Confirmed in a primary source 14 Sept 2026 Infrastructure

An Intel spin-out raised 205 million dollars and teamed up with Qualcomm to attack the part of Nvidia's business nobody has copied.

Cornelis Networks, the company built from Intel's Omni-Path networking business, announced on Monday that it has raised about 205 million dollars and launched an architecture it calls Active Compute Fabric, which puts programmable computing inside the network itself rather than treating the network as plumbing between chips. Alongside it, Qualcomm announced a collaboration on rack-scale inference, with Qualcomm's data-centre chief Tony Pialis appearing with Cornelis chief executive Lisa Spelman at the AI Infra Summit in Santa Clara.

The point of attack is specific: Nvidia's hardest advantage is not the chip, it is NVLink, the private network that makes a rack of chips behave as one machine. The caveat is that this is early. The current switch, CN5000, ships now; the next one, CN6000, is only sampling with customers and broadens in the fourth quarter, and the 50 percent traffic reduction Cornelis cites comes from pre-production simulation, not from a customer's data centre.

Until now

Challenges to Nvidia were framed as challenges to the accelerator, and the rack-level interconnect that makes many chips behave as one machine had no credible open alternative with a large backer attached.

Now

Treat the interconnect as the contested layer. An open standards stack, UALink and ESUN, now has funded silicon behind it and a top-five chip company doing joint validation, which is the first coordinated attempt on the part of Nvidia's position that is not the chip.

The argument every challenger to Nvidia has lost is not about the chip. AMD has a competitive accelerator and Google, Amazon and now Qualcomm build their own. What none of them has had is the thing that turns 72 chips into one machine, which is Nvidia's private interconnect, and this is why a rack of Nvidia hardware is bought as a rack rather than assembled from parts. Cornelis Networks, which is what is left of Intel's Omni-Path business after Intel sold it in 2020, raised about 205 million dollars on Monday and announced a design that goes at exactly that layer, with an unusual idea attached. Instead of moving data faster between chips, it does work on the data while it is moving: assembling cache, routing tokens to the right expert in a mixture-of-experts model, and performing the AllReduce step inside the switch rather than making the chips do it. The company's own description is that the payload does not arrive the way it left.

What makes it more than a funding round is the company standing next to it. Qualcomm has spent this year moving into data-centre inference, carded here on September 8 when it announced custom silicon work with Amazon, and its data-centre chief is on stage with Cornelis on Tuesday to talk about rack-scale inference and joint validation. A chip vendor without an interconnect and an interconnect vendor without a chip have an obvious reason to find each other, and the standards they are building to, UALink and ESUN, are the open answer to NVLink. The honest limits are that the CN6000 is still sampling, the 50 percent traffic figure comes from simulation rather than a deployment, and no joint product, volume or date has been named. Treat the interconnect as the contested layer: an open standards stack now has funded silicon behind it and a top-five chip company doing joint validation, which is the first coordinated attempt on the part of Nvidia's position that is not the chip.

How it got here

2020
Intel sells its Omni-Path Architecture networking business; Cornelis Networks is formed around it.
Sep 08
Qualcomm and Amazon announce custom silicon work for data-centre inference, carded in Edition 016 with the falsifier that neither has disclosed deployed capacity or a named instance family.
Sep 14
Cornelis announces about 205 million dollars of funding led by IAG Capital Partners and launches Active Compute Fabric, built to UALink and ESUN, with the CN5000 shipping and the CN6000 sampling for fourth-quarter availability.
Sep 14
Qualcomm and Cornelis announce a rack-scale inference collaboration and joint validation, with Tony Pialis and Lisa Spelman keynoting together on September 15. The same day the semiconductor index has its worst session since July.
Who has to react
Nvidia, whose rack-level lock-in is the target and whose shares fell 3.36 percent the same day; Broadcom and Marvell, who sell the switching silicon this competes with; the UALink and ESUN consortium members, who now have a funded implementer to point at; Amazon, already working with Qualcomm on inference silicon and short an interconnect; and every hyperscaler negotiating a 2027 rack purchase, who has just acquired a second thing to ask about.

The evidence

ObservedCornelis Networks announced approximately 205 million dollars of new funding on September 14, 2026, with IAG Capital Partners identified by SiliconANGLE as the lead investor; the round letter, valuation and full syndicate were not disclosed. The announcement coincided with the AI Infra Summit in Santa Clara, which runs September 15 to 17.

