PDF 原檔：Optical Networking 光通 (大文章) _gs_original.pdf
原始內容
rolaman Researcr
EQUITY RESEARCH | April 17, 2026 | 2:11AM HKT
 _gs_001.png)
The next mega trend in AI infrastructure
Networking unlocks computing capability for single AI chips, connecting multiple chips (working together), enabling seamless data exchange and low latency, and driving AI to the next level. Along with an AI infrastructure ramp-up and rising computing power per rack, we expect all configurations to enjoy strong growth ahead, opening a further 9x TAM unlock to US$154bn. Specifically, we discover: (1) 16x / 45x dollar content increase in scale out / scale up, (2) 13x larger TAM for optics expanding from scale out to scale up, (3) 10x larger value market for pluggable optical modules in scale out from 2H25 AI server model to 2H27 AI server model. In this report, we analyze: (1) dollar content across different configurations, (2) Scale-out and Scale-up market opportunities, and (3) Component contributions across copper cables, pluggable optical modules, CPO, and PCB midplanes .
Allen Chang +852 2978-2930 allen.k.chang@gs.com Goldman Sachs (Asia) L.L.C.
Verena Jeng +852 2978-1681 verena.jeng@gs.com Goldman Sachs (Asia) L.L.C.
James Schneider +1 212 357-2929 jim.schneider@gs.com Goldman Sachs & Co. LLC
Mark Delaney +1 212 357-0535 mark.delaney@gs.com Goldman Sachs & Co. LLC
Ryo Harada +81 3 4587-9865 ryo.harada@gs.com Goldman Sachs Japan Co., Ltd.
Goldman Sachs does and seeks to do business with companies covered in its research reports. As a result, investors should be aware that the firm may have a conflict of interest that could a ffect the objectivity of this report. Investors should consider this report as only a single factor in making their investment decision. For Reg AC certi fication and other important disclosures, see the Disclosure Appendix, or go to www.gs.com/research/hedge.html. Analysts employed by non-US affi l ates are not registered/qualified as research analysts with FINRA in the U.S.
Goldman Sachs (Asia) L.L.C.,
Contributing Authors
Allen Chang +852-2978-2930 allen.k.chang@gs.com Goldman Sachs (Asia) L.L.C.
Mark Delaney, CFA +1(212)357-0535 mark.delaney@gs.com Goldman Sachs & Co. LLC
Michael Ng, CFA +1(212)902-8618 michael.ng@gs.com Goldman Sachs & Co. LLC
Christian Hinderaker, CFA
+44(20)7774-7366 christian.hinderaker@gs.com Goldman Sachs International
Atsushi Ikeda +81(3)4587-9940 atsushi.ikeda@gs.com Goldman Sachs Japan Co., Ltd.
Ryan Huang, CFA +886 2 2730-4084 ryan.huang@gs.com Goldman Sachs (Asia) L.L.C.,
Taipei Branch
Ting Song +852-2978-6466 ting.song@gs.com Goldman Sachs (Asia) L.L.C.
Al Wang +886(2)2730-4081 al.wang@gs.com Goldman Sachs (Asia) L.L.C., Taipei Branch
Yuri Izumikawa
+81(3)4587-3643 yuri.x.izumikawa@gs.com Goldman Sachs Japan Co., Ltd.
Goldman Sachs & Co. LLC
Makoto Takahara
Verena Jeng +852-2978-1681 verena.jeng@gs.com Goldman Sachs (Asia) L.L.C.
Ryo Harada +81(3)4587-9865 ryo.harada@gs.com Goldman Sachs Japan Co., Ltd.
Katherine Murphy +1(212)902-1151 katherine.a.murphy@gs.com Goldman Sachs & Co. LLC
Anmol Makkar +1(212)357-1366 anmol.makkar@gs.com Goldman Sachs & Co. LLC
Evelyn Yu +886 2 2730-4187 evelyn.yu@gs.com Goldman Sachs (Asia) L.L.C., Taipei Branch
Daiki Takayama +81(3)4587-9870 daiki.takayama@gs.com Goldman Sachs Japan Co., Ltd.
Yifan Hu +852-2978-0996 yifan.hu@gs.com Goldman Sachs (Asia) L.L.C.
Zorayda Montemayor +1(212)357-6403 zorayda.montemayor@gs.com Goldman Sachs & Co. LLC
Makoto Takahara +81(3)4587-4270 makoto.takahara@gs.com Goldman Sachs Japan Co., Ltd.
Goldman Sachs International
James Schneider, Ph.D. +1(212)357-2929 jim.schneider@gs.com
Goldman Sachs & Co. LLC
Chao Wang +886(2)2730-4195 kuan-chao.wang@gs.com Goldman Sachs (Asia) L.L.C., Taipei Branch
Daniela Costa +44(20)7774-8354 daniela.costa@gs.com Goldman Sachs International
Alexander Duval +44(20)7552-2995 alexander.duval@gs.com
Goldman Sachs International
Bruce Lu +886 2 2730-4185 bruce.lu@gs.com Goldman Sachs (Asia) L.L.C., Taipei Branch
Xuan Zhang +852-2978-1478 xuan.zhang@gs.com Goldman Sachs (Asia) L.L.C.
Hiroki Muramatsu +81(3)4587-9872 hiroki.muramatsu@gs.com Goldman Sachs Japan Co., Ltd.
Tommaso Nocchi +44(20)7552-2232 tommaso.nocchi@gs.com
Goldman Sachs International
Scale out: Adding more equipment and connecting them through switching technologies, a widely used way of network expansion. Nowadays, AI clusters support scale out connections of 100k+ GPUs.
Scale up: Adding more GPUs and computing resources within the same piece of equipment, typically within the same server rack. Nowadays there are scale-up expansions that connect across racks, or the so-called supernodes, where the networking speed across racks are optimized to close to the connections within the same rack.
Networking is the next frontier in AI infrastructure, poised to enhance computing capability through seamless data exchange and low latency. While an investor concern is on networking configuration replacing another, we expect all configuration will enjoy strong growth. Specifically:
- Dollar content increase by 16x / 45x in Scale Out / Scale Up per computing unit n from GB300 NVL72 (per computing unit means 72 GPUs per rack to reach NVL72) to Rubin Ultra NVL576 (per computing unit means 72 GPUs per rack, and 8 racks together to reach NVL576), with opportunities across pluggable optical modules, optical engines in CPO, copper cables, and PCB midplanes.
- A 13x larger addressable market for optical modules / optical engines expanding n from scale out (e.g. GB300 NVL72) to scale up (e.g. Rubin Ultra NVL576 level 2 scale up via CPO) per computing unit.
- A 10x larger value market for pluggable optical modules in scale out per n computing unit from GB300 NVL72 to Rubin Ultra NVL576, even with a 29% CPO penetration rate. The numbers of pluggable optical module (1.6T equivalent) per computing unit would increase from 216 units in GB300 NVL72 to 2.5k units in Rubin Ultra NVL576.
We expect the aggregate dollar content per computing unit across scale up and scale out to increase by 29x from US$315k in GB300 NVL72 to US$9.4bn in Rubin Ultra NVL576, and assuming the numbers of racks through the full product cycle are 48k racks for GB300 NVL72, and 16.5k computing units for Rubin Ultra NVL576, the aggregate value TAM across scale up and scale out would increase by 9x from US$15bn in GB300 NVL72 (mainly in 2026 ) to US$154bn in Rubin Ultra NVL576 (mainly in 2028 ). Among the US$154bn value TAM, 69% goes to scale up, or US$106bn, and CPO contributes US$91bn , or 59% of the US$154bn value TAM, assuming CPO at 29% penetration rate in scale out.
