Recurrent / state-based / event-driven models (RNNs, SSMs like Mamba, RWKV, xLSTM, spiking/neuromorphic) in production and research, sorted by the structural reason a transformer cannot follow. Every item lists real products/models, the headline number, and a source. Full dataset (117 use cases, 84 companies): raw/.
Maturity: [deployed] in production · [piloted] trials / pre-production · [research] lab result.
TL;DR
- Claim: recurrent / state-based / event-driven models are already in production wherever a transformer structurally cannot go.
- 6 reasons it is the only fit: hard latency floor · microwatt always-on · unbounded context · sparse events · on-device adaptation · efficiency at frontier scale.
- Both ends, deployed: edge (Syntiant 280 µW, Prophesee event cameras, Medtronic adaptive DBS) and datacenter (Nemotron-H beats Qwen-2.5-72B, Jamba 256K context, Titans >2M tokens).
- Scope: 117 use cases, 84 companies, each sourced.
At a glance
| Argument (why only recurrence works) | Flagship |
|---|---|
| 1. Hard real-time latency floor | Medtronic adaptive DBS, FDA/CE 2025 |
| 2. Microwatt always-on energy | Syntiant 280 µW (100M+ units shipped) |
| 3. Unbounded context, bounded state | Titans >2M tokens; NetMamba 60× faster IDS |
| 4. Sparse-event compute | Prophesee event cameras (Sony IMX636) |
| 5. Continuous on-device adaptation | Jamba 3B, 250K context on a MacBook |
| 6. Efficiency at frontier scale | Nemotron-H beats Qwen-2.5-72B at 3× throughput |
| (cross-cutting) Bio & health | Medtronic DBS · iRhythm ECG (1.1M+ patients) |
1. Hard real-time latency floor
Decision within one sensor-to-action cycle: no batching, no KV-cache prefill.
- Servo & motion control
[deployed]— sub-millisecond loops; O(N) KV-prefill per step can't fit. EdgeDRNN 10M-param GRU at <0.5 ms, 2.42 W (2002.03197); Synapticon shipped 15,000 AI servo drives (2025). - Legged & manipulation robots
[deployed]— policies need 100-500 ms proprioceptive history. xLSTM (LRAM) matched transformer accuracy on 432 tasks at linear-time inference (ICML 2025); ANYmal LSTM controller. - Prosthetics & exoskeletons
[piloted]— gait phase over hundreds of samples, ~20 ms terrain response. AMPRO3 + EdgeDRNN (transfemoral leg); AltAi SNN EMG hand. - BCI / neural decoding
[deployed]— motor intent over hundreds of ms; quadratic attention over 10,000+ channels is infeasible live. Neuralink N1 (1024 ch), Synchron Stentrode, BrainGate (LSTM decoders). - Adaptive deep brain stimulation
[deployed]— beta-band power tracked, stimulation adjusted within ms on a µW implant. Medtronic Percept RC + BrainSense, FDA Feb / CE Jan 2025 (Medtronic). - Real-time voice & dialogue
[deployed]— duplex speech, live transcription, conversation under ~500 ms; a growing KV-cache forces recompute. DuplexMamba (2502.11123); Google Meet S2ST (research.google); AssemblyAI streaming, AppTek; Gemini 3.1 Flash Live (source). - RF / radar electronic warfare
[piloted]— integrate Doppler / PRI over time at constant cost per sample. DeepSig OmniSIG (US Army PC-C5); BrainChip + Raytheon AFRL micro-Doppler; DSTL + SynSense + Prophesee MOU. - Adaptive optics
[piloted]— predict turbulence one frame ahead at 500-2000 Hz. LSTM cuts bandwidth error 78.3 nm → 19.9-40.0 nm RMS (JATIS).
2. Microwatt always-on energy
Runs 24/7 on a coin cell; idle power must track activity, not wall-clock.
