Salience Labs: the $125M push from optical-switch demos to production

Salience Labs has announced a $125 million Series B to scale optical circuit switches for AI data centres. The important change is funding for manufacturing and deployment after a 64-port demonstration, rather than proof that the whole roadmap is ready for production. For infrastructure teams, the next question is how an optical path fits the complete network and workload.

The company's announcement is dated October 14, 2026, the same day as The Next Web's report. Bullhound Capital led the round, with AMD Ventures and Intel Capital among the participants. The British Business Bank separately confirms its £10 million investment. The release announces completion without separately dating the legal closing.

Three stages of optical-switch development

An optical circuit switch establishes paths for light between ports. It addresses data movement between computing resources, rather than performing the model's calculations itself. Salience's public material describes several stages that should remain distinct:

• OC-32M: A 32-port module and rack-mounted evaluation platform described on the product page

• 64-port OCS: A demonstration at ECOC 2026, confirmed in the funding announcement

• 128- and 256-port OCS: Future solutions that the new funding is intended to enable

The September 20 event notice preceded ECOC's September 21–23 exhibition in Málaga. Demonstrating a larger switch is a technical milestone; it does not by itself establish production volume, delivery dates or customer deployment results. A port count also needs a topology before it can be translated into a number of connected GPUs.

Path latency and reconfiguration measure different operations

Salience's OC-32M specification lists 10 nanoseconds of input-to-output latency and less than 300 microseconds to reconfigure the switch. Both are manufacturer specifications. The first describes passage through a selected path; the second concerns changing the connection arrangement. Treating them as interchangeable would misstate the design.

For a workload using a stable optical path, the transit figure helps characterize that part of the network. For a design that frequently reallocates connections, the reconfiguration interval and the work coordinating that change also matter. Neither number alone measures a complete request's latency or tokens delivered to a user.

The page also lists power below one watt per port. That is a component-level claim, not the electricity consumption of a complete rack. TechKili has not benchmarked these figures or measured end-to-end AI performance.

Scaling manufacture is a separate engineering task

The funding builds on a February 25 agreement with Tower Semiconductor, which described moving from development into pre-production using Tower's silicon-photonics platforms. The October release adds resources for volume manufacture and a larger Silicon Valley presence, including a planned second West Coast location.

For an infrastructure evaluation, useful follow-up evidence would connect the switch to a documented topology, link requirements, traffic pattern and full-system measurements. Manufacturing output and qualified deployment results would answer different questions from a demonstration or financing total. The round makes that next stage better funded; it does not supply those results in advance.

Sources

• Salience Labs' October 14 funding announcement

• British Business Bank's investment confirmation

• OC-32M product and evaluation platform

• 64-port ECOC demonstration notice

• Tower's February manufacturing partnership

• The Next Web's October 14 report