OFC Day 2

Same party, different venue.

I’m at the Optical Fiber Communications conference again this year, in San Francisco. No sponsor, but I couldn’t bear to miss the fun.  And it hasn’t disappointed.

This conference is huge. Something like 16,000 people, between the conference and the show floor. The show floor is huge. It fills the lower levels of two buildings plus under the street between them. I wanted to take a photo to show how big, but there is no vantage point to do that. I’ve gotten my step (11,600 of them) in today, and will tomorrow and Thursday as well.

Most of my focus has been on the high-speed PON front. Since the informal theme is really AI data centers (AI = accelerated infrastructure, as in GPU and NPU accelerators), it was hard to focus on my space. Some take-aways:

• The cadence of everything in the AI space is insane, while the challenges keep getting harder. Vendors are having to anticipate products by two or three cycles, before their customers know what they’re going to need. Question: is this really healthy?
•  AI data centers are rapidly moving to 800 G(bps) optics for scale out, while copper still wins for scale-up… but cluster sizes are growing, and that’s not sustainable forever. 1.6 T is all over the show floor, and 3.2 T is on the horizon. To support that, 200G lanes have become a reality. Pradeep Sidhu (founder of Juniper, now at Microsoft via acquisition) mad a case that progress should stop at 200G and innovation focus elsewhere… a plea must have fallen on deaf ears.
• Market Watch was all about data centers again this year.  Our friends at Omdia and Dell’Oro are forecasting even more explosive growth in capex and optical modules, with even shorter generational cycles. Everybody is anxious about supply chain, tariffs and geopolitical insanity.
• Nvidia gave a short presentation. The narrative is that as reasoning advances, inferencing — which every organization has to do, versus training, which is limited to maybe 20 companies—becomes more compute intensive. Inferencing infrastructure is going to look more like training infrastructure — ergo, larger clusters, more xPUs, more networking, more optical. Also, the current pluggable optics paradigm is running into scaling trouble, and the time for co-packaged optics is coming—fast. Thus the CPO switch they introduced at GTC last week.
• The equities analysts had a mixed story.  Investors are caught up in the enthusiasm, but concerned about how sustainable it is (I share that concern). AI hyperscaler  capex levels have always been volatile, thus lots of uncertainty about when and how deep the next correction will be. Supply chain, geopolitics, and trade wars are weighing on confidence. CPO looks to be a big disruptor, and is likely to move valuations.

Back on my preferred topic. Very High-Speed PON is a big topic in the research conference, hardly at all on the show floor. The debate over IM-DD vs coherent continues, and there were lab results from both camps. Professor Roberto Guadino (Polytechnico di Torino and de-facto leader of European PON research efforts in academia) gave a great tutorial on the state of play, technical challenges and solutions for both IM-DD and coherent plus his own group’s research. For PON technology geeks like me, a feast. In that and subsequent sessions, there was the usual parade of grad students and industry researchers showing off their results. I haven’t digested it all, so will write about take-aways later. The general sense is that VHSP is very much a research project, and the commercial interests are letting the engineers play for now; thus, talking to people, there weren’t the kinds of hardened positions that we normally see by the time a technology hits the standards committees.  Of course, with a standard 5 years away, that’s a good thing.

I had been anxious that CableLabs CPON activity would conflict with ITU-T and dilute the components and systems market to the point of unviability. I turns out that this is not the plan.They are going ahead developing their spec, and planning to present it when ITU-T Q2/15 holds its beauty contest… but hope that it contributes to the group effort rather than expecting it to be adopted whole, and intending to move with the industry consensus.

I was barely able to scratch the surface on the show floor. First observations:

• Nokia was showing off their new 25GS-PON residential ONTs; it seems that a few of their customers are introducing 10 Gbps+ service for the bragging rights.
• I am taking a pulse on 50G PON components and modules. Thus far, it’s a mixed bag. I spoke with one major vendor (not naming names here) who is sitting on completed designs for 25- and 50G PON opto-electronics and mixed signal parts, but isn’t introducing them until they’re convinced there’s a market. Another vendor with similar kit, MA/COM, is shipping, and just came out with an improved avalanche photodiode (APD) for improved receiver sensitivity and an externally modulated laser + semiconductor optical amplifier (EML+SOA). Two other vendors I talked to also are innovating on higher gain, higher bandwidth APDs.
• My friends at PICAdvanced were showing off an NG- PON2 transceiver with 4-channel bonding to yield a roughly 38 Gbps PON. Despite years of work on their NG-PON2 photonic IC (PIC), they’re still shipping modules based on discrete parts. That involves the channel bonded SFP+. Once again, the PIC is a year out, as it has been for more years than I care to count. They are also talking about but not yet ready to demo a 4x50 G TWDM PON PIC-based OLT and ONT transceivers. They’re also showing an NG-PON2 ONT stick (including the MAC/TC and Ethernet bridge) as their next generation. I’ve been a cheerleader for TWDM PON for something like 17 years, and have been  waiting for volume shipments of PICs to drive costs down enough to make it economic. Of course I’m disappointed that it’s been so long in coming, especially in light of my rosy predictions back in 2018. In the meanwhile, operators have been deploying XGS-PON in volume with declining prices, 25GS-PON is a fairly low-impact option for Nokia customers and 50G PON is launching. Operators (other than Verizon) who were serious about NG-PON2 deployments have backed down. Whether the presumable lower priced PIC-based NG-PON2 transceivers return to favor seems doubtful. However, if pricing comes in below 25- and 50G transceivers, there might still be a shot.

To be continued.