Ex-Intel Veterans Raise $100M to Fix AI Data Centers

The most expensive piece of equipment in an AI data center might be the one that never does anything at all.

"The most expensive idle asset in a data center is a GPU, CPU, or an accelerator waiting on the network," says Ed Doe, chief executive of Delos Data. His company, founded by Intel veterans and backed by former Intel CEO Pat Gelsinger, just raised more than $100 million on a thesis that sounds almost heretical in Silicon Valley: in the age of AI, the processor is no longer the bottleneck. The network is.

Delos Data, a Palo Alto startup, announced Tuesday that it had closed the round to build chips and software that move data faster inside AI data centers. Backers included Matrix Partners, Playground Global, Socratic Partners, Capricorn's Technology Impact Fund, Matter Venture Partners, and IAG, plus industry veterans from compute, hyperscale systems, networking silicon, and optical connectivity.

It is a bet that the AI infrastructure race is entering its second act — and that the winners of the first act, the chipmakers, are about to hand the stage to the companies that connect them.

An original TechPulse USA editorial illustration of a vast AI data center hall at night, with glowing processor chips linked by brilliant light streams forming a luminous network fabric. Credit: Original illustration by TechPulse USA.

The raise: over $100 million for the network, not the compute

Reuters reported Tuesday that the San Francisco–datelined round brings Delos Data's funding to more than $100 million. The company said it will use the capital to grow its software and hardware engineering teams and accelerate product development and sales.

What makes the raise notable is not just its size but its investor list. Matrix Partners led the validation from the venture side, with general partner Stan Reiss framing the company as "attacking the part of the AI stack where the value is now getting decided." Playground Global — the VC firm where former Intel CEO Pat Gelsinger is now a general partner — also joined, a reunion of sorts: the Delos Data founders spent their formative years at Intel, the company Gelsinger once ran.

Who they are: Intel veterans building for the post-training world

Delos Data is led by CEO and co-founder Ed Doe and CTO and co-founder Dan Daly, both veterans of Intel. The company describes itself as an AI inference company — and that word, inference, is doing a lot of work in the pitch.

For the first phase of the AI boom, data centers were built almost exclusively for training: cramming thousands of Nvidia GPUs together with Nvidia's proprietary networking technology and letting them chew through internet-scale datasets. That world is still enormous. But the industry's center of gravity has shifted toward agentic AI — systems that don't just answer one question but chain together dozens of reasoning steps, tool calls, and model invocations on behalf of a user.

Those workloads move data in fundamentally different ways. "Inference workloads move data in a way that today's interconnect was never designed to serve," Doe said. "This funding positions the company to expand our team and offerings to solve the data transfer problem holistically by optimizing the network as the core of the AI system."

Nonstop AI: 10x lower latency, 10x higher efficiency

Delos Data's answer is a product family it calls Delos Nonstop AI™, built around a new Nonstop AI Reference Architecture. The lineup announced alongside the raise includes:

• Delos Nonstop AI™ Data Interface — a data-movement interface the company claims delivers 10x lower latency and 10x higher efficiency than existing approaches;
• Delos Nonstop AI™ Clusters — composable cluster hardware announced earlier this year;
• Delos Nonstop AI™ Server — the server building block, also announced earlier this year.

The architecture's headline promise is flexibility: customers can compose clusters from their own mixture of hardware, models, and topologies — Nvidia chips alongside AMD, Cerebras, or anything else — all operating together in one domain. In an industry terrified of vendor lock-in, that is a feature designed to be heard.

Performance claims of "10x" should always be treated as company-reported until independently benchmarked. But the direction of travel is well-documented: as AI workloads shift from training to inference, the fraction of time chips spend waiting for data — not computing — has become the industry's open wound.

Gelsinger's verdict: "The network is the AI"

The most quotable validation came from Gelsinger, whose journey from Intel CEO to venture partner at Playground Global has made him one of the most watched figures in semiconductors. Speaking about the investment, he put it bluntly:

"I can have all the computers in the world, and if I can't get these devices communicating effectively with one another, I've got lots of hot hardware, burning watts and sitting here twiddling their thumbs," Gelsinger told Reuters.

In the company's announcement he went further: "Adding compute to an inference problem is not enough. AI and inference performance is a network problem now — the network is the AI. Delos started from the workload and let it dictate the architecture all the way down to the interface silicon. That discipline is what turns an interconnect idea into infrastructure the industry can actually deploy."