Connection summary (2025)
Trays
(1) Blackwell network in details
Compute tray x 18 per rack
72 GPUs, 36 CPUs, 18 Switch ASICs, 72 NICs
Switch tray x 9 per rack
Copper cable : GPU ratio = 1: 72
Exhibit 1: Nvidia GB300 AI rack network summary (2025)
Compute Tray
TOR Switch
GPU |GPU|GPU
CPU
CX-I
CX-I
8
BlueField 3
CX-8 supports 800Gb/s scale-out bandwidth per GPU
Switch Tray
NVLink 5
NVLink 5
Switch
Switch
NVLink 5 supports 7.2Tb /s
(unidirectional) scale-up bandwidth per GPU
Front
1.6T
Scale-out
Optical transceiver: GPU ratio = 1:3 (3-layer 1.67) or 1: 2 (2 layer 1.67))
(3-layer, 256 Racks, 18,432 GPUs)
Spine Switch (1.6T) x 128
Switch specification: 72 ports at 1.6T
• Downlink: 72 ports x 128 = 9,216
 _gs_002.png)
Scale out switch spec refer to Quantum-X800.
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
• Computer tray × 10
Switch tray x9
Computer tray x 8
Connection summary (2026E)
Trays
(2) Rubin network in details
Compute tray x 18 per rack
Switch tray x 9 per rack
72 GPUs, 36 CPUs, 36 Switch ASICs, 144 NICs
Copper cable : GPU ratio = 1: 72
Exhibit 2: Nvidia Vera Rubin AI rack network summary (2026E)
Compute Tray
GPU
GPU
GPU
CPU
BlueField 4
CX-
CX-
CX-
CX-
CX-
CX-
CX-9 NIC supports 1.6Tb/s scale-out bandwidth per GPU
Switch Tray
(28.8Tb/s)
NVLink 6
Switch
NVLink 6
NVLink 6
Switch
NVLink 6
Switch
Switch
NVLink 6 supports 14.4Tb /s
(unidirectional) scale-up
GPU
TOR Switch
1.6T
Scale-out
Optical transceiver: GPU ratio = 1:6 (3-layer 1.6T) or 1: 4 (2 layer 1.6T)|
CPO scenario: Optical transceiver: GPU ratio = 1:4 (3-layer 1.6T) or 1: 2 (2 layer 1.6T)
(3-layer, 256 Racks, 18,432 GPUs)
Spine Switch (1.6T) x 228
Switch specification: 64 ports at 1.6T
• Downlink: 64 ports x 288 = 18,432
 _gs_003.png)
bandwidth per GPU
Scale out switch spec refer to Spectrum-6. The data in this exhibit is inferred by GS based on technology roadmap of Nvidia and our supply chain check.
Source: Company data, Goldman Sachs Global Investment Research
(3) Rubin Ultra network in details
Blades
Scale-up (Kyber NVL144)
Trays
Exhibit 3: Nvidia Rubin Ultra AI rack network summary (2027E)
Copper flyover cables
(144 GPUs per Rack)
(4 GPUs per blade)
Compute Blade
CX-9CX-9|
CX-9 CX-9
Blue
Field-4
CX-9 CX-9
CX-9 CX-9
CX-9 NIC supports 1.6Tb/s scale-out bandwidth per
GPU
Switch Blade
NVLink
10.0
Switch
NVLink
Switch
NVLink
Switch
NVLink
Switch
NVLink
Switch
Witten
NVLink 7 supports 14.4Tb /s|
(unidirectional) scale-up bandwidth per GPU
Copper cable
Optical fiber
→ TOR switch
(4 GPUs per Tray)
Compute Tray
GPU
CX-
Cx-
 _gs_004.png)
The data in this exhibit is inferred by GS based on technology roadmap of Nvidia and our supply chain check.
Source: Company data, Goldman Sachs Global Investment Research
IDS)
GPU
GPU
GPU
TOR Switch
TOR Switch
 _gs_005.png)
CPO version
Scale-up (Oberon NVL576)
576 GPUs, 288 CPUs, 864 Switch ASICs, 1, 152 NICs
Scale up connection: L1 copper cables + L2 CPO
(72 GPUs per Rack x 8 Racks)
( 4 ) Scale up / Scale out TAM and EPS implications
Exhibit 5: Spec Table: We outline four possible rack structure for the upcoming GPU platforms (GSe)
The connection architecture and specifications below are based on Nvidia technology roadmap, supply chain check and our inference based on the networking connections
 _gs_006.png)
| Connection assumptions | Vera Rubin Spec A Vera Rubin Spec B Rubin Ultra Spec A Rubin Ultra Spec B | Vera Rubin Spec A Vera Rubin Spec B Rubin Ultra Spec A Rubin Ultra Spec B | Vera Rubin Spec A Vera Rubin Spec B Rubin Ultra Spec A Rubin Ultra Spec B | Vera Rubin Spec A Vera Rubin Spec B Rubin Ultra Spec A Rubin Ultra Spec B | Vera Rubin Spec A Vera Rubin Spec B Rubin Ultra Spec A Rubin Ultra Spec B |
|---|---|---|---|---|---|
| Model | GB300 NVL72 | Vera Rubin NVL 72 | Vera Rubin NVL 72 (25% CPO scale out) | Rubin Ultra NVL 144 (29% CPO scale out) (29% | Rubin Ultra NVL 576 CPO scale out) Remarks |
| Shipment period | 2H25-2026 | 2H26-2027 | 2H26-2027 | 2H27-2028 | 2H27-2028 cable + CPO |
| Scale up | Copper cable | Copper cable | Copper cable | PCB Midplane | Copper |
| Scale out | Optical modules (1.6T) | Optical modules (1.6T) | CPO TOR Switch (1.6T) + Optical module (1.6T) | CPO TOR Switch (3.2T) + Optical module (3.2T) CPO | TOR Switch (3.2T) + Optical module (3.2T) |
| # of GPU package | 72 18 | 72 36 | 72 | 144 72 | 72 * For Rubin Ultra NVL576, we show per rack |
| # of NV Switch ASIC # of NIC | 518 | 36 | 108 data in the table, despite it's a 8-rack scale up | ||
| Scale up | 72 | 144 | 288 | 144 system | |
| Scale up Bandwidth (unidirectional, Tb/s) Total Switching Copper cable (Tb/s) | 1,037 cable | 1,037 | 144 1,037 | 2,074 4,147 - | |
| Capacity (Tb/s) backplane bandwidth Notes # of copper cables (units) Bandwidth of each copper cable (Gb/s) | 518 100% copper 5,184 200 - | 2,074 1,037 100% copper cable 5,184 200 - | 2,074 1,037 100% copper cable 5,184 200 - | - L1 - - 2,074 connection | 2,074 6,221 2,074 scale up within rack Total bandwidth = # of copper cales x |
| PCB midplane bandwidth (Tb/s) Notes # of PCB Midplane (units) Bandwidth of each PCB (Gb/s) CPO/ NPO bandwidth (Tb/s) Notes # of optical fiber (units) Bandwidth of each optical fiber # of optical engine (units) Data rate of Optical engine (Tb/s) | - 58 346 58 72 0.8 3 216 100% 1.6 3.0 - - - | 100% 1.6 6.0 - - - - - | 1.5 4.5 | 0.9 2.1 | 3.2 0.9 capacity in scale-out network x %in CPO Data rate each optical engine (3.2T) |
| Attach rate summary Optical module : GPU - scale out Optical engine : GPU - scale up Optical engine : GPU - scale out | 3.0 - - | 6.0 - - | - | 3.2 | 2.1 |
| GPU uplink (Tb/s) # of GPU (units) GPU uplink per chip (Tb/s) # of layer module # of optical module (units) Optical module penetration (GSe) Data rate of optical module (Tb/s) Attach rate (Optical module ports: NPO # of CPO ports (units) | - - - | 1.5 | - | 4.5 | |
