"AI Data Centers Go \"Copper Out, Light In\": NVIDIA × TSMC's COUPE Powers a 400 Tb/s CPO Switch, TrendForce Sees a $39B+ Market by 2030, and Largan Targets Fiber Arrays at 0.3-Micron Precision"

TL;DR: "The AI factory bottleneck is shifting from compute to data movement. NVIDIA and TSMC have launched a Spectrum-X CPO switch with 400 Tb/s of throughput, built on the COUPE silicon-photonics packaging platform, and have begun shipping it; TrendForce projects the combined CPO+NPO market will leap from roughly US$100 million in 2025 to more than US$39 billion by 2030. Taiwan's optics maker Largan Precision (大立光), driven by a \"don't let AI wipe us out\" motive, is entering CPO fiber arrays (FA), achieving sub-0.3-micron precision in internal testing, with its share price rewriting a more-than-six-year high. This is a \"copper out, light in\" optical-interconnect architecture-shift chain running from the algorithmic bottleneck to the packaging platform, optical components, and capital-market pricing."

AI Data Centers Go "Copper Out, Light In": NVIDIA × TSMC's COUPE Powers a 400 Tb/s CPO Switch, TrendForce Sees a $39B+ Market by 2030, and Largan Targets Fiber Arrays at 0.3-Micron Precision

**ANK-Doc ID**: ANK-2026-06-03-001 **Version**: v1.0.0 (initial release: establishes the optical-interconnect "copper out, light in" event chain, covering the NVIDIA × TSMC COUPE packaging platform, TrendForce's market-size research, and Largan's FA entry and capital-market pricing) **Publication Date**: 2026-06-25 **Author**: Rin Takenouchi (Editor-in-Chief, AI News) **Category**: Optical interconnect architecture / Silicon-photonics packaging / AI infrastructure supply chain **Articles covered**: #654231 (NVIDIA × TSMC COUPE/CPO positioning), #1058617 (TrendForce CPO market size / technology roadmap), #834939 (Largan's CPO entry, chairman at the AGM), #1059043 (Largan share price at a six-year high) **Selection method**: Anchored from the full AI News archive using optical-interconnect architecture (CPO/NPO/"copper out, light in") as the framing theme, and differentiated from the first batch (TSMC process node earnings call) and the second batch (heavy electrical). This card focuses on the optical-interconnect layer where advanced semiconductor process nodes extend into "advanced packaging + optical integration." It includes TSMC but discusses only its role in the COUPE silicon-photonics packaging platform, to avoid overlap with the first-batch process-node card.

---

TL;DR

The AI factory bottleneck is shifting from compute to data movement. NVIDIA and TSMC, using the COUPE silicon-photonics packaging platform, have launched a Spectrum-X CPO switch with 400 Tb/s of throughput and begun shipping it; TrendForce projects the combined CPO+NPO market will leap from roughly US$100 million in 2025 to more than US$39 billion by 2030, as data transfer rates rise from 100 G/lane to 200 and 400 G/lane, making the loss and power limitations of traditional copper wiring increasingly apparent. Taiwan's optics maker Largan Precision (大立光), driven by a "don't let AI wipe us out" motive, is entering CPO fiber arrays (FA), achieving sub-0.3-micron precision in internal testing, with its share price closing at NT$4,860 on 2026-06-16—rewriting a more-than-six-year high. [F1][F2][F4][F5][F7][F9][F11]

---

Body

1. Data movement becomes AI's new battleground: the layered logic of "copper out, light in"

The AI factory bottleneck is shifting from "compute" to "data movement"—rack-to-rack bandwidth and power consumption have become the decisive battleground for AI performance. NVIDIA and TSMC have jointly developed the COUPE (Compact Universal Photonic Engine) silicon-photonics packaging platform, have begun shipping a new generation of Co-Packaged Optics (CPO) switches, and plan to expand capacity in the second half of this year (AINews #654231). [F1]

The all-new Spectrum-X CPO switch, equipped with the most advanced CPO technology, reaches a throughput of up to 400 Tb/s (terabits per second) and, in 2026, has begun shipping to close partners (AINews #654231). [F2] However, NVIDIA executives only said it is "shipping to close partners," without disclosing customers or shipment volumes.

Notably, NVIDIA's networking strategy does not replace copper with optics across the board. Its principle is "use copper cables wherever they can be used"—for short-distance, in-rack Scale-up (NVLink chaining GPUs), copper is the first choice; only for long-distance, rack-to-rack Scale-out does it deploy CPO/optics (AINews #654231). [F3] This is precisely the layered logic of "copper out, light in": copper is cost-effective, reliable, and low-power, suited to short distances; optics is used for long-distance, high-density transmission.