- Wake-word & keyword spotting
[deployed]— ~1 mW budget rules out continuous attention. Syntiant NDP100/120/250 @ 280 µW (100M+ units), Aspinity AML100 (<40 µW), BrainChip Akida Pico, SynSense Xylo-Audio 2 (<500 µW), Innatera Pulsar (~400 µW), Polyn NeuroVoice, Femtosense SPU-001; 65 nm SNN idles 75 nW (PMC8329666). - Hearing aids & hearables
[piloted]— 1-5 mW device budget. Spiking-FullSubNet ~3 orders of magnitude lower power than DNN speech enhancement (IEEE TNNLS 2025); SynSense Xylo-Audio 2, Syntiant NDP earbud designs. - Predictive maintenance
[deployed]— vibration/acoustic monitoring on multi-year-battery nodes. Intel Loihi recurrent SNN 0.0032 J/inference vs 11.3 J on x86, >97% accuracy (2506.13416); EdgeSpike ran 64 nodes on a railway viaduct 7 months; Innatera Pulsar+42T, SynSense Xylo-IMU, BrainChip Akida Edge AI Box. - Presence / radar / UWB
[piloted]— micro-Doppler integrated over seconds. Innatera Pulsar + Socionext radar / Aria Sensing UWB (~600 µW, 500× lower energy than an MCU); Aspinity AML100. - Energy-harvesting IoT
[deployed]— µW to low-mW harvested power. Event-driven SNNs idle at 75 nW; STM32U3, Ambiq Apollo510, BrainChip AkidaTag (Nordic nRF5340 + AKD1500).
3. Unbounded context, bounded state
Consume the stream once; per-step memory and compute stay O(1) however long the session runs.
- Long / infinite-context LLMs
[deployed]— a 1M-token Llama-3-8B needs ~137 GB KV-cache (over one 80 GB A100); an SSM holds 2-4 GB (Flow AI / Introl). KV-cache wall is now priced (Gemini 2× >200K, GPT >272K, GPT-4 128K 64×; Jamba 8× vs Mixtral, 32× vs Llama-2-7B). Nemotron 3 → 1M, RULER 86.34 vs 23.43 (WP); RWKV-X 64K passkey to 1M (2504.21463). - Test-time / agentic memory
[research]— Titans >2M effective context, beats GPT-4 on BABILong (2501.00663); MEM1 +~35 pts on multi-session tasks at near-constant memory (2506.15841); StateX lifts 64K needle 26%→42% (2509.22630). - The data deluge (consume once)
[deployed]— NetMamba 60× faster than ET-BERT at 99%+ IDS accuracy (2405.11449); financial CryptoMamba-SSM / FinMamba (2502.06707); fraud/AML Interleaved Sequence RNNs (USPTO #12373670); telemetry Selector.ai, NASA LSTM (1802.04431); 5G/6G SpikingRx 7.6-9× lower energy (2409.05610); satellite onboard Loihi 2 (2401.06911), BrainChip Akida; radio-telescope SKA ~700 TB/day, UWA pipeline 100 mW (2412.06124); particle physics ALICE Run 3, Mamba jet reconstruction (2506.18336); nanopore basecalling Oxford Nanopore Dorado, Readfish/BOSS. - Persistent human-AI memory
[deployed]— context rot degrades all 18 frontier models with fill (Chroma). Microsoft 365 Copilot Memory (GA Jul 2025), Mem0 ($24M Series A), Letta/MemGPT (1M+ agents); companions Replika (40M+), Character.AI, Xiaoice (660M) (market ~50M users, >$580M, Grand View); tutoring TASA (2511.15163), Khanmigo, Carnegie MATHia; co-pilots GitHub Copilot Memory, Letta Code, MS Durable Task Scheduler (100M+ weekly).
4. Sparse-event compute
Event sensors fire only on change; compute is proportional to events, not frame rate.
