---
title: "AI Hyperconnected Era: Racing to 'Light'?"
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/topics/41423680.md"
description: "Since late 2022, when ChatGPT burst onto the scene, AI has unleashed wave after wave of semiconductor supercycle opportunities. From compute (GPUs) and storage to coordination and scheduling (CPUs), the buildout has already minted a string of $1tn market-cap companies.If there is one remaining pillar of AI infrastructure poised to produce the next $1tn champion, Dolphin Research’s top call is AI-era high-speed interconnects. If compute solves AI’s 'IQ' and storage its 'memory'..."
datetime: "2026-06-03T09:59:59.000Z"
locales:
  - [en](https://longbridge.com/en/topics/41423680.md)
  - [zh-CN](https://longbridge.com/zh-CN/topics/41423680.md)
  - [zh-HK](https://longbridge.com/zh-HK/topics/41423680.md)
author: "[Dolphin Research](https://longbridge.com/en/news/dolphin.md)"
---

# AI Hyperconnected Era: Racing to 'Light'?

Since ChatGPT burst onto the scene in late 2022, AI has unleashed wave after wave of semiconductor super-cycles across compute (GPU), storage, and orchestration (CPU). It has also helped mint multiple trillion-dollar companies.

If there is one AI infrastructure segment still waiting for its own trillion-dollar breakout, our top pick is the AI era's super connectivity. If compute solves AI's IQ and memory solves its recall, then interconnect must rocket short- and long-term memory in and out of the compute core at extreme speed.

Borrowing Jensen Huang's framing, as compute and memory bottlenecks ease while energy remains a decade-long challenge, the next chokepoint is high-speed networking for the AI era. Legacy cloud-era networks cannot meet Agentic AI demands for moving trillions of parameters, MoE models, and sparsely activated workloads.

This note continues our look at the optical-electrical shift behind AI network speed, focusing on CPO. We break down CPO along the following lines:

**1) What is CPO, and can it truly replace traditional copper links?**

**2) Can it fully displace mainstream pluggable optical modules?**

**3) Under this trend, how will competitive dynamics evolve across the value chain?**

**In this piece, we first address foundational questions across the supply chain.**

**Main text:**

**I. What is CPO?**

In a traditional data center, an optical module converts incoming optical signals to electrical signals for the system, or transforms internal electrical signals into optical signals for fiber transmission. It acts as both bridge and translator during data transfer.

Functionally, CPO (Co-Packaged Optics) embeds the roles of a conventional optical module, but differs in two key ways.

**1) Structure**

Pluggable optics resemble an RJ45-style port, whereas CPO integrates the optical engine that performs E/O conversion directly with the chip (primarily the switch ASIC) on the same package substrate or interposer. This is a fundamentally different integration approach.

**2) Use cases**

Pluggables are typically used for inter-rack links (Scale-out). CPO can serve both inter-rack and intra-rack (Scale-up) interconnects, replacing pluggables between racks and copper inside the rack.

**Illustration: Pluggables vs. CPO**

![IMG_256](https://pub.pbkrs.com/uploads/2026/ed92b6c3d9d145a4adcfd2a58d699946?x-oss-process=style/lg)

**Source: GTC 2025, Dolphin Research**

We have seen both $NVIDIA(NVDA.US) and $Broadcom(AVGO.US) aggressively promote their CPO switch solutions. Why the intense focus? As data center compute scales, bandwidth demand is exploding, and clusters are getting much larger. Legacy links become bottlenecks along the way.

**1) Bandwidth limits**

For inter-rack, the switch faceplate area is fixed while pluggable form factors struggle to shrink, capping the number of ports and constraining aggregate bandwidth. Pluggables today top out at 1.6 Tbps per module and roughly 51.2 Tbps per switch, with a potential path to 3.2 Tbps modules and 102.4 Tbps switches, which is near the practical ceiling for pluggables.