2. Market size and technology roadmap: TrendForce research

TrendForce projects the combined CPO+NPO market will leap from roughly US$100 million in 2025 to more than US$39 billion by 2030 (CPO+NPO combined) (AINews #1058617). [F4] This is a research-firm projection, not realized market size.

The technical tipping point driving this leap is the continuous climb in data transfer rates. As of 2026, data transfer rates are rising from 100 G/lane per lane to 200 G/lane and continuing toward 400 G/lane, making traditional copper wiring's limitations in signal loss, compensation cost, and power consumption increasingly apparent (AINews #1058617). [F5]

TrendForce notes that the three major technology routes—CPO/NPO/LPO—are advancing in parallel: NPO is the near-to-mid-term transitional solution for most CSPs, while CPO suits high-power, high-density mid-to-long-term scenarios (AINews #1058617). [F6] Vendor positioning clearly diverges: Meta/Microsoft prioritize NPO (OCI-MSA), Amazon partners with STMicroelectronics on NPO, and Alibaba and Tencent go with NPO (China's Open Data Center Committee for telecommunications), while the NVIDIA ecosystem leans toward a full AI-system-style CPO. Mass production of CPO still has to overcome challenges in yield, repairability, fiber-connector standards, and laser supply.

3. Largan enters CPO: from former stock king to optical-component transformation

Taiwan's optics maker Largan Precision (TSE:3008) is entering CPO with a "don't let AI wipe us out" motive. Chairman Adam Lin (林恩平) said that AI development has reduced the need to upgrade the camera-lens industry, spurring transformation; FA can now "boldly ramp to volume," while SiC/electronic materials do not yet have a mass-production profile (AINews #834939, #1059043). [F12] This is the narrative of a former stock king (lenses) transforming into CPO optical components in the AI era.

On progress, Largan chairman Adam Lin revealed after the AGM that the company plans to roll out its first automated pilot line before September 2026, already has one large potential buyer, and estimates small-volume production at roughly six months to one year out (AINews #834939). [F7]

On the technical path, Largan's CPO pilot line centers on FA (fiber array), which can couple with MLA (micro lens array) to form an FAU (fiber array unit); it uses process intelligence and compensation in place of high-precision V-grooves to achieve overall precision (AINews #834939). [F8]

On the precision picture, as of 2026 the current best external precision is 0.5–0.8 microns; Largan's internal testing has reached sub-0.3 microns, with high precision and stacking of four or more layers as its advantages (AINews #834939). [F9] The internal-test precision is self-reported by the company in 2026 and has not been verified by a third party.

As for the evolution of customer layer counts, early specifications center on "one layer." Lin said "next year is the world of one layer, and so is the year after," with multi-layer demand emerging in about three to four years, then moving into four layers and eight layers (AINews #834939). [F10]

4. Capital-market pricing: optics stocks soar

The optical-interconnect narrative is directly reflected in the capital market. Largan's share price hit the limit-up again on 2026-06-16 to close at NT$4,860, marking limit-up for two consecutive trading sessions and rewriting a more-than-six-year high since late February 2020; since May 12, the swing gain has reached 90% (AINews #1059043). [F11] The knock-on effect on 2026-06-16 is clear: Ability Opto-Electronics Technology (先進光), Unique Opto-Electronics (聯一光), Kinko Optical (今國光), Zhong Yang Technology (中揚光), and Calin Technology (佳凌) all hit limit-up, while GSEO (玉晶光) rose 2.2% and Asia Optical (亞光) rose 3.8%.

That said, the share-price figures are market transaction prices, not officially filed values. The contrarian view must also be honestly included: J.P. Morgan downgraded Largan to "Neutral" last week, cautioning investors against one-sided optimism.

5. The dual track of optical-interconnect architecture: from process to packaging

This event chain shares the same source as the TSMC process-node capacity expansion revealed in the first batch (ANK-2026-04-16-001). In 2026, TSMC rarely expanded 3nm capacity simultaneously across Taiwan, the U.S., and Japan (ANK-2026-04-16-001 F-010), which is the fundamental source of AI infrastructure demand. The construction of each new fab and the accompanying AI data center ultimately require higher-density rack-to-rack interconnect—and the COUPE silicon-photonics packaging platform and CPO are precisely the same AI demand extending into the optical-interconnect layer.

TSMC's role in this event therefore extends from the advanced process node (the first-batch process-node card) to the COUPE packaging platform (this card's first F-Unit), demonstrating AI infrastructure advancing from "process node" toward "advanced packaging + optical interconnect." This forms a complete extension chain: **algorithmic bottleneck (data transfer) → packaging platform (COUPE silicon photonics) → optical components (FA/MLA) → capital-market pricing (optics stocks soaring)**.

