Cornelis Networks
Sells switches, NICs/SuperNICs (HFIs), cables and fabric-management software as an end-to-end interconnect, direct and through server OEMs (Lenovo, Dell, HPE, Supermicro); revenue is hardware-led with attached software/support, supplemented by non-dilutive US government R&D awards.
Cornelis has never officially disclosed a post-money valuation. Only the current ~$174M CB Insights third-party estimate is a real anchor; the 2020 and 2022 points are conservative models pegged to disclosed round sizes ($29M Series B, ~$85-139M cumulative) and are NOT company-confirmed. A $20M convertible note (Dec 2024) sits between the 2022 and 2025 points but does not itself price the equity, so it is omitted from the trail. Treat the trajectory as directional, not precise.
Earnings, margins, COGS & capex
Private company with no published financials. Cumulative equity + convertible funding is reported by third parties between ~$85M (Tracxn) and ~$131-139M (CB Insights / PitchBook); disclosed events include a $29M Series B (Nov 2022, led by IAG Capital), a $20M convertible note (Dec 2024) and a ~$15-20M raise (Apr 2025; SEC Form D reported $15M of a combined equity/debt offering). This is materially undercapitalized versus the AI-networking incumbents (NVIDIA, Broadcom, Arista), so Cornelis leans heavily on non-dilutive US-government funding -- an $18M DOE NNSA next-generation networking contract (powering the 952-node 'Lynx' cluster at Lawrence Livermore National Laboratory) and IARPA's AGILE program (with Georgia Tech) -- to co-fund its ASIC roadmap. Commercial traction is early: the CN5000 400G product began shipping June 2025 with broad OEM availability from Q3 2025, and the CN6000 800G SuperNIC was unveiled Nov 2025; unit economics, revenue scale and burn are undisclosed.
Revenue trend
Margins
n/a
n/a
COGS structure
not disclosed. Structurally: fabless ASIC + switch/NIC hardware BOM (foundry-fabricated silicon, optics/cables, PCB, assembly), so COGS is dominated by third-party silicon and optics rather than owned fabrication.
Capex
not disclosed. Low physical capex (no fabs); the true 'investment' is capitalizable/expensed R&D and mask/NRE spend for successive ASIC generations (CN5000 -> CN6000 -> CN7000).
Latest earnings
not applicable
not disclosed
- Headcount (May 2026)
- ~239 employees (Tracxn; other sources lower)
- CN5000 first shipments
- June 2025; broad OEM availability Q3 2025
- CN6000 800G SuperNIC
- unveiled Nov 2025; single platform supporting Omni-Path, RoCEv2 and Ultra Ethernet; up to 1.6B messages/sec
- Anchor DOE deployment
- 952-node 'Lynx' cluster at LLNL for NNSA (CN5000); $18M DOE NNSA R&D contract
- CN5000 scale
- up to ~500,000 endpoints; sub-microsecond latency; up to 1.2B messages/sec
- Claimed HPC edge vs InfiniBand NDR
- ~2x message rate, ~35% lower latency, up to 30% faster simulations (vendor benchmark)
Growth drivers
- AI back-end/scale-out networking spend as clusters scale to 100k+ accelerators, where the network is a rising share of cluster cost
- Demand for a credible non-NVIDIA fabric alternative from hyperscalers, neoclouds and sovereign/government buyers seeking supplier diversity
- US government HPC programs (DOE NNSA next-gen networking contract, IARPA AGILE) providing non-dilutive funding and anchor deployments
- OEM design-ins (Lenovo EveryScale, plus Dell/HPE/Supermicro availability) that give reach without a large direct sales force
- Ultra Ethernet Consortium roadmap (CN6000 800G SuperNIC unveiled Nov 2025; CN7000 1.6T planned) aligning Omni-Path strengths with an open-Ethernet standard buyers already trust
Bull & bear
A credibly-differentiated, government-anchored fabric with real production heritage, riding the largest spend wave in datacenter networking as buyers actively seek a non-NVIDIA alternative -- an asymmetric, cheap option on AI-interconnect diversification and a plausible acquisition target.
- Omni-Path is not vaporware -- it has a decade-plus lineage and live HPC deployments, so the CN5000 launch is an upgrade of a proven fabric, not a science project
- DOE NNSA (an $18M contract + the LLNL 'Lynx' cluster) and IARPA AGILE funding give both cash and third-party technical validation few startups can claim
- The AI back-end network is structurally growing and buyers (hyperscalers, neoclouds, sovereigns) explicitly want a second source to NVIDIA -- Cornelis is one of the very few independent full-fabric options
- OEM partnerships (Lenovo EveryScale, plus Dell/HPE/Supermicro) provide distribution leverage disproportionate to headcount
- Ultra Ethernet alignment (CN6000 unveiled Nov 2025, CN7000 planned) converts a proprietary-fabric risk into standards tailwind while keeping Omni-Path's latency/collective advantages
- At an estimated ~$174M valuation it is a cheap call option; a single strategic acquirer (AMD, Intel, an OEM, or a networking vendor) could value the IP and team well above that
A sub-scale challenger fighting NVIDIA and Broadcom in a capital-intensive ASIC race on a fraction of their budgets, with undisclosed/likely-small revenue, government-lumpy demand, and the risk that Ultra Ethernet commoditizes its exact edge before it reaches scale.
