The promise sounds almost too good to be true: smart glasses that actually last all day. Not six hours. Not "up to eight under ideal conditions." All day.
Mosaic SoC, a Zurich-based spinout from ETH Zürich, thinks it has a shot. The semiconductor startup announced a $3.8 million pre-seed round on April 30, led by Founderful with participation from the Kick Foundation. The timing is no accident. Smart glasses shipments surged 110% year-over-year in the first half of 2025, according to Counterpoint Research, and the category is only heating up.
What Mosaic is building—a purpose-built perception chip that integrates SLAM (simultaneous localization and mapping), object detection, and neural inference into silicon designed for continuous operation—addresses perhaps the most stubborn engineering challenge in consumer wearables. The question is whether a pre-seed-stage startup can carve out space in a market already dominated by giants like Qualcomm and Meta, both of whom have spent years and untold millions chasing the same goal.
Why Now?
The smart glasses market is having a moment. IDC expects the category to drive the majority of extended reality growth in 2026, with shipments forecast to climb 33.5% after jumping 44.4% in 2025. Meta commands 72.2% of the XR market as of 2025, largely on the strength of its Ray-Ban collaboration, according to IDC data. The company has refreshed its Ray-Ban Meta lineup, launched its first Oakley-branded smart glasses, and introduced a display-equipped variant—the Ray-Ban Display—in September 2025.
But here's the rub: that display model, which remains "extremely limited" outside the U.S. due to inventory constraints, gets about six hours of mixed-use battery life according to reviews published between January and February 2026. Six hours. For a device that needs a smartphone tether to offload heavy compute.
Display-less models fare slightly better, claiming up to eight hours under certain conditions, though real-world use varies—sometimes considerably. The core challenge is brutally simple: packing spatial awareness, generative AI, and genuine all-day endurance into eyewear-class form factors requires rethinking the entire compute stack.
The market seems willing to overlook early limitations, at least for now. Waveguide-equipped smart glasses—those with optical displays—saw shipments surge dramatically in the second half of 2025, per Counterpoint data published April 8, 2026. TrendForce projects roughly 950,000 display-equipped glasses units will ship in 2026, a figure Meta has revised upward twice already, according to a January 30 report.
Google's Android XR platform, announced in December 2024, has pulled Samsung into the fray; the Korean giant confirmed smart glasses for 2026 in reports published early this year. Qualcomm demonstrated on-glass generative AI in a June 2025 demo, showing real-time inference running directly on eyewear, and positioned its Snapdragon Wear Elite platform as the engine for "the next wave of AI wearables" in March 2026 materials. Even Solos launched camera-enabled AI glasses at CES 2026 for $299, underscoring the price pressure building in the category.
The inflection point, it seems, is here.
The Physics Problem
The promise of spatial intelligence—devices that understand where they are, what they're looking at, and how to respond—runs headlong into physics. SLAM algorithms, object detection pipelines, and neural network inference are computationally expensive. Run them continuously on conventional architectures and battery life collapses, often more dramatically than product teams anticipate.
Event-based vision sensors offer one potential escape hatch. These cameras capture changes in a scene rather than full frames, reducing data volume and power draw to milliwatt or even microwatt levels. Prophesee's GenX320 sensor, used in 7invensun's aSee wearable eye-tracker, operates in modes consuming tens of microwatts, according to product briefs published in 2025 and 2026.
Academic work reinforces the potential. A November 2025 arXiv paper described battery-powered pupil tracking on a neuromorphic SoC consuming less than 5 milliwatts per eye at 100 Hz. Another 2025 study detailed an event-camera gesture system, Helios 2.0, running under 20 milliwatts. These are encouraging numbers—if they can be replicated outside the lab.
Yet sensors are only part of the equation. The compute stack matters just as much, perhaps more. Heterogeneous system-on-chip designs—mixing general-purpose cores with dedicated accelerators for vision, SLAM, and neural inference—have proliferated. Syntiant's NDP250, showcased in 2024 and 2025, targets always-on audio and vision processing; the company claimed more than 100 million deployments across its NDP family as of June 2025. Himax unveiled WiseEye and WiseGuard "milliwatt-class" always-on AI sensing at CES 2026. Ambient Scientific's GPX10 Pro, sampling in 2025 with volume production targeted for the first quarter of 2026, markets itself as delivering "100× MCU performance" for always-on AI, though independent benchmarks remain sparse as of May 2026.
Research continues to push boundaries. A February 2026 arXiv paper on spiking neural networks reported up to 847 giga-operations per second per watt on edge hardware, with latency measured in milliseconds. The IEEE Low-Power Computer Vision Challenge, analyzed in an April 21 paper, continues to set new efficiency records. These are research benchmarks, not product specifications. But they signal where the field is heading.
The ETH Pedigree
Mosaic SoC's founders bring credentials in exactly this domain. Dr. Alfio Di Mauro, the CEO, completed a PhD focused on SoC architectures for event-driven computing at ETH Zürich. Dr. Moritz Scherer, the CTO, researched hardware-software co-design for energy-efficient neural network inference at the extreme edge, also at ETH.