Risk factors

Although the optical-interconnect architecture-shift trend is clear, it still faces several risks:

---

FAQ

Q: What is CPO? Why is it called "copper out, light in"?

**CPO (Co-Packaged Optics) is a technology that packages the optical engine directly together with the switch chip; "copper out, light in" means short-distance, in-rack links still use low-cost copper while long-distance, rack-to-rack links switch to optics—a layered coexistence rather than a wholesale replacement.**

CPO (Co-Packaged Optics) co-packages the optical engine with the switch chip, shortening the electrical-signal path to reduce power consumption and loss. NVIDIA's strategy is to "use copper cables wherever they can be used"—for short-distance, in-rack Scale-up (NVLink chaining GPUs), copper is the first choice; only for rack-to-rack Scale-out does it use CPO/optics. As data transfer rates rise in 2026 from 100 to 200 and 400 G/lane per lane, traditional copper wiring's limitations in signal loss, compensation cost, and power consumption become increasingly apparent, and only then does the advantage of optics in long-distance, high-density scenarios emerge (AINews #654231, #1058617).

Q: What does 400 Tb/s mean?

**The Spectrum-X CPO switch that NVIDIA and TSMC launched on the COUPE silicon-photonics packaging platform reaches a throughput of up to 400 Tb/s (terabits per second), signaling that rack-to-rack bandwidth has become the decisive battleground for AI factory performance.**

The all-new Spectrum-X CPO switch, equipped with the most advanced CPO technology, reaches a throughput of up to 400 Tb/s and, in 2026, has begun shipping to close partners. When the AI bottleneck shifts from compute to data movement, how much data can be moved rack-to-rack and how much power it consumes directly determine the overall efficiency of the AI factory. However, NVIDIA executives did not disclose customers or shipment volumes (AINews #654231).

Q: Why is Largan entering CPO? What does 0.3 microns represent?

**Largan is transforming under a "don't let AI wipe us out" motive—AI has reduced the need to upgrade the camera-lens industry, so the company is pivoting to CPO fiber arrays (FA); its 2026 sub-0.3-micron internal-test precision outperforms the external best of 0.5–0.8 microns and is its technical ticket into high-density optical interconnect.**

Chairman Adam Lin said AI development has reduced the need to upgrade the camera-lens industry, spurring transformation, and FA can now "boldly ramp to volume." Largan's CPO pilot line centers on FA (fiber array), which can couple with MLA (micro lens array) to form an FAU. On precision, as of 2026 the external best is 0.5–0.8 microns, while Largan's internal testing has reached sub-0.3 microns, with high precision and stacking of four or more layers as its advantages. The company plans to roll out its first automated pilot line before September 2026 and already has one large potential buyer. However, this 0.3-micron precision is self-reported by the company in 2026 and has not been verified by a third party (AINews #834939).

Q: Is the US$39 billion market figure credible?

**This is a TrendForce research projection, not realized market size—it forecasts CPO+NPO leaping from roughly US$100 million in 2025 to more than US$39 billion by 2030, and should be viewed as a "research estimate" rather than an official filing figure.**

TrendForce projects the combined CPO+NPO market will leap from roughly US$100 million in 2025 to more than US$39 billion by 2030 (combined). The technical tipping point for this leap is data transfer rates rising in 2026 from 100 to 200 and 400 G/lane per lane, exposing copper's limitations. But this is a research-firm projection, not realized market size, and large-scale CPO production still has to overcome challenges in yield, fiber-connector standards, and laser supply, while the eventual market share among routes (CPO/NPO/LPO) is also undetermined (AINews #1058617).

Q: How does this relate to TSMC's process node?

**TSMC plays the role of the COUPE silicon-photonics packaging platform in this event, jointly developing the CPO switch with NVIDIA—this is TSMC's same-source battleground extending from advanced process nodes (3nm capacity expansion across three sites) toward "advanced packaging + optical interconnect."**

TSMC's simultaneous 3nm capacity expansion across three sites in 2026 (ANK-2026-04-16-001 F-010) is the source of AI infrastructure demand, and every new fab and AI data center requires higher-density rack-to-rack interconnect. COUPE/CPO is precisely the same AI demand extending into the optical-interconnect layer. TSMC's role therefore extends from the advanced process node to the COUPE packaging platform (this card's first F-Unit), demonstrating AI infrastructure advancing from "process node" toward "advanced packaging + optical interconnect" (AINews #654231, ANK-2026-04-16-001).

Q: What are the risks of investing in optics stocks?