| - - - - - | - - | - - - - | 2 - 144 1.6 | 5,184 400 - - - 1.6 Bandwidth of each cable / 2 (unidirectional) Total bandwidth = # of PCB Midplane x speed (800G/ 1.6T/ 3.2T) | |
| - - | - - - - | - - - - - | PCB all to all | - 1,037 among racks Bandwidth of each PCB | |
| - | |||||
| - | 1,036,800 L2 | scale up 5,184 Total bandwidth = # of copper cales x | |||
| (Gb/s) | - | - - - | - - - | 400 324 3.2 Bandwidth of each cable / 2 Total bandwidth = # of optical engine x Bandwidth of each optical engine | |
| - | (unidirectional) | ||||
| Tb/s) | 115 691 115 72 1.6 | 115 691 115 | 115 691 Assuming 3-layer network 115 72 GPU uplink = # of GPU per rack x GPU uplink | ||
| Scale out Scale out Bandwidth (unidirectional, Total Switching Capacity (Tb/s) | 432 | 72 1.6 3 324 | 230 1,382 230 | 3.2 2.1 63 29% scale-out network x %in optical module / Data rate of each optical module (1.6T/ 3.2T) # of CPO ports = Total switch capacity in scale-out network x %in CPO / Data rate of | |
| Attach rate (Optical engine: GPU) | 1.6 | / | |||
| 4.5 108 25% 1.6 1.5 | 307 71% 3.2 2.1 125 | ||||
| 75% 1.6 | |||||
| 63 # of CPO optical engine = Total switch | |||||
| 153 71% # of optical module = Total switch capacity | |||||
| Optical | |||||
| 3 | 3 | 3 | |||
| in | |||||
| CPO/ | |||||
| GPU) | |||||
| CPO penetration (GSe) | 29% 3.2 | 3.2 each CPO port (1.6T/ 3.2T) | |||
| Data rate of CPO (Tb/s) | - | 0.9 | |||
| Attach rate (CPO ports: GPU) Among which: | 108 | 0.9 | |||
| # of optical engine (units) | 125 | ||||
| Data rate of Optical engine | |||||
| (Tb/s) | |||||
| 0.9 | 0.9 | ||||
| Scale out sum | |||||
| 3.0 | 6.0 | 6.0 | 3.0 | ||
| 3.0 |
(1) Shipment period is based on the Nvidia's product roadmap; (2) # of GPU/ NV Switch ASIC/ NIC per rack are based on Nvidia's product roadmap, industry check and our calculation based on the announced specifications; (3) Bandwidth and switching capacity are based on Nvidia's product roadmap.
Source: Company data, Goldman Sachs Global Investment Research
Following the 2026 GTC , where a clear technology roadmap of the next generations of GPU, server rack, and networking solutions are outlined, we present an in-depth analysis of the specification changes for the next two generations of GPU platforms (Vera Rubin and Rubin Ultra). By integrating insights from the company's announcement and our industry checks with the supply chain, we intend to delineate the inter-connection architecture of the forthcoming server racks (Exhibit 5), breakdown the dollar content for each networking component (Exhibit 6), forecast the TAM opportunity for each
component (Exhibit 8) and assess the EPS impact for our key coverage names. With these, we aim to address the frequently asked investor question about how to evaluate the market size resulting from complex server networking upgrades ahead.
The next generation racks: Networking technology continues to evolve in both scale out (across racks / computing unit) and scale up (within one computing unit). Based on the GTC 2026 announcement, we discuss the major four configurations in Vera Rubin and Rubin Ultra platform, to evaluate the TAM opportunities and EPS impact across the supply chain. Separately, we consider the CPO penetration rate in scale-out connections in our analysis (Exhibit 5), which is based on industry checks and our optimistic perspective on CPO adoption. As CPO is a next-generation technology, its pace of adoption will significantly influence the connection TAM opportunity, especially given its dollar content per rack. Detailed specification assumptions are shown in Exhibit 5 and below.
- Vera Rubin Spec A - NVL72. Scale up adopts copper cable cartridge, i.e. the n Oberon rack. Scale out adopts pluggable optical module connections.
- Vera Rubin Spec B -NVL72 (CPO scale out) . Scale up adopts copper cable n cartridge, i.e. the Oberon rack. Scale out could use CPO or pluggable optical module, and we assume 25% CPO penetration rate.
- Rubin Ultra Spec A - NVL144 (CPO scale up). Scale up adopts PCB midplane for n connection, i.e. the Kyber rack. Scale out could use CPO or pluggable optical module, and we assume 29% CPO penetration rate.
- Rubin Ultra Spec B - NVL576 (CPO scale up). 72 GPU per rack, and eight racks as a n full NVL576 (one computing unit). Scale up adopts copper cable cartridge for the first layer connection (i.e. Oberon rack design for the connections within the server rack) and CPO for the second layer connection (i.e. the connections between the eight racks). Scale out could use CPO or pluggable optical module, and we assume 29% CPO penetration rate.
Based on above, we derive: (1) dollar content per rack across different configurations, (2) value TAM opportunities (dollar content per rack * number of server racks), and (3) implication to supply chain EPS. Key takeaways: (1) Rubin Ultra brings dollar content increases across scale up and scale out compared to GB300, (2) Rubin Ultra scale up brings dollar content increases across copper cables, PCB midplane, and CPO switch, (3) pluggable optical modules would continue to increase dollar content increase per rack in scale out compared to GB300 despite our assumption of 25-29% CPO switch penetration rate, (4) CPO appears costly considering 3D packaging, high integration, and the need to upgrade devices to semiconductor processing levels; however, the total cost of ownership (TCO) is attractive especially in high bandwidth requirement (e.g. 6.4T, 12.8T) considering the constraints of pluggable optical module and the power efficiency and energy savings, and (5) On the EPS opportunity, we see most suppliers benefiting from strong EPS contribution from a single configuration that is higher than their 2025 full year EPS, given the AI server racks shipment ramp up and specification upgrades driving the dollar content increase.