**2) Signal integrity limits**

Inside the rack, higher line rates over copper face severe attenuation and distortion over distance, shrinking usable reach. Copper cables can support up to 1.8 TB/s (e.g., NVDA's NVLink copper) within ~2 m, yet per-GPU bandwidth needs are pushing toward 3.6 TB/s.

**3) Thermal and power limits**

Rising speeds drive up link power, while thermal management gets harder. Given US data centers' power constraints, power draw translates directly into cost pressure. CPO can materially ease these issues; NVDA cites ~3.5x better power efficiency.

**II. What are the data transfer scenarios inside a data center?**

We segment the routing choices across scenarios and layers as follows. See below for Scale-out vs. Scale-up examples.

**Illustration: Scale-out and Scale-up**

![IMG_256](https://pub.pbkrs.com/uploads/2026/ea31fbc11118bc4f53223b92614908dc?x-oss-process=style/lg)

**Source: NADDOD, Dolphin Research**

**1) Scale-up: intra-rack interconnect**

This covers intra-rack, especially intra-server links among CPU, GPU, NICs, DDR, and drives. Today these are predominantly copper: PCIe slots and memory slots with PCB traces, SATA and other copper cables. CPO could disrupt these incumbent schemes.

**2) Scale-out: inter-rack interconnect**

This covers links among racks or between servers and switches. These use optics as the medium, primarily fiber plus pluggable modules. CPO is a key trend here and is progressing faster than intra-rack use.

3) Beyond that are data center-to-data center and external links, which are outside the scope of this note.

From hyperscaler roadmaps, CPO is currently aimed at inter-rack first, with intra-rack adoption likely later. This sequencing reflects readiness and ROI.

**III. CPO is still early in rollout. What are the main bottlenecks?**

**1) Advanced packaging maturity**

CPO diverges sharply from traditional pluggable manufacturing. While conventional optoelectronics resemble broader photonics components and modules, CPO mounts the optical engine on the substrate or interposer, relying on advanced packaging such as CoWoS. Unlike typical advanced packaging, CPO must heterogeneously integrate electronic ICs with photonic ICs, requiring hybrid bonding such as $Taiwan Semiconductor(TSM.US) COUPE.

The challenge: these processes are hard, and both NVDA and AVGO depend on TSMC's limited capacity. Upstream constraints may also emerge across optical coupling tools, hybrid bonders, testers, and ABF substrates. Yield for such heterogenous integration remains suboptimal, inflating costs vs. pluggables, and will take time to improve.

**2) Serviceability**

Pluggables are easy to swap. CPO co-packages the opto-electronics with the substrate/interposer and even the chip, making field service and maintenance much harder. Designing for fault tolerance and building operational redundancy can mitigate this.

**3) Thermal management**

High-density co-pack of optical engines with chips raises local hotspots, potentially exceeding laser limits. More capable cooling solutions are needed, adding cost and complexity.

**4) Standardization**

With NVDA, AVGO and others racing out proprietary end-to-end CPO switch stacks, industry-wide interface and packaging standards are still emerging. Absent standards, coordinated R&D, manufacturing, and deployment are harder, slowing commercialization. Time and ecosystem alignment will be required.

Ultimately, CPO must win on total cost at scale. There are also alternative paths, both more advanced and more incremental. How do they relate? We outline the landscape next.

**IV. Comparing technical paths**

**1) CPO**

Co-Packaged Optics integrates the optical engine with a chip on the same substrate, typically a switch ASIC, though in principle it could be a GPU or other compute die. This is the primary definition used here.

**2) NPO**

Near-Packaged Optics stops short of co-packaging on the same substrate or interposer, instead placing optics and ASIC on the same PCB. In China, $Alibaba(BABA.US) and Huawei are advancing NPO as a capacity-aware compromise where advanced packaging is constrained. It could be the near-term mainstream domestically, limiting NVDA penetration in that market.