- The resource asymmetry is brutal: ~$85-139M lifetime raised versus competitors spending that much in days of R&D; interconnect ASICs demand sustained multi-hundred-million investment per generation
- No disclosed revenue traction -- CN5000 only began shipping mid-2025 and CN6000 was only unveiled in late 2025, with no evidence of hyperscale design wins at volume
- NVIDIA bundles networking with GPUs and owns the software stack (NCCL/CUDA); buying accelerators effectively pulls in NVIDIA's fabric, squeezing independents
- Ultra Ethernet, once mature, gives every Ethernet vendor lossless/congestion features -- eroding Cornelis's core differentiation and turning its moat into table stakes
- Dependence on government/HPC revenue is narrow, cyclical, and procurement-gated -- hard to build a venture-scale growth story on alone
- Funding overhang: the last disclosed rounds were small (~$15-20M), and needs a large growth round to keep pace on the roadmap; failure to raise risks a down round or distressed sale of the IP
What it is worth
Private, pre-IPO; no public price. Reference points only: third-party estimated valuation ~$174M (CB Insights) and cumulative funding reported ~$85M (Tracxn) to ~$131-139M (CB Insights / PitchBook), with disclosed capital events including a $29M Series B (Nov 2022), a $20M convertible note (Dec 2024) and a ~$15-20M raise (Apr 2025; Form D reported $15M of a combined equity/debt offering). No official post-money disclosed.
Out-spent on the ASIC roadmap and squeezed by Ultra Ethernet commoditization; forced into a down round or distressed IP sale below the current estimated mark.
Continues as a government/HPC-anchored niche fabric vendor with modest commercial CN5000/CN6000 traction; valuation roughly tracks its last small rounds absent a large growth raise.
Wins meaningful independent-fabric share and/or is acquired at a strategic premium (IP + team + government relationships) well above the ~$174M estimate as buyers pay up for NVIDIA alternatives.
Value is essentially a call option on (a) capturing share of the fast-growing non-NVIDIA AI-fabric market and (b) strategic-acquisition optionality for the Omni-Path IP and team. Any equity exposure is illiquid, accreditation-gated secondary only, and highly binary. Not a buy/sell recommendation; not financial advice.
SWOT
Strengths
- Mature, battle-tested fabric IP — Omni-Path descends from Intel's HPC interconnect and QLogic/TrueScale InfiniBand lineage, with real production deployments -- not a from-scratch startup stack
- Genuinely differentiated architecture for HPC/AI collectives — credit-based flow control, fine-grained adaptive routing, lossless/congestion-managed transport tuned for MPI and all-reduce patterns
- Anchor US-government demand and non-dilutive funding (DOE NNSA $18M contract + LLNL 'Lynx' deployment, IARPA AGILE) that both funds R&D and validates the technology
- OEM distribution via Lenovo/Dell/HPE/Supermicro extends reach without incumbent-scale sales spend
- Positioned as the 'independent' fabric -- attractive to buyers deliberately avoiding NVIDIA lock-in
Weaknesses
- Tiny capitalization (~$85-139M raised) versus competitors with tens of billions in R&D -- a structural resource gap in an ASIC arms race
- No disclosed revenue scale; commercial adoption of CN5000 is still early and unproven at hyperscale
- Heavy dependence on government/HPC niche revenue, which is lumpy and procurement-cycle-bound
- Ecosystem/software maturity gap versus NVIDIA's CUDA+NCCL+InfiniBand full stack and the broad Ethernet toolchain
- Roadmap execution risk — must convert the CN6000 (800G) Nov-2025 unveil into volume shipments and deliver CN7000 (1.6T) on time against far better-funded competitors
Opportunities
- AI back-end network is one of the fastest-growing datacenter spend lines; even a small share is large relative to Cornelis's size
- Ultra Ethernet Consortium standardization lets Cornelis ride an open standard rather than fight the whole Ethernet ecosystem
- Sovereign-AI and US-government buildouts explicitly want non-NVIDIA, domestically-aligned suppliers
- Potential strategic acquisition target for a server OEM, an AI-chip challenger (AMD, Intel), or a networking vendor wanting a differentiated fabric
- Neocloud / specialized AI datacenter operators seeking cost- and vendor-diversified fabrics
Threats
- NVIDIA's dominance across both InfiniBand (Quantum) and AI Ethernet (Spectrum-X), bundled with GPUs -- the gravitational competitor
- Broadcom's Tomahawk/Jericho scale and the broad move of AI networking toward merchant Ethernet silicon
- Ultra Ethernet commoditizing the exact congestion/lossless advantages Cornelis sells, eroding differentiation over time
- Funding risk — if a large growth round doesn't materialize, Cornelis could be out-spent or forced into a distressed sale
- Customer concentration in government/HPC leaves it exposed to budget and procurement shifts
Moats, dependencies & bottlenecks
Moats
adaptive routing, credit-based lossless transport tuned for collectives) Real and defensible today for MPI/HPC and AI all-reduce; durability threatened as Ultra Ethernet standardizes similar capabilities.
