White-Hair Stock God Serenity: OpenAI and Amazon’s new network architecture “will not weaken the demand for optical modules”

Recently, some people on Wall Street have claimed that OpenAI’s MRC and Amazon’s (AMZN) RNG—two new network architectures—will “flatten” data centers, reducing the number of optical modules used per single XPU and cutting potential optical transceiver demand by as much as 40% to 50%. “White-haired stock god” Serenity, an overseas analyst, stepped in to rebut that view, saying that even if the number of optical modules per single XPU declines, the explosive growth in the overall AI compute scale will still push total optical communications demand upward. He also expressed his stance on the CPO laser roadmap, saying he favors ultra-high-power continuous-wave lasers as the long-term winners.
(Background recap: OpenAI teams up with NVIDIA, AMD, and Microsoft to launch the “MRC network protocol,” which completely solves the 100,000-plus GPU training traffic jam)
(Additional background: AI stock god Serenity locks onto “NVIDIA’s optical upstream bottleneck,” after Ayar Labs announced it is joining NVLink Fusion)

Table of contents

Toggle

• The shorts bet the number of switch layers will be cut
• Serenity counter-calculates that more XPU installs only increase demand
• VCSEL is just a stopgap; UHP CW is the endgame

Key summary

• The shorts argue that OpenAI MRC and Amazon RNG will flatten the network, estimating optical transceiver demand could drop by 40% to 50%
• Serenity rebuts: even if optical modules per single XPU decline, the scale of XPU deployments will rise sharply, so overall optical module demand is still expected to grow
• On the laser roadmap, Serenity expects VCSEL to serve as the NPO short-reach transition, while UHP CW is the long-term CPO solution, aligning with Broadcom’s position

When the market starts worrying that AI network architectures might “in turn” kill optical communications demand, Serenity—dubbed the “white-haired stock god”—chooses to stand on the opposite side of the shorts. He addressed discussions about optimizations to OpenAI’s MRC network architecture and whether Amazon (AMZN) RNG architecture could lower optical module demand, saying that fewer optical modules used per single XPU doesn’t mean the total will be less.

Regarding $AMZN RNG + OpenAI’s MRC and VCSELs.

In what appears to be an indirect response on B. Riley's $AAOI sell report regarding optimizations that "flattens networks and cuts transceiver TAM 40-50%".

Rosenblatt TLDR: They've already modeled for this, even though each XPU… pic.twitter.com/A41ceTqKwJ

— Serenity (@aleabitoreddit) July 10, 2026

The fuse for this debate was an equities investment bank: B. Riley downgraded optical transceiver maker Applied Optoelectronics (AAOI) on the grounds of “flattening network risk.”

The shorts view MRC and RNG as removing several layers of switches from data centers, shrinking optical links accordingly.

The shorts bet that the number of switch layers will be cut

First, we need to introduce the differences between these two network architectures. OpenAI’s MRC (Multipath Reliable Connection) breaks an 800Gb/s network card into eight 100Gb/s lanes, each connected to different switches, enabling more than 100,000 GPUs to be strung together with just two layers of switches—while saving cost and power. This protocol was jointly introduced by OpenAI together with AMD, Broadcom, Intel, Microsoft, and NVIDIA.

Amazon’s RNG (Resilient Network Graph) is more direct: it uses passive optical ShuffleBox to replace traditional multi-stage aggregation switches, cutting active networking equipment by more than 60% in one move. With fewer layers and fewer switches, the optical modules each XPU needs to carry naturally decrease as well. Based on this, B. Riley estimates that optical transceiver demand could be cut by 40% to 50%. The logic of the shorts is not wrong, but they only calculated the numerator and ignored the denominator.

Serenity counter-calculates: more XPU installs keep increasing demand

Serenity’s rebuttal is simple: the denominator is exploding. He points out that as the number of XPU deployments rises significantly, even if each single XPU needs fewer optical modules, the multiplied total optical module demand is still expected to grow. In other words, the amount saved by the architecture gets more than “bought back” by the expansion of compute scale.

This is also a consistent move from Serenity: when others see demand potentially collapsing, he counts whether bottlenecks across the entire supply chain still exist. Market data backs his side: the AI optical module market size is expected to grow from $16.5 billion in 2025 to $26 billion in 2026. Amazon itself has also signed multi-billion-dollar long-term optical fiber supply agreements with Corning, backing long-term demand for data centers. If demand truly gets cut, the one that should be worried most—Amazon—is the one stockpiling.

VCSEL is just a transition; UHP CW is the endgame

Beyond total volumes, Serenity also commented on CPO (co-packaged optics, bundling the optical engine and switching chip together) laser roadmaps. He believes VCSEL (vertical-cavity surface-emitting laser) is better suited as a transition technology for NPO (near-packaged optics) and short-reach interconnect stages, because its advantages are low cost, maturity, and being deployable right now. But as we extend to the CPO generation, ultra-high-power continuous-wave lasers (UHP CW) are the better solution.

This assessment aligns with Broadcom’s earlier public statements. Broadcom positioned CPO as the “North Star” for energy efficiency and bandwidth, while treating the VCSEL-based NPO as a “bridge you can deploy now.” Broadcom has also demonstrated 200G VCSEL, EML, and CW laser technologies. VCSEL is the bridge to get you going first; UHP CW is the destination on the far side of the bridge.

Back to the original question: will the architecture cause each single XPU to use fewer optical modules? It will. But as long as AI is still crazily stacking server racks, the optical communications path is still widening—not narrowing. The shorts bet on efficiency; Serenity bets on scale. What comes next is what matters: the actual pull rate for 1.6T optical modules in 2026 to 2027—whose model is more accurate, and the shipment numbers will speak for themselves.

This is not investment advice.

Common questions

Why are OpenAI’s MRC and Amazon’s RNG architectures considered to reduce optical module demand?

Both architectures are about “flattening” data center networks. MRC lets two switch layers connect 100,000 GPUs, while RNG uses passive optical ShuffleBox to replace multi-layer switches, cutting more than 60% of active equipment. With fewer layers and switches, the optical modules required per single XPU decline, so B. Riley estimates demand could fall by 40% to 50%.

Why does Serenity think optical communications demand will still grow?

Serenity thinks the key is the total number of XPU deployments. Even if optical modules per single XPU decrease, the rapid expansion of AI compute scale will drive a surge in the number of XPU units, so the multiplied overall demand is still expected to rise. The AI optical module market is estimated to grow from $16.5 billion in 2025 to $26 billion in 2026, supporting his view.