At an ICML 2025 workshop on on-device learning and federated agents, a presenter showed a tiny demo that landed harder than any grand keynote: a smartwatch that quietly relearns your gait after an ankle sprain. No cloud call. No privacy pop-up. The model adapts locally during your day; overnight, a federated pass distills what's useful into anonymized updates that make everyone else's watch a little smarter—without exporting anyone's raw data. It was almost boring in how ordinary it felt, and that's precisely why it mattered. The future shows up when the obvious things get smart and stay private.
We've spent the last decade routing intimate moments through distant servers. That bought us scale, but it also introduced latency, fragility, and a standing privacy tax we've come to accept as normal. What's changing now is the substrate. Small, competent language models and perception models can run on phones, watches, vehicles, and clinical devices; federated and collaborative learning can move updates instead of data; and a new generation of power-frugal silicon and inference IP is making "always-on, always-learning" a product default rather than a research stunt. Put differently: intelligence is migrating back to where the data is born.
Our investment thesis follows that migration. First, we're backing compute that sips power. The constraint is the point. When you can deliver sub-watt inference with tight, predictable latency, whole categories of products change shape: hearables that transcribe and translate on-ear, cameras that understand scenes on-sensor, meters that forecast at the pole. You don't need a datacenter when the world itself is instrumented with tiny, attentive models. The economics flip: instead of paying in bandwidth and battery, you pay once in silicon and get privacy and responsiveness for free.
Second, we're investing in inference IP that the rest of the world's chips can adopt. Most teams building phones, cars, appliances, or industrial gear don't want to run an internal AI research lab; they want building blocks that drop into their SoCs and "just work" under hard power, thermal, and safety budgets. Hardened kernels, memory-aware schedulers, on-device guardrails, and rollback logic turn edge inference from a bespoke art into a reliable library. When AI behaves like any other well-engineered subsystem—pegged to predictable performance and certification paths—you unlock adoption at manufacturing scale.
Third, we're leaning into time-series intelligence that is native to the device. Not every problem needs a chatty generalist; many need a quiet specialist. Detecting an arrhythmia from wrist biosignals, flagging bearing wear from vibration, filtering speech robustly in noise—these are streaming tasks that reward causal features, compact adapters, and rock-solid reliability more than maximal parameter counts. Done right, they improve outcomes in places where bandwidth is scarce, privacy is non-negotiable, and failure modes are costly.
What does this unlock in the near term? Personal copilots that don't depend on a backend and therefore feel instantaneous and trustworthy; field tools that keep working in dead zones and sync when they can; bedside devices that monitor and suggest without exporting PHI; industrial systems that predict and react at the edge so a truck roll or a turbine restart is the exception, not the plan. The common thread is presence. Cloud AI delivered insight; edge-native AI delivers presence—the sense that your tools are here with you, attentive, and respectful.
To get there, a few things must be non-negotiable. Developer experience has to compress from quarters to hours: from model to bitstream to device with clear, automated steps. Update hygiene has to be built-in—safe on-device learning, robust rollback, remote attestation—so products evolve without surprising users or regulators. Fleet orchestration needs to respect physics: federated schedules that don't crater batteries or bandwidth, policies that keep local behavior local. And certification can't be an afterthought; medical, automotive, and safety standards require evidence packages that show not only performance but containment.
Our ask to founders working in this stack is simple because the environments are not: show us your latency, your energy per inference, and your plan for drift—how you learn, and just as importantly, how you unlearn—at the edge. If you can demonstrate those numbers on a modest dev board with a noisy, real-world dataset, we're already speaking the same language. The rest—markets, distribution, partnerships—gets easier when the technical truths are that crisp.
That small smartwatch demo at ICML wasn't flashy. It was inevitable. Intelligence that stays put—near its data, near its user, near the moment—wins on privacy, reliability, and feel. That's the arc we're investing behind: a world full of small, capable agents that don't ask permission to be helpful, because they don't need to leave your side to do it.
