Cornelis Raised $205M to Fight InfiniBand. The libfabric Pool Is 6.

On September 14, 2026, Cornelis Networks closed a $205M Series C led by IAG Capital Partners to scale its CN5000 and CN6000 Omni-Path products, launch an Active Compute Fabric aimed straight at Nvidia InfiniBand, and sample an 800G SuperNIC. If you have been asked to pipeline against that raise, the problem is not budget. The problem is that the US pool of engineers who have actually shipped libfabric code is small enough to fit in a single email thread.

The pool is 6 self-identified, a few hundred real, and already mostly inside Cornelis

In Refolk's index of professional profiles, exactly 6 people in the United States list libfabric as a skill. The InfiniBand+RDMA adjacent pool is 28. Cornelis already employs an estimated 160 to 175 engineers. The math of this raise is that Cornelis has to convert InfiniBand and RoCE talent, not hire libfabric free agents, because the free agents do not exist.

libfabric, for anyone newer to this stack, is the user-space Open Fabrics Interfaces API maintained by the OFI Working Group inside the OpenFabrics Alliance. It is what sits between MPI, SHMEM, or an AI framework and the actual fabric hardware, whether that is Omni-Path, InfiniBand, or Slingshot. Shipping it in production requires a lineage that traces back through Intel's OPA group, Cray, Cisco, UNH InterOperability Lab, and the DOE labs that co-authored the original 2015 paper.

6

The implication for anyone running Cornelis Networks hiring, or recruiting for Enfabrica, Cerebras, or anyone else building a lossless fabric: the top of your funnel is not "libfabric engineers." It is "engineers one hop from libfabric engineers," and that is a very different search.

The money is going at scale-up networking, not more fabric software

Cornelis raised $205M to scale production of CN5000 and CN6000, launch the Active Compute Fabric, and announce a strategy collaboration with Qualcomm Technologies. That mix tells you the 2027 hiring wave is silicon and systems, not another round of libfabric contributors.

The round brings total raised to $298M across five rounds against reported revenue of roughly $20M. A few load-bearing anchors from the announcement and public filings:

• Lead investor: IAG Capital Partners, announcement dated September 14, 2026.
• Headline products: CN5000, CN6000, and an 800G SuperNIC inside the Active Compute Fabric.
• Active deployments: Texas Advanced Computing Center (TACC) and the U.S. Department of Energy, which in practice means ANL, LANL, ORNL, and Sandia are referenceable.
• Strategic partner: Qualcomm Technologies, announced jointly, which means Cornelis is also now competing for Qualcomm-adjacent networking talent.
• Leadership: Lisa Spelman (CEO, ex-Intel Xeon GM), Philip Murphy (co-founder/COO).

Watch the June 2026 hire of ASIC Verification Engineer Joao Salas. That is the shape of the next wave of reqs: SerDes, verification, packaging, firmware. The libfabric pool is the headline, but it is not where most of this $205M lands.

What the Refolk index actually shows

Everything in the table below comes from Refolk's index of professional profiles across GitHub, LinkedIn, and the open web, run against US-only filters.

Two things to notice. First, the skill-tag count is a floor. Plenty of ex-Intel OPA engineers do not list libfabric on their profile even though they wrote parts of it. The realistic ceiling, counting unlabeled Intel Hillsboro alumni and DOE lab staff, is a few hundred. Second, the InfiniBand pool is Oracle- and Microsoft-heavy, not Nvidia-heavy. In the indexed set, Oracle alone surfaces more InfiniBand+RDMA engineers (5) than Nvidia (1).

If your sourcing playbook was "poach from Nvidia Networking," invert it. The realistic feeder pools are OCI's HPC team in Seattle and Azure HPC/AI in Redmond.

The five pools that actually exist

There is no national libfabric market. There are five overlapping pools, and every real Cornelis hire will come from one of them.

1. Ex-Intel OPA Hillsboro. The original Omni-Path team. Scattered across Hillsboro DCAI networking, with a long tail that bounced to Nvidia (James Dinan), AWS Annapurna, and smaller fabric startups. This is the richest vein and the one Cornelis has mined hardest since the 2020 spinout.
2. DOE national labs. Howard Pritchard at LANL. Yanfei Guo and Ken Raffenetti at ANL. ORNL, Sandia, NERSC/LBNL. GS-14 and GS-15 scientists running two to three times under private-sector market. They move for mission-aligned startups, and Cornelis's DOE deployments give it a pitch no InfiniBand recruiter can match.
3. Oracle OCI HPC in Seattle. The single densest cluster of InfiniBand+RDMA practitioners in the Refolk index, with 5 surfaced profiles and more unlabeled.
4. Microsoft Azure HPC in Redmond. Smaller than OCI in this index (2) but a long-running InfiniBand shop with real lossless-fabric production experience.
5. Ex-Rockport Networks and ex-Mellanox diaspora. Rockport shuttered in 2023 and the engineers landed at Cerebras, Enfabrica, and various AI infra startups. Ex-Mellanox folks who left Nvidia after the Spectrum-X ramp are available in Yokneam and Austin.

Expand the top-of-funnel with contributors to the ofiwg/libfabric GitHub repo and anyone who presented at the OpenFabrics Alliance Workshop, SC25, or ISC in the last three cycles. The contributor graph on that repo is a direct sourcing list, and it reads as the maintainer lineage of the standard, not a resume database.

