Investors
Backed by Draper Associates since 2017. The capital funds the step from validated array to production tape-out, delivers partner-optimised arrays with tailored memristors as the commercial proof point, and carries existing engagements through qualification. It does not fund the discovery of whether the physics works.
Early institutional backer
Nine years of deep-tech engineering
Arrays optimised for partner solutions
How the business is built
CyberSwarm designs neuromorphic computing hardware and licenses it as intellectual property. Revenue arrives in two phases, and they are sized differently.
Phase one
Modules built on our own arrays, sold into defence, industrial and automotive applications where the power envelope leaves no alternative, alongside engineering fees during co-development. This is specialised hardware, not volume consumer silicon. It produces revenue while the architecture is still being proven at scale, and it puts working systems into customers’ hands, which is what makes the licensing conversation credible later.
Phase two
The ARM model. Once reference designs exist, two things are licensable: the architecture itself, and the memristor devices, which other manufacturers fabricate under licence. Revenue becomes a per-unit royalty on licensee volume, with no inventory and no fab risk on our side. The relevant denominator stops being our own production capacity and becomes the connected-device base.
The two phases are not alternatives. Hardware revenue during the proving period is what allows the licensing transition to be made from a position of demonstrated deployment rather than from a specification.
Revenue architecture
| Stream | Timing | Character | Margin profile |
|---|---|---|---|
| Engineering and co-development | Immediate | Milestone-based; funds integration engineering | Cost recovery plus |
| Hardware supply | Near term | Framework agreements, deployment-driven | Product margin |
| Licensing royalties | Post-integration | Per unit, at licensee volume | Near-zero marginal cost |
| Support and toolchain | Ongoing | Annual, attached to licences | High |
The structure is deliberate. Engineering fees and hardware supply generate revenue during the period when royalties cannot yet exist, bridging the qualification-cycle gap. Royalties are the terminal value.
Why licensing rather than manufacturing
Investment thesis
Most AI-hardware companies compete on architecture on top of a foundry process anyone can buy. CyberSwarm’s advantage begins one level lower, in a patented IGZO resistive element and the array built from it. That is expensive to replicate and difficult to design around.
For a decade, edge AI was a nice-to-have because models were small enough to fit or workloads could be pushed to a data centre. Physical AI removes both escapes: a robot, a vehicle or an implant has a fixed power budget and no acceptable round-trip. The architecture that wins there is the one that stopped moving weights.
We are not building a training accelerator, and we are not asking anyone to abandon their GPU cluster. We sit at the deployment end of the same pipeline, in devices those clusters will never be installed in, which makes the incumbents distribution partners rather than obstacles.
The reason novel hardware fails commercially is almost never the hardware. It is that nobody can deploy to it. Compiler, runtime, calibrated simulator and PyTorch-compatible SDK are built alongside the device precisely so that adoption does not require faith.
European defence, industrial and healthcare buyers increasingly need compute that is not dependent on a foreign cloud. An architecture where inference physically cannot leave the device is a strong answer to a question procurement is now required to ask.
Founded 2017, filing since 2018, granted patents in 2024 and 2025. Current funding accelerates the step from validated architecture to production silicon and first licences, not the discovery of whether the physics works.
Market
Every figure below is published by the analyst house named beside it. We do not model our own market.
$36.2B
Edge AI chip market, 2026
Grand View Research. Projected to $291.8B by 2033, 34.7% CAGR
$20.3B
Neuromorphic computing by 2030
Grand View Research, from $5.3B in 2023, 19.9% CAGR
$96B
Edge AI chipset market, 2031
ABI Research, from $34.4B in 2026
Third-party estimates of edge AI silicon vary by an order of magnitude depending on what is counted, which reflects disagreement about definitions rather than about demand. We cite them for orientation and do not rely on any of them.
The denominator that matters to a licensing business is devices, not wafers. The installed base of connected devices passed 18.4 billion units in 2025 and is forecast to exceed 38 billion by 2030. We address the subset that must adapt locally: automotive sensing and ADAS nodes, defence and autonomous platforms, industrial sensing and predictive maintenance, and robotics.
Traction
Engagements are described by status, without naming counterparties, and without a weighted pipeline figure. The same record, in more detail, is on the projects page.
