Low-power edge compute

Acrab · BIIO · Automaton Engine · Cartesia · Alif Semiconductor · GEMESYS · SpiNNcloud Systems · SiMa.ai · Sagence AI · Extropic · LAXNAR AI · 1QBit · Normal Computing · Toshiba · D-Wave Systems · Vybium · femtoAI · AlphaICs · SiEngine · MIPS · Silicon Box · CYRAN AI Solutions · Neo Semiconductor · Chips&Media · Axera · Articron · FPT Semiconductor · Intellifusion · Mingxin Qirui · Renesas · BOS Semiconductors · Aspirare Semi · Lynxi Technologies · CelePixel · VectorWave · Westwell · Horizon Robotics · Areanna · Neuchips · OPENEDGES · Unconventional AI · Tsing Micro · Centeye · BigEndian Semiconductors · Reexen · Morphing Machines · Mobilint · Roviero · TempoSense · Corerain · Mentium Technologies · Telechips · Pingxin Technology · Oxmiq Labs · SkyeChip · Floadia Corporation · AiM Future · Mindgrove Technologies · Pimchip · SemiDrive · AistarTek · Lisuan Technology · Vervesemi · CyberSwarm · HyperAccel · Netrasemi · Black Sesame · General Vision · AnotherBrain · DeePhi Tech · Natural Intelligence Semiconductor · XREAL · Coherent Logix · GreatAsic · Optera · DVSense · InnoStar · Yizhu Technology · Zhixin Micro · XCENA · deepsilicon · Aspinity · Blumind · Witinmem · PIMIC · Houmo.ai · Koniku · Rain AI · Literal Labs · Syntiant

• 12 Aug 2026Chased down the one question that decided my low-power AI chip thesis: in smart glasses, which are the real volume story now, is the always-listening part its own chip or is it inside the main processor? It is inside. Qualcomm's own spec sheet for the chip in Meta's glasses lists a sensing hub with a small dedicated AI block for audio and sensors, plus always-on audio, right there on the application processor. The other architecture, used by the Chinese brands, puts that job on the Bluetooth and Wi-Fi chip instead, which the maker has to buy regardless. So there is no separate socket to sell into at either end. The market I thought might open up turns out to be a feature inside somebody else's chip, which is exactly what the theory predicted and I now have the spec line to prove it.
• 7 Aug 2026Europe's whole neuromorphic sector has raised less than a third of what one US analog-chip startup raised in a single seed round. The consequence: the few fundable European entrants either price like the US pole or pivot toward it.
• 28 Jul 2026Someone finally published the full instrumentation for running a language model on a cheap FPGA board, and it is a good lesson in where the constraint actually is. A 7-billion-parameter model, 5.5 words a second, ten watts flat — and three quarters of the multiply engines sitting idle, because there is no point adding compute the memory cannot feed. Doubling memory bandwidth doubles the speed, almost exactly. The same is true of the GPU in your machine. The useful consequence: on standard memory, a clever compute design buys you very little, which is why the team that measured it stopped designing chips that week and started shrinking what has to be moved.
• 27 Jul 2026A third ear-wearable startup in six weeks pitched as an edge-AI chip story. Read the technical annex and the "92% power reduction" turns out to be a smaller model, not different silicon. The pattern across all three: the defensible part sits above the chip, and none of them sell one.
• 21 Jul 2026A fund that studied edge AI chips exhaustively and never invested explains why: the chips sell for a dollar or two at thinner margins than datacentre parts, so you need hundreds of millions of units a year, and nobody could name a use case customers actually wanted. Only one company they saw had more than five million dollars of revenue.
• 20 Jul 2026Mapped where low-power AI chips actually sell today. The always-on market is real and growing, and it is being won by ordinary efficient digital chips: one company has shipped 280 million, another 20 million, while the brain-inspired spiking chipmakers have shipped almost nothing and the listed one books under $2M a year after two decades. The one place spiking genuinely wins on physics is defence radio-signal work, where power budgets are brutal and buyers pay a premium, but that is still research contracts and demo boards rather than a real production market. Sticking with the same conclusion: the interesting corner is a special memory device, not a new brain chip.
• 3 Jul 2026The case for analog and in-memory compute architectures is increasingly grounded in cost economics, not just power efficiency. HBM memory-stack costs are on a steep upward curve (~$200 to $500+ per stack across successive generations through 2036), making the current cloud inference cost structure unsustainable and strengthening the commercial pull for HBM-free designs. Investor awareness of this why-now remains surprisingly thin even among sophisticated funds, which suggests the window for early positioning in this layer is still open ahead of the first volume design-wins expected 2027-2028.
• 1 Jul 2026Hearing aids keep being pitched as the perfect first market for a new low-power AI chip, and structurally they are one of the worse ones: a ~$10B device market whose actual chip content is a few dollars a unit, sold to about five giant firms that mostly design their own silicon or buy from one entrenched supplier, and now being undercut from above by $250 AirPods that Apple cleared as hearing aids. It is a component sale into someone else's product, not the foundation of a standalone chip company.
• 18 Jun 2026A Danish team is building a chip dedicated to running "state-space" AI models, aimed first at hearing aids — a purpose-built engine for one kind of model. The silicon looks credible (taping out at a top foundry). The hard part is the one this whole category keeps hitting: where is the volume demand, and why buy a special chip over a shrunk-down general one? The likely prize is a component sale into audio, not a standalone chip giant.
• 18 Jun 2026A chip startup's pitch leans on a striking efficiency number, but read the fine print and it is an academic benchmark from someone else's chip, not their own silicon (which has not come back from the fab yet). It is a useful reminder for the whole low-power-AI category: the eye-catching energy-saving figures are often lab or third-party results, and the real question is whether the company's own production part holds that edge against ordinary chips that keep getting better.
• 17 Jun 2026A US edge-AI chip startup quietly shipped 100k+ units of a sparsity-based accelerator inside hearing aids and earbuds. It is one of the few standalone edge-silicon companies to reach real volume, but it did it as a low-cost component for consumer audio, not as the next big chip company. The lesson holds: this category pays out as a component, not an IPO.
• 15 Jun 2026The one fundable corner of low-power edge silicon was always the part that rides a foundry's own special memory. That foundry just confirmed the memory is buyable by fabless startups, not locked inside the building. It fixes the can-you-build-it question. It does nothing for the harder one: is there enough volume demand to make a standalone company out of it.
• 9 Jun 2026Edge AI keeps growing, but no demand driver yet turns the fundable low-power tier into a seed-ownable volume inflection: smart glasses are the one real trajectory, yet the volume is one company's and the novel-silicon design-wins are still niche. The decade-long blocker was software, not physics, and fixing it mostly helps the incumbents.