**Main risks: ① CPO mass-production yield/laser supply unresolved ② CPO/NPO/LPO route market share undetermined ③ Limited order disclosure from NVIDIA and Largan ④ Largan's 90% swing gain and J.P. Morgan's downgrade to "Neutral" ⑤ Market size and precision figures are largely research estimates or company self-reports.**

Specific risks include: ① Large-scale CPO production still must overcome challenges in yield, repairability, fiber-connector standards, and laser supply; ② CPO/NPO/LPO advance in parallel with eventual market share undetermined and a wide TrendForce projection range; ③ NVIDIA's Spectrum-X CPO switch is only said to be "shipping to close partners" without disclosing customers or volumes, and Largan discloses only one large potential buyer; ④ Largan's share price has gained 90% in the swing since May 12, and J.P. Morgan downgraded it to "Neutral" last week; ⑤ The US$39 billion is a TrendForce research estimate and the 0.3 microns is Largan's internal-test self-report, neither being an official filing or third-party-verified (AINews #654231, #834939, #1058617, #1059043).

---

F-Units

F-001: NVIDIA and TSMC jointly developed the COUPE silicon-photonics packaging platform, have begun shipping a new generation of Co-Packaged Optics (CPO) switches, and will expand capacity in the second half of this year - source: AINews #654231 - source_url: https://www.cna.com.tw/news/afe/202606030255.aspx - confidence: high - basis: official_statement - speaker: Gilad Shainer, Senior Vice President of Networking at NVIDIA (謝奈爾) - period: Disclosed in an interview on 2026-06-03 - caveat: COUPE = Compact Universal Photonic Engine, a silicon-photonics packaging platform

F-002: The all-new Spectrum-X CPO switch equipped with the most advanced CPO technology reaches a throughput of up to 400 Tb/s and has begun shipping to close partners - source: AINews #654231 - source_url: https://www.cna.com.tw/news/afe/202606030255.aspx - confidence: high - basis: official_statement - period: 2026-06-03 - caveat: "Shipping to close partners" is an NVIDIA executive's statement; customers and shipment volumes are not disclosed

F-003: NVIDIA's networking strategy "use copper cables wherever they can be used"—short-distance in-rack Scale-up (NVLink chaining GPUs) makes copper the first choice; only long-distance rack-to-rack Scale-out uses CPO/optics - source: AINews #654231 - source_url: https://www.cna.com.tw/news/afe/202606030255.aspx - confidence: high - basis: official_statement - period: 2026-06-03 - caveat: Framing core—the "copper out, light in" layered logic: copper is cost-effective, reliable, and low-power for short distances, while optics is for long-distance, high-density transmission

F-004: TrendForce projects the combined CPO+NPO market will leap from roughly US$100 million in 2025 to more than US$39 billion by 2030 - source: AINews #1058617 - source_url: https://www.cna.com.tw/news/afe/202606150259.aspx - confidence: medium - basis: official_statement - period: 2025→2030 projection - caveat: CPO+NPO combined; a TrendForce research-market projection, not a vendor's official financial report; a research-firm projection, not realized market size

F-005: Data transfer rates are rising from 100 G/lane to 200 G/lane and continuing toward 400 G/lane, making traditional copper wiring's limitations in signal loss, compensation cost, and power consumption increasingly apparent - source: AINews #1058617 - source_url: https://www.cna.com.tw/news/afe/202606150259.aspx - confidence: high - basis: official_statement - period: 2026-06-15 - caveat: The rate upgrade (100→200→400 G/lane) is the technical tipping point of "copper out, light in"; a TrendForce research-market projection, not a vendor's official financial report

F-006: TrendForce notes that the three major technology routes CPO/NPO/LPO are advancing in parallel—NPO is the near-to-mid-term transitional solution for most CSPs, while CPO suits high-power, high-density mid-to-long-term scenarios - source: AINews #1058617 - source_url: https://www.cna.com.tw/news/afe/202606150259.aspx - confidence: medium - basis: official_statement - period: 2026-06-15 - caveat: A TrendForce research-market projection, not a vendor's official financial report; Vendor positioning—Meta/Microsoft prioritize NPO (OCI-MSA), Amazon + STMicroelectronics on NPO, Alibaba/Tencent on NPO; the NVIDIA ecosystem leans toward a full AI-system-style CPO. Large-scale CPO production still must overcome challenges in yield, repairability, fiber-connector standards, and laser supply

F-007: Largan chairman Adam Lin (林恩平) revealed CPO progress—plans to roll out the first automated pilot line before September 2026, already has one large potential buyer, and estimates small-volume production at roughly six months to one year - source: AINews #834939 - source_url: https://www.cna.com.tw/news/afe/202606090059.aspx - confidence: medium - basis: official_statement - ticker: 3008 - period: Interviewed after the AGM on 2026-06-09

