Video Summary - 엔비디아 다음 알파는 GPU가 아닙니다 | CPO·광통신 11개 종목 전수공개 [GTC 타이베이 3편]

Scientific / technical concepts & discoveries mentioned

AI data center network bottleneck
- As AI clusters scale, the limitation shifts from GPU compute to data transfer, including:
  - GPU-to-GPU communication
  - communication within labs/racks (intra–data center networks)
  - communication between data centers (scale-out / inter–network traffic)
  - request/response data flows
  - model parallelization (splitting a model across multiple GPUs)
  - related routing and back-and-forth data movement
- Growing need for network bandwidth and power efficiency, especially for:
  - long-context inference
  - reasoning
Electrical interconnect limits → optical communication
- Electrical signaling over longer distances degrades due to:
  - signal attenuation (“signal drops”)
  - higher power consumption (“fever worsens” / heat)
  - deteriorating signal quality
- Light (optical communication) is suggested to enable longer, fast, complex interconnects.
CPO (Co-packaged Optics)
- Definition: Co-packaging and Optics—integrating optical components with the switch chip in a single package.
- Motivation:
  - replace “pluggable transceivers” (power-hungry, less efficient) with integrated optics
  - reduce power/heat and improve overall efficiency
- Key idea: integrate the optical engine into the switch package to shorten effective distance and improve thermal/power performance.
Spectrum-X Ethernet Photonics (NVIDIA)
- NVIDIA claims CPO Spectrum-X Ethernet Photonics has entered production phase.
- Commercial availability is targeted for 2H 2026.
- Interpretation: “production” is not immediate large revenue—rather it indicates supply chain readiness and early adoption, with broader scaling later.
- Product specifications mentioned (some subtitle units are garbled):
  - Max bandwidth: 409.6 (unit text unclear)
  - ~6× network power efficiency vs existing pluggable transceivers (subtitle says “fluxable transceivers”)
  - 5 nm / 5-wire AI application runtime (unclear meaning/units)
  - Deployment time: ~1,300 (unclear units)
Integrated optical-engine architecture
- 32 silicon photonics engines per compact package.
- Each engine contains:
  - 16 transmitters + 16 receivers
- Engine throughput:
  - 3.2 T/s (subtitle formatting unclear)
- Implication: in CPO, yield and assembly automation become critical bottlenecks:
  - one failed engine can compromise an entire package.
ELS (External Laser Source)
- Definition: laser light source located in an external module rather than inside the switch package.
- Reason: switch internals are heat-sensitive, so moving the laser outside can improve:
  - lifespan
  - replaceability
  - maintainability
- NVIDIA architecture usage stated:
  - 16 ELS for a single-switch (single-A) chip
  - 64 ELS for a four-switch (quad-A) chip
- Consequence: as the number of switch chips increases, demand rises for:
  - ELS units/lasers
  - fiber attachments
  - optical connectors
Optical-fiber attachment & packaging ecosystem constraints
- CPO depends on more than optics, including:
  - fiber attach technology (attaching fibers to chips)
  - precision connectors
  - specialized optical fibers maintaining polarization
  - microoptics
  - detachable connectors
- These create manufacturing yield bottlenecks.
- Packaging/testing complexity also matters, including:
  - multichip packaging (combining multiple chips in one package)
  - wafer bumping
  - wafer sorting
  - assembly testing
- As systems scale, packaging/testing must scale too.
Quantitative demand estimation for optical ports (Vera Rubin)
- A calculation estimates optical communication demand from a system called Vera Rubin:
  - “Vera Rubin MVL contains four” (optical-related items)
  - “Rubin GPU 1 contains two” (subtitles ambiguous)
- Uses an NVIDIA-stated assumption:
  - 1.6 TB/s of external scale-out network bandwidth per GPU
- Then estimates:
  - total bandwidth for “one area” as 115.2 TB/s
  - converted into “800G optical ports” (cited as 800 GB/s per port)
  - result: 144 ports (115.2 / 0.8)
- Extrapolation:
  - hyperscalers worldwide are said to have “tens of thousands” of Vera Rubin systems
  - implying tens of millions of 800G ports
- Conclusion framing:
  - optical components + CPO could become a “supercycle” after the prior GPU memory cycle (HBM).
Elasticity of optical-component stocks vs NVIDIA
- NVIDIA benefits are argued to be less price-sensitive due to its scale and massive market capitalization.
- Smaller optical-component suppliers are described as having higher stock price elasticity relative to shipment growth, with bidirectional risk:
  - upside if adoption accelerates
  - downside if schedules slip, yields drop, or technology is replaced

Lists / methodologies outlined (stock-analysis framework & verification)

1) Four-stage framework for evaluating “CPO theme” stocks

Stage 1: Official NVIDIA exposure (highest reliability)
- Companies explicitly named by NVIDIA (the “11 companies”).
Stage 2: “Yangsan flood damage” (evidence-backed volume)
- Example: Foxconn with reported production target.
Stage 3: Indirect options (adjacent exposure, weaker revenue visibility)
- Examples: AAOI (appears as “Applied Auto Electronics” in subtitles) and Four Technologies.
- Criteria:
  - proximity to CPO/optical comms
  - but not on NVIDIA’s official list
  - with mass production/sales evidence potentially lacking.
Stage 4: Simple theme stocks
- “Story stocks” with “CPO” attached but unconfirmed:
  - actual customers
  - mass production
  - yield
  - volume
- Warned against grouping these with higher-reliability stages.

2) Counter-scenarios (conditions under which the CPO thesis could be wrong)

Five conditions:

Foxconn CPO switch mass production delayed in 3Q 2026 or misses 10,000 units/year target significantly.
NVIDIA-related revenue/impact at Lumentum and Coherent does not appear in earnings guidance until 2H 2026.
Existing “1.6T” optics persist longer than expected or are bypassed by LPO
- LPO described as Linear Drive Pluggable Optics (subtitle garbling)
- positioned as an intermediate technology that improves power efficiency without fully moving to CPO.
For indirect stocks (e.g., AAOI):
- hyperscalers/A-type camp approval does not lead to orders
- by 2027, full-scale mass production volume does not increase sequentially.
The 11 officially named companies fail to achieve meaningful revenue, making the market structure entrenched with weak monetization.

3) Competitive landscape tracking (not limited to NVIDIA)

Competing/parallel efforts mentioned:

Broadcom pushing CPO with a “Tomahawk series” (subtitle garbled as “CP with Tomaok series”)
Marvell for optical signal processing (optical DSPs, custom silicon)
Cisco using “Silicon One” for enterprise/cloud markets
Big-tech proprietary approaches advancing proprietary optical interconnects, including:
- Google TPU
- Meta MTIA
- AWS “Trainium” (spelled “Treinium”)

Implication: optical communication demand may extend beyond NVIDIA GPU ecosystems, affecting how suppliers’ customer exposure should be weighted.

Researchers / sources featured (named individuals not provided)

No individual researchers are listed. The content focuses on companies and industrial players. Featured sources include:

Additional named competitors / ecosystem players

NVIDIA
Broadcom
Marvell
Cisco
Google
Meta
AWS
Coherent (along with Lumentum as a major optical/laser exposure pair)
Applied Optoelectronics (appears as “Applied Auto Electronics”)
Four Technologies (subtitle: “Four Technologies”)
Google TPU, Meta MTIA, AWS Trainium (named as proprietary AI hardware platforms)

엔비디아 다음 알파는 GPU가 아닙니다 | CPO·광통신 11개 종목 전수공개 [GTC 타이베이 3편]

Key takeaways