Exhibit 6: Dollar content per rack: GB300, Vera Rubin and Rubin Ultra
Volume assumptions are from the spec table (Exhibit 5), ASP assumptions are based on spec table and our industry checks
| Dollar content | GB300 NVL72 | Vera Rubin Spec A Vera Rubin NVL 72 | Vera Rubin Spec B Vera Rubin NVL 72 | Rubin Ultra Spec A Rubin Ultra NVL 144 | Rubin Ultra Spec B Rubin Ultra NVL 576 |
|---|---|---|---|---|---|
| Shipment period | 2H25-2026 | 2H26-2027 | 2H26-2027 | 2H27-2028 | 2H27-2028 |
| Scale up | Copper cable | Copper cable | Copper cable | PCB Midplane | Copper cable + CPO |
| Scale out | Optical modules (1.6T) | Optical modules (1.6T) | CPO TOR Switch (1.6T) + Optical module (1.6T) | CPO TOR Switch (3.2T) + Optical module (3.2T) | CPO TOR Switch (3.2T) + Optical module (3.2T) |
| # of GPU package | 72 | 72 | 72 | 144 | 72 |
| # of NV Switch ASIC | 18 | 36 | 36 | 430 | 430 |
| # of NIC | 72 | 144 | 144 | 288 | 144 |
| Networking costs per rack (US$ k) | 315 | 489 | 504 | 1,113 | 1,169 |
| Scale-up | 140 | 140 | 140 | 381 | 803 |
| Scale-out | 175 | 349 | 364 | 732 | 366 |
| Scale up dollar content per rack (US$ k) | 140 | 140 | 140 | 381 | 803 |
| Copper cable - backplane (US$ k) | 93 | 93 | 93 | - | 156 |
| Volume (unit) | 5,184 | 5,184 | 5,184 | - | 5,184 |
| ASP (US$) | 18 | 18 | 18 | - | 30 |
| Copper cable - flyover in switch tray (US$ k) | 47 | 47 | 47 | 156 | 78 |
| Volume (unit) | 5,184 | 5,184 | 5,184 | 10,368 | 5,184 |
| ASP (US$) | 9 | 9 | 9 | 15 | 15 |
| PCB midplane (US$ k) | - | - | - | 225 | - |
| Volume (unit) | - | - | - | 2 | - |
| ASP (US$) | - | - | - | 112,500 | - |
| Optical engine &FAU (CPO/ NPO, US$ k) | - | - | - | - | 324 |
| Volume (unit) | - | - | - | - | 324 |
| ASP (US$) | - | - | - | - | 1,000 |
| ELS (CPO/ NPO, US$ k) | - | - | - | - | 65 |
| Volume (unit) | - | - | - | - | 162 |
| ASP (US$) | - | - | - | - | 400 |
| Fiber cable and MPO(CPO/ NPO, US$ k) | - | - | - | - | 156 |
| Volume (unit) | - | - | - | - | 5,184 |
| ASP (US$) | - | - | - | - | 30 |
| Shufflebox (US$ k) | - | - | - | - | 25 |
| Volume (unit) | - | - | - | - | 36 |
| ASP (US$) | - | - | - | - | 700 |
| Scale out dollar content per rack (US$ k) | 175 | 349 | 364 | 732 | 366 |
| Optical module (US$ k) | 173 | 346 | 259 | 491 | 245 |
| Volume (unit) | 216 | 432 | 324 | 307 | 153 |
| ASP (US$) | 800 | 800 | 800 | 1,600 | 1,600 |
| Optical engine &FAU (CPO/ NPO, US$ k) | - | - | 86 | 200 | 100 |
| Volume (unit) | - | - | 108 | 125 | 63 |
| ASP (US$) | - | - | 800 | 1,600 | 1,600 |
| ELS (CPO/ NPO, US$ k) | - | - | 11 | 25 | 13 |
| Volume (unit) | - | - | 27 | 63 | 31 |
| ASP (US$) | - | - | 400 | 400 | 400 |
| Fiber cable and MPO(CPO/ NPO, US$ k) | 2 | 4 | 4 | 6 | 3 |
| Volume (unit) | 108 | 216 | 216 | 216 | 108 |
| ASP (US$) | 18 | 18 | 18 | 30 | 30 |
| Shufflebox (US$ k) | - | - | 4 | 10 | 5 |
| Volume (unit) | - | - | 6 | 14 | 7 |
| ASP (US$) | 700 | 700 | 700 | 700 |
Source: Company data, Goldman Sachs Global Investment Research
Exhibit 7: CPO switch specification comparison
| Nvidia | Nvidia | Broadcom | Broadcom | |
|---|---|---|---|---|
| Quantum-X Photonics | Spectrum-X Photonics | Tomahawk 5 - Bailly | Tomahawk 6 - Davisson | |
| Network | InfiniBand | Ethernet | Ethernet | Ethernet |
| Switching capacity | 115.2Tb/s | 409.6 Tb/s | 51.2 Tb/s | 102.4 Tb/s |
| # of Switch ASIC | 4 | 4 | 1 | 1 |
| Switch ASIC | 28.8Tb/s | 102.4Tb/s | 51.2Tb/s | 102.4Tb/s |
| Ports | 144 x 800G | 512 x 800G | 128 x 400G | 64 x 1.6T / 128 x 800G |
| Speed | 200G/lane SerDes | 200G/lane SerDes | 100G/lane SerDes | 200G/lane SerDes |
| External laser source | 18 | 64 | TBA | TBA |
| Optical engines | 72 (1.6T) | 128 (3.2T) | 8 (6.4T) | 16 (6.4T) |
| Size | 4U | 5U | 4U | TBA |
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
How cro works in data transmission
Laser
Optical Engine
Exhibit 8: Scale up / Scale out TAM opportunities
Laser
Optical Engine
PIC
The TAM opportunity is based on the connection content value of each rack (Exhbit 6) multiple by GSe shipment of Vera Rubin / Rubin Ultra racks
MUX/
| Value TAM | GB300 NVL72 | Vera Rubin Spec A Vera Rubin NVL 72 | Vera Rubin Spec B Vera Rubin NVL 72 (25% CPO scale out) | Rubin Ultra Spec A Rubin Ultra NVL 144 (29% CPO scale out) | Rubin Ultra Spec B Rubin Ultra NVL 576 (29% CPO scale out) |
|---|---|---|---|---|---|
| Scale up | Copper cable | Copper cable | Copper cable | PCB Midplane | Copper cable + CPO |
| Scale out | Optical modules (1.6T) | Optical modules (1.6T) | CPO TOR Switch (1.6T) + Optical module (1.6T) | CPO TOR Switch (3.2T) + Optical module (3.2T) | CPO TOR Switch (3.2T) + Optical module (3.2T) |
| Scale up + Scale out TAM (US$m) | 15,070 | 28,291 | 29,158 | 73,458 | 154,313 |
| Copper cable - backplane | 4,468 | 5,393 | 5,393 | - | 20,529 |
| Copper cable - flyover in switch tray | 2,234 | 2,697 | 2,697 | 10,264 | 10,264 |
| PCB Midplane | - | - | - | 14,850 | - |
| Optical module | 8,274 | 19,976 | 14,982 | 32,390 | 32,390 |
| Optical engine & FAU | - | - | 4,994 | 13,230 | 55,998 |
| ELS | - | - | 624 | 1,654 | 10,207 |
| Fiber cable andMPO | 93 | 225 | 225 | 428 | 20,956 |
| Shufflebox | - | - | 243 | 643 | 3,970 |
| Scale up TAM (US$ m) | 6,702 | 8,090 | 8,090 | 25,114 | 105,970 |
| Copper cable - backplane | 4,468 | 5,393 | 5,393 | - | 20,529 |
| Copper cable - flyover in switch tray | 2,234 | 2,697 | 2,697 | 10,264 | 10,264 |
| PCB Midplane | - | - | - | 14,850 | - |
| Optical engine & FAU | - | - | - | - | 42,768 |
| ELS | - | - | - | - | 8,554 |
| Fiber cable andMPO | - | - | - | - | 20,529 |
| Shufflebox | - | - | - | - | 3,326 |
| Mix% | 100% | 100% | 100% | 100% | 100% |
| Copper cable - backplane | 67% | 67% | 67% | 0% | 19% |
| Copper cable - flyover in switch tray | 33% | 33% | 33% | 41% | 10% |
| PCB Midplane | 0% | 0% | 0% | 59% | 0% |
| Optical engine & FAU | 0% | 0% | 0% | 0% | 40% |
| ELS | 0% | 0% | 0% | 0% | 8% |
| Fiber cable andMPO | 0% | 0% | 0% | 0% | 19% |
| Shufflebox | 0% | 0% | 0% | 0% | 3% |
| Scale out TAM (US$ m) | 8,367 | 20,200 | 21,067 | 48,344 | 48,344 |
| Optical module | 8,274 | 19,976 | 14,982 | 32,390 | 32,390 |
| Optical engine & FAU | - | - | 4,994 | 13,230 | 13,230 |
| ELS | - | - | 624 | 1,654 | 1,654 |
| Fiber cable andMPO | 93 | 225 | 225 | 428 | 428 |
| Shufflebox | - | - | 243 | 643 | 643 |
| Mix% | 100% | 100% | 100% | 100% | 100% |
| Optical module | 99% | 99% | 71% | 67% | 67% |
| Optical engine & FAU | 0% | 0% | 24% | 27% | 27% |
| ELS | 0% | 0% | 3% | 3% | 3% |
| Fiber cable andMPO | 1% | 1% | 1% | 1% | 1% |
| Shufflebox | 0% | 0% | 1% | 1% | 1% |
| Assumption: GPU Rack forecast (GSe, k units, full lifecycle) | Assumption: GPU Rack forecast (GSe, k units, full lifecycle) | ||||
| GB300 Vera Rubin | 48 | 58 | 58 | ||
| Rubin Ultra | 66 | 132 |
Waveguide
(1) Spec A come with lower connection value while Spec B come with higher connection value; (2) Read more about GPU rack shipment in our Global Server Model
Source: Company data, Goldman Sachs Global Investment Research
Exhibit 9: How CPO works in data transmission
 _gs_007.png)
Source: Company data, Compiled by Goldman Sachs Global Investment Research
Scale up racks
→ TOR|
Exhibit 10: Our high end estimate imply a CPO TAM totalling US$97bn in 2026-28E
High-end based on GB300, Vera Rubin/Rubin Ultra Spec B; Low-end based on GB300, Vera Rubin/ Rubin Ultra Spec A
PCB midplanes
CPO/ NPO scale up