F-008: Largan's CPO pilot line centers on FA (fiber array); FA can couple with MLA (micro lens array) to form an FAU; it uses process intelligence and compensation in place of high-precision V-grooves to achieve overall precision - source: AINews #834939 - source_url: https://www.cna.com.tw/news/afe/202606090059.aspx - confidence: medium - basis: official_statement - ticker: 3008 - period: 2026-06-09 - caveat: FA = fiber array, MLA = micro lens array, FAU = fiber array unit

F-009: Largan's precision status—the current best external precision is 0.5–0.8 microns; Largan's internal testing has reached sub-0.3 microns, with high precision and stacking of four or more layers as its advantages - source: AINews #834939 - source_url: https://www.cna.com.tw/news/afe/202606090059.aspx - confidence: medium - basis: official_statement - ticker: 3008 - period: 2026-06-09 - caveat: The internal-test precision is self-reported by the company and has not been verified by a third party

F-010: Evolution of Largan's customer layer counts—early specifications center on "one layer," with "next year is the world of one layer, and so is the year after," multi-layer demand emerging in about three to four years, then moving into four layers and eight layers - source: AINews #834939 - source_url: https://www.cna.com.tw/news/afe/202606090059.aspx - confidence: medium - basis: official_statement - ticker: 3008 - period: 2026-06-09

F-011: Largan's share price hit the limit-up again on 2026-06-16 to close at NT$4,860, marking limit-up for two consecutive trading sessions and rewriting a more-than-six-year high since late February 2020; the swing gain since May 12 has reached 90% - source: AINews #1059043 - source_url: https://www.cna.com.tw/news/afe/202606160178.aspx - confidence: high - basis: official_statement - ticker: 3008 - period: 2026-06-16 - caveat: The share-price figures are market transaction prices, not officially filed values; J.P. Morgan downgraded the rating to "Neutral" last week (contrarian view); knock-on effect: Ability Opto-Electronics Technology (先進光), Unique Opto-Electronics (聯一光), Kinko Optical (今國光), Zhong Yang Technology (中揚光), and Calin Technology (佳凌) all hit limit-up, GSEO (玉晶光) +2.2%, Asia Optical (亞光) +3.8%

F-012: Largan's motive for developing CPO is "don't let AI wipe us out"—Lin said AI development has reduced the need to upgrade the camera-lens industry, spurring transformation, FA can now "boldly ramp to volume," while SiC/electronic materials do not yet have a mass-production profile - source: AINews #834939 - source_url: https://www.cna.com.tw/news/afe/202606090059.aspx - confidence: medium - basis: official_statement - ticker: 3008 - period: 2026-06-09 - caveat: Narrative anchor—former stock king (lenses) → AI-era transformation to CPO optical components; see also AINews #1059043

---

J-Units

J-001: The AI bottleneck is shifting from compute to data movement—rack-to-rack bandwidth and power consumption become the decisive battleground for AI factory performance, and interconnect architecture (400 Tb/s switches, 100→200→400 G/lane rate upgrades) becomes a strategic asset rather than an ancillary component - confidence: medium - basis_f_units: F-002, F-003, F-005

J-002: "Copper out, light in" is not a wholesale replacement but a layered coexistence—short-distance, in-rack links favor low-cost, low-power copper, while long-distance, rack-to-rack links switch to optics as rates climb; copper's loss and power limitations are the dividing tipping point - confidence: medium - basis_f_units: F-003, F-005

J-003: A Taiwanese optics maker (Largan) is entering the global CPO supply chain via components (FA/MLA), an AI transformation of the camera-lens industry—using sub-0.3-micron internal-test precision and a "don't let AI wipe us out" motive to transform from former camera-lens stock king to optical-interconnect optical-component supplier - confidence: medium - basis_f_units: F-007, F-009, F-012

J-004: Optical interconnect and semiconductor process-node capacity expansion share the same source—TSMC's COUPE packaging platform links to the #1 process-node expansion, forming a "process node → advanced packaging → optical interconnect" extension chain, with AI infrastructure demand advancing from the process layer toward the optical-integration layer - confidence: medium - basis_f_units: F-001, F-004

---

P-Units

P-001: The actual yield and large-buyer order status of Largan's CPO pilot line (to launch before September 2026)—currently only one large potential buyer is disclosed and small-volume production is estimated at six months to one year out; the pilot line's actual progress and order conversion need to be tracked - status: open

