AGI will require far more energy than the world can easily supply today.

The bottleneck is not just better models, more data, or more chips. It is power. Training frontier models, serving inference at scale, and deploying intelligence into the physical world all require enormous amounts of electricity. If AI keeps improving, energy becomes one of the main constraints on progress.

This matters because AGI will not stay inside the data center. It will be used across robotics, manufacturing, defense, science, and industrial systems. That means the energy demand is not just for training runs. It is for continuous inference, always-on infrastructure, and the electrification of the physical economy that AI will help automate.

We think one of the most important opportunities in the next decade is building the energy stack for an AGI world: generation, storage, transmission, power management, and new systems that can support much larger compute and industrial loads.

We are interested in startups that can help produce, store, and deliver far more energy for a world with AGI.

Most spacecraft in LEO still run on solar panels and batteries. That stack works, but it has real limits: no generation during eclipse, degradation over time, and surface-area, mass, and deployment tradeoffs. As spacecraft become more power-hungry, that architecture starts to look like a constraint rather than a default.

We’re interested in startups building the next generation of power infrastructure for spacecraft in LEO. That could mean alternative generation, better storage tightly coupled with generation, beamed power, peak-power architectures, or other systems that materially improve availability, endurance, and mission flexibility. The wedge is not “replace solar for everything.” It is to enable classes of spacecraft that need far more reliable and abundant power than the traditional satellite stack was built for.

The tailwind is emerging in-space compute. SpaceX asked the FCC for approval for up to 1 million solar-powered AI data-center satellites and others are also beginning to push toward orbital data-center infrastructure, pointing to a future where spacecraft are designed around much larger power demands than before. Whether or not those exact visions scale, they point in the same direction: orbital systems are being designed around much larger power budgets than before.

If you’re building power systems, storage, beamed energy, or adjacent infrastructure for high-power spacecraft in LEO, we’d love to hear from you.

The global manufacturing system was built for a different era — one defined by cheap labor, stable geopolitics, and long supply chains. That model is breaking down, but reshoring remains too expensive and labor-intensive under the traditional factory model.

A new model is emerging: autonomous factories. These are manufacturing systems built from the ground up around robotics, AI, and software rather than human labor. Robots perform machining, assembly, inspection, and material movement. AI systems manage scheduling, monitor quality, coordinate workflows, and continuously improve performance across the factory. The result is a production environment that can run with far less labor, adapt more quickly, and operate competitively in high-cost regions.

This is becoming possible because multiple tailwinds are converging at once. Robotics has improved enough to automate a much wider range of industrial tasks with greater precision and reliability. At the same time, advances in AI are making it possible to coordinate increasingly complex systems, turning factories into programmable, optimizable environments rather than fixed physical workflows. And as geopolitical instability exposes the fragility of offshore manufacturing, the demand for resilient, domestic production capacity is becoming far more urgent.

The opportunity is not just to automate pieces of the factory. It is to rebuild manufacturing around a new operating model entirely — one that is more resilient, more scalable, and less dependent on fragile global labor and logistics networks.

We are especially interested in teams starting with high-value, strategically important categories such as defense hardware, energy systems, robotics components, and critical materials, where the economic and national importance of domestic manufacturing is highest. These markets can serve as the wedge for a much broader transformation of industrial production.

Today's deployed robots are frozen. They ship with a fixed model and never improve from their own experience. Every customer site, every edge case, every successful recovery is wasted as a training signal. Meanwhile, the data they need most to improve only exists in deployment: real contact, real human interaction, real environmental variation that no simulation can replicate.

This creates a problem. Teams that delay deployment wait for data that only deployment can provide. Teams that deploy but aren’t learning from the accumulated experience. The teams that will pull ahead are the ones that are able to treat deployment as a learning system: robots that actively get better from their own operation.

This means two things happening together. First, operational infrastructure that makes early, narrow deployment safe and productive: human-in-the-loop monitoring, intervention systems, and data pipelines that turn every hour of operation into durable training signals. Second, learning infrastructure that turns that signal into improved behavior: reinforcement learning systems with real safety bounds, no idealized resets, and adaptation cycles that an integrator's operations team can manage without RL researchers on staff.

