Who is funded, who got bought, and where our layer sits.
Names are facts, not endorsements; funding figures are as reported in the linked sources. The picture that emerges: the deployment layer is being pulled into single-silicon stacks, and the governed release layer for mixed fleets is unowned.
Four tiers, one gap.
Silicon, accelerators, runtimes
- NVIDIA Jetson Orin / Thor; NXP i.MX95 Neutron; Qualcomm Dragonwing / Snapdragon
- Axelera, DEEPX, SiMa.ai, Quadric; Hailo (Microchip acquisition pending)
- TensorRT Edge-LLM, eIQ, GenieX; llama.cpp, vLLM, ExecuTorch, LiteRT
- Runtime optimizers: General Instinct, Wafer, Liquid AI models
signed release
policy gate
audit · rollback
Edge platforms, OTA, cloud planes
- Edge Impulse (Qualcomm), ZEDEDA, Avassa, Barbara, CTHINGS.CO, CraftifAI
- OTA stacks: RAUC, SWUpdate, Mender, JetPack OTA
- Cloud planes: Vertex AI, Amazon Bedrock, Microsoft Foundry
Partners and channels, not competitors.
| Player | Status, 2026 | Note |
|---|---|---|
| NVIDIA Jetson Orin / Thor | Incumbent for GPU edge; TensorRT Edge-LLM, JetPack 7.x, Inception program. | ByteHub's validating targets. |
| NXP i.MX95 (Neutron NPU) | eIQ GenAI Flow v3.0 (Mar 2026); ONNX Runtime Neutron execution provider for INT4 LLMs. | ByteHub's low-power industrial target. |
| Qualcomm Dragonwing / Snapdragon | Acquired Edge Impulse (170k+ developers) and Nexa AI (Jul 2026 → GenieX runtime); Arduino is a Qualcomm company (Nexa AI; New Market Pitch). | The vertical-integration pattern: silicon vendors buying the deployment layer for their own chips. |
| Axelera AI, DEEPX, SiMa.ai, Quadric | Independent first tier; SiMa.ai's "Palette Neat" agentic development environment; Microchip agreed to acquire Hailo (not closed). | Future execution-profile targets; SiMa shows the developer-tooling arms race. |
Overlap on "make it run", not on "govern the release".
| Player | Round / status | What they sell | Overlap with ByteHub |
|---|---|---|---|
| General Instinct | YC Spring 2026, two founders (YesPress) | Instinct Edge: model + target device + latency budget → optimized offline runtime for Jetson, mobile NPUs, ARM, Apple Neural Engine, Snapdragon; open-source InstinctRazor compression. | Runtime generation only. No fleet lifecycle, signing, policy, or cloud gate. |
| Wafer | $40M Series A, Sep 2026 (Marathon, Chemistry, AMD Ventures) (Wafer) | Continuous inference optimization across model, engine, kernels, and hardware. | Datacenter-first; optimization, not device governance. |
| Liquid AI | LFM2.5 family; QAD Q4 checkpoints | Small agentic models for phones and edge devices. | Model supplier; a candidate artifact for ByteHub to package. |
| Vendor runtimes | TensorRT Edge-LLM, eIQ, GenieX | Fast local inference on the vendor's silicon. | The substrate ByteHub builds on. |
Adjacent: partner where possible, displace where necessary.
| Player | Round / status | What they sell | Gap ByteHub fills |
|---|---|---|---|
| CraftifAI | $3M seed, Feb 2026 (Ankur Capital) (The AI World) | GenAI / agentic workflow to build embedded software and optimize, quantize, and deploy edge models; GStreamer, ROS2, Android. | Development-time acceleration, not signed release governance for fleets. |
| CTHINGS.CO | €1.8M seed, Jul 2026 (Navivo) (TMC Insight) | Orchestra: distributed edge infrastructure automation. | Infrastructure orchestration; model-aware evaluation and policy missing. |
| ZEDEDA, Avassa, Barbara | Established | OS, container, and application lifecycle for edge nodes. | Container-level, not model artifact + policy + audit. |
| Edge Impulse (Qualcomm) | Acquired | Embedded ML development platform (TinyML lineage). | Training and deploy tooling owned by one silicon vendor. |
| OTA stacks: RAUC, SWUpdate, Mender, JetPack OTA | Mature | OS, firmware, and container updates. | ByteHub extends this discipline to model artifacts. |
Cloud planes — Vertex AI / Gemini, Amazon Bedrock, Microsoft Foundry — all offer frontier reasoning, evaluation, guardrails, and identity. None ships the device-side policy gate, minimized-context escalation, or signed edge artifacts. They are the provider-aware cloud half of the ByteHub design.
Where ByteHub sits.
Cloud AI platforms give model APIs. Silicon SDKs give accelerated inference. OTA tools update the OS. Agent wrappers give chat. ByteHub is the model lifecycle and policy layer between AI capability and real-world device operations: signed, target-compatible model artifacts; staged rollout and rollback; a default-off policy gate to frontier cloud; per-tenant audit.
| Threat | Honest answer |
|---|---|
| Silicon vendors integrate the deployment layer (the Qualcomm pattern) | Vendor tools stay single-silicon. Industrial fleets mix i.MX, Jetson, Android, and x86. ByteHub is cross-silicon and provider-aware by design. |
| OTA vendors add "model" as an artifact type | They lack model-aware evaluation, execution profiles, and cloud policy routing. Partner where possible. |
| Cloud vendors push edge agents (Foundry / Arc, IoT Operations, Bedrock at the edge) | Useful cloud half. The device-side gate and offline behaviour still need an owner. |
| Buyers build it in-house | That is today's status quo — and the reason design partners exist. |
Watchlist, updated quarterly: Qualcomm GenieX roadmap · Microchip–Hailo close · SiMa.ai Palette adoption · Jetson Thor T4000 pricing · NXP eIQ Neutron LLM model list · LFM / Gemma / Qwen small-model releases · EU CRA enforcement dates.
Missing a player? Tell us.
This map is only useful if it is complete. We add players and correct rounds as sources surface.