**Illustration: Integration spectrum from pluggable to NPO to CPO (substrate/interposer) to OIO**

![IMG_256](https://pub.pbkrs.com/uploads/2026/27d09669d1d2e5898d3f4301f782bd86?x-oss-process=style/lg)

**Source: ASE, Dolphin Research**

**3) OIO**

Optical I/O is a step beyond CPO for compute. It co-packages, or even monolithically integrates, the optical engine with the compute die, targeting intra-rack and intra-server links.

**Illustration: Pluggable vs. CPO vs. OIO**

![IMG_256](https://pub.pbkrs.com/uploads/2026/753ea8a1c454743a4dabeeb01ed0c277?x-oss-process=style/lg)

**Source: TSMC, Openlight, Dolphin Research**

**Revisiting data center architecture:**

Servers handle compute with GPUs and CPUs, plus memory and storage. Switches handle networking among servers and to the outside via ASICs that perform packet switching. Storage is typically distributed within servers today.

Within this setup, it is intuitive why CPO starts with switch ASICs. Switches are the interchanges carrying the heaviest bandwidth and port density, and they face the toughest power constraints. That is where CPO delivers the most urgent value.

**4) CPC**

Co-Packaged Copper integrates high-speed copper connectors onto the package substrate. It has clear cost advantages but cannot overcome copper's bandwidth and attenuation limits, restricting scope to some intra-rack links among GPU/CPU nodes, switches, and storage. NVDA still uses copper intra-rack today, but is likely to migrate to optical.

**5) LPO**

Linear-Drive Pluggable Optics is a slimmed pluggable that removes the DSP/CDR, retaining and strengthening analog Driver and TIA for direct drive. In short, it drops the power-hungry DSP and error correction, relying on beefed-up analog front-ends and letting the switch ASIC drive the laser directly.

**Illustration: Traditional vs. LPO**

![IMG_256](https://pub.pbkrs.com/uploads/2026/ab9c8a43eacf93e7018c6e4c53988daa?x-oss-process=style/lg)

![IMG_256](https://pub.pbkrs.com/uploads/2026/dce633b2d7958512ad35d7acf4b71cde?x-oss-process=style/lg)

**Source: Bryon Moyer, Semiconductor Engineering, Dolphin Research**

The trade-off: PCB traces still introduce loss while signal quality requirements rise, limiting reach. At higher rates (1.6T and beyond), SI challenges become acute. Simpler structure comes with performance sacrifice.

Bottom line: NPO, CPC, and LPO are pragmatic interim paths. But as speeds rise and clusters scale, these compromises hit ceilings. CPO is the next-gen architecture the industry must cross.

**6) What about OCS? Does it threaten CPO?**

Optical Circuit Switches avoid O/E/O conversion entirely by establishing physical light paths using optical switch matrices. Think of arrays of micromirrors redirecting beams under control signals. At first glance, this bypasses the need for CPO in switches, but reality is more nuanced.

**How are switches deployed in the data center?**

**(1) On the motherboard:** GPUs perform core compute, then pass data to CPUs, which forward to NICs (with ASICs), or GPUs can send to NICs directly. These steps occur on one board or within a single server.

**(2) Within the rack:** Server traffic then goes to a rack switch. Multiple servers interconnect at high speed, with a top-of-rack (ToR) switch handling external communications and exchanges with other racks. All of this is intra-rack.

**(3) Across racks:** A data center comprises many racks coordinated by Spine switches. Spines manage connectivity among Leaf switches and to the outside, forming the backbone of the fabric.

**Illustration: Spine and Leaf in a DC fabric**

![IMG_256](https://pub.pbkrs.com/uploads/2026/0ef53687b944c76ee2514baf0cac82dd?x-oss-process=style/lg)

**Source: Bryon Moyer, Semiconductor Engineering, Dolphin Research**

**OCS mainly targets the Spine layer.**

First, Spine switches are expensive and power-hungry, making them prime candidates for alternatives. Second, OCS only forwards signals like a mirror; traditional switches also parse packets, inspect headers, and decide routing. Thus, OCS can stand in for Spines, but replacing Leafs would require adding packet-processing elements such as SmartNICs, complicating the architecture.

