One hundred nanoseconds. That is the latency Delos Data claims for the data interface at the centre of its new architecture — roughly an order of magnitude below the microsecond a standard network interface card needs to move the same packet.

On September 15, the Palo Alto company said it raised more than $100 million to build it. The investors who signed the round — Matrix, Playground, Socratic Partners, Capricorn's Technology Impact Fund, Matter Venture Partners and IAG, alongside operators from the compute, hyperscale and optical-connectivity industries — are not funding a faster chip. They are funding a wager on where the next AI bottleneck sits.

🎯
The AI bottleneck is moving from compute to interconnect, and private capital is following it.

A fabric built for homogeneous training clusters leaves expensive accelerators idle once inference mixes vendors and memory types.

The durable advantage may belong to whoever keeps heterogeneous chips talking, not to whoever ships the fastest single chip.

What Delos Nonstop AI Actually Does

Delos Data sells a data interface. The company's Nonstop AI architecture composes graphics processors, XPUs, CPUs, memory and storage into one resilient data domain, so each component keeps talking to the others at low latency even when the cluster is a mix of makes.

The product ships in several form factors: a chip for customers designing their own boards, a half-height PCIe card, and a full-height server card. The company also shows a server appliance, the Nonstop AI Server, which accepts up to four third-party graphics cards, and a development and test part the company calls Morpheus.

$100M+ Delos Data funding

September 2026 round

Announced September 15 and led by Matrix and Playground, with capital earmarked for engineering and early deployments. · Reuters, 2026

The engineering claim behind the architecture is narrow and testable. The company says its interface bridges two data semantics — accelerator memory on one side, flash and CPU sockets on the other — without an intermediate translation step. Fewer copies means fewer stalls. Fewer stalls means the expensive part of the cluster spends more of its life doing arithmetic.

The Cost of a Waiting Accelerator

Training taught the industry to count floating-point operations per second. Inference changes the unit of account. A model serving an agent request spends most of its wall-clock time moving state and waiting, not multiplying matrices.

The most expensive idle asset in a data centre is a GPU, CPU or an accelerator waiting on the network. Inference workloads move data in a way that today's interconnect was never designed to serve.— Ed Doe, CEO and co-founder, Delos Data

That framing explains the latency headline. It puts its interface latency around 100 nanoseconds, against microseconds for a conventional network card. If an accelerator can round-trip data ten times faster, it can be shared across more work before it becomes the constraint.

100ns interface latency

Data-interface latency claim

Roughly ten times below the microsecond a standard NIC takes, by the company's own measurement. · Electronics Weekly, 2026

Why can't operators simply add more GPUs?

Adding accelerators helps only if they can be fed. Once a cluster is large enough, the fabric sets the ceiling, and extra chips sit idle behind a saturated link. Spending on interconnect can raise the utilization of hardware already paid for.

A $100M Round Inside an Interconnect Wave

The company is one entry in a wider rotation of capital. Optical-interconnect maker Ayar Labs has raised hundreds of millions of dollars in 2026. Cornelis, spun out of Intel's old Omni-Path networking business, raised $205 million and partnered with Qualcomm. Coherent-DSP startup Celero closed a $275 million Series C, and Olix raised $312 million for a photonic inference platform.

As we wrote in September, Celero's $275 million round rested on the same diagnosis. The bottleneck in AI infrastructure is shifting away from the processor and toward whatever connects processors.

DimensionCompute-firstNetwork-first
Unit of scarcityFLOPsData movement
Best fitHomogeneous trainingMixed inference
Failure modeUnderused acceleratorsFabric lock-in
Who owns the ceilingChip vendorFabric vendor

Comparison drawn from Delos Data and competitor positioning, September 2026

The Delos round is smaller than the largest of those, which shapes how it should be read. A $100 million raise buys engineering time and early customer deployments, not a supply chain. The company says the money will grow its software and hardware teams and accelerate development and sales — the sequence a component vendor needs before volume.

What Has to Be True

Three conditions decide whether the interconnect thesis pays off for the company specifically.

The three conditions the thesis needs

Heterogeneity keeps spreading. Inference clusters must keep mixing vendors and memory types for an independent fabric layer to matter.

Latency beats lock-in. Buyers must weight utilization gains above the simplicity of one vendor's integrated stack.

Delos reaches design wins early. Interfaces freeze, and a startup has a short window to be designed into production clusters before standards settle.

Each condition has a counterweight. Nvidia's own fabric and the Ethernet-based scale-up standards now forming could absorb the problem and leave the independent layer without a market. A hyperscaler weighing a second supplier against operational simplicity may still choose the integrated stack. The company also competes for engineering talent with far better-funded firms.

None of that makes the round irrational. It makes it specific. The case for Delos is the case for a layer that only becomes valuable when AI infrastructure stops being uniform.

Does the interconnect layer become its own market by 2028?

🔮
By 2028, at least one independent interconnect vendor will appear in a top-five hyperscaler's production inference cluster. Horizon: two years.

Probability: 60% — the same utilization math pushing capital into optics and switching applies to whoever owns the data interface.

✅ Arguments for

Interconnect funding is compounding across optics, switching and interfaces.

Inference keeps mixing chip types, which rewards a vendor-neutral fabric.

Confirmation criteria: a named independent fabric in a hyperscaler production cluster.

❌ Arguments against

Integrated vendors can bundle fabric with silicon and close the gap.

Standards bodies may settle the interface before an independent layer scales.

Disconfirmation criteria: scale-up standards adopted without a neutral interface tier.

Scenarios

🟢 Optimistic scenario (25%)

Heterogeneous inference clusters become the default, and a neutral interface becomes standard equipment.

Implications: The company and its peers graduate from component vendors to a distinct layer.

🟡 Base-case scenario (55%)

Interconnect vendors win slots inside mixed clusters while integrated stacks keep the largest accounts.

Implications: A real but contested market, with pricing pressure from bundled silicon.

🔴 Pessimistic scenario (20%)

Standards freeze around incumbent interfaces and buyers default to one vendor's stack.

Implications: Independent fabric startups are absorbed or become features.
📊
Key signals to track

Design wins named in hyperscaler inference clusters

Whether scale-up standards add a neutral interface tier

Follow-on rounds at higher valuations across the interconnect group

Utilization figures published by large inference operators
Delos Data, a chip startup founded by Intel veterans, raises $100 million for AI networks
Reuters' account of the round and the founders' diagnosis that data movement, not compute, is the emerging constraint.
Reuters
Primary news source for the funding event, used for the round size, investors and the Intel-veteran founding team.
Delos raises $100m for faster datacentre networking ICs
Trade-press detail on the Nonstop AI approach and the 100-nanosecond latency claim against a conventional network card.
Electronics Weekly
Carries the technical framing and the direct quote from the chief executive that anchors this analysis.
Delos Data Raises $100M+ to Attack Agentic AI Data Bottleneck
Specialist networking coverage that places the round inside the wider interconnect wave alongside Ayar Labs, Cornelis and Celero.
Converge Digest
Used for the competitive context that shows the company is one of several bets on the same bottleneck.