| Value TAM range (high-end, Spec B) | 2026E | 2027E | 2028E |
|---|---|---|---|
| Scale up TAM (US$ m) | 7,228 | 30,697 | 82,837 |
| Scale out TAM (US$ m) | 10,921 | 27,023 | 39,835 |
| CPO TAM (US$ m) | 1,024 | 24,840 | 70,881 |
| Optical engine & FAU | 864 | 16,269 | 43,858 |
| ELS | 108 | 2,763 | 7,961 |
| Fiber cable and MPO | 10 | 4,734 | 15,965 |
| Shufflebox | 42 | 1,074 | 3,096 |
| Value TAM range (low-end, Spec A) | 2026E | 2027E | 2028E |
| Scale up TAM (US$ m) | 7,228 | 12,321 | 20,358 |
| Scale out TAM (US$ m) | 10,771 | 26,408 | 39,733 |
| CPO TAM (US$ m) | - | 3,557 | 12,093 |
| Optical engine & FAU | - | 3,007 | 10,223 |
| ELS | - | 376 | 1,278 |
| Fiber cable and MPO | - | 28 | 96 |
| Shufflebox | - | 146 | 497 |
(1) CPO TAM refers to CPO components applied in both scale up and scale out connections; (2) By year estimate is based on the Server rack allocation among years and the full lifecycle TAM in Exhibit 7; (3) High end is based on Spec B and Low end on Spec A of each model
Source: Company data, Goldman Sachs Global Investment Research
Exhibit 12: Connections in data center: Scale-up vs. Scale out
 _gs_008.png)
TOR Switch: Top of rack switch; EoR Switch: End of row switch
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
Value TAM by year
Scale out clusters
AEC
•AOC
- DAC
Exhibit 11: Our high end estimate imply a demand for 110k scale out CPO switch in 2028E
High-end based on GB300, Vera Rubin/Rubin Ultra Spec B; Low-end based on GB300, Vera Rubin/ Rubin Ultra Spec A
Volume TAM by year
| Volume TAM range (high-end, Spec B) | 2026E | 2027E | 2028E |
|---|---|---|---|
| CPO TAM (k units) | |||
| Optical engine & FAU | 1,080 | 16,027 | 40,172 |
| ELS | 270 | 6,907 | 19,902 |
| Fiber cable and MPO | 540 | 158,674 | 532,330 |
| Shufflebox | 60 | 1,535 | 4,423 |
| Key data | |||
| Nvidia AI rack shipment (k units) | 50 | 77 | 121 |
| Scale out CPO penetration | 5% | 25% | 29% |
| Scale out CPO switch (k units*) | 15 | 88 | 110 |
| Volume TAM range (low-end, Spec A) | 2026E | 2027E | 2028E |
| CPO TAM (k units) | |||
| Optical engine & FAU | - | 1,879 | 6,389 |
| ELS | - | 940 | 3,195 |
| Fiber cable and MPO | - | 940 | 3,195 |
| Shufflebox | - | 209 | 710 |
| Key data | |||
| Nvidia AI rack shipment (k units) | 50 | 77 | 121 |
| Scale out CPO penetration | 0% | 11% | 27% |
| Scale out CPO switch (k units*) | - | 26 | 89 |
(1) CPO TAM refers to CPO components applied in both scale up and scale out connections; (2) assume 72 optical engines per scale out switch (*); (3) By year estimate is based on the Server rack allocation among years and the full lifecycle TAM in Exhibit 7; (4) High end is based on Spec B and Low end on Spec A of each model
Source: Company data, Goldman Sachs Global Investment Research
( 5 ) Choice of connection: PCB vs. Copper vs. Optics
Exhibit 14: Copper vs. Optics vs. PCB
 _gs_009.png)
Source: Company data
Copper cables and PCBs are commonly used materials for short-distance connections. PCBs handle the short connections within servers, while copper cables connect inside or between servers. Both offer lower cost and power consumption in short distances; however, signal quality degrades rapidly over longer distances or higher speeds, leading to optical fiber connections to stand out for long distance or high speed interconnections. AI data centers are being designed for faster bandwidth, larger scale, easier deployment and lower costs. There are many different options for AI data center connections (as we list in Exhibit 15), and connection technologies change over time. Major trends recently include (1) copper cables evolving from DAC to ACC to AEC, toward longer distance connection, (2) PCBs evolving from intra-tray use to intra-rack connections, and (3) optics evolving from scale out to scale up, with AOC and CPO switches, covering shorter distance connection.
Exhibit 15: From PCB to Fiber: different types of connections for different distances
 _gs_010.png)
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
Scale up configurations: (1) From short to longer distances: In general, PCB traces are for connections within a server tray, DAC copper cables for connections within a server rack (e.g. GB200's copper cartridge) and AEC copper cables for connections among server racks (e.g. AmazonTrn2-Ultra64). AEC cables come with retimers that drives the connection distance to reach 5-30m (vs. <3m of DACs). (2) PCB midplane: Server trays and switch trays could be connected directly to a PCB midplane, which is likely to feature M9 CCL materials and 78 or more layers. The initial costs could be higher, though the midplane could be easier to assemble and more competitive in costs-to-performance ratio for higher speed requirements (e.g. Rubin Ultra NVL144). (3) Optics show strength in longer interconnections and larger bandwidth, which Google uses in part of its scale-up 3D torus. To meet the larger bandwidth demands and shorten connection distances, Optics connections are expanding from pluggable optical modules to onboard optics (NPO), and co-packaged optics (CPO), which could be seen in scale up configurations (e.g. Rubin Ultra NVL576 level 2 scale up). Scale out configurations: in general, optics are for AI cluster scale out given the requirement on high speed and larger interconnection size.
Exhibit 16: Connection solutions of mainstream AI servers
 _gs_011.png)
| Model | Year | Scale-up | Scale-out | |
|---|---|---|---|---|
| GB200 | 2024 | Copper | Optical (400G, 800G) | |
| Nvidia | GB300 | 2025 | Copper | Optical (800G, 1.6T) |
| VR200 (1.6T) | 2026 | Copper | Optical (800G, 1.6T) | |
| Rubin Ultra | 2027 | Copper, PCB, Optical | Optical (1.6T, 3.2T) | |
| V6 | 2024 | Copper, Optical (800G) | Optical (800G) | |
| V7 | 2025 | Copper, Optical (1.6T) | Optical (1.6T) | |
| V8e | 2026 | Copper, Optical | Optical | |
| V8p | 2026 | Copper, Optical | Optical | |
| Amazon | Trainium 2 | 2025 | Copper | Copper, Optical (400G) |
| Amazon | Trainium 3 (Teton 3 PDS/ MAX) | 2026 | Copper | Copper, Optical |
| Amazon | Trainium 4 | 2027 | Copper | Copper, Optical |
| Meta | MITA-T V1 (Minerva) | 2026 | Copper | Optical (800G) |
| Huawei | Cloud Matrix 384 | 2025 | Copper, Optical | Optical |
400G/ 800G/ 1.6T/ 3.2T refers to the fastest data rate in the network, while there can be lower speed ports in the network
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
( 6 ) Optics: speed upgrade continues
Exhibit 17: Roadmap of data center connection speed
 _gs_012.png)
Source: Company data, Goldman Sachs Global Investment Research
We expect the speed migration to continue through 2028E, from 800G to 1.6T in 2026, and toward 3.2T and above in the following years, and followed by China Cloud speed migration to lengthen the migration cycle. To meet the increasing demand on bandwidth, power consumption, and miniaturization, the form of optical connection is evolving (Exhibit 18): (1) pluggable optical module are shifting towards silicon photonics from EML, which come with higher integration, lower cost and reduced reliance on laser supply, (2) expanding from pluggable optical module to onboard optics (NPO) and co-packaged optics (CPO), covering short distance connection with high bandwidth and better power efficiency. CPO will initially be integrated with switch ASIC, followed by XPUs (GPU, CPU, ASICs, etc.).