**The picture:**

Today, NVDA's Quantum X800-Q3450 and AVGO's Tomahawk 6 - Davisson represent CPO-based Spine switches, while $Alphabet - C(GOOG.US) is pushing OCS also at the Spine layer. There is direct competition here. In the end-state, OCS may replace Spines, but high-volume Leafs still need E/O near the ASIC, and servers need optical links among boards and between compute and NIC ASICs. CPO and OCS will likely be complementary.

**V. What are the value-chain components?**

**(A) CPO principles and architecture**

CPO is essentially an upgraded optical engine for E/O conversion. It comprises the following parts.

**1) Photonic circuits**

(i) Modulators: encode electrical 0/1 into optical signals by controlling light amplitude and waveform. (ii) Photodetectors: PDs convert optical signals into electrical signals. (iii) Waveguides: on-chip light paths akin to microscopic fibers.

**2) Electronic circuits**

(i) Driver: amplifies weak incoming electrical signals from the switch or server to drive the laser via the modulator. (ii) TIA: amplifies the tiny current output from the PD into a voltage signal suitable for downstream processing.

**3) Light source (laser)**

Modulators do not emit light; they control it. A laser provides the light to be modulated.

**Illustration: Optical engine block diagram**

![IMG_256](https://pub.pbkrs.com/uploads/2026/711e96fd0f20d6cb0abfad2c5564f11e?x-oss-process=style/lg)

**Source: Zong Zeguo et al., '400G FR4 Silicon Photonics Transceiver', Dolphin Research**

Two more elements matter. **4) DSP and CDR** repair signals: DSP compensates physical impairments, while CDR recovers clean clocks and retimes data (often integrated in the DSP). Both CPO and LPO remove the power-hungry, high-cost, high-latency DSP from the optical engine, but CPO shifts DSP-like functions into the switch ASIC and integrates CDR into high-speed SerDes, whereas LPO relies on stronger analog front-ends.

What are high-speed SerDes? They serialize parallel on-die data into high-speed lanes and deserialize incoming serial streams back into parallel data inside the ASIC. They are central to modern IO.

**(B) The CPO supply chain**

**1) The CPO system**

The optical engine includes the photonic and electronic sections above, and together with the ASIC forms the core of a CPO switch. Who delivers this? Unlike stand-alone pluggables from specialists such as $Zhongji Innolight(300308.SZ), $Eoptolink(300502.SZ), and Coherent, CPO will be led by system and silicon owners.

We expect value to accrue as follows. **(i) Platform and switch silicon owners:** NVDA/Google/AVGO/$Marvell Tech(MRVL.US) define architecture and standards and sell full systems. **(ii) Foundry/OSAT:** TSMC/$Advanced Semiconductor Engineering(ASX.US)/$Amkor Tech(AMKR.US) provide wafer fab, opto-integration, and advanced packaging. **(iii) Component vendors:** $Coherent Corp.(COHR.US)/$Lumentum(LITE.US) continue supplying optoelectronic devices. **(iv) Traditional module makers:** Zhongji Xuchuang/Xinyisheng and peers supply NPO, LPO, and serviceable CPO optical engines during transition.

**2) Key components beyond the optical engine**

**(i) Lasers**

CPO integrates E/O elements but not lasers today, so external lasers are still required. CPO also demands far higher laser power (3–4x), with stricter performance and reliability, lifting content value. Two routes are relevant.

**1) EML lasers:** The traditional path integrates laser and modulator, suited for ≥200G and longer reach, and is dominated by Lumentum, II-VI (Coherent), Sumitomo. **2) CW lasers:** An emerging path that fully disaggregates the laser, offering cost/power benefits and better alignment with CPO. Chinese vendors such as Source Photonics (CN), Accelink, and Changguang Huaxin have mass-produced 70 mW/100 mW parts and secured sizable orders.

