AMD and Intel Both Just Backed the Same Tiny Optical-Networking Startup. Nvidia Didn't.
Salience Labs, a four-year-old photonics startup based outside Oxford, England, closed a $125 million Series B on October 14 — a respectable but hardly record-setting sum in a week when AI infrastructure deals routinely clear ten figures. What makes the round worth a second look is who wrote the checks. AMD Ventures and Intel Capital both joined as new investors, putting two chipmakers that compete head-on for every GPU and CPU socket in the data center on the same cap table, backing the same networking bet. Nvidia, which has scattered money across several rival optical-interconnect startups, was notably absent.
The round, led by Bullhound Capital — a London-based firm that also backs AI chip startup d-Matrix and quantum computing firm Oxford Quantum Circuits — was oversubscribed, according to the company. Other new participants included SBI Group, high-frequency trading firm Hudson River Trading, and the British Business Bank, alongside returning investors Applied Ventures, Oxford Science Enterprises, Cambridge Innovation Capital, and Silicon Catalyst Ventures. Dealroom's funding database puts Salience's total capital raised to date at roughly $167 million, meaning the company had banked only around $40 million across all prior rounds before this one more than tripled its war chest.
A networking bottleneck, not a chip bottleneck
Salience's pitch is that the next constraint on AI training and inference isn't transistor density — it's how many GPUs can be wired together before latency and power consumption make adding more pointless. Today's "scale-up" domains, the tightly coupled clusters of GPUs that act as a single logical accelerator, cap out around 72 chips using conventional copper and electrical switching. Salience builds all-optical circuit switches designed to push that ceiling higher by routing light rather than electrical signals between GPUs, which the company argues cuts power draw and preserves latency as cluster sizes grow. It demonstrated a 64-port version of the switch — which it calls an industry first for silicon-photonics-based optical circuit switching — in September, following an initial 32-port unit earlier in the year, and is already engineering 128- and 256-port successors.
Vaysh Kewada, CEO and co-founder, Salience Labs: "There is a fundamental network limitation on the size of the scale-up domain... if you look at who has actually shown silicon, working silicon, it becomes a dwindling number and a handful."
Why two rivals, and not the market leader
The strategic logic behind AMD's and Intel's checks is narrower than it might look. Both companies trail Nvidia badly in the data-center accelerator market, and Nvidia's NVLink interconnect — a proprietary technology tying its own GPUs together — is a meaningful part of why customers who buy one Nvidia chip tend to buy the whole rack. A credible, vendor-neutral optical switching layer gives AMD and Intel a path to compete on networking without having to replicate Nvidia's closed interconnect stack from scratch, and it lets them split the R&D cost of doing so with a third party rather than each other. Patrick Rundell, a director at AMD Ventures, framed the bet in infrastructure terms: AI systems, he said, are "increasingly being designed at rack scale, where networking has a direct impact on system performance." Intel Capital's Srini Ananth went further, calling photonics "quickly becoming essential to scaling AI infrastructure."
AMD's own roadmap suggests the company is hedging rather than converting outright: it has said its upcoming MI500 GPUs will support both copper and optical connectivity when they ship in 2027, rather than committing fully to light-based networking. That makes the Salience investment look less like a bet the technology has already won and more like insurance against being locked out if it does. Nvidia's absence from the round is conspicuous given that the chipmaker has invested in several of Salience's rivals in the crowded optical-interconnect field, including Ayar Labs and Lumentum; Salience executives note that despite a growing roster of companies chasing optical switching, few besides Salience have shown functioning silicon rather than roadmaps.
The market Salience is chasing
The bet is sized against a market research firm Dell'Oro Group forecasts could be worth close to $1 trillion in cumulative spending on AI back-end data-center switches between 2026 and 2030, with scale-up networking — Salience's specific target — accounting for more than half of that by the end of the decade. That figure helps explain why a $125 million check from two of the industry's largest balance sheets is a rational hedge rather than a speculative flier: even a small share of a trillion-dollar buildout justifies venture-scale bets on multiple unproven approaches at once.
Salience says the new capital will fund faster deployment of its switches, scale up manufacturing with partners Keysight Technologies and Tower Semiconductor, and expand its U.S. footprint with a second office in Silicon Valley, alongside further international expansion. Callum Stewart, a Bullhound Capital principal joining Salience's board as part of the deal, called it backing for "this transformational company" — standard deal-closing language, but one that undersells the more interesting story sitting inside the investor list: when the industry's two also-ran chipmakers quietly agree on where to place a joint bet against the market leader's moat, that alignment is itself a data point about where the AI buildout's next bottleneck is expected to bite.