OCP Global Summit 2026 | Lumotive, iPronics | James Burch, Vivek Sarathy | Fast Switching OCS for AI and Hyperscale Data Centers

3 days ago

6 min read

1. Why These Two Companies

Both presenters come from outside the 3D MEMS camp. Burch is OCS Engineering Lead at Lumotive, and Sarathy is Vice President, Products at iPronics. According to public information (not from this talk), Lumotive builds metasurface beam-steering chips originally used in lidar, while iPronics builds programmable silicon photonic circuits.

The author list goes beyond them. The slides name eight authors from iPronics, Ciena, Lumotive, Intel, Carnegie Mellon University, Lumentum (two) and Tripple Stone. The content draws on the OCP OCS work stream, building on the OCS white paper published in April 2026 and a second fast-switching white paper that is still to come.

In ECOC2026 | 95% of OCS Shipments Are Google's Business we covered how concentrated this market is. The question this talk tries to answer: can OCS move beyond Google and become something many more data centers can use? The speakers say yes, but only if switching gets much faster and the surrounding ecosystem keeps pace.

2. The Starting Point: MEMS Proved OCS Works, and Also Limits It

The first claim on the slides is that OCS is proven at scale: by eliminating O-E-O conversion, Google achieved a 30% cost and 40% power improvement, and the speaker added that latency also improved.

▲ OCS is proven at scale, but the millisecond reconfiguration of 3D MEMS limits its use cases, and CPO makes the need more urgent (Image source: Lumotive, iPronics | 2026 OCP Global Summit presentation)

The problem is in the second point: 3D MEMS dominates OCS, but it reconfigures in milliseconds, which suits only slowly changing workloads. Burch credited 3D MEMS with very low insertion loss, yet milliseconds remain its ceiling.

The third point matters even more. Because sub-millisecond switching has never been broadly available, fast-lock transceivers, control planes and orchestration software have had no reason to be optimized. It is a chicken-and-egg problem: the switch is slow, so nothing around it needs to be fast; nothing around it is fast, so a faster switch alone does not help.

The slide ends with one more push: CPO makes the need more urgent, because there will be more optics, closer to the servers, changing more often.

3. Breaking Down the Technologies: Small Things Switch Faster

The speakers split fast-switching technologies into two families: free-space metasurfaces and in-plane integrated photonics. The core principle fits in one sentence: small things switch faster than big things.

Technology	Reconfiguration time
Metasurface (free-space)	~50 μs
Integrated photonics MEMS	1–20 μs
Integrated photonics thermo-optic	10–100 μs
Integrated photonics electro-optic	ns–μs (tens of GHz), ns (30–56 GHz), sub-ns (100+ GHz to THz)

Metasurface: sub-wavelength pixels only a few hundred nanometers wide switch in tens of microseconds. It remains a free-space architecture, so systems look much like existing 3D MEMS, which is good news for current OCS chassis designs.
Integrated MEMS: an in-plane waveguide is pushed toward another waveguide, giving a cross or bar state.
Thermo-optic: the most common silicon photonics approach today, using micro-heaters with ring resonators or Mach-Zehnder interferometers.
Electro-optic: uses carrier-concentration changes or the Pockels effect in thin-film lithium niobate, reaching sub-nanosecond switching.

The speaker stressed that "integrated photonics" is a broad term covering silicon photonics, thin-film lithium niobate, barium titanate, indium phosphide and heterogeneous integration.

▲ Fast-switching technologies compared: metasurface ~50 μs, integrated MEMS 1–20 μs, thermo-optic 10–100 μs, electro-optic down to sub-nanosecond (Image source: Lumotive, iPronics | 2026 OCP Global Summit presentation)

One line of small print on that table deserves attention: radix, cycle endurance and insertion loss are outside the comparison. Low insertion loss is exactly the 3D MEMS strength the speaker acknowledged. Comparing only speed means comparing the incumbent on its weakest metric.

▲ On a log time axis, fast-switching technologies sit between microseconds and sub-nanoseconds while 3D MEMS sits at milliseconds, yet end-to-end time is still bounded by transceiver relock and the control plane (Image source: Simple Tech Trend original; data from Lumotive, iPronics 2026 OCP Global Summit presentation)

4. Breaking Down the Challenges: The Switch Got Faster, Three Other Things Did Not

The benefits-and-challenges slide has an attractive left column: higher network utilization, real-time topology adaptation, less overbuild, support for MoE and disaggregated inference, and less reliance on electrical packet switching. The right column is what matters:

Transceivers: after every reconfiguration they must reacquire clock and data, restore equalization and stabilize the link. The speaker said lock time must improve by "orders of magnitude".
Control plane: scheduling, routing, telemetry, monitoring and troubleshooting all need to keep microsecond pace.
Endurance: the speaker noted that if switching is three or six orders of magnitude faster, lifetime switching cycles rise accordingly.