**Illustration: EML vs. CW lasers**

![IMG_256](https://pub.pbkrs.com/uploads/2026/eb1d2e0088437bf793ea152c5c00b610?x-oss-process=style/lg)

**Source: Sumitomo Electric, Dolphin Research**

Next are four fiber components, seldom used in pluggable-centric designs.

**(ii) FAU (Fiber Array Unit):** precisely mounts fibers to align with on-chip waveguides. This enables high-precision coupling.

**Illustration: Fiber Array Unit**

![IMG_256](https://pub.pbkrs.com/uploads/2026/a2d990116930ff158981111791f07f3c?x-oss-process=style/lg)

**Source: Corning, Dolphin Research**

**(iii) PMF (Polarization Maintaining Fiber):** preserves optical polarization state along the path. This avoids polarization-induced loss.

**(iv) Fiber Shuffle:** reorders fiber positions for dense systems to match complex backplane layouts. It simplifies assembly in high-density environments.

**Illustration: Fiber Shuffle**

![IMG_256](https://pub.pbkrs.com/uploads/2026/1ef48ef47e4b61a2bb571a873ebda5bc?x-oss-process=style/lg)

**Source: Hyoptic, Dolphin Research**

**(v) MPO (Multi-Fiber Push On) connectors:** multi-fiber connectors for parallel fiber links. They enable high-density fan-in/fan-out.

**Illustration: MPO ports**

![IMG_256](https://pub.pbkrs.com/uploads/2026/6215abc8265dbb50e5bf86754e0b9c71?x-oss-process=style/lg)

**Source: Senko, US Conec, Dolphin Research**

Why are these rare in traditional modules? First, pluggables accept fibers via standardized receptacles, but CPO requires precise fiber-to-waveguide coupling, hence FAUs. Second, direct-modulated pluggables are insensitive to polarization and PMF used to be too costly, but CPO's external laser supply makes polarization critical, hence PMF. Third, pluggables typically need only two fibers (Tx/Rx), manageable by hand, whereas CPO must route many fibers to the backplane, necessitating Fiber Shuffles. Lastly, CPO at 400G and above may need 8 or even 16 parallel fibers within tight faceplate real estate, making MPO essential.

We will analyze market sizing and investable opportunities across the CPO stack in our next piece.

**<End\>**

### Related Stocks

- [NVDA.US](https://longbridge.com/en/quote/NVDA.US.md)
- [AVGO.US](https://longbridge.com/en/quote/AVGO.US.md)
- [300308.CN](https://longbridge.com/en/quote/300308.CN.md)
- [300502.CN](https://longbridge.com/en/quote/300502.CN.md)
- [AMKR.US](https://longbridge.com/en/quote/AMKR.US.md)
- [COHR.US](https://longbridge.com/en/quote/COHR.US.md)
- [LITE.US](https://longbridge.com/en/quote/LITE.US.md)
- [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md)
- [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md)
- [MRVL.US](https://longbridge.com/en/quote/MRVL.US.md)
- [ASX.US](https://longbridge.com/en/quote/ASX.US.md)
- [TSM.US](https://longbridge.com/en/quote/TSM.US.md)
- [BABA.US](https://longbridge.com/en/quote/BABA.US.md)
- [89988.HK](https://longbridge.com/en/quote/89988.HK.md)
- [09988.HK](https://longbridge.com/en/quote/09988.HK.md)
- [HBBD.SG](https://longbridge.com/en/quote/HBBD.SG.md)

## Comments (3)

- **Usagi的成长历险记 · 2026-06-03T12:08:02.000Z**: This is what you call professional, having a clear description of the data center in mind
- **ljndme · 2026-06-03T10:44:36.000Z**: Bug hunting: optical fiber is not light.
- **石来运转 · 2026-06-03T10:10:11.000Z**: Professional.