Salience Labs raises $125M for all-optical AI data center switches
The AI buildout's bottleneck moved from the processor to the wire, and today's money followed it: a British photonics startup pulled in a nine-figure round with both Intel and AMD on the cap table, while two smaller bets went to the less glamorous end of the agent economy.
Salience Labs raised an oversubscribed $125 million Series B led by Bullhound Capital, with Intel Capital and AMD Ventures both participating — two chipmakers that compete for the same accelerator sockets, funding the same switch. SBI Group, Hudson River Trading and the British Business Bank joined as new investors, and Applied Ventures, Oxford Science Enterprises, Cambridge Innovation Capital and Silicon Catalyst Ventures returned. The Oxford company builds optical circuit switching for AI data centers: instead of converting light to electricity and back at every hop, its silicon-photonics switch establishes a direct optical path between endpoints, cutting the conversion overhead and power draw that racks of electrically switched fabric impose. Salience demonstrated a 64-port all-optical switch at ECOC 2026 in September, which it calls an industry first for silicon photonics, and says 128- and 256-port versions are in development. A 32-port product, the OC-32M, shipped earlier this year with spec-sheet figures of 10-nanosecond latency and under one watt per port — company numbers, not independently measured.
The strategic read is in who signed the check. Intel and AMD are not optical networking vendors; they sell the accelerators whose clusters stall when the fabric can't feed them. Once thousands of GPUs have to exchange data on every step, the network stops being plumbing and starts setting how much of the hardware's throughput a customer can actually bill for — which is why the round's language is about token economics, not bandwidth. It also lands in the same slipstream as every other interconnect raise of the past month, including the optical-link money we covered when Ayar Labs added $150 million as copper interconnects hit the wall. Caveats worth keeping: nobody has published independent numbers on the 64-port switch, and neither Intel nor AMD has committed to buying one — an investment is not a design win.
Harmonic Security raised a $50 million Series B led by Crosspoint Capital Partners, taking its total funding to $76 million. The San Francisco and London company sits between employees and their AI tools: small language models inspect each interaction in context and return a verdict in milliseconds, on the theory that a static rule set can't tell a developer pasting a database schema from the same developer pasting live credentials. It says more than 100 companies are customers and it has expanded scope from browser-based chatbots to desktop AI apps and agentic workflows. The company's previous round was a $17.5 million Series A led by Next47 in October 2024 — a roughly threefold step up on the size, though no valuation was disclosed this time.
Groundcover launched what it calls agent-to-agent observability, letting outside coding agents — Claude, Cursor, Kiro and Codex — reach into production telemetry. Agents that know which evidence they need can pull logs, traces and Kubernetes events through the company's Model Context Protocol tools; agents that only know something is broken hand the job off as an asynchronous task, get a job ID back immediately, and retrieve a root cause analysis or a draft dashboard later. It runs on the company's bring-your-own-cloud architecture with node-based pricing, so retaining more data at higher fidelity doesn't inflate the bill. Groundcover has raised $160 million to date, including a $100 million Series C led by One Peak in July, and the capability shipped earlier this year — today's news is the packaging across three surfaces, not new technology. The bet underneath it is the interesting part: if every coding agent needs a production-context provider, the observability vendor stops being a dashboard and becomes shared memory for the whole agent fleet.
What to watch: whether any hyperscaler freezes an interconnect choice around optical switching before the 128- and 256-port parts prove out. Cluster designs lock years ahead of the qualification runs.
If the network is now what decides how much of a GPU bill a customer can justify, should labs and hyperscalers be judged on accelerator counts at all? Tell us in the comments.