Endurance can be quantified. The following is an STT estimate, not a slide figure: a year has about 31.5 million seconds. Reconfiguring once per second means about 32 million cycles a year; once per millisecond, about 32 billion; once per 100 μs, about 320 billion. No company in this talk disclosed cycle-endurance data for fast switches.

▲ End-to-end reconfiguration time is the sum of switch, transceiver relock and control plane; speeding up only the switch leaves the total dominated by the slowest term (Image source: Simple Tech Trend original; data from Lumotive, iPronics 2026 OCP Global Summit presentation)

Put simply: the speaker himself said transceiver relock is already slower than MEMS switching, so taking the switch from milliseconds to microseconds only makes the transceiver and control plane a more visible bottleneck.

5. Four Use Cases, From Near to Far

Sarathy covered the use cases, laid out on the slides as an evolution path:

Use case	Problem	Outcome
1 Traffic engineering	Traffic hotspots between rack pairs	Higher network utilization
2 AI workload adaptation	Communication patterns change across training phases	Better GPU efficiency
3 Fast optical fabrics	Coordinating rapid optical path changes	Latency and power gains
4 Scalable optical fabrics	Scaling connectivity across large AI clusters	Flexible, scalable infrastructure

▲ Four use cases for fast OCS: from traffic engineering and workload adaptation to fine-grained scheduling and scalable optical networks (Image source: Lumotive, iPronics | 2026 OCP Global Summit presentation)

The wording itself signals maturity. Use cases 1 and 2 are framed as problem, solution, outcome; use case 3 switches to problem, opportunity, potential outcome, and uses "could". The further along the list, the more it reads like a research topic rather than a product spec.

STT's view: use case 1 is most likely to land first. Detecting hot rack pairs, offloading large flows onto optical circuits and adjusting as traffic shifts does not require microsecond cycles, so modest improvements to existing transceivers and control planes are enough. Use case 3, which switches in real time per workload and per data flow to each GPU, needs all three challenges solved at once.

6. Risks and Counterarguments

No quantified gains. The only benefit figures in the deck are Google's 30%/40%, and those came from millisecond MEMS. There is no number for how much GPU utilization faster switching would buy.
Speed-only comparison. Radix, insertion loss and endurance are explicitly excluded, and low insertion loss is a 3D MEMS strength the speaker acknowledged. Whether fast technologies can match on radix and loss is the precondition for replacing MEMS.
Electrical packet switching keeps improving. In Slow-and-Wide Was Not Killed by Physics but by 60,000 Fibers (in Chinese) we argued that optical network challenges are often at the system level rather than the component level; if switch ASICs keep lowering energy per bit, the room for "eliminating O-E-O" shrinks.
The CPO argument cuts both ways. Optics closer to servers does call for more flexible optical networks, but CPO's own adoption timeline is still debated, see CPO Is No Longer Crying Wolf.
The speakers have a stake. Both presenters' companies are betting on non-MEMS technology, so a "MEMS is too slow" narrative works in their favor.

7. What It Means for Taiwan's Supply Chain

No Taiwanese company appears in the slides or the author list, so this section covers structure only, without pairing companies.
A change in manufacturing model. 3D MEMS OCS is precision optical assembly; integrated-photonics switches are wafer processing plus packaging. If fast OCS moves onto chips, value shifts toward foundry and advanced packaging, a structural strength of Taiwan, but there is no order signal yet.
Transceivers are the first link called out. "Fast-lock transceivers" is listed among the three main challenges, which means a new spec requirement for optical module and DSP suppliers; watch whether the fast-switching white paper turns relock time into a concrete number.

8. Summary

The most honest part of this talk is its admission that the switch is not the only bottleneck. 3D MEMS has proven the value of OCS at Google, but milliseconds gave the surrounding ecosystem no reason to get faster; pushing the switch to microseconds or nanoseconds only makes the transceiver and control plane a more visible bottleneck.

So over the next 12 months the thing to watch is not whose switch is fastest, but three things: whether the fast-switching white paper defines end-to-end reconfiguration time and sets targets, whether any company discloses cycle-endurance data, and whether use case 1 produces the first deployment outside Google.

The fast-OCS race will not be won by the fastest switch, but by whoever first defines the moment a reconfiguration is complete.

This article is for technology and industry trend analysis only and does not constitute investment advice.