- Industrial counting & inspection
[deployed]— event pixels fire at the instant of change, no frame buffer. Prophesee IMX636HD (Sony) + IDS uEye EVS (Mar 2025), General Vision NeuroMem, SynSense Speck. - Automotive ADAS
[piloted]— a pedestrian can cross the danger zone inside one frame. PMRVT 7.72 ms at 48.7% mAP on Gen1, 3× lower latency than non-recurrent (2212.05598); Sony-Prophesee IMX636, Terranet VoxelFlow (Mercedes-Benz), Continental. - In-cabin driver monitoring
[piloted]— always-on DMS at ~50 mW can't run an RGB CNN per frame. SynSense Speck (BMW trial 2024), Xperi DTS; EU mandates DMS in new vehicles by 2025. - AR/VR gaze / foveated rendering
[piloted]— saccades finish in <10 ms, under a 120 fps frame. Zinn Labs DK1 (Prophesee GenX320) sub-1° at 120 Hz; CVPR 2025 best 1.61 px (2504.18249). - Drone / UAV avoidance
[research]— a 10 m/s drone crosses 1 m in 100 ms. ColibriUAV DVS-SNN 3.5 ms event-to-action at 10.7 mW; Loihi async perception 4.3% of GPU energy (2305.18371). - Event optical flow / SLAM
[research]— P-SSE SSM encoder (CVPR 2025), linear-time, +8.5% DSEC-Flow EPE (UZH RPG). - Event-based video surveillance
[piloted]— at 5-10% scene activity, 10-20× lower energy than frame-based. Prophesee GenX320/IMX636 (1,200+ customers), BrainChip AKD1500, SpiNNcloud SpiNNaker2 (Sandia, Jun 2025). - Space situational awareness
[piloted]— objects cross a telescope FOV in tens of ms at single-watt budgets. DARPA FENCE (Raytheon MWIR demo Apr 2026), Neurobus (ESA/CNES/Airbus).
5. Continuous on-device adaptation
Personalize / learn on device, where a growing KV-cache and full backward pass don't fit, and data must stay local.
- On-device personalization & continual learning
[deployed]— always-on background inference on a 4-16 GB NPU. RWKV v5 Eagle shipped in RWKV.cpp (blog.rwkv.com). - Local LLMs on phones / laptops / NPUs
[deployed]— unified memory is 4-16 GB, no HBM; a KV-cache overflows in seconds, an SSM state is kilobytes. SSMs reach 220K context in 24 GB VRAM where Llama/Qwen top out near 65-73K (2507.12442); RWKV Android/iOS (RWKV7-G1 2.9B 33 tok/s mobile GPU, 7B 16.39 tok/s ARM, int8 14B holds 3 GB at any length, 2412.10856); Jamba Reasoning 3B 250K context on a MacBook (AI21); NPU kernels XAMBA up to 4.8× (2502.06924), eMamba 48.6× lower energy (ACM TECS); Cartesia Sonic on Apple M-series (Cartesia); toolchain llama.cpp / MLX / RWKV on Qualcomm QNN, Apple CoreML, MediaTek (RWKV). - Privacy / sovereignty
[deployed]— on-device inference is the structural answer to GDPR and the EU AI Act (GPAI rules binding Aug 2025). $80B sovereign-cloud IaaS spend in 2026, 61% of Western European CIOs shifting to local AI (Gartner).
6. Efficiency at frontier scale
At datacenter scale the quadratic-attention + KV-cache wall caps context, throughput, and cost; linear-time recurrence removes it.
