Skip to content
Today
Sections

Build Independently confirmed2 publishers 3 min read Published Updated

Delos Data raises more than $100M to keep inference running when a link drops

Delos Data's Nonstop AI Reference Architecture is meant to hold a mixed domain of GPUs, CPUs, memory and storage together through hardware and link failures. The 10x figures attached to it describe silicon the company still calls planned.

The Engineer · Build desk

Send a correction

What happened

  • Delos Data of Palo Alto says it has raised more than $100 million, and it launched its Nonstop AI Reference Architecture on September 15th with the stated goal of keeping inference running through hardware and link failures.
  • The announced set is Nonstop AI Clusters, a disaggregated Nonstop AI Server that mixes processors and switches, the Nonstop AI Data Interface and Mosaic software.
  • The Data Interface, the piece meant to link compute, memory and storage, is planned silicon, and Delos Data credits it with 10x lower latency and 10x higher efficiency.
  • Named investors include Matrix, Playground Global, Socratic Partners, Capricorn's Technology Impact Fund, Matter Venture Partners and IAG, plus unnamed backers from the computing and networking industries.
  • A May 27th, 2026 release described 1,000-GPU scale-up domains as practical and 10,000 GPUs as potentially possible. No independent party has validated either figure.

Why it matters

  • Who has to decide decision A buyer weighing the Data Interface is pricing a design target for a part not yet built, so the evaluation has to wait on delivery of a benchmark with a stated baseline.
  • What it limits constraint Delos Data does not sell the accelerators or the models, so its revenue depends on customers agreeing to buy a separate data path between parts they already source elsewhere.
  • Who's exposed exposure With the press release as the sole source for the money and the product, anyone repeating the 10x figures or the 300x demand projection is passing along a company number that no test configuration or named author stands behind.

A stateless request that meets a broken link costs one retry. Delos Data defines agentic inference as persistent workloads that run across multiple devices and do not stop after one request [12]. Those workloads hold state in the memory of the devices they occupy. A failed device or a failed link takes part of that state with it, so keeping the job alive means the surviving devices can still reach or reconstruct what the dead one held [3].

The 10x latency figure describes a part Delos Data calls planned silicon [14]. That is a design target, and design targets do not automatically transfer. For the number to say anything about a buyer's cluster, the baseline has to be named: latency of which path, measured between which endpoints, at which message size, against which interface in place today. Efficiency needs a unit too. Delos Data's press release is the primary source for the financing and the product claims. It gives no test configuration, no independent benchmark and no named production customer, and it does not say how much was raised, how the round was structured, what the company is worth or who led it [6][17].

The stated benefits of the cluster and server designs are reduced cost per token, improved tokens per watt and increased tokens per second [19]. All three move with the model, the batch size and the sequence length. Comparing them across two vendors means running the same workload on both.

The financing announcement cites an unattributed industry projection of roughly 300x growth in token demand by 2030, with agentic inference accounting for most of the increase [16]. Take 2026 as the baseline and that works out to about 4.2x a year compounded, since 300 to the one-fourth power is 4.16 [21].

"The most expensive idle asset in a data center is a GPU, CPU or an accelerator waiting on the network," Doe said in the September 15th announcement [7]. Doe ran Barefoot Networks, which Intel acquired in 2019, and Barefoot's programmable Tofino Ethernet switch became part of Intel's networking portfolio [8]. Daly came into Intel through its 2011 acquisition of Fulcrum Microsystems, where he was technical director of software and systems and worked on software control planes for high-speed, low-latency switch hardware, according to an Open Networking Foundation biography [9]. Control-plane software for low-latency switching sits close to the problem of holding a distributed job together when a link goes away, and both founders have spent much of their careers on systems that move data between processors [10].

Wen Hsieh, founding managing partner at Matter Venture Partners, described Delos Data as having built software, followed by an AI server [15]. The capital goes to expanding software and hardware engineering teams and to product development and sales [5]. Until the Data Interface silicon exists, the testable claim is the software and cluster architecture: whether a mixed domain of accelerators, memory and storage keeps a persistent inference job running when a device or a link fails [13][11].

What to watch

  • A published test configuration and baseline for the Data Interface's 10x latency and efficiency figures.
  • A named production customer running a persistent inference job through a device or link failure on Nonstop AI Clusters.
  • Any demonstration of the 1,000-GPU scale-up domain by a party other than Delos Data.

Clarity's read

What the record supports and how the coverage leans. The claims behind it follow.

Reality

Evidence42
Adoption15
Hype gap+34
Incentives82
Confidence55
Why these scores

Claim ledger

Ranked by verification strength, evidence, and original report placement.