The customer for this is defined and waiting. Hundreds of robotics integrators serve specific verticals (i.e. logistics, food processing, agriculture, electronics assembly) and every one of them needs their deployed systems to improve over time. A picking system that learns a faster grasp strategy from its own production data. A mobile robot that discovers more efficient routes after months in a facility. They will pay for this. They cannot build it themselves.

We are interested in teams building the combined deployment and learning infrastructure that lets robots improve from real-world operation - spanning safe early deployment, data capture, and production reinforcement learning that integrators can use without in-house ML teams.

Rare earths and critical minerals are upstream of nearly everything that matters: robotics, AI hardware, energy systems, satellites, defense, and advanced manufacturing. But today, the supply chain for these materials is concentrated, fragile, and geopolitically exposed.

The opportunity is not just to mine more. It is to rebuild the full stack: exploration, extraction, separation, refining, and downstream production. The real bottleneck is turning raw deposits into reliable, high-purity material at scale.

This is becoming more possible because the tools have improved. Better sensing, Earth observation, autonomous equipment, and AI-driven process control can make mining faster, safer, cleaner, and more efficient. What used to be slow, manual, and fragmented can now become far more automated and integrated.

We are interested in teams building new platforms for critical minerals and rare earths that can secure supply, reduce dependency on fragile foreign infrastructure, and strengthen industrial sovereignty.

Most leading AI systems today have two distinct phases: training, then deployment. You pre-train, mid-train, post-train, freeze the weights, and deploy. The model at inference is static. It never learns from its own experience.

Humans and animals work fundamentally differently. From the moment a baby is born, it is simultaneously deployed and actively learning. There is no separation. Every interaction is both an output and a learning signal. Day 1 is training. Day 10,000 is still training. The system never freezes.

When you hire an employee, they improve every day. They learn your preferences, your workflows, your edge cases. They get smarter as you teach them. The most transformative AI systems will work the same way: models that continuously learn and adapt to their owners, getting better with every interaction.

This is the architectural gap between current frontier AI and true intelligence. Artificial General Intelligence is valuable, but what we need is Artificial General Learners. Systems that are always deployed, always learning, always improving from experience.

This is the most important unsolved problem in AI, and solving it will unlock AI's ability to truly accelerate humanity.

We are looking for founding teams with strong theses toward architectural changes, online learning infrastructure, new environments that facilitate continuous learning with minimal sink states, or product and business model innovations that create more effective teaching interactions for deployed AI systems.

Reliable lunar-to-Earth data links remain expensive, bespoke, and tightly coupled to a small set of government architectures. Existing solutions prioritize flagship missions and high-power assets, leaving commercial lunar payloads, secondary missions, and distributed surface systems underserved. 

We are interested in teams building practical, scalable RF/optical communications systems for Moon-to-Earth links; covering terminals, ground integration, and network architectures. Solutions may trade peak throughput for reliability, cost, or deployability, but must work with realistic mass, power, and pointing constraints. 

The goal is to make lunar data return routine, affordable, and operationally simple, not heroic.

A robot can identify a wine glass with 99% accuracy and still shatter it. That is the state of the field.

The problem is not sensors. Affordable, high-resolution tactile hardware exists and is improving rapidly. The problem is that robotics has no ecosystem for tactile knowledge. There is no world model for touch. No pretrained representations for contact, friction, compliance, or material properties. No standard way to capture, store, share, or learn from tactile experience across tasks and platforms.

Vision succeeded not just because cameras got better, but because the field built shared data, representations, and benchmarks - an infrastructure for visual understanding that any team could build on. Tactile and force understanding has none of this. Every team that needs contact intelligence starts from scratch, for one task, on one robot, and throws it away when the project ends.

This gap matters because the failures that block deployment are overwhelmingly contact failures. Too much force on a fragile object. Too little grip on a slippery surface. No awareness of compliance when a material deforms. These are not perception problems. They are understanding problems that require representations rich enough to be learned from, reasoned over, and transferred.