- SSM / hybrid LLMs vs transformers
[deployed]— Mamba-2-Hybrid 8B beats an 8B Transformer on all 12 benchmarks (+2.65 avg), up to 8× generation (2406.07887); Mamba-3 half-state, ~7× cheaper/token at 64K (2603.15569); Jamba 1.5 256K at ~9 GB KV, 2.5×, 65.4 Arena Hard (AI21); Nemotron-H 56B beats Qwen-2.5-72B 9/17, up to 3×, 47B runs 1M in FP4 on one RTX 5090 (2504.03624); Nemotron 3 Super 7.5× over Qwen3.5-122B (NVIDIA), Nano 2 6× over Qwen3-8B (2508.14444); Falcon-Mamba-7B matches Llama 3.1 8B, constant memory to 130K (HF), Falcon-H1R-7B 83.1% AIME-2025, 68.6 LiveCodeBench (Falcon); RWKV-7 2.9B 3B-class multilingual, exceeds the TC⁰ ceiling (2503.14456); Zamba2-2.7B 2× TTFT / 27% less memory, ZAYA1-8B 89.6 HMMT (Zyphra); Codestral Mamba 7B 75% HumanEval at 256K (Mistral); IBM Granite 4.0 / Bamba 9:1 ratio, >70% RAM cut, Bamba 2.5×, ISO 42001 (IBM); Hunyuan-TurboS 56B/560B, 7th LMSYS Arena 1356 (2505.15431); xLSTM-7B (NXAI, EU) (2510.02228). - Depth-recurrent reasoning beats scale
[research]— HRM 27M → 40.3% ARC-AGI-1 vs o3-mini-high 34.5% (2506.21734); TRM 7M → 44.6% vs DeepSeek-R1 671B 15.8%, won ARC Prize 2025 (2510.04871); GRAM 10M → 52.0% ARC-AGI-1, 97.0% Sudoku-Extreme (2605.19376); Huginn-0125 3.5B reaches ~50B-equivalent compute at 32 recurrence steps (2502.05171). - The economics
[deployed]— inference is 80-90% of AI energy (MIT Tech Review); AI data-center electricity set to ~double to 945 TWh by 2030 (IEA); at 64K vs 2K tokens attention FLOPs rise ~1,024×, Mamba ~5× throughput (2312.00752); HBM sold out through 2026, capex >$400B (2025) → ~$700B (2026), KV elimination cuts per-request HBM ~10× (OpenReview); reasoning models burn 33-39 Wh/long prompt vs ~0.45 Wh, 86× (2505.09598); EU electricity ~$95/MWh vs ~$45 US (CNBC), xLSTM validated on AMD ROCm (AMD); ~600× token-price decline since 2020, mostly architectural (2603.28576).
Bio & health (cross-cutting)
Continuous, on-device, low-power biosignals with adaptation; spans latency, µW, and on-device-learning arguments.
- EEG seizure detection
[piloted]— response within a ~4 ms sample on a 300-400 µW implant/wearable. SynSense Xylo (<375 µW), BrainChip Akida + Onsor NEXA glasses (>95% accuracy). - Long-term ECG arrhythmia
[deployed]— 14-day coin-cell patch, O(1) per beat. iRhythm Zio (>1.1M patients; 64% of actionable arrhythmias missed after 48 h Holter), CardioLike-Net 5.85 mJ/inference. - Wearable sleep staging
[research]— NeuroSleep +7.5% accuracy at 45.8% lower compute than a dense transformer (2602.15888). - SSM biosignal models on MCUs
[research]— FEMBA runs a Mamba EEG model on a GAP9 RISC-V MCU at 75 mJ/inference, ~2 MB, 27× fewer FLOPs (2603.26716). - SIGINT / RF anomaly
[piloted]— SWaP-constrained airborne/satellite platforms can't carry GPUs. DARPA INTACT ($120M through 2027) Phase-2 awards (Nov 2025) to Intel, IBM, SynSense, Applied Brain Research.
Neuromorphic and event-driven hardware landscape
European players marked (EU).