camel lo. cro lechorosy mig duons
Ethernet Switch occallro
arate
Optical
CDR
Substrate
Engine
Exhibit 18: Optics technology migrations
 _gs_013.png)
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
( 7 ) CPO with Switch kick o ff in 2026
 _gs_014.png)
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
CPO expands optics to cover short distance connection and higher bandwidth. It places optical engines as close as possible to the chips, shortening the electrical paths from several centermeters to millimeter level and lower power consumption. The shorter transmission path could also reduce latency and save DSPs and retimers, along with the power consumption by these devices. The higher integration also brings smaller size. On other hand, CPO requires supply chain technology migration rather than single devices upgrade, which could take time to develop; in addition, the co-packaging leads to higher maintenance costs: in a pluggable optical module, if the optical engine fails, one replaces the optical module, and the switch system remains intact, while in (1) onboard optics / NPO, it would need to replace the switch PCB, (2) CPO with the switch, failure would
Mone
Pluggable transceivers
OIO
EML-based → Silicon Photonics
(2024)
1) -
Interposer
Accelerator
Pluggable
HBM
affect the switch ASIC, and (3) CPO with XPU, failure would affect the XPU (e.g. GPU, CPU, NPU, etc.). The lifecycle of PIC and EIC are different, with PIC are more delicate, leading to the design of pluggable optical module. As a result, we expect (1) pluggable optical module would co-exist with NPO / CPO, and on continuous speed migration toward 3.2T, (2) CPO would be more attractive to clients in short distance and higher bandwidth that pluggable optical module could not achieve, and (3) pluggable optical module suppliers would also enjoy new optics devices opportunities in NPO / CPO, such as optical engine, FAU, ELS module, etc.
Exhibit 20: CPO: key development of major players
| Key players | Progress | Progress details | Highlights |
|---|---|---|---|
| Nvidia | - CPO Switch commercially avaiable in 2026 (Scale out) | Mar-2025: Announced CPO switch (Quantum-X InfiniBand, Spectrum-X Ethernet) Early 2026: Commercial availability of CPO switch | Adopt MRM(Micro Ring Modulator) technology, achieving higher density and efficiency |
| Broadcom | - Davisson (102.4T) CPO switch sampling in Oct 2025 (Scale-out and scale-up) | Mar-2022 : world's first 25.6T CPO Demo June-2023: 51.2T CPO sampling Mar-2024: Bailly (51.2T) CPO switch deliverd to customers Oct-2025: Davisson (102.4T) CPO switch delivered to customers | Adopt MZM (Mach-Zehnder Modulator), which is more matured, while also developing MRM |
| Marvell | - CPO ethernet swith sampling in 2027 (Scale-out) - Developing CPO for XPUs (Scale-up) | Feb-2026: Acquired Celestial AI, a startup focusing on CPO for XPUs 2027: CPO (204T) Ethernet switch sampling | CPO solution to combine with custom XPUs for CSPs |
| Ranovus x Mediatek | Announced co-developed CPO for ASIC in 2024 | Mar-2024: Announced Odin CPO solutions (6.4T) collaborating with Mediatek's ASIC platform | Targeting CPO technology on XPU |
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
Laser +
Photo diode
•IlInN O IIAC
Nro switchby nagle
- Input coupler
• Modulator
Exhibit 22: CPO Switch BoM breakdown, by value (GSe)
CPO switch BoM (Quantum-X Photonics)
| US$ | # | ASP | Value |
|---|---|---|---|
| Switch ASIC | 4 | 3,000 | 12,000 |
| Optical engines (1.6T) | 72 | 450 | 32,400 |
| FAU | 72 | 50 | 3,600 |
| ELS | 18 | 400 | 7,200 |
| Among which: CWlaser (300mw) | 144 | 30 | 4,320 |
| Shuffle box | 1 | 2,500 | 2,500 |
| MPO connectors/ cables | 144 | 40.0 | 5,760 |
| Single mode Fiber | 1,152 | 11 | 12,343 |
| BoM | 75,803 | ||
| Markup | 62,220 | ||
| Selling price | 130,000 |
ASP estimates are based on industry checks
Source: Company data, Goldman Sachs Global Investment Research
Exhibit 24: NPO Switch by Ruijie
 _gs_015.png)
Picture from Ruijie
Source: Company data
Exhibit 23: CPO Switch BoM breakdown, by % (GSe)
 _gs_016.png)
Source: Company data, Goldman Sachs Global Investment Research
Exhibit 25: NPO Switch by Ruijie
 _gs_017.png)
Picture from Ruijie
Source: Company data
( 8 ) Pluggable optical modules: SiPh-based in expansion
Exhibit 26: We expect SiPh's penetration in datacom market to increase from 6% in 1Q24 to 46% in 4Q28E
 _gs_018.png)
Global Optical Module TAM.
Source: Company data, Goldman Sachs Global Investment Research
Exhibit 27: Inside a Silicon photonics chip
 _gs_019.png)
Source: Company data
Electrical Signal
Digital
Optical Signal
TOSA
ROSA +
TIA
Laser
Housing
CW Laser
Light
Optical Signal
Analog ww.