It is a remarkable line from a man who spent four decades in the compute business. And it echoes a line of thinking gaining traction across the industry: that the next trillion dollars of AI infrastructure value will accrue not to whoever makes the fastest chip, but to whoever keeps those chips fed.

Why now: the data center is getting complicated

Delos Data's timing rests on three converging trends that Reuters and other outlets have documented this week:

First, heterogeneity. Early AI data centers were essentially Nvidia monocultures. Today, Cerebras Systems and AMD have gained real traction, Nvidia itself now offers multiple kinds of computing chips, and enterprises run a mix of accelerators. Someone has to make that mix communicate at full speed.

Second, the agentic shift. "We don't know what the next infrastructure architecture is going to be for agentic AI," CTO Dan Daly told Reuters. "But we do know that we can provide the quickest, fastest way to move that data around." Agentic workloads scale across devices in ways that stress interconnects far more than training jobs did.

Third, the spending question. Wall Street expects AI data center spending to approach $800 billion in 2026, and investors are growing nervous about whether that pace holds. When capital gets cautious, efficiency stops being a nice-to-have: squeezing more output from existing hardware becomes the industry's priority, and that is exactly what faster interconnects promise.

An original TechPulse USA editorial illustration — close-up of a futuristic silicon chip with golden circuit traces and light streams racing through glowing interconnect bridges. Credit: Original illustration by TechPulse USA.

The competition: a crowded, high-stakes lane

Delos Data is not entering an empty field. Nvidia's proprietary networking stack — built around its InfiniBand and Spectrum products — is the incumbent, and Nvidia has every incentive to keep it that way. The Ultra Ethernet Consortium, backed by a who's who of the industry, is pushing open networking standards. European startups like Axelera AI, which announced new AI-factory supply deals this week, are attacking adjacent inference problems.

Delos Data's differentiator, as Reiss put it, is starting from the workload rather than the silicon: "For years the assumption was that more compute buys more output. That relationship has broken down at inference scale, and the team that fixes the data transfer problem captures an enormous amount of the value the industry is currently leaving on the floor."

The open question is go-to-market. Interconnect technology lives or dies on adoption by hyperscalers and large enterprises — a long, relationship-driven sales cycle. The company says the new funding will accelerate both product development and sales, a pairing that suggests it knows the technology is only half the battle.

The money trail: why $100M says the thesis is investable now

Venture capital has a way of narrating its own history: first it funded the model labs, then the cloud GPU providers, then the application layer. The Delos Data round belongs to a newer chapter — the unglamorous infrastructure that determines whether the previous chapters' investments actually pay off.

The investor composition tells that story. Matrix Partners brings classic enterprise-infrastructure investing muscle; Playground Global brings deep semiconductor expertise and Gelsinger himself; the participation of specialist funds like Capricorn's Technology Impact Fund and Matter Venture Partners, plus what the company describes as industry luminaries from compute, hyperscale, networking silicon, infrastructure software, and optical connectivity, reads like a deliberate assembly of every discipline the problem touches.

It is also a sign of how the AI funding landscape has matured. Two years ago, $100 million for interconnect technology would have been a curiosity next to the multi-billion-dollar model raises. Today, with AI capex nearing $800 billion a year and investors questioning whether the spending will continue, efficiency technology is the growth story — the thing that makes the rest of the spend defensible.

Whether Delos Data becomes a category leader or a strategic acquisition target will depend on execution, but the round itself is a data point: the smart money is now flowing to the plumbing, not just the palaces.

What to watch next

• Customer proof: The company has not yet named deployment customers. Hyperscaler design wins would transform this from a thesis into a trajectory.
• Independent benchmarks: The 10x latency and efficiency claims need third-party verification before enterprises will bet data centers on them.
• The Nvidia question: Open, mix-and-match networking is a direct challenge to Nvidia's vertically integrated stack. How the market's 800-pound gorilla responds will shape everyone's timeline.
• Agentic workloads: If AI agents become the dominant data center workload — as Daly and others argue — the interconnect thesis gets stronger every quarter.

The AI industry spent four years learning that compute is king. Delos Data's $100 million raise is a bet that the crown is passing — to the wires between the kings.

---

Sources: Reuters (Sept. 15, 2026, reporting by Stephen Nellis), FinSMEs (Sept. 15, 2026), company announcement. Performance figures (10x latency/efficiency) are company-reported.