The pattern is by now familiar in artificial intelligence: a startup working on an unglamorous layer of the stack raises money from some of the most glamorous names in finance. Bloomberg reported Tuesday that Celero Communications has completed a $275 million funding round at a valuation above $3 billion, with Atreides Management, Valor Equity Partners and Alphabet’s CapitalG leading the investment.
Celero’s business is the connective tissue of AI computing. The company makes high-speed interconnect chips, the components that move data between the processors inside a data center, the wiring that the industry’s metaphors describe as the capillaries feeding the brain of an AI cluster. The product is less visible than the accelerators themselves, but no AI training run happens without it.
The funding reflects a shift in where the industry’s bottlenecks have moved. For the first years of the AI boom, the constraint was the accelerators themselves, the graphics processors that perform the calculations behind large models, and companies that secured them had an advantage. That constraint has eased as production has scaled, and the industry has discovered that moving data between thousands of processors is as hard as doing the calculations.
The technical problem is unforgiving. Training a large model requires constant communication between the processors in a cluster, with each stage of the work depending on data from every other stage. The connections between chips, once an afterthought in data center design, now determine how fast a cluster can run and how much of its expensive computing power sits idle waiting for data. Interconnect performance has become a direct multiplier on the economics of AI.
The market’s response has been a wave of investment. Networking specialists that once served the telecom and enterprise markets have found their products in demand from AI builders, and incumbents with decades of interconnect experience, such as Broadcom and Marvell, have seen their AI businesses grow rapidly. The funding round for Celero shows the appetite extending to startups attacking the problem with new architectures.
The competitive field is unusually structured. Nvidia, the dominant maker of AI accelerators, has built its own interconnect into its systems, using proprietary links that tie its processors together with performance competitors struggle to match. The industry’s response has been an effort to create open standards, with consortiums of chip makers and cloud providers developing alternatives that would let customers mix hardware from different suppliers. The outcome of that standards war will determine the shape of the interconnect market for years.
Celero’s positioning in that contest was not detailed in the announcement, and the company has been selective about its technical disclosures. What the funding round says is that investors believe there is room for new entrants in a market where the largest customers are desperate for alternatives to dependence on a single supplier. The presence of CapitalG, Alphabet’s investment arm, signals that the computing giants see the interconnect layer as strategically important enough to back challengers.
The financial backers represent an unusual mix. Atreides Management, the investment firm of venture capitalist Gavin Baker, has become one of the most active technology investors of the AI era, while Valor Equity Partners, known for its work with growth companies, and CapitalG bring both capital and connections. Their participation values Celero at more than $3 billion before the company has established the kind of revenue base that such valuations usually require, a bet on the market rather than the financials alone.
The money will go toward the industry’s most expensive problem: developing chips that work with the next generations of accelerators. Interconnect technology must be designed in step with the processors it serves, anticipating the speeds and architectures of hardware that has not been built. The engineering cycle is long, the customers are few and exacting, and the penalty for a design that does not align with the industry’s direction is total.
The startup’s investors are also betting on a structural feature of the market: concentration. The number of customers that buy interconnect at scale is small, dominated by the handful of companies that build the largest AI clusters, and serving them well can generate enormous revenue from a short customer list. That is the model that made the incumbents valuable, and the newcomers are being funded on the assumption that the largest builders want enough suppliers to keep their options open.
The investment wave in interconnect reflects a broader realization about the AI buildout. The industry has spent trillions of dollars of planned capital on processors, data centers and power, and the value of that spending depends on systems that can actually work together. The companies that make the pieces between the famous components have become indispensable, and their investors are betting that indispensability translates into durable revenue. Whether Celero becomes one of those companies will depend on the standards contest and the quality of its engineering, but Tuesday’s round is evidence that the market for the industry’s plumbing has arrived.