Government relationships + non-dilutive funding (DOE NNSA contract + LLNL deployment, IARPA AGILE) Sticky, high-trust demand and cash, but narrow and procurement-cycle-dependent.
Existing HPC sites are natural upgrade customers; not large enough to lock the AI market.
Dell/HPE/Supermicro) Distribution leverage, but OEMs multi-source and can favor NVIDIA/Broadcom fabrics.
Dependencies
Fabless; ASIC fabrication and optical component availability/pricing are outside its control.
Customer/financing concentration A pillar of both revenue and R&D funding; budget/policy shifts hit hard.
Reach depends on OEMs choosing to promote Cornelis fabric over incumbents.
Roadmap execution assumes further large raises in a competitive, capital-intensive category; last disclosed rounds were small.
Standards/ecosystem A double-edged dependency -- tailwind for adoption, but standardization also commoditizes its differentiation.
Strategic/technology Origin of the technology and an investor; not an operational dependency today.
Advantages
- Production-proven fabric heritage rather than a greenfield stack
- Architecture genuinely optimized for HPC/AI collective operations and low-latency, lossless transport
- Independent, non-NVIDIA positioning valued by supplier-diversity-seeking buyers
- Access to non-dilutive US-government R&D funding and anchor deployments (DOE NNSA, IARPA)
- OEM distribution partnerships that punch above its headcount
Weaknesses
- Severely sub-scale capitalization for the category
- Undisclosed and likely-small commercial revenue; unproven at AI hyperscale
- Narrow, government/HPC-concentrated demand base
- Differentiation exposed to erosion by Ultra Ethernet standardization
- Ongoing funding-overhang / financing risk
Bottlenecks
- Capital: far smaller war chest than NVIDIA/Broadcom in a per-generation-expensive ASIC race
- Software/ecosystem depth versus NVIDIA's integrated CUDA+NCCL+InfiniBand stack
- Sales scale and hyperscale reference customers — disclosed large wins are government/HPC (e.g. LLNL 'Lynx'), not commercial hyperscale
- Roadmap cadence: must ramp CN6000 (800G) into volume and deliver CN7000 (1.6T) on schedule to stay relevant
- Reliance on lumpy government procurement cycles for anchor revenue
Top signals & trends
Top signals
Moves from roadmap to revenue-generating product.
Extends the roadmap into merchant-Ethernet territory; still must ramp to volume.
Cash, a named anchor deployment, and validation from demanding buyers.
Distribution and credibility beyond direct sales.
Adoption tailwind but also a commoditization vector for its moat.
Funding not keeping pace with the category's capital intensity.
Suggests limited disclosed traction or a small last round; third-party estimate only.
Trends
Expands the addressable market Cornelis targets.
Directly favors independent fabric suppliers.
Validates Ethernet direction but pits Cornelis against merchant-silicon scale.
Commoditizes capabilities that today differentiate Omni-Path.
Raises odds Cornelis is acquired for its IP and team.
Ecosystem & competitor graph
Suppliers feed the company; customers pull from it. Line thickness shows the strength of each tie (supply-chain dependency, customer earnings contribution). Hover to isolate a tie.
typical for fabless ASICs) Likely fabrication partner for Cornelis silicon (not company-confirmed).
Origin of Omni-Path IP and an equity backer via Intel Capital.
Optical transceiver / cable and PCB/EMS suppliers Merchant optics and manufacturing partners for switches, NICs and cabling.
US Department of Energy / NNSA (Lawrence Livermore National Laboratory) Anchor HPC customer and non-dilutive funder -- $18M R&D contract and the 952-node 'Lynx' CN5000 cluster.
IARPA / US intelligence community (AGILE program, with Georgia Tech) R&D funding for novel computer architectures; Cornelis is a named research collaborator.
Universities and government/academic HPC centers Traditional Omni-Path installed base and upgrade path.
Enterprise/commercial HPC and emerging AI-cluster operators Targeted commercial buyers for CN5000/CN6000 and roadmap parts.
Dominant AI-fabric incumbent, bundles networking with GPUs and owns the software stack.
Leading merchant AI-Ethernet silicon powering most non-NVIDIA fabrics.
Leading AI/datacenter Ethernet switching systems vendor (Broadcom-based).
Silicon One and AI datacenter switching; broad enterprise reach.
Custom AI silicon, DCI/optics and networking components.
Building AI networking to pair with Instinct GPUs; also a potential acquirer.
Technology origin and Intel Capital backer; competes in adjacent networking.