Why Cornelis is effectively hiring against itself

With 160 to 175 engineers in-house and only 6 self-identified libfabric specialists loose in the US, more than 95% of the country's pre-trained libfabric talent is already on Cornelis's payroll. The $205M mostly funds engineers Cornelis has to convert from InfiniBand, RoCE, or UCX, not people it can hire ready-to-ship.

The mechanism is a decade of Intel exclusivity. Omni-Path was Intel-only hardware from 2015 to the 2020 Cornelis spinout. No other vendor shipped a production OPA fabric, which means no outside training pipeline ever formed. Anyone who has not touched Intel's OPA code since 2020 has stale skills. Combine that with the fact that libfabric's lead architect of record (Sean Hefty) and OFI 2.0 driver (Jianxin Xiong) are both still at Intel, and you get an industry where the standard is written by a handful of people inside two organizations.

The practical consequence for an InfiniBand recruiter working this market: assume a long ramp on any hire whose last fabric was InfiniBand or RoCE, price in a signing bonus that reflects that ramp, and build your evaluation loop around systems depth rather than specific OPA fluency. The engineers who will matter in 2027 are the ones who can rewrite a provider under OFI, not the ones who have memorized the current one.

Cornelis already employs 27 times more OPA-fluent engineers than exist as free agents in the entire United States.

Where to source for the 800G SuperNIC wave

The 800G SuperNIC pivots Cornelis's hiring profile from fabric software to SerDes and ASIC verification, which is a different pool entirely and a different sourcing strategy.

Treat it as a silicon req stack. The feeder pools are different, the geography shifts, and the libfabric lineage barely matters. Where to look:

• Broadcom Tomahawk Ultra team in San Jose and Austin.
• Marvell switching and SerDes groups.
• AMD Pensando DPU and NIC engineers.
• Nvidia Networking (ex-Mellanox) in Austin and Yokneam, specifically ASIC verification and physical design.
• Enfabrica and Cerebras, both of which absorbed ex-Rockport silicon engineers.
• AWS Annapurna Labs in Cupertino and Austin for ex-Intel NIC and SmartNIC hires.

Portland is still the center of gravity because the ex-Intel OPA team and the Hillsboro silicon ecosystem are both there, but the Rochester, MN site (ex-IBM silicon) is going to pull more of this hiring than its headcount suggests. Wayne, PA stays administrative and executive.

Three sourcing moves for the next 60 days

Three moves that will matter more than any outbound sequence you can write.

1. Mine the ofiwg/libfabric contributor graph weekly. Every commit author with a US email domain and a non-Intel, non-Cornelis affiliation is a lead. There are not many. That is the point.
2. Pre-wire the DOE labs before SC26. Howard Pritchard's team at LANL, Yanfei Guo and Ken Raffenetti at ANL, and the ORNL and Sandia HPC groups will be at the Supercomputing conference. GS-14 and GS-15 scientists are the single most under-priced segment in this market, and Cornelis's DOE references give them a credible landing zone.
3. Build an Oracle OCI HPC and Azure HPC watch list. These are the two densest InfiniBand clusters in the index, and both have seen enough internal churn in 2026 to make outbound productive. Expect a meaningful OPA ramp and price accordingly.

For a pool this shallow, tooling matters. HPC networking sourcing used to mean maintaining a hand-curated spreadsheet of 300 names and refreshing it after every SC. That spreadsheet is still useful, but it is brittle and it misses the contributors who never show up in search. Running the same query through Refolk, with "ex-Intel OPA, active in ofiwg, open to Portland or Minnesota" as the plain-English prompt, surfaces the same list plus the GitHub-only engineers who never updated their LinkedIn. That overlap is where the next five hires live.

FAQ

How many US engineers actually have libfabric experience?

Six self-identify in Refolk's index of professional profiles. The realistic ceiling, counting unlabeled ex-Intel OPA engineers and DOE lab staff who shipped the standard without listing it as a skill, is a few hundred. Treat 6 as the floor for self-identification and build your pipeline around the ex-Intel Hillsboro, DOE lab, Oracle OCI, and Azure HPC pools where the skill is embedded but not advertised.

Should I poach from Nvidia Networking for Cornelis Networks hiring?

Less than you think. In Refolk's index, Oracle alone surfaces five InfiniBand+RDMA engineers in the US to Nvidia's one, and Microsoft Azure adds another two. The denser feeder pools for an InfiniBand recruiter working the Cornelis side of the market are OCI HPC in Seattle and Azure HPC in Redmond, with the ex-Mellanox Austin and Yokneam diaspora as a secondary layer.

What is the realistic ramp for an InfiniBand engineer hired into an Omni-Path role?

Price in a long one. Omni-Path was Intel-only hardware from 2015 until the 2020 Cornelis spinout, which means no third-party training pipeline exists and the standard is still driven by a handful of people inside Intel and the DOE labs. Any hire whose last fabric was InfiniBand or RoCE needs real runway to become productive on OPA internals, and your comp and evaluation loop should reflect that.

Where will the 800G SuperNIC hires come from?

Mostly outside the libfabric pool. The SuperNIC pivots the hiring profile toward SerDes, ASIC verification, and packaging, which means Broadcom Tomahawk Ultra, Marvell, AMD Pensando, AWS Annapurna, and the ex-Mellanox Austin team are the realistic feeder pools. Portland stays the center of gravity, but expect the Rochester, MN site to pull a disproportionate share of the silicon reqs.