We expect the adoption of silicon photonics in optical transceiver modules to grow from 6% in 1Q24 to 45% in 4Q28 (Exhibit 26). Silicon Photonics provides advantage over traditional discrete optical transceivers due to (1) a higher level of integration and smaller size, (2) lower power consumption, and (3) lower costs. These advantages will become more prevalent when the industry migrates to faster speeds. Nevertheless, the EML would also co-exist in AI data center, especially for long distance transmission given CW lasers require high power to deliver similar performance to EML in long distances. In short distance, some clients would still choose EML considering it is a long-established technology with a longer track record, and the reliability of the network is critical for AI computing, and GPU remains the major cost contributor, making the cost reduction from the light source less significant. The price of like for like products declines as scale ramps up, similar to many other technology components; nevertheless, the speed migration toward 800G / 1.6T / 3.2T would continue to drive the blended ASP expansion. Gross margin of optical module suppliers will ramp to 48%-55%, driven by product mix migration, as higher speed products would shift towards SiPh-based optical modules, benefiting from lower laser costs. Electrical Signal
Exhibit 28: 800G: SiPh has 26% BoM advantage and 15% price advantage
800G Optical module BoM
| 800G BoM (US$) | EML | EML | Silicon Photonics | Silicon Photonics | Diff |
|---|---|---|---|---|---|
| Components | # | Value | # | Value | Value |
| TOSA (excl. laser, driver) | 1 | 15 | - | - | (15) |
| Laser | 8 x 100G EML | 96 | 4 x 70mw | 16 | (80) |
| Driver | 2 | 20 | 2 | 20 | - |
| ROSA (excl. TIA) | 1 | 20 | - | - | (20) |
| TIA | 2 | 20 | 2 | 20 | - |
| Silicon Photonics chip | - | - | 2 | 40 | 40 |
| DSP | 1 | 80 | 1 | 80 | - |
| PCBA | 1 | 30 | 1 | 25 | (5) |
| Others | 29 | 29 | - | ||
| Total BoM | 310 | 230 | -80 | ||
| Cost advantage of SiPh | -26% | ||||
| ASP | 430 | 365 | -65 | ||
| GM | 28% | 37% | -23% | ||
| Price advantage of SiPh | -15% |
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
Exhibit 30: EML-based optical transceivers: more discrete devices inside
 _gs_020.png)
Source: Company data
Exhibit 29: 1.6T: SiPh has 32% BoM advantage and 20% price advantage for 1.6T applications
1.6T Optical module BoM
| 1.6T BoM (US$) | EML | EML | Silicon Photonics | Silicon Photonics | Diff |
|---|---|---|---|---|---|
| Components | # | Value | # | Value | Value |
| TOSA (excl. laser, driver) | 1 | 15 | - | - | (15) |
| Laser | 8 x 200G EML | 160 | 4 x 70mw | 16 | (144) |
| Driver | 2 | 30 | 2 | 30 | - |
| ROSA (excl. TIA) | 1 | 60 | - | - | (60) |
| TIA | 2 | 30 | 2 | 30 | - |
| Silicon Photonics chip | - | - | 2 | 70 | 70 |
| DSP | 1 | 130 | 1 | 130 | - |
| PCBA | 1 | 35 | 1 | 25 | (10) |
| Others | 40 | 40 | - | ||
| Total BoM | 500 | 341 | -159 | ||
| Cost advantage of SiPh | -32% | ||||
| ASP | 1,000 | 800 | -200 | ||
| GM | 50% | 57% | 21% | ||
| Price advantage of SiPh | -20% |
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
Exhibit 31: Silicon photonic optical transceivers: higher integration with simplified structure
 _gs_021.png)
Source: Company data
SiPho
un photome Upucal Hanscelers. How udla tanellosion works
SiPh transceiver module
Laser
Silicon Photonic chip
Exhibit 32: Silicon photonic optical transceivers: how data transmission works
TIA
Photo detector
 _gs_022.png)
Source: Company data, Compiled by Goldman Sachs Global Investment Research
( 9 ) Light sourcing: Supply chain and Shortage
Supply tightness remains in 2026, across both EML and CW lasers, due to (1) strong demand driven by AI servers ramp up, speed migration, and optical connections expansion, (2) InP substrate supply constraints, which are used in both EML and CW lasers, and across PD (receivers) and LD (transmitters), and under geopolitical tension / export controls by the China government, and (3) it takes time for capacity expansion and ramp up across the supply chain. Nevertheless, we see continuous capacity expansion plans from the supply chain, such as: VPEC plans to expands InP MOCVD from 60 units to 64 units in 2H26, both Landmark and YJ Semi commits on significant capacity expansion in 2026, Lumentum plans to expand capacity by 40% from CY3Q25 to CY2Q26, and likely to bring more capacities online in 2026, and Coherent has also committed to doubling the capacity. Overall, we expect light source supply to remain tight through 2027, and could turn more balanced in 2H28 post supply chain capacity expansion, or a slowdown in AI servers specification upgrades as the industry shifts to inferencing from training, or ease of geopolitical tension / export controls by the China government.
2025
2025
Q1
Very tight
2025
2025
2026 2026
2026
Q2E
VCSEL: Vertical-Cavity Surface-Emitting Laser
2026
Q3E
Q4E
2027
Q1E
2027
Q2E
2027
Q3E
2027
Q4E
Tight
Exhibit 34: Light source supply
 _gs_023.png)
Source: Goldman Sachs Global Investment Research
Light sourcing in CPO could also vary: CW lasers stand out as the most widely adopted light source in optical modules, and could serve longer-distance connection, fulfilling both scale out and scale up. VCSEL also benefits from technology readiness and has higher energy efficiency (lower power consumption) than CW laser solutions. Although its effective reach is shorter than that of CW laser, the distance is sufficient for scale up. MicroLED is another potential choice in scale up, with high energy efficiency, and low latency, while the technology readiness is lower compared to CW lasers and VSCEL.
Exhibit 36: Different choices of laser for CPO
| SiPh + CWLaser | VCSEL | MicroLED | |
|---|---|---|---|
| Energy efficiency | Lower | High | High |
| Effective reach (meter) | >1km | <100m | <20m |
| Latency | Low to Medium | Low to Medium | Low |
| Cost per bit | Medium | Low | Low |
| Technology readiness | High | High | Low |
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
2028
Q1E
2028
Q2E
2028
2028
Q3E
Q4E
Balance
cambre so: nuge shetwork soluton Opular mouble allach laleal let/ 10101< 1o layer network (ouue)
Domain-1
x32
x32
Domain-32
6-1 C-2
x32
Rail-8
5-32
Core
800G
( 10 ) Attach rate: Pluggable optical modules to GPU / ASIC in scale out
Spine
8006 ]
Leaf
400G |
Server x32
S-31
S-32
Exhibit 38: Optical transceiver attach rate
L-1
12l
Srv
| Model | Year | Attach rate (GSe) | Data rate | |
|---|---|---|---|---|
| Nvidia | GB200 | 2024 1: 2~3 | 800G | |
| Nvidia | GB300 | 2025 | 1: 2~3 | 1.6T |
| Nvidia | VR200 | 2026 | 1: 4~6 | 1.6T |
| Nvidia | Rubin Ultra | 2027 | TBA | 3.2T (GSe) |
| V6 | 2024 | 1:4 | 800G | |
| V7 | 2025 | 1:4 | 1.6T | |
| Amazon | Trainium 2 | 2025 | 1:4 | 400G |
| Amazon | Trainium 3 (Teton 3 PDS/ MAX) | 2026 | 1:4 | TBA |
| Meta | MITA-T V1 (Minerva) | 2026 | 1:8~12 | 800G (GSe) |
| Huawei | Cloud Matrix 384 | 2025 | 1:18 | 400G |
| Biren | BR20x | 2026 | TBA | 400G/800G |
400G/ 800G/ 1.6T/ 3.2T refers to the fastest data rate in the network, while there can be lower speed ports in the network
Source: Company data, Compiled by Goldman Sachs Global Investment Research
The Chip-to-optical module attach ratio depends on (1) network structure : a large cluster with 10 thousand GPUs requires three layers of network, while smaller clusters can be supported by two layers. (2) the adoption % of optics in connections. Nvidia's upcoming VR200 racks will come with an attach ratio of 1: 4~6 (1.6T), doubling comparing to GB300's 1: 2~3 (1.6T). The increase of optical module usage is driven by the increase in port rate on the server side (i.e. VR200's compute tray support 1.6T per GPU, while GB300's compute tray is at 800G per GPU), while the data rate at the networking side remains the same at 1.6T. The Optical module attach rate of Google and Amazon ASIC are at similar level of 1:4, as they adopt copper cables and OCS switch in scale-up/ scale-out networking. Meta's ASIC can come with a higher attach ratio of 1:8~ 1:12 (800G), per our industry check, given their complicated DSF (Disaggregated Scheduled Fabric) network design. Similarly, Huawei's Cloud Matrix 384 comes with attach ratio of 1:18 as it adopts all optics connection. Other Chinese GPU/ ASIC other than Huawei usually use a standardized 8 GPU per server design, and 1:4~6 attach ratio (800G) is commonly seen.