Their prior work includes Siracusa, a 16-nanometer RISC-V extended reality SoC with an at-MRAM neural engine published in 2023 and 2024; Kraken, a 22-nanometer multi-sensor fusion SoC for nano-UAV visual processing from 2022; and Marsellus, an AI-IoT SoC with two-to-eight-bit deep neural network acceleration from 2023. It's the kind of resume that tends to get investor attention in the chip world.
The company describes its product as an integrated hardware and software platform for spatial intelligence—SLAM, object detection, and neural inference baked into silicon. No performance metrics have been disclosed publicly, which is standard at this stage. Jon Peddie Research, in a note published last week, positioned Mosaic "at the intersection of AI and vision" and noted advisory ties to ETH and the PULP (Parallel Ultra Low Power) platform, a longstanding research initiative in ultra-efficient processors.
The funding round gives Mosaic runway to move from research prototypes to production-ready chips. Pre-seed rounds of this size in the semiconductor space typically support tape-out costs—the process of finalizing a chip design for manufacturing—along with early customer engagements and team expansion. Mosaic lists openings for digital design engineers, embedded software developers, and application engineers on its website, signals of a push toward silicon and software validation.
A Crowded, Fragmented Field
Mosaic enters what might generously be called a competitive landscape, though "competitive" may overstate the clarity of the market. Chip companies serving wearables and edge AI span a wide spectrum, each staking out different territory.
Qualcomm's AR2 Gen 1 platform, launched in November 2022 and integrated into products through 2025 and 2026, powers many smart glasses today. The company demonstrated on-glass generative AI in 2025 and positions its broader "personal AI" vision—a mesh of phones, watches, glasses, and earbuds—as central to its 6G and agentic AI strategy, per company blogs published throughout 2025 and early 2026.
BrainChip's Akida neuromorphic processor showcased event-based vision with Prophesee at Embedded World 2025 and CES 2026, targeting always-on wearable and edge applications. Perceive's Ergo 2, covered by TechInsights in March 2026, supports transformers on low-power budgets; earlier product launches claimed sub-100-milliwatt compute power, though the Ergo 2 targets roughly 400 milliwatts.
Software-focused companies like SLAMcore, which has partnered with Arm and Qualcomm, provide visual SLAM stacks for developers. Sevensense, acquired by ABB in January 2024, integrates visual SLAM into autonomous mobile robots—a reminder that spatial intelligence extends well beyond wearables. ABB materials from 2025 cite mobile robot market growth around 20% CAGR through 2026, and NVIDIA's Jetson Thor platform, configurable from 40 to 130 watts, addresses higher-compute robotics with Blackwell-based modules shipping in 2025 and 2026.
The fragmentation reflects the early stage of the market. Display-less smart glasses, which IDC expects to lead 2026 growth, offload heavy compute to smartphones but still need on-device, always-on perception for context—exactly Mosaic's positioning. Display-equipped models demand even more local processing. Robotics and automotive applications have different power envelopes entirely.
Purpose-built silicon for each segment remains an open question.
What Comes Next
Edge AI chip reports published in early 2026 forecast continued growth through the next decade, with smartphones dominating volumes today and robotics driving increasing AI-per-device requirements. Omdia data from late 2025 and early 2026, covered in February, noted that semiconductor revenues passed $1 trillion in 2026 for the first time, with AI as the primary driver and edge AI inference chips highlighted as a growth area.
An Edge AI & Vision Alliance piece from November 2025 emphasized the stakes: 41.6 billion connected IoT devices projected by 2025 generating 79 zettabytes of data annually, per IDC, with energy constraints, privacy concerns, and latency all favoring on-device processing.
AR Insider's March 17 spatial computing revenue outlook projects an inflection point in 2026 and 2027 as major OEMs expand their smart glasses catalogs. Meta leadership, in remarks covered during a late January earnings call, has framed AI glasses as the future of computing. Qualcomm executives, in statements from late 2025 and early 2026, emphasize personal AI across a device mesh and the importance of on-device perception and battery efficiency.
Yet the path forward has obstacles beyond the technical. Privacy scrutiny intensified in early 2026, with over 70 advocacy organizations opposing facial recognition on Meta's smart glasses, according to TechRadar coverage from January and February. The EU AI Act's core rules take effect August 2, with staggered timelines extending to 2027 and beyond. California's CPPA final regulations on automated decision-making technology and cybersecurity audits took effect January 1. Illinois BIPA settlements, including a $4.25 million Lytx settlement covering 2016 to 2025 and approved last year, signal ongoing litigation risk around biometric data in wearables.
The Open Question
Mosaic SoC's bet is straightforward: purpose-built perception silicon, tightly integrated with software, can solve the always-on problem at a power budget that makes all-day wearables practical. The founders have published extensively on ultra-low-power vision SoCs. The market is surging, perhaps faster than anyone expected.
But the gap between research benchmarks and shipping products is wide, and the graveyard of promising chip startups is crowded. Incumbents like Qualcomm, Meta, and Google have deep pockets, established customer relationships, and ecosystem lock-in. Mosaic's $3.8 million gives the team a shot at proving its architecture in silicon—at demonstrating that the years of academic work can translate into something manufacturers will actually design into products.
Whether that translates to design wins in a market moving this fast remains to be seen. The founders have the technical chops. The timing looks right. The question, as always in semiconductors, is execution. And in a field where power budgets are measured in milliwatts and battery life in hours, the margin for error is thinner than the glasses themselves.