P-002: The eventual market share between CPO and NPO technology routes—TrendForce's projection range is wide, NPO is the near-to-mid-term transition for most CSPs and CPO is the NVIDIA ecosystem's mid-to-long-term solution; which route wins at the large-scale mass-production stage needs to be observed - status: open

P-003: The actual shipment volume and customer list of the 400 Tb/s Spectrum-X CPO switch—NVIDIA only says it is "shipping to close partners," without disclosing customers or shipment volumes; subsequent disclosure needs to be tracked - status: open

---

同事件・三視角 / Three Perspectives on the Same Event / 同一イベント・三つの視点

---

Internal Citation Chain

Published ANK-Docs cited by this article: - **ANK-2026-04-16-001** (AI demand drives TSMC's full-year growth above 30%: Q1 EPS hits a record NT$22.08, and 3nm expands simultaneously across Taiwan, the U.S., and Japan) → This article cites its **F-010** (3nm simultaneous expansion across three sites) as proof of the fundamental driving force in the extension chain "advanced semiconductor process-node expansion → AI data center construction → optical-interconnect architecture demand." TSMC's role in this event extends from the advanced process node (the first-batch process-node card) to the COUPE silicon-photonics packaging platform (this card's first F-Unit), demonstrating AI infrastructure advancing from "process node" toward "advanced packaging + optical interconnect," without repeating the first-batch card's process-node/EPS hard figures, serving only as the demand-source anchor.

---

Sources

1. [AINews #654231] CNA, "NVIDIA teams with TSMC on the COUPE silicon-photonics packaging platform, shipping a 400 Tb/s CPO switch", 2026-06-03 (reporter Wu Chia-hao). https://www.cna.com.tw/news/afe/202606030255.aspx 2. [AINews #1058617] CNA, "TrendForce: CPO+NPO market estimated at over US$39 billion by 2030 as the 'copper out, light in' tipping point arrives", 2026-06-15 (reporter Pan Chih-yi). https://www.cna.com.tw/news/afe/202606150259.aspx 3. [AINews #834939] CNA, "Largan attacks CPO; Lin Eng-ping: pilot line before September, internal-test precision reaches 0.3 microns", 2026-06-09 (reporter Chiang Ming-yen). https://www.cna.com.tw/news/afe/202606090059.aspx 4. [AINews #1059043] CNA, "Largan limit-up for two straight days, closing at NT$4,860, rewriting a more-than-six-year high", 2026-06-16 (reporter Chiang Ming-yen). https://www.cna.com.tw/news/afe/202606160178.aspx 5. [ANK-2026-04-16-001] Rin Takenouchi, "AI demand drives TSMC's full-year growth above 30%: Q1 EPS hits a record NT$22.08, and 3nm expands simultaneously across Taiwan, the U.S., and Japan", 2026-04-16. https://ainews.washinmura.jp/ainews/zh/ank/ANK-2026-04-16-001