Exhibit 39: Ruijie's network solution: optical module attach rate at 1:4/ 1:6 for 2-/ 3-layer network (800G)
 _gs_024.png)
Exhibit from Ruijie
Source: Company data
L-1
S-2
t-2
8-32
x8
S-1
L-t c-32
L1-2
C-32
Rail
Mother Board
Injection module
(Unn laser diode sourcel
( 11 ) Optical Circuit Switch: Achieving all optics networking Dichroic splitter
ASIC
OSFP
Monitor light
(850nm)|
Signal light
Pluggable
Optical
Dynamic
Exhibit 40: Google's Palomar OCS based on MEMS technology
2D
lens altdy
201
lens dridy
 _gs_025.png)
Source: Company data, Goldman Sachs Global Investment Research
Exhibit 41: Optical modules are not needed on OCS switch, while still needed on server motherboards
 _gs_026.png)
Source: Company data, Compiled by Goldman Sachs Global Investment Research
As more optical connections are adopted in data centers, OCS (Optical Circuit Switch) will be an alternative option. Compared to traditional data center switches, which require optical-electrical-optical transformation, the OCS switch is completely based on optical signals , which allows for higher bandwidth, power efficiency, and scalability. The OCS switch provides an analog light path from the input fiber to the output fiber, and it passes through 800G/ 1.6T/ 3.2T lights in the same way. Therefore, by deploying the OCS switch, AI data center does not need to replace their switch when upgrading to faster data rate, and a same OCS switch can support different data rate at the same time. This feature makes it an optimal solution for today's fast changing AI clusters. Key progress: (1) Innolight targets SiPh OCS in 2027, (2) Robotechnik announced that they obtained an order of ' fully automated OCS (Optical Switch) packaging line' from its Europe customer. The order is worth EUR 7.7m, (3) Lumentum's optical circuit switches (OCS) backlog reached beyond $400 million in Feb 2026, (4) Coherent: per management, the company has engaged with over 10 customers on OCS as of Feb 2026. Shipments and backlog include 64x64 systems and 320x320 systems, and expected OCS revenue to grow sequentially in the coming quarters after Feb 2026.
Exhibit 42: The development timeline of OCS
| Development timeline of OCS | Development timeline of OCS |
|---|---|
| 2015 | Google launched the Apollo OCS project to develop the OCS switch |
| 2023 | Google announced TPU v4 is the first supercomputer to deploy OCS, with 4096 chips interconnected by 48 internally-developed OCS |
| 2025 | Google's TPU v7 SuperPod adopts OCS to interconnect 9,216 chips OCP announced new OCS project, with participants including Lumenutm, iPronics, Google, Nvidia, Coherent, Microsoft , etc. |
| 2026 | Lumentum announced that it's scaling rapidly to meet extraordinary customer demand that has already driven our backlog well beyond $400 million Coherent expects OCS revenue to ramp through this year and next, with over 10 customers engagements now |
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
2D MEMS array
20 MEMs arrayi
OCS
Exhibit 43: OCS: major technology options
| MEMS | LC/LCoS | Piezo/ DLBS | SiPh | |
|---|---|---|---|---|
| What does it mean | MEMS Mirror or Micro-electromechanical systems | Liquid crystal | Direct Light Beam-Steering | Silicon Photonics |
| How does it work | Control the deflection of the mirrors by applying voltage, thereby changing the propagation path of the light. | Beam deflection utilizing the electro-optic effect of liquid crystals and the cascading of crystal wedges. | Two collimator arrays are placed face to face to form a switch matrix. Electric field drive the the collimators to shift and tilt, aligning corresponding ports and achieve optical switching. | Constructing defined optical path matrix on a silicon-based chip, allowing optical signals to be transmitted along a predetermined path. |
| Advantages and Disadvantages | Matured solution with mass production track record | Pros: Higher reliability as there's no mechanical moving parts, low voltage and low power consumption | Pros: More reliable and Lower loss than MEMS Cons: hard to support very large port numbers | Pros: Fast switching time. Potential low cost for mass production Cons: High insertion loss |
Progress
Switching time
Reliability
Drive voltage
Insertion Loss
Crosstalk
Mass production, now the major solutoin
Medium
(<100ms)
Low
~ 100V
Low
(~ 3dB)
Low switching time is slow
Mass production
Slow
(>100ms)
High
<= 10V
Low
(~ 4dB)
Low
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
Exhibit 44: ASP of OCS vs. traditional switch
 _gs_027.png)
As of Apr 2026
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
Exhibit 45: OCS switch by Accelink
 _gs_028.png)
Source: Company data
In qualification
Medium
(<100ms)
High
~ 10V
Low
(~ 2.5dB)
High
In qualification
Low
(nanosecond)
High
Low
High
(~ 6dB)
High
Cons:
( 12 ) Technology adoption: pace varies on infrastructure
Exhibit 46: Technology adoption by CSP
| Nvidia | Microsoft | Meta | Amazon | Oracle | Chinese CSP | ||
|---|---|---|---|---|---|---|---|
| Optical module | |||||||
| 800G | √ | √ | √ | √ | √ | √ | √ |
| 1.6T | √ | √ | O (2027) | O (2026) | O (2027) | O (2027) | O |
| 3.2T | O (2028) | O (2028) | |||||
| CPO | |||||||
| Scale up | O | O | |||||
| Scale out | O | O | O | ||||
| OCS | |||||||
| Scale up or scale out | √ |
Note:
√: adopted
Source: Company data, Data compiled by Goldman Sachs Global Investment Research
While the networking technology migration is a certain trend, the pace of adoption varies depending on the clients' data center infrastructure readiness and their specific design needs. Key factors that affect technology migrations include: (1) Depreciation and utilisation of older facilities. If current generation capacities have not yet been fully depericiated, the financial pressure of moving quickly to the next generation can be significant. (2) New infrastructure readiness. Sometimes the technology migration is slowed by the construction time for new buildings, power grids and heat dissipation infrastructure. (3) Waiting for cost reduction . When a new technology is put into use, the cost can be substantially higher than in the later stage where it enters mass production. (4) Uncertain technology direction. When the technologies are still developing and there are multiple potential directions, heavy early investment could pose a risk in an evolving technology landscape.
O: See progress in adoption
Scale up
GPU
Scale out
GPU
GPU
GPU
Scale across
GPU
Appendix: Scale up vs. S c ale out vs. Scale across
GPU
GPU
Adding more resources
To the same equipment
Adding new equipment
GPU
GPU
GPU
GPU
GPU
GPU
GPU
GPU
Exhibit 47: Illustration of scale up, scale out and scale across connections
 _gs_029.png)
Source: Company data, Goldman Sachs Global Investment Research
Given the increasing need for high speed AI computing, the industry has been trying to connect more GPUs to expand the size of AI clusters. There are three major ways of expansion: (1) Scale up : adding more GPUs and computing resources within the same piece of equipment, typically within the same server rack (e.g. Nvidia's Vera Rubin rack connects 72 GPUs in a rack); nowadays there are scale-up expansions that connect across racks, or the so-called supernodes, where the networking speed across racks are optimized to close to the connections within the same rack; (2) Scale out: adding more equipment and connecting them through switching technologies, a traditional way of network expansion. Nowadays, AI clusters support scale out connections of 100k+ GPUs; (3) Scale across: connecting servers across data centers in different locations; Nvidia introduced their solutions for scale-across networking via in-house Ethernet switch and NIC (network interface controller) products.
DAC
ACC
Copper Cable
Glass core
Speed of limit = 204,190 km/sec
Copper Cable
Appendix: Illustration of di ff erent connection products Redriver
Exhibit 50: DAC/ ACC/ AEC/ ADC: widely used in scale up and scale out connections Pluggable
Glass core fiber vs. Hollow core fiber
Pluggable
Pluggable
AEC
Pluggable
Hollow core fiber
Copper (DAC/ AAC/ AEC) vs. Optical cables (ADC)
ADC
DSP
Pluggable
Optical Cable
DSP
 _gs_030.png)
Micron
Source: Company data
Exhibit 51: Hollow core fiber: potentially for scale-across connections
Hollow core fiber support extra low latency comparing to traditional fiber
 _gs_031.png)
Source: Company data
Pluggable