  1. [1]

    Palo Alto, California-based Delos Data says it has raised more than $100 million.

    Reported Supported Source: Delos Data press release, via runtimewire.comView cited source
  2. [2]

    Delos Data launched its Nonstop AI Reference Architecture on September 15th.

    Reported SupportedView cited source
  3. [3]

    Delos Data says its Nonstop AI Reference Architecture lets customers connect a mixture of GPUs, CPUs, other accelerators, memory and storage within one data domain while keeping inference workloads running through hardware and link failures.

    Reported SupportedView cited source
  4. [4]

    Backers include Matrix, Playground Global, Socratic Partners, Capricorn's Technology Impact Fund, Matter Venture Partners and IAG, along with unnamed investors from the computing and networking industries.

    Reported SupportedView cited source
  5. [5]

    Delos Data plans to use the capital to expand its software and hardware engineering teams and accelerate product development and sales.

    Reported SupportedView cited source
  6. [6]

    Delos Data's press release is the primary source for the financing and product claims, and the available sources do not identify the exact financing amount, round structure, valuation, lead investor or total funding to date.

    Reported SupportedView cited source
  7. [7]

    "The most expensive idle asset in a data center is a GPU, CPU or an accelerator waiting on the network," Doe said in the September 15th announcement.

    Reported Supported Source: Ed Doe, Delos Data co-founderView cited source
  8. [8]

    Doe held senior networking roles at Intel, including leadership of Barefoot Networks, which Intel acquired in 2019; Barefoot's programmable Tofino Ethernet switch technology became part of Intel's networking portfolio.

    Reported SupportedView cited source
  9. [9]

    An Open Networking Foundation biography of Dan Daly says he joined Intel through its 2011 acquisition of Fulcrum Microsystems, having joined Fulcrum in 2001 as technical director of software and systems, working on software control planes for high-speed, low-latency switch hardware.

    Reported Supported Source: Open Networking Foundation biographyView cited source
  10. [10]

    Doe and Daly have spent much of their careers working on systems that move data between processors.

    Reported SupportedView cited source
  11. [11]

    Delos Data's position is that the processors, models and storage systems can vary while Delos Data supplies the architecture that connects them and keeps work moving when a component fails.

    Reported SupportedView cited source
  12. [12]

    Agentic inference, in Delos Data's framing, consists of persistent workloads that run across multiple devices instead of handling one request and stopping.

    Reported SupportedView cited source
  13. [13]

    Delos Data's announced product set includes Nonstop AI Clusters, a resilient cluster architecture for agentic inference; Nonstop AI Server, a disaggregated scale-up server architecture intended to let customers mix processors and switches; Nonstop AI Data Interface; and Mosaic software.

    Reported SupportedView cited source
  14. [14]

    Nonstop AI Data Interface is planned silicon for linking compute, memory and storage.

    Reported SupportedView cited source
  15. [15]

    In the September financing announcement, Matter Venture Partners founding managing partner Wen Hsieh described Delos Data as having built software, followed by an AI server.

    Reported Supported Source: Wen Hsieh, Matter Venture PartnersView cited source
  16. [16]

    The funding announcement cites an industry projection of roughly 300x growth in token demand by 2030, with agentic inference accounting for most of that increase; the announcement does not identify the projection's author.

    Reported InsufficientView cited source
  17. [17]

    Delos Data says its Data Interface delivers 10x lower latency and 10x higher efficiency; the reports include no independent benchmark, test configuration or named customer result supporting those figures.

    Reported InsufficientView cited source
  18. [18]

    A May 27th, 2026, Delos Data release announced the cluster architecture and server design, which Delos Data says are intended to scale across thousands of GPUs, CPUs and accelerators.

    Reported InsufficientView cited source
  19. [19]

    Delos Data listed reduced cost per token, improved tokens per watt and increased tokens per second as intended benefits.

    Reported InsufficientView cited source
  20. [20]

    Delos Data material described 1,000-GPU scale-up domains as practical and 10,000 GPUs as potentially possible, and those figures have not been independently validated.

    Reported InsufficientView cited source
  21. [21]

    Roughly 300x token demand growth by 2030 works out to about 4.2x a year compounded if the projection's baseline year is 2026.

    Derived Insufficient

Sources

2 independent publishers whose own reporting we read for this story.

  1. mezha.net

    1 article · September 15, 2026

    Стартап ветеранів Intel залучив $100 млн для прискорення обміну даними в ШІ-дата-центрах
  2. runtimewire.com

    1 article · September 15, 2026

    Delos Data raises $100M-plus to keep AI inference running through failures

Topics and entities

Topics

Entities