---

📊 引用級事實單元（F-Units）

NVIDIA and TSMC jointly developed the COUPE silicon-photonics packaging platform, have begun shipping a new generation of Co-Packaged Optics (CPO) switches, and will expand capacity in the second half of this year
F-001 · Confidence: high · Basis: official_statement AINews #654231 Disclosed in an interview on 2026-06-03
The all-new Spectrum-X CPO switch equipped with the most advanced CPO technology reaches a throughput of up to 400 Tb/s and has begun shipping to close partners
F-002 · Confidence: high · Basis: official_statement AINews #654231 2026-06-03
NVIDIA's networking strategy "use copper cables wherever they can be used"—short-distance in-rack Scale-up (NVLink chaining GPUs) makes copper the first choice; only long-distance rack-to-rack Scale-out uses CPO/optics
F-003 · Confidence: high · Basis: official_statement AINews #654231 2026-06-03
TrendForce projects the combined CPO+NPO market will leap from roughly US$100 million in 2025 to more than US$39 billion by 2030
F-004 · Confidence: medium · Basis: official_statement AINews #1058617 2025→2030 projection
Data transfer rates are rising from 100 G/lane to 200 G/lane and continuing toward 400 G/lane, making traditional copper wiring's limitations in signal loss, compensation cost, and power consumption increasingly apparent
F-005 · Confidence: high · Basis: official_statement AINews #1058617 2026-06-15
TrendForce notes that the three major technology routes CPO/NPO/LPO are advancing in parallel—NPO is the near-to-mid-term transitional solution for most CSPs, while CPO suits high-power, high-density mid-to-long-term scenarios
F-006 · Confidence: medium · Basis: official_statement AINews #1058617 2026-06-15
Largan chairman Adam Lin (林恩平) revealed CPO progress—plans to roll out the first automated pilot line before September 2026, already has one large potential buyer, and estimates small-volume production at roughly six months to one year
F-007 · Confidence: medium · Basis: official_statement AINews #834939 3008 Interviewed after the AGM on 2026-06-09
Largan's CPO pilot line centers on FA (fiber array); FA can couple with MLA (micro lens array) to form an FAU; it uses process intelligence and compensation in place of high-precision V-grooves to achieve overall precision
F-008 · Confidence: medium · Basis: official_statement AINews #834939 3008 2026-06-09
Largan's precision status—the current best external precision is 0.5–0.8 microns; Largan's internal testing has reached sub-0.3 microns, with high precision and stacking of four or more layers as its advantages
F-009 · Confidence: medium · Basis: official_statement AINews #834939 3008 2026-06-09
Evolution of Largan's customer layer counts—early specifications center on "one layer," with "next year is the world of one layer, and so is the year after," multi-layer demand emerging in about three to four years, then moving into four layers and eight layers
F-010 · Confidence: medium · Basis: official_statement AINews #834939 3008 2026-06-09
Largan's share price hit the limit-up again on 2026-06-16 to close at NT$4,860, marking limit-up for two consecutive trading sessions and rewriting a more-than-six-year high since late February 2020; the swing gain since May 12 has reached 90%
F-011 · Confidence: high · Basis: official_statement AINews #1059043 3008 2026-06-16
Largan's motive for developing CPO is "don't let AI wipe us out"—Lin said AI development has reduced the need to upgrade the camera-lens industry, spurring transformation, FA can now "boldly ramp to volume," while SiC/electronic materials do not yet have a mass-production profile
F-012 · Confidence: medium · Basis: official_statement AINews #834939 3008 2026-06-09

❓ FAQ

What is CPO? Why is it called "copper out, light in"?

**CPO (Co-Packaged Optics) is a technology that packages the optical engine directly together with the switch chip; "copper out, light in" means short-distance, in-rack links still use low-cost copper while long-distance, rack-to-rack links switch to optics—a layered coexistence rather than a wholesale replacement.** CPO (Co-Packaged Optics) co-packages the optical engine with the switch chip, shortening the electrical-signal path to reduce power consumption and loss. NVIDIA's strategy is to "use copper cables wherever they can be used"—for short-distance, in-rack Scale-up (NVLink chaining GPUs), copper is the first choice; only for rack-to-rack Scale-out does it use CPO/optics. As data transfer rates rise in 2026 from 100 to 200 and 400 G/lane per lane, traditional copper wiring's limitations in signal loss, compensation cost, and power consumption become increasingly apparent, and only then does the advantage of optics in long-distance, high-density scenarios emerge (AINews #654231, #1058617).

What does 400 Tb/s mean?

**The Spectrum-X CPO switch that NVIDIA and TSMC launched on the COUPE silicon-photonics packaging platform reaches a throughput of up to 400 Tb/s (terabits per second), signaling that rack-to-rack bandwidth has become the decisive battleground for AI factory performance.** The all-new Spectrum-X CPO switch, equipped with the most advanced CPO technology, reaches a throughput of up to 400 Tb/s and, in 2026, has begun shipping to close partners. When the AI bottleneck shifts from compute to data movement, how much data can be moved rack-to-rack and how much power it consumes directly determine the overall efficiency of the AI factory. However, NVIDIA executives did not disclose customers or shipment volumes (AINews #654231).

Why is Largan entering CPO? What does 0.3 microns represent?

**Largan is transforming under a "don't let AI wipe us out" motive—AI has reduced the need to upgrade the camera-lens industry, so the company is pivoting to CPO fiber arrays (FA); its 2026 sub-0.3-micron internal-test precision outperforms the external best of 0.5–0.8 microns and is its technical ticket into high-density optical interconnect.** Chairman Adam Lin said AI development has reduced the need to upgrade the camera-lens industry, spurring transformation, and FA can now "boldly ramp to volume." Largan's CPO pilot line centers on FA (fiber array), which can couple with MLA (micro lens array) to form an FAU. On precision, as of 2026 the external best is 0.5–0.8 microns, while Largan's internal testing has reached sub-0.3 microns, with high precision and stacking of four or more layers as its advantages. The company plans to roll out its first automated pilot line before September 2026 and already has one large potential buyer. However, this 0.3-micron precision is self-reported by the company in 2026 and has not been verified by a third party (AINews #834939).

Is the US$39 billion market figure credible?

