NVIDA’S GTC event got all the attention last week, and, hey, it’s the biggest company in the world. NVIDIA probably deserves more attention.

But the REAL action in the AI world at the OFC 2026 conference in Los Angeles. Stocks soared and swooned with each days presentations (and after late nights at the 70th floor bar at the downtown Los Angeles Ritz Carlton or the scenic Cara Cara bar at the Proper DTLA). Optical stocks have roared — doubling and tripling over the last year — proving that GPUs are OLD news. Optical is the story for 2026. That story 👇.

You can find prior notes and the full research archive at https://epistrophy.beehiiv.com.

As always, I’m focused on three things:
1) Technology-driven change;
2) the latest in innovation and startup trends, and;
3) stock fraud.

Companies Discussed

In This Note:

1. Companies Discussed

2. When Elephants Fell In ❤️ With Lasers

3. Tweet O’ The Week

4. 📆 of Epistrophy Events

5. Availability This Week

When Elephants Fell In ❤️ With Lasers

Why last week’s OFC 2026 Conference will move markets for years to come.

When I tell my grandkids about last week’s OFC 2026 conferences, I’ll tell them about the elephants.

In optical networking, “elephant loads” are large, sustained data flows—often gigabytes to terabytes—that dominate network capacity, in contrast to short, latency-sensitive “mice,” (a concept originated by Nick McKeown in “The Nature of Data Center Traffic: Measurements & Analysis.”) These flows now define AI infrastructure demand, where distributed training drives continuous, high-throughput communication among thousands of GPUs, concentrating demand onto 400G and 800G links. Because elephant loads are predictable and long-lived, they are reshaping network design: operators prioritize throughput, manage congestion to protect smaller flows, and sometimes establish dedicated optical circuits to bypass packet overhead. The result will be a new kind of network engineered for steady, industrial-scale data movement—elevating optical interconnect from a supporting role to a central constraint in modern data center design.

Elephany loads were all the talk of OFC 2026  in Los Angeles, held March 15-19. I don’t know when I’ve ever heard a more bullish conference. I heard about elephants in every briefing and every exectutive speech. We had extensive technical briefings with executives at Applied Optoelectronics (AAOI: NASDAQ), Cisco (CSCO: NASDAQ), Coherent (COHR: NYSE), Fabrinet (FN: NYSE) and Lumentum (LITE: NASDAQ) as well as international players like Aixtron SE (AIXXF:NASDAQ) and private companies like NewPhotonics. The mood was far beyond healthy demand. It bordered on rabid industrial mobilization. 

Booth chatter, formal presentations and side meetings all converged on the same point: compute has outrun connectivity, hyperscalers are trying to close the gap and every credible supplier of lasers, modulators, transceivers and optical switching now talks in the language of scarcity, line yields and floor space. OFC’s record attendence reflected that urgency. 

AI is driving a new sort of network demand. AI clusters need three different optical jobs at once. They need scale-out links between racks. They increasingly need scale-up links inside dense compute domains where electrical reach, power and signal integrity begin to fail. They need scale-across links between buildings and campuses. 

Those domains are moving from 100G-per-lane and 200G-per-lane assumptions toward 400G-per-lane road maps. Once that happens, the optical bill of materials rises almost everywhere at once: more lasers, more DSP-adjacent photonics, more fiber management, more packaging and in several architectures, more external light sources and optical circuit switching. 

Lumentum’s OFC launch captured the direction. It showed a 1.6T DR4 OSFP prototype using four 400G differential EMLs, a 16-channel DWDM ultra-high-power laser source for co-packaged optics and a VCSEL-based optical scale-up platform built around 1060nm arrays aimed at rack-level AI interconnects. 

400G-per-lane is not a cosmetic upgrade. It is a physics problem. Coherent used OFC to show 400G-per-lane PAM4 optical links for emerging 3.2T pluggables using both differential EMLs and a silicon-photonics implementation based on its silicon PN-junction Mach-Zehnder modulator. Lumentum, for its part, called its 400G differential EML module a stepping-stone to 3.2T. The industry’s road map is now explicit: 1.6T becomes the production ramp, 3.2T becomes the next engineering target and 6.4T – which we heard talk of for the first time – is an emerging goal. 

The second technical shift is architectural. For years, the optical story in data centers centered on pluggables at the faceplate. OFC 2026 made clear that the discussion has moved inward, toward co-packaged optics, external laser sources and optical circuit switching. Coherent leaned into CPO with multiple demonstrations across its vertical stack. Lumentum pitched ultra-high-power and super-high-power 1310nm lasers, including a device delivering more than 1 watt at 25 degrees Celsius, specifically for CPO and silicon-photonics architectures. The logic is unforgiving. As switch ASIC bandwidth rises, the cost of hauling electrical signals to the front panel rises with it in power, heat and complexity. Moving optics closer to the ASIC cuts that penalty. It also changes who captures the value, because the winning supplier must combine laser performance, photonic integration, packaging and reliability at scale.

