{"slug": "femtoai-says-200000-spu-chips-are-in-market-as-marshall-signs-on", "title": "femtoAI says 200,000 SPU chips are in market as Marshall signs on", "summary": "FemtoAI said more than 200,000 of its SPU-001 inference chips are in market as the San Bruno, California, chipmaker expands work with Marshall, Samsung-linked ABOV Semiconductor and smart-glasses vendors. Co-founder and CEO Sam Fok also reported 5x growth during the first half of 2026, though the metric was not defined. The deployment count, disclosed in a July 22 announcement, provides the clearest measure yet of commercial progress for the Stanford spinout aiming to run AI on devices that cannot accommodate a GPU or persistent cloud connection.", "body_md": "[femtoAI](https://femto.ai/?ref=runtimewire) said more than 200,000 of its SPU-001 inference chips are in market as the San Bruno, California, chipmaker expands work with Marshall, Samsung-linked appliance supplier ABOV Semiconductor and smart-glasses vendors. The figure, disclosed in a [July 22 announcement distributed by PR Newswire](https://www.prnewswire.com/news-releases/femtoai-unlocks-5x-growth-with-full-stack-ai-inference-platform-302832353.html?ref=runtimewire), gives the clearest measure yet of commercial progress for a Stanford spinout trying to put useful AI into devices that cannot accommodate a GPU or a persistent cloud connection.\n\nSam Fok, femtoAI's co-founder and CEO, also reported 5x growth during the first half of 2026. femtoAI did not define the metric. Revenue, bookings, customer count, unit shipments and developer activity would each describe a materially different business, so the multiple cannot be used as a financial growth rate.\n\nThe deployment count provides firmer context. femtoAI reported shipping more than 100,000 SPU units by the end of 2025. Its latest figure says more than 200,000 chips are \"in market,\" wording that does not establish how many were shipped during the six months through June or how many are operating in finished consumer products. It does show that Fok's long-running bet on sparse computing has moved beyond laboratory prototypes and evaluation boards.\n\n### From brain-inspired research to commercial silicon\n\nFok founded femtoAI, then called Femtosense, in 2018 with CTO Alex Neckar and Chief AI Officer Scott Reid. The three came from Stanford University's Brains in Silicon Lab, where researchers studied computing architectures modeled on the brain's ability to perform useful work with limited energy.\n\nFok and Neckar had been developing that thesis for years before forming femtoAI. In 2012, the pair won a [$100,000 Qualcomm Innovation Fellowship](https://engineering.stanford.edu/news/stanford-engineering-team-wins-100000-qualcomm-innovation-fellowship?ref=runtimewire) for neuromorphic computing research that used digitally connected silicon neurons. They later worked on Braindrop, a programmable mixed-signal neuromorphic system designed around sparse representations.\n\nThe commercial insight was to keep sparsity while rebuilding the surrounding architecture for conventional AI models and manufacturing. In 2022, when femtoAI announced its Series A, Fok described the company's thesis as stripping out unnecessary computation so AI could run on lower-power devices.\n\nThat approach appears in femtoAI's current stack: pre-optimized models, developer software and custom silicon designed together. The [software toolkit](https://femto.ai/products/software-and-tools/?ref=runtimewire) lets engineers prune, quantize, simulate and compile models for the SPU, while the [SPU-001 accelerator](https://femto.ai/products/ai-accelerator/?ref=runtimewire) provides 1 MB of physical on-chip SRAM on a 22-nanometer process. femtoAI describes the design as enabling 10x model capacity when sparsity is applied and says speech, audio and one-dimensional data workloads can run below one milliwatt.\n\nThe distinction matters because sparsity can lose much of its benefit when model optimization and hardware execution are handled separately. femtoAI's \"dual sparsity\" design removes low-value weights from models and skips zero-valued operations in silicon. That reduces storage and computation at the same time, provided the model can be compressed without an unacceptable loss in accuracy.\n\nfemtoAI claims up to 100x lower power or energy consumption and up to 10x lower memory use. Those figures appear throughout femtoAI's own [technical materials](https://femto.ai/technology/?ref=runtimewire), and the July announcement does not include an independent benchmark, baseline processor, test workload or accuracy comparison. The gains should therefore be read as femtoAI's measurements rather than general performance guarantees.\n\n### Marshall gives femtoAI a consumer-audio route\n\nThe newest commercial name is Marshall. femtoAI says the audio brand will use its technology across premium headphone and earbud families, extending earlier audio work with hearing-aid supplier NewSound. The announcement does not identify a Marshall product, shipment date or unit commitment.