Roadmap
Eight parallel workstreams, board-approved. Three run today with current headcount; five commence on funding. Together they take the architecture from a validated array to production silicon and the first royalty-bearing agreements.
01
Running
The physical manifold proof, in software and electronics hardware. The 4 × 4 module is operational with full interface electronics and software control of array input and output.
02
Running
Field application support for co-development partners. Active against the automotive engagement in testing.
03
Started
The compiler and model-mapping layer for licensees. The workstream the company names as its genuine execution risk.
04
Commences on funding
The architecture is developed and validated; this workstream takes it to manufacture. Mask sets, design tooling, IP blocks, test and validation, and the foundry engagement, culminating in tape-out.
05
Commences on funding
Fabrication, characterisation, and commercial proof of concept for different types of arrays with memristors optimised for partners’ solutions. Array fabrication is demonstrated; integration and interconnection with the electronics begin on funding.
06
Commences on funding
Entity, staffing, manufacturing readiness, for the initial hardware phase and for defence-adjacent work.
07
Commences on funding
Export-control programme, CFIUS, ITAR posture assessment. ECCN 3E001 classification is complete; the programme commences on funding.
08
Commences on funding
To approximately 51 people, weighted toward device engineering, circuit design, toolchain and field application engineering, and sequenced to remove single points of dependency first.
Post-funding, the roadmap moves from single-array proof to multi-array systems, a productised licensable IP core with full toolchain support, and qualification cycles in automotive and defence.
The commercial proof point
The stage at which the architecture transitions from demonstrating physical principles to delivering arrays with memristors tailored to partner specifications. Five milestones lead to it, each with the dependency it waits on.
The gate is the same in every engagement: a customer-relevant inference workload, executed on arrays optimised for partner solutions at the target power envelope, witnessed by the counterparty. That artefact moves a discussion to co-development, and co-development to a licence. It is what this funding supports.
Competitive landscape
| Category | Approach | Where we differ |
|---|---|---|
| GPU and general accelerators | Digital, memory and compute separated | We remove the data-movement energy cost architecturally |
| Edge NPU and digital ASIC | Digital, fixed-function, optimised dataflow | Same architectural energy floor; we operate below it |
| Analogue in-memory startups | Crossbar-based, Euclidean representation | Non-Euclidean geometry implemented in hardware |
| IP licensors | Digital cores, mature toolchains | We compete on a design point they cannot reach; they compete on ecosystem maturity |
Where we are behind: toolchain maturity and ecosystem. Established IP licensors have decades of compiler, tooling and developer support. The toolchain workstream addresses this, and it is a genuine execution risk rather than a formality.
Where the moat sits: the combination of granted patents, process trade secrets held outside the filings, export-control compliance capability, and the geometric approach itself. Any one is replicable in time. The combination is not easily.
Risks and mitigations
The central near-term risk. Production silicon is the gate everything else passes through.
What is done
Foundry partner selected; the largest single allocation of capital; staged validation before mask-set commitment.
Automotive and defence convert on the customer’s timeline, not ours.
What is done
Engineering fees and hardware supply bridge the gap; multiple verticals with independent cycles.
Behind incumbent IP licensors on developer ecosystem.
What is done
A dedicated workstream; early licensees co-develop against real workloads.
Organisational risk at this growth rate is real, and key-person concentration is a real exposure at current scale.
What is done
Sequenced hiring; a two-site structure; documentation and deliberate redundancy in the device and circuit teams.
A novel process at commercial scale, in the initial hardware phase.
What is done
An established foundry partner; a conservative array scaling path rather than a single large jump.
ECCN 3E001. An operating requirement and a screening constraint on counterparties.
What is done
A funded compliance programme; counterparty screening; US operational control and domestic hiring intent. It is also a barrier to entry for competitors without a compliance function.
The pipeline is weighted to a small number of partners.
What is done
A non-exclusive channel structure; field-limited licensing prevents foreclosure of a market.
Large players can fund parallel research.
What is done
Patent position established; process trade secrets undisclosed; the geometric approach is not an obvious adjacency.
Backers
since 2017
Silicon Valley
“Neuromorphic computing is not meant to replace digital systems. It represents a new computational layer between traditional computing and quantum. Designed for problems that neither solves efficiently.”
Diligence
Investor relations reaches the founding team directly at mihai@cyber-swarm.net.