| Company | Chip / product | Target markets |
|---|---|---|
| Innatera (EU, NL) | Pulsar neuromorphic MCU (SNN + RISC-V + CNN/DSP, <$5) | Smart home, wearables, industrial IoT, radar presence/gesture |
| SynSense (EU, DE/CH + CN) | Speck event-vision SoC; Xylo-Audio/IMU/Bio | Edge vision, hearables, automotive DMS, BCI, industrial IoT |
| Prophesee (EU, FR) | Metavision IMX636HD (Sony), GenX320, EVK4 | Automotive ADAS, industrial inspection, AR/VR eye tracking, robotics |
| iniVation (EU, CH) | DAVIS346, Aeveon adaptive event sensor | SLAM, robotics, space, scientific imaging, research |
| SpiNNcloud (EU, DE) | SpiNNaker2 (22nm, 153 ARM cores/chip) | Neuromorphic HPC, brain simulation, sovereign GenAI data centers |
| GEMESYS (EU, DE) | ReRAM memristor analog AI (on-device training) | Edge IoT, energy grid, cochlear implants; sovereign AI |
| SEMRON (EU, DE) | CapRAM 3D-stacked memcapacitor analog AI | Edge LLM inference, data centers; sovereign AI |
| Neurobus (EU, FR) | Neuromorphic edge software + integration | Space situational awareness, drone/satellite autonomy, defense |
| Opteran (EU, UK) | Opteran Mind (insect-brain software) | Cobots, drones, mining robots, space rovers (ESA/Airbus) |
| NXAI (EU, AT) | xLSTM | Edge and on-premise frontier LLMs, robotics, industrial |
| Applied Brain Research (CA) | TSP1 LMU/SSM accelerator; Nengo | Edge ASR, voice interfaces, biosignal, robotic control |
| Intel (US) | Loihi 2 / Hala Point (1.15B neurons) | Research, HPC, robotics, brain simulation, national security |
| IBM (US) | NorthPole; Granite 4.0 (Mamba hybrid) | Edge/datacenter inference, enterprise long-context |
| BrainChip (AU) | Akida AKD1000/1500/Pico, AkidaTag | Edge AIoT, industrial, automotive, healthcare, defense/space |
| Syntiant (US) | NDP100/115/120/250 (280 µW) | Earbuds, hearing aids, smart home, always-on voice/motion |
| Femtosense / Polyn / Aspinity (US) | SPU-001, NASP, AML100 | Hearables, VAD/KWS, smart home, always-on audio |
| Mythic (US) | Analog APU (17 fJ/MAC) | Automotive (Honda JDA), datacenter LLM inference, defense |
| Cartesia (US) | Sonic on-device SSM (Albert Gu) | Real-time voice, on-device assistants, Apple Silicon |
| Tsinghua / Lynxi (CN) | Tianjic / TianjicX hybrid ANN+SNN | Autonomous robots, domestic IoT, logistics |
Europe holds a deep bench (Innatera, SynSense, Prophesee, iniVation, SpiNNcloud, GEMESYS, SEMRON, Neurobus, Opteran, NXAI), aligned with the European Commission 2023 White Paper linking brain-inspired compute to semiconductor sovereignty.
Why now
- Energy. Data-centre electricity +17% in 2025 (AI centres +50%), AI power projected to triple by 2030 (IEA); EU electricity ~$95/MWh, ~2× the US (CNBC).
- Data scale. 21.1B connected IoT devices in 2025 generate 79.4 ZB/year (IoT Analytics).
- Regulation. GDPR, the EU AI Act (GPAI rules binding Aug 2025), China PIPL push inference on-device.
- Markets. Neuromorphic chip revenue ~$50M (2025) → ~$185M by 2030 (~30% CAGR); EU neuromorphic market $2.28B (2025) → $10.9B (2034), Germany 27.4% share (Market Data Forecast); TinyML $1.4-2.0B (19-22% CAGR); event cameras $5.1-8.0B (14-18% CAGR); on-device AI projected $185B by 2035.
- Capital & IP. Neuromorphic patent filings +401% in 2025. Funding: Unconventional AI $475M seed at $4.5B; Liquid AI $250M Series A (AMD) at $2.35B; Mythic $125M + Honda JDA; Innatera $21M; SpiNNcloud €10M; GEMESYS €8.6M; SEMRON €7.3M. Horizon Europe put €1.5B into neuromorphic research in 2025.
- Public deployments (one year). Medtronic adaptive DBS (FDA/CE 2025), iRhythm Zio Japan (May 2025), IDS uEye EVS (Mar 2025), Google Meet S2ST (May 2025), Microsoft Copilot Memory GA (Jul 2025).