\n\nThat relationship fits SPU-001's first-generation strengths. Speech enhancement, call cleanup, voice commands and sound detection can provide visible product features without the thermal and battery costs of running larger processors continuously. femtoAI launched ClaraCall 3.0 and ClaraCall 3.0-S, two call-quality enhancement models, through its developer portal alongside the commercial update.\n\nSmart glasses offer another route into high-volume consumer electronics. femtoAI says six next-generation glasses vendors have selected its platform, though Orka is the only customer named. Those devices need always-on audio, sensor and vision processing inside frames with little room for batteries or heat dissipation.\n\nfemtoAI faces established competition for those sockets. [Syntiant](https://www.syntiant.com/press-release/syntiant-delivers-full-stack-embodied-ai-to-real-world-users-at-ces-2026-bringing-intelligence-to-where-people-live-work-and-play/?ref=runtimewire) sells low-power processors, models and sensors for many of the same audio, smart-home and wearable uses, and says it has deployed more than 100 million pieces of purpose-built silicon and machine-learning models. femtoAI's 200,000-chip count remains small by that comparison. Its opportunity rests on customers deciding that sparsity-aware hardware and model tooling produce enough power, memory or integration savings to justify another chip architecture.\n\n### Appliance deals test whether the stack can scale\n\nA second distribution path runs through ABOV Semiconductor, which supplies microcontrollers for appliances, remote controls and home automation. In February, femtoAI and ABOV signed a [five-year agreement](https://femto.ai/press_releases/femtoai-and-abov-deliver-ultra-efficient-edge-ai-for-consumer-electronics/?ref=runtimewire) to package SPU-001 alongside ABOV microcontrollers and femtoAI software for language interfaces and predictive diagnostics.\n\nThe July announcement describes that work as involving Samsung appliances, though it does not specify models, shipment volumes or Samsung's contractual role. The ABOV agreement is strategically important because an accelerator sold alongside an established microcontroller can enter designs through a supplier that consumer-electronics manufacturers already use.\n\nDeveloper adoption is part of the same strategy. femtoAI says nearly half of its customers are building their own algorithms or models in its developer portal. The denominator is undisclosed, leaving the scale of that activity unclear. Supporting customer-created workloads could still widen SPU-001 beyond femtoAI's initial catalog of audio models into robotics, vision denoising, industrial fault monitoring and energy management.\n\nfemtoAI says a next-generation SPU is under development, without releasing specifications or a delivery schedule. The first chip is optimized around constrained audio and sensor workloads. A successor that handles more demanding workloads would put femtoAI into broader markets with higher performance requirements and better-funded competitors.\n\n### Patient capital behind a long hardware build\n\nThe commercialization effort has taken eight years. In September 2022, femtoAI raised an [$8 million Series A](https://www.businesswire.com/news/home/20220928005770/en/Edge-AI-Innovator-Femtosense-Raises-%248-Million-Series-A-Funding?ref=runtimewire) led by [Fine Structure Ventures](https://finestructure.vc/?ref=runtimewire), with participation from [J2 Ventures](https://www.j2vp.com/?ref=runtimewire), [Quest Venture Partners](https://www.questvp.com/?ref=runtimewire), [Amino Capital](https://www.aminocapital.com/?ref=runtimewire), [Gaingels](https://gaingels.com/?ref=runtimewire) and [SV Pacific Ventures](https://svpventures.com/?ref=runtimewire), among others. TechCrunch reported at the time that the financing valued femtoAI at $27 million and brought its total raised to $11 million.\n\nfemtoAI's current investor list also includes [Kleiner Perkins](https://www.kleinerperkins.com/?ref=runtimewire), Vanedge Capital and [Calm Ventures](https://www.calmvc.com/?ref=runtimewire).\n\nFok has spent the years since Stanford converting a neuromorphic research idea into silicon that can survive consumer-product economics. Marshall, ABOV and the unnamed smart-glasses customers give femtoAI routes to larger deployments. The next proof point will come from named products and repeatable shipment volumes, alongside a growth metric that customers and investors can evaluate.", "url": "https://wpnews.pro/news/femtoai-says-200000-spu-chips-are-in-market-as-marshall-signs-on", "canonical_source": "https://runtimewire.com/article/femtoai-200000-spu-chips-marshall-edge-ai", "published_at": "2026-07-22 16:54:15+00:00", "updated_at": "2026-07-22 17:09:14.189193+00:00", "lang": "en", "topics": ["ai-chips", "ai-infrastructure", "ai-products", "artificial-intelligence"], "entities": ["femtoAI", "SPU-001", "Marshall", "ABOV Semiconductor", "Sam Fok", "Stanford University", "Qualcomm"], "alternates": {"html": "https://wpnews.pro/news/femtoai-says-200000-spu-chips-are-in-market-as-marshall-signs-on", "markdown": "https://wpnews.pro/news/femtoai-says-200000-spu-chips-are-in-market-as-marshall-signs-on.md", "text": "https://wpnews.pro/news/femtoai-says-200000-spu-chips-are-in-market-as-marshall-signs-on.txt", "jsonld": "https://wpnews.pro/news/femtoai-says-200000-spu-chips-are-in-market-as-marshall-signs-on.jsonld"}}