The third shift is switching itself. AI traffic patterns increasingly reward low-latency optical fabrics that can reconfigure high-bandwidth paths for elephant loads instead of forcing everything through a pure packet-switching model. Coherent is pushing liquid-crystal optical circuit switches in configurations up to 512x512, stressing lower voltage and higher reliability than Micro-Electro-Mechanical Systems-based approaches. Lumentum is pushing “MEMS”-based Optical Circuit Switching (“OCS” to this acronym-loving industry) rooted in wavelength selective switching and talking like a company that believes optical switching will become a material line of business, not a science project. 

In an extremely bullish presentation, Lumentum CEO Michael Hurlston said “we are only at the starting line” for OCS and CPO, adding that OCS backlog was already beyond $400 million. The company’s OFC presentations and recent earnings materials tie that claim to a larger thesis: circuit-switched optical fabrics consume a fraction of the power of electrical switches and may fit non-blocking AI networks better as agentic workloads spread and traffic becomes less predictable. 

That is where the bull case gets harder, and more durable. Demand is no longer asking the supply chain for “more optics.” It is asking for more indium phosphide, more epitaxy, more laser output, more automated transceiver assembly and more qualified U.S. and allied-country capacity. Applied Optoelectronics described one response in blunt manufacturing terms. The company says it’s fastly expanding its US production facilities after a warrant agreement valued at up to $4 billion in product purchases over ten years investment from customer Amazon.com (AMZN:NASDAQ). CFO Stefan Murry talked about a 210,000-square-foot Sugar Land, TX expansion that it says is central to onshoring AI-transceiver production. In our briefing, management walked through a far more aggressive automation story than most investors appreciate: zero-human-contact “Phase 3” manufacturing, custom equipment across 14 of 17 major process steps, labor-hour reductions above 90%, cycle-time cuts above 35% and defect levels below 50 parts per million for 800G transceivers. Even if one discounts the promotional edge, the demand from customers was more than believable.

Fabrinet’s briefing supplied the companion piece. It is the contract-manufacturing side of the same boom, and its numbers looked less like cyclical optics and more like a capacity-constrained industrial. Fabrinet reported $1.13 billion in fiscal second-quarter revenue, up about 36% from a year earlier, and said construction of Building 10, a 2 million-square-foot facility, remains on track for completion by the end of calendar 2026, with an initial 250,000 square feet targeted by midyear. In our session, management argued that five months of operating margin could pay for the entire project and described a fixed-cost structure that makes the expansion unusually low risk. Seamus Grady (a big fan of the Drill Down podcast!) also said Fabrinet has three CPO engagements, including one with Nvidia (NVDA: NASDAQ), and called its relationship with Lumentum “the best it’s ever been.” That is the kind of sentence people say when the purchase orders are already changing shape. 

Cisco’s role in this week’s story is subtler but important. Cisco is both a systems company and a live read on how optical demand propagates into routers, switching and coherent pluggables. Its OFC page stressed end-to-end AI networking across optical, optics, silicon, systems and software. Cisco’s January said it took $1.3 billion of hyperscaler AI orders in fiscal Q1, with optics balanced against Silicon One-based systems, and forecast more than $3 billion of AI infrastructure revenue in fiscal 2026. 

The deepest implication is competitive. Coherent, Lumentum and AOI are all trying to move from being “optics suppliers” toward being indispensable bottleneck owners. Coherent wants to span materials, devices and systems. Lumentum wants to own high-power lasers, indium phosphide, OCS and pieces of scale-up. AOI wants investors to believe that vertical integration plus automation can close the cost gap with Asia while offering supply-chain security. Fabrinet wants to monetize all of it as the trusted builder. That is why this OFC felt so bullish. The value pool is broadening from transceiver vendors to the people who control laser output, packaging yield, switch architecture and production capacity. This market is ceasing to reward only clever component design. It is starting to reward industrial depth.

The takeaway was clear as the no-longer-smoggy Los Angeles sky: the next two years will bring 1.6T into volume, push 3.2T into credible early deployment, drag optical technology deeper into scale-up networks and make optical circuit switching a live capital-allocation debate inside every serious AI infrastructure vendor. By the end of the decade, more of the AI machine will be optical than today’s numbers imply. 

The conference did not feel euphoric in the old manner of optics booms. It felt disciplined, capacity-minded and grounded in orders, floor plans and thermal budgets. That is a more dangerous kind of bullishness for skeptics, who may find themselves trampled by the elephants. 

Tweet O’ The Week

📆 of Epistrophy Events

Availability This Week

I’ll be in Madrid all week for a client meeting and a look at the AI buildout from Europe. Use the time zone difference if you can — I’m happy to talk over stories and offer quotes if it’s of use. But email might be spotty. Text me!

Written reports are available to clients, with video summaries on YouTube, and of course our popular summaries of the summaries on Instagram, TikTok, and YouTube Shorts.