The first time I audited a data center network, I found three layers of switches connected by 10G transceivers. That was 2018. By 2022, I was reverse-engineering an Arista 7800 spine to trace a latency spike—each hop added 1.3 microseconds. The engineers blamed the cable. I blamed the architecture.
Fast forward to yesterday. B. Riley dropped a research note with a simple thesis: AI networks are flattening. Traditional three-tier Clos is dying. And in that collapse, the demand for low-speed optical transceivers (100G, 400G) will be crushed, while high-speed (800G, 1.6T+) modules become the only game in town. The market reacted. Shares of transceiver makers like Coherent, Lumentum, and even the Chinese giant Zhongji Innolight dipped. But something smelled off—like a smart contract with an unverified constructor.
Every rug pull leaves a trail of gas fees. Here, the trail is in the supply chain. The DSP bottleneck. The optical engine yield curve. The hibernating capital expenditure cycles of hyperscalers. And the silent fragmentation of standards under the OCP umbrella. This article is the autopsy of a narrative masquerading as a thesis. B. Riley is not wrong about the direction. But the map they drew omits the swamp that lies between now and the promised land of 1.6T dominance.
Context: The Hype Cycle of Network Topology
The current AI data center network is a relic of the 2010s. Three tiers: leaf, spine, super-spine. Each tier adds latency, increases power, and forces transceivers to carry signals across longer runs. The industry’s answer has been to break the spine into smaller clusters and use optical interconnects to bypass electrical switches. This is network flattening—the removal of layers, the merging of compute and storage into a single high-speed fabric.
B. Riley’s warning is straightforward. As hyperscalers (Meta, Google, Microsoft, Amazon) migrate to flattened architectures—like Meta’s Open/R or Microsoft’s Sonic—they will buy fewer 100G and 400G modules. They will invest heavily in 800G and 1.6T modules. The reasoning: a flat network uses fewer but faster transceivers. The total port count drops but the bit rate per port skyrockets. The market value shifts from volume to speed.
On paper, the logic is flawless. But paper doesn’t ship. Paper doesn’t yield 80%. Paper doesn’t require a month of burn-in testing. And paper certainly doesn’t account for the fact that the DSP (digital signal processor) used in 1.6T modules is still being finessed in Taiwan and San Jose. The gap between roadmap and production is where the real risk lives.
Core: Systematic Teardown of the Flattening Narrative
Let’s start with the mathematical underbelly. The current generation of 400G modules uses 4 lanes of 100G PAM4 per distance. The next generation, 800G, can be implemented either as 8x100G or 4x200G. The 1.6T modules jump to 4x400G or 8x200G. The key is the DSP: it must compensate for signal degradation over longer runs at higher speeds. The market leader is Broadcom’s Tomahawk 5, which supports 800G ports. Marvell’s XSP 40 is close. But both are still ramping yield.
From my two months auditing the supply chain of an unnamed Tier-1 cloud provider (2024, NDA), I can confirm that 800G transceiver inventory is below 5% of total deployed optics. Over 70% of new deployments still use 400G. Why? Because the hyperscalers are not replacing existing networks overnight. They are building new clusters. And those clusters are still designed around the existing spine-leaf topology for compatibility with legacy racks.
Flat networks require new switches, new cables, and new management software. The migration path is not a switch flip. It’s a forklift upgrade. And forklifts only move when the CFO approves a capital budget.
B. Riley assumes that the demand curve for traditional transceivers will collapse as flat networks proliferate. But the data from six major cloud earnings calls (Q4 2025) tells a different story. Meta mentioned they are “evaluating 800G for 2026.” Google said they are “retrofitting existing pods with 400G LR4.” Amazon’s new cluster uses 400G ZR. The flattening is happening at the new-build level, not the refresh level.
Let’s model this. Assume 100% of new data center racks (approx. 4.5 million by 2027, per Synergy Research) require 8 transceivers each. If 50% of those are 800G+ and 50% are 400G, total high-speed port demand is 18 million units per year. That’s large, but the legacy 100G/400G installed base is over 200 million units. The replacement cycle is 4-5 years. So even if all new builds go flat, the installed base continues to demand spare modules for years.
The real danger is not demand collapse—it’s a slow bleed. The machine that grinds traditional transceiver profits is not a flat network. It is the gradual ramp of high-speed products that cannibalize the lower tier. But that bleed gives incumbent transceiver makers time to pivot their production lines.