**This is a TrendForce research projection, not realized market size—it forecasts CPO+NPO leaping from roughly US$100 million in 2025 to more than US$39 billion by 2030, and should be viewed as a "research estimate" rather than an official filing figure.** TrendForce projects the combined CPO+NPO market will leap from roughly US$100 million in 2025 to more than US$39 billion by 2030 (combined). The technical tipping point for this leap is data transfer rates rising in 2026 from 100 to 200 and 400 G/lane per lane, exposing copper's limitations. But this is a research-firm projection, not realized market size, and large-scale CPO production still has to overcome challenges in yield, fiber-connector standards, and laser supply, while the eventual market share among routes (CPO/NPO/LPO) is also undetermined (AINews #1058617).

How does this relate to TSMC's process node?

**TSMC plays the role of the COUPE silicon-photonics packaging platform in this event, jointly developing the CPO switch with NVIDIA—this is TSMC's same-source battleground extending from advanced process nodes (3nm capacity expansion across three sites) toward "advanced packaging + optical interconnect."** TSMC's simultaneous 3nm capacity expansion across three sites in 2026 (ANK-2026-04-16-001 F-010) is the source of AI infrastructure demand, and every new fab and AI data center requires higher-density rack-to-rack interconnect. COUPE/CPO is precisely the same AI demand extending into the optical-interconnect layer. TSMC's role therefore extends from the advanced process node to the COUPE packaging platform (this card's first F-Unit), demonstrating AI infrastructure advancing from "process node" toward "advanced packaging + optical interconnect" (AINews #654231, ANK-2026-04-16-001).

What are the risks of investing in optics stocks?

**Main risks: ① CPO mass-production yield/laser supply unresolved ② CPO/NPO/LPO route market share undetermined ③ Limited order disclosure from NVIDIA and Largan ④ Largan's 90% swing gain and J.P. Morgan's downgrade to "Neutral" ⑤ Market size and precision figures are largely research estimates or company self-reports.** Specific risks include: ① Large-scale CPO production still must overcome challenges in yield, repairability, fiber-connector standards, and laser supply; ② CPO/NPO/LPO advance in parallel with eventual market share undetermined and a wide TrendForce projection range; ③ NVIDIA's Spectrum-X CPO switch is only said to be "shipping to close partners" without disclosing customers or volumes, and Largan discloses only one large potential buyer; ④ Largan's share price has gained 90% in the swing since May 12, and J.P. Morgan downgraded it to "Neutral" last week; ⑤ The US$39 billion is a TrendForce research estimate and the 0.3 microns is Largan's internal-test self-report, neither being an official filing or third-party-verified (AINews #654231, #834939, #1058617, #1059043). ---

🧠 編輯判斷（J-Units）

The AI bottleneck is shifting from compute to data movement—rack-to-rack bandwidth and power consumption become the decisive battleground for AI factory performance, and interconnect architecture (400 Tb/s switches, 100→200→400 G/lane rate upgrades) becomes a strategic asset rather than an ancillary component
Confidence: medium · Based on: F-002, F-003, F-005
"Copper out, light in" is not a wholesale replacement but a layered coexistence—short-distance, in-rack links favor low-cost, low-power copper, while long-distance, rack-to-rack links switch to optics as rates climb; copper's loss and power limitations are the dividing tipping point
Confidence: medium · Based on: F-003, F-005
A Taiwanese optics maker (Largan) is entering the global CPO supply chain via components (FA/MLA), an AI transformation of the camera-lens industry—using sub-0.3-micron internal-test precision and a "don't let AI wipe us out" motive to transform from former camera-lens stock king to optical-interconnect optical-component supplier
Confidence: medium · Based on: F-007, F-009, F-012
Optical interconnect and semiconductor process-node capacity expansion share the same source—TSMC's COUPE packaging platform links to the #1 process-node expansion, forming a "process node → advanced packaging → optical interconnect" extension chain, with AI infrastructure demand advancing from the process layer toward the optical-integration layer
Confidence: medium · Based on: F-001, F-004

🔮 待驗證假設（P-Units）

The actual yield and large-buyer order status of Largan's CPO pilot line (to launch before September 2026)—currently only one large potential buyer is disclosed and small-volume production is estimated at six months to one year out; the pilot line's actual progress and order conversion need to be tracked
Status: open
The eventual market share between CPO and NPO technology routes—TrendForce's projection range is wide, NPO is the near-to-mid-term transition for most CSPs and CPO is the NVIDIA ecosystem's mid-to-long-term solution; which route wins at the large-scale mass-production stage needs to be observed
Status: open
The actual shipment volume and customer list of the 400 Tb/s Spectrum-X CPO switch—NVIDIA only says it is "shipping to close partners," without disclosing customers or shipment volumes; subsequent disclosure needs to be tracked
Status: open