Delos Data Raises $100 Million to Fix the Networking Bottleneck Inside AI Data Centers

Founded by Intel veterans, Delos Data raised $100 million to build chips and software designed to move data faster across increasingly diverse AI compute clusters.

The most expensive AI chips are useful only when data can move between them quickly enough.

That networking problem is becoming more complicated as data centers adopt a wider mix of accelerators. Reuters reported on September 15 that Delos Data, founded by Intel veterans, raised $100 million to develop networking chips and software for AI infrastructure.

As AI hardware diversifies, the network connecting accelerators may become as strategically important as the accelerators themselves.

From uniform clusters to mixed compute

The first phase of the generative AI boom was heavily centered on Nvidia GPUs and Nvidia networking. The market is now becoming more heterogeneous.

AMD and specialized accelerator companies are gaining attention, while different workloads may favor different hardware. Inference for agents can also create traffic patterns that differ from large training jobs.

Delos Data is betting that data-center operators need networking technology flexible enough to connect this changing mix of compute.

Data movement is expensive

Modern AI systems split work across many chips. If those chips spend too much time waiting for data, expensive hardware sits underutilized.

That wastes both energy and capital. Faster interconnects can increase the effective performance of an entire cluster without changing the underlying processors.

In AI infrastructure, utilization can matter almost as much as raw chip performance.

The startup opportunity

Delos Data’s round included investors such as Matrix Partners and Playground, according to Reuters.

The company enters a market where incumbents have deep hardware expertise, but rapid architecture changes can create openings for new designs.

Why developers should care

Most application developers never interact directly with data-center networking. But infrastructure efficiency eventually affects the price and availability of AI APIs.

If providers can use accelerators more efficiently, inference costs can fall and model capacity can become easier to scale.

What to watch

The key question is whether Delos can turn its technology into production deployments inside major AI clusters.

Hardware startups face long qualification cycles, manufacturing complexity and strong incumbents. But the size of AI infrastructure spending means even a narrow bottleneck can support a substantial business.

The next AI infrastructure breakthroughs may come not only from faster chips, but from making thousands of different chips work together efficiently.