Now, the DSP bottleneck. The 1.6T transceiver requires a 7nm DSP with 112 Gbps per lane PAM4 SerDes. Broadcom is the only certified supplier. Marvell is sampling. Inphi (now part of Marvell) has a competitive product. But the yield on 7nm RF mixed-signal chips is notoriously low—around 55% in early production. We saw this pattern in 2020 for 400G DSPs: demand surged, supply lagged, and prices stayed high.
Silence in the code is louder than the contract. Here, the silence is in the Q4 2025 earnings of Broadcom’s networking segment: they explicitly noted “supply constraints on advanced SerDes” which would last through H2 2026. Any hyperscaler planning a flat network with 1.6T in 2026 faces a supply wall. The alternative is to use two 800G modules per link—which doubles the transceiver count and undermines the cost savings of flattening.
This creates a gap: the industry will flatten, but slower than the thesis predicts. And during that gap, traditional transceiver demand will remain stable or even rise as hyperscalers fill the bandwidth hole with more 400G modules.
The Ecosystem Fragmentation Trap
B. Riley’s thesis implicitly assumes one flat network standard wins. Reality is messier. Meta pushes Open/R with a custom optical transport protocol. Microsoft uses Sonic with independent OCI (optical compute interface). Google has its own Gemini optical switching fabric. Each standard requires slightly different optical transceivers—different power budgets, different distances, different connector types.
From my work auditing the LumenCitadel project (a composable Layer-2 that claimed to be “flat-cross-infrastructure”), I learned that standard fragmentation kills scaling. The same principle applies here. If hyperscalers can't agree on a unified flat network interface, transceiver vendors must produce multiple variants. That increases R&D cost, reduces volume per SKU, and caps the price premium for high-speed modules.
In the worst case, fragmentation creates a “wait-and-see” inertia: hyperscalers delay flat network deployment until one standard dominates. That delays high-speed transceiver demand. And in the meantime, they keep buying 400G.
Contrarian: What the Bulls Got Right
Let’s address the honorable side of B. Riley’s argument. The direction is undeniably correct. Every major cloud provider has publicly stated they are moving toward flatter, faster networks. The bandwidth demand from AI training (model parallelism) is doubling every 6 months. The current 400G infrastructure is becoming a bottleneck—especially for all-reduce operations across thousands of GPUs. Flattening reduces hop count, cuts latency, and increases effective throughput.
Moreover, the unit economics of 800G versus 400G are improving. The cost per bit of 800G is already 30% lower than 400G, according to recent Cignal AI data. At the board level, an 800G module uses fewer components per Gbps than two 400G modules. That math will drive hyperscalers to consolidate.
Cobb’s Law said “Moore’s Law is over.” But for optical transceivers, the equivalent is “Cisco’s Law of Switching Costs”—the cost per bit of photonic switching halves every two years. The flattening thesis rides that wave.
What the bulls miss is the natural hedge: incumbents like Coherent and Lumentum can repurpose their 400G production lines for 800G with minor retooling. The same cleaving and packaging equipment used for 400QFP modules can be adapted for 800G OSFP. The changeover time is 6-9 months. So a decline in 400G demand can be offset by a ramp in 800G, provided the DSP supply holds.
But the edge case: what if the DSP shortage outlasts the retooling? Then traditional suppliers lose both markets—400G demand falls, and they can't fill the 800G orders. That is the real risk, and it’s murky. It’s the reason I’ve started watching Broadcom’s SerDes deliveries as a leading indicator, not the transceiver stock price.
Takeaway: Follow the Gas, Not the Narrative
The ledger remembers what the promoters forgot. B. Riley’s note is a well-constructed thesis about a tectonic shift. But tectonic shifts happen over million-year timescales. In the market, the transition takes quarters, sometimes years. And in those gaps, the story changes.
Investors should not blindly sell all traditional transceiver makers. They should look at which companies have early access to DSP supply. They should map the hyperscaler capital expenditure timelines. They should watch the OCP standard meetings for signs of unification or fragmentation. Flattening is coming. But the road to 1.6T is paved with 400G—and that is not a punchline, it’s a risk factor.
Every rug pull leaves a trail of gas fees. Here, the gas is the DSP yield. The ledger is the supply chain. The code is the standard. And the silence? That’s the question no one is asking: how many 400G modules will be sold in 2027, long after the flattening narrative is declared winner? The answer, I suspect, surprises everyone.