The whole observation, as recorded

The product is the Active Compute Fabric, described as an open architecture that integrates programmable compute into the network fabric so that operations are performed on data in transit: assembling key-value cache data, coordinating expert dispatch for mixture-of-experts models, and accelerating collective operations such as AllReduce. The company's chief marketing officer summarised it as the payload not arriving the way it left. Pre-production simulations are cited as showing up to a 50 percent reduction in network traffic; no customer measurement, bandwidth figure or latency figure was published. Product status: the CN5000 switch is shipping now and the CN6000 is sampling with customers with expanded availability expected in the fourth quarter of 2026. On standards, HPCwire reports Cornelis building its scale-up interconnect around UALink and ESUN, the open alternatives to Nvidia's proprietary NVLink. On Qualcomm: the two companies announced a collaboration covering rack-scale inference across scale-up and scale-out environments, with Tony Pialis, executive vice president and general manager for data centre at Qualcomm Technologies, joining Cornelis chief executive Lisa Spelman for a September 15 keynote, and the companies describing joint validation work toward future rack-scale AI data-centre designs. No contract value, volume commitment or shipping date for a joint product was disclosed. Spelman, quoted: AI infrastructure is reaching a point where faster endpoints alone are not enough, and the fabric has to become an active part of the compute system. Cornelis is the company formed from Intel's Omni-Path Architecture business, sold in 2020. Context on the same day: the Philadelphia semiconductor index had its worst session since July.

Established

Cornelis Networks announced about 205 million dollars of funding on September 14, 2026, with IAG Capital Partners reported as lead investor.
It launched the Active Compute Fabric architecture, performing operations on data in transit including key-value cache assembly, expert dispatch and AllReduce acceleration, and says its scale-up interconnect is built to UALink and ESUN.
Qualcomm and Cornelis announced a collaboration on rack-scale inference, with Qualcomm's data-centre general manager Tony Pialis keynoting alongside Cornelis chief executive Lisa Spelman on September 15.
The CN5000 switch is shipping and the CN6000 is sampling with customers, with expanded availability expected in the fourth quarter of 2026.

Not established

That the performance claim holds. The 50 percent network traffic reduction comes from pre-production simulation, with no customer deployment, bandwidth figure or latency figure published.
That Qualcomm has committed anything. No contract value, volume, revenue share or joint product date was disclosed, and joint validation is not a purchase order.
That any hyperscaler will buy it. No named customer, design win or deployment was announced.
That this reaches Nvidia's position. Nvidia ships integrated racks at volume today; this is a component vendor and a chip partner with a sampling product.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityModerate

PersistenceYear

Cost is medium. About 205 million dollars is real money committed by investors who can lose it, and building switching silicon is capital-intensive and slow, but this is a private round of undisclosed structure rather than a public commitment by a large company, and no valuation was given. Commitment is medium: Cornelis is fully committed, because the funding and the product are its whole business, but Qualcomm's side is a collaboration and joint validation with no contract value, volume or product date attached, which is the cheapest form of partnership a chip company can enter. Reversibility is moderate; the money is spent and the silicon is taped out, but a collaboration announced from a conference stage can end without anyone issuing a statement. Persistence is a year, tied to the CN6000 availability window in the fourth quarter and to whether a rack-scale design appears in the 2027 buying cycle.

What to watch

strengthens A named hyperscaler or large AI operator discloses a deployment using Cornelis fabric, or Qualcomm and Cornelis announce a joint rack-scale product with a shipping date, by June 30, 2027.
weakens The CN6000 does not reach the stated expanded availability in the fourth quarter of 2026, or no customer-measured performance figure for Active Compute Fabric is published, by March 31, 2027.

The words

Interconnect: The wiring and switching that lets many chips inside a rack exchange data fast enough to work on one problem together. Modern AI training is limited by this as much as by the chips.
NVLink: Nvidia's own high-speed interconnect. Because it is proprietary, a rack of Nvidia chips is hard to replicate with parts from anyone else, which is where much of Nvidia's advantage sits.
UALink and ESUN: Open industry standards for connecting accelerators inside a rack, created by a consortium as an alternative to NVLink. Cornelis says its scale-up product is built to them.
Scale-up and scale-out: Scale-up means making one rack behave as a single larger machine; scale-out means linking many racks together. Nvidia is strongest at the first, which is what this announcement targets.
AllReduce: The step in training where every chip has to share its results with every other chip and agree on a combined answer. It dominates network traffic, which is why doing it inside the switch matters.

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.