — revealing the real price range under the 200G EML duopoly and how large SiPh's cost advantage is in the 1.6T era. 🔒 [Buy the full edition to view the complete BOM comparison table]
Chapter 5: The OCS Ecosystem — Who Is Reshaping This "Web of Light"?
The rise of OCS is not a solo act but a full mobilization from "system architecture" down to "chip foundry."
5.1 System Definers: Google + NVIDIA
Google: the high priest of all-optical. From in-house OCS (Palomar) → defining the 3D torus topology → in-house manufacturing via Celestica → tightly binding the network to the TPU. 2026 OCS demand ~15,000 units.
NVIDIA: the pragmatic follower. Spectrum-X is still mainly electronic switching, but NVIDIA is actively researching OCS integration in the OCP white paper and the ACTINA paper. Once NVIDIA puts OCS into its reference architecture, the TAM explodes.
5.2 The OCS Duo: Lumentum + Coherent
Lumentum (LITE): The MEMS Leader
Core products: R300 (300×300 ports), R64 (64×64 ports)
InP capacity already 8× its 2023 level; fifth InP fab in production
NVIDIA has prepaid to lock up most InP capacity + options
In talks with other hyperscalers on long-term "prepayment + price guarantee" contracts = a replay of the HBM lock-up playbook
Coherent (COHR): The New Liquid-Crystal OCS Challenger
Bets not on MEMS but on digital liquid crystal (DLC)
Product line: DLX in 64×64, 320×320 and 512×512 configurations
Already shipping, with 7 customers in trials
5.3 The Coherent-Lite Alliance: Marvell + Nokia + Ciena + Cisco
Company | Role |
Marvell | Supplies the core component: the Aquila 1.6T Coherent-Lite DSP |
Nokia | Long-haul coherent optics + DCI integration |
Ciena | DCI leader; acquired Nubis Communications in Oct 2025 |
Cisco (Acacia) | Deepest coherent-optics technology bench; FY26 AI infra order guidance of US$9 billion |
5.4 Three Things Where Taiwan Can Win: CW Laser + FAU + InP
Taiwan's real foothold is not DSP but CW lasers, FAU and InP upstream materials.
Segment | Controlled By | Opportunity for Taiwan |
DSP (the optical module's brain) | Broadcom, Marvell | Minimal |
Switch ASIC | Broadcom, NVIDIA, TSMC | Mainly a foundry role |
CW Laser/EML | Lumentum, Coherent | LandMark, WIN Semiconductors, Luxnet ✓ |
FAU/fiber | Corning, TFC | Browave, FOCI ✓ |
InP upstream | Sumitomo, AXT | LandMark long-term supply lock-in ✓ |
High-density fiber cable | (emerging) | EZconn exclusive to Google ✓ |
🔒 The full edition includes each Taiwanese vendor's detailed foothold, capacity, customer ties, catalysts and risk list.
🔒 Chapter 6: Stock Playbook (2026-2028) — Full Edition Exclusive
This chapter is paid-edition exclusive content. It does not issue buy/sell ratings; it only provides structured view cards on positioning / catalysts / risks / tracking signals. Stocks covered (13 in total) US stocks (2) Lumentum (LITE) — Halfway to the North Star: 1.6T fab fully loaded, OCS H2 backlog, scale-up + CPO boom in 2027 Coherent (COHR) — A shovel seller with two weapons: IDM structure + DLX liquid-crystal OCS route Taiwan stocks (11) EZconn (6442) ★ TOP PICK — The purest Google-chain proxy (Google is 90% of revenue; Torpedo cable + patch panel) LandMark Optoelectronics (3081) — The scarcity of non-China InP; 5-year long-term contract with Sumitomo Electric WIN Semiconductors (3105) — The hidden beneficiary of Lumentum's CW outsourcing Browave (3163) — Shuttle Box, MPO connectors TFC Communication (TFC) — 40-50% share of Google FAU Luxnet (4979) — Core of the Marvell supply chain FOCI (3363) — FAU and fiber jumpers Unimicron (3037) — OCS PCB (38-layer) + sole supplier of NVIDIA CPO substrate VPEC (2455) — PD epitaxy foundry GCS Holdings (4991) — PD components TrueLight (3234) — eBeam process backup What each stock view card contains
Dimension | What's Disclosed |
Core positioning | 🔒 |
2026 structural catalysts | 🔒 |
Capacity expansion figures | 🔒 |
Key risks (3) | 🔒 |
Tracking signals (what to watch next quarter) | 🔒 |
Valuation reference | 🔒 |
🔒 [Buy the full edition to view complete view cards for all 13 stocks]
🔒 Chapter 7: Three 2026-2028 Timelines — Full Edition Exclusive
Short term (2H 2026): Lumentum + 1.6T twin engines 🔒 Full shipment rhythm + catalyst timeline Mid term (2027): CPO + scale-up boom 🔒 TSMC COUPE mass-production timeline + Anthropic Phase 2 deployment rhythm Long term (2028-2030): the generational shift from MEMS → silicon photonics OCS 🔒 The key years for Taiwan to secure "silicon photonics PIC foundry" positions The entire optical communications industry will go through a generational shift in 2028-2030. When technology nodes tighten all at once, the winner structure gets redrawn. 🔒 [Buy the full edition to view the complete timelines]
Conclusion: When Optics Becomes the Highway of New Compute
After the whole report, three sentences to close:
1. Optical communications is no longer a "cyclical bet" — it is a "capacity-expansion bet." The boom-bust cycle once driven by telecom carriers is dead; the hyperscalers' US$2 trillion capex backlog locks in the next three years of demand. Whether you can keep up depends entirely on execution — capacity, yield, manufacturing and supply-chain management.
2. OCS is not just Google's story; it is the next-generation backbone of the entire data center industry. From the OCP white paper to China MIIT's "Millisecond Computing Access" special action, from Microsoft ProjecToR to NVIDIA Beyond CLOS, from Lumentum R300 to Coherent DLX, from ACTINA to MixNet — this trajectory is now irreversible.
3. Taiwan's real foothold is the double scarcity of "upstream components + non-China identity." Not DSP, not switch ASICs, but CW lasers, FAU and InP. EZconn, LandMark Optoelectronics, WIN Semiconductors, Browave, TFC — the story for these five is not over yet.
One last line for readers:
When Google published Jupiter Evolving in 2022, most of the industry saw it as "Google using something no one else can build." Three years later, in 2026, the OCP used a white paper to tell us: everyone else is getting ready to build it too. For Taiwanese vendors, the window between "watching this technology" and "preparing this capacity" may be only 18 to 24 months.
Don't miss this window.
🔓 Want the Full Edition?
The full edition includes the following content not disclosed in the preview:
Full Edition Exclusive | Content |
🎯 STT golden ratio formula 1 : ? : ? : 1/??? | The codebook for reading Google's order volumes |
📐 Cube-position memory formula | One sentence to calculate how many optical links a TPU at any position in the 4×4×4 cube needs |
💰 OCS system $??k BOM | Seven cost items broken down one by one; the 80% gross margin revealed |
📊 2026-2028 quantified market size | Complete TAM estimates for DAC / optical modules / OCS / fiber |
🔬 800G / 1.6T BOM comparison | Item-by-item EML vs SiPh cost comparison |
🎯 Complete view cards for 13 stocks | Including the full analysis of TOP PICK EZconn |
⏰ Three 2026-2028 timelines | Complete short / mid / long-term investment rhythm |
Within 24 hours of payment, the full PDF edition will be sent to your inbox.
Comments