The opportunity is to build the data and representation layer for tactile intelligence: collection pipelines, shared datasets, learned representations for contact and force that work across robots and tasks. Whoever builds this becomes the backbone of physical manipulation - the way pretrained vision models became the backbone of computer vision.

We are interested in teams building the data infrastructure and learned representations that make tactile and force understanding a transferable, reusable capability for robotics.

There is a unique moment in history to rethink most hardware products from first principles.

Legacy incumbents built tractors, drones, forklifts, excavators, generators, trucks, and boats, and think of their market as a market for that mechanical form factor. But that framing will soon become obsolete. The correct way to think about these products is as vessels for AGI. Each one is a different body designed for a different use case, but all will be inhabited by human-level intelligence to complete their tasks.

A tractor company will not be selling a tractor. They will be selling human intelligence embedded in a tractor's body to carry out farming tasks. A drone company will not be selling drones, but drones with superintelligent pilots. Once you see the future of hardware through this lens, you realize every major product category needs to be rebuilt from scratch by founders who understand this.

The technical breakthroughs that enable this future are converging. Improving multimodal foundation models allow intelligence to inhabit these vessels. Satellite connectivity gives them permanent access to maximum intelligence and tokens, anywhere on earth, removing the constraint of what can run at the edge.

Improving global supply chains lets startups match legacy OEM quality at lower costs, and AI lets small teams speedrun the hardware knowledge gaps that used to take decades. New distribution channels through short-form video and founder-led communities let insurgent brands scale in ways incumbents don't understand.

This is how American startups build the next generation of globally dominant hardware companies: by recognizing that every product category is about to be inhabited by intelligence, and building for that future now.

We want to fund technical founders building AI-native hardware companies across every major product category.

The US has a $4.6 trillion infrastructure maintenance backlog. Globally, the number is almost incomprehensible. Bridges, pipelines, buildings, roads, power lines, rail, water systems. These are all degrading faster than human crews can inspect and repair them.

The current model is broken in a specific way: maintenance is episodic. An inspector visits a bridge every two years. A building gets a roof replacement every twenty. A pipeline is checked after a leak is reported. Between inspections, damage accumulates silently. Small problems become large failures. Repair costs grow exponentially. The result is a world where physical infrastructure is in perpetual decline, interrupted by expensive, disruptive fix cycles.

Continuous robotic maintenance inverts this model. Instead of periodic inspection and deferred repair, robotic systems monitor structures in real time and perform small, frequent repairs before damage compounds. A crack is sealed the week it appears, not the year it becomes critical. Corrosion is treated at first detection. Wear is compensated continuously rather than replaced catastrophically.

This extends beyond the built environment. Natural systems: forests, coastlines, waterways and soil suffer the same problem: degradation outpaces the human capacity to restore. Robotic systems that can operate persistently in remote, harsh, or hazardous environments make continuous stewardship of both built and natural infrastructure feasible at a scale that human labor never could.

The pieces exist. Inspection robots, autonomous drones, and remote sensing are already deployed. What is missing is the integration into closed-loop systems that detect, decide, and repair, not just report. The opportunity is end-to-end: from continuous sensing through automated diagnosis to physical intervention, operating across structures that range from bridges to ecosystems.

We are interested in teams building robotic systems for continuous physical maintenance -  closed-loop platforms that detect degradation and act on it, extending the lifespan of infrastructure and natural systems rather than letting them decay between human visits.

Historically, this industry has fallen short not because of the spacecraft or RF architecture itself, but because these systems required ground sensors in remote locations to adopt new communication standards. Many early approaches achieved basic connectivity, but failed on power efficiency, latency, reliability, or economics at scale.

It is now 2026, not 2016. The industry and underlying technology have matured significantly over the past decade. Companies are now pushing to place massive compute capacity in space, and SpaceX has reportedly acquired $20B of cellular and wireless spectrum to support future direct-to-cell services.

The best opportunity is to build a reliable system architecture that delivers responsive connectivity to sensors using their existing communication chips, while respecting realistic power budgets and sound system-level design.

Humans have an entire institutional infrastructure for learning—preschools for motor skills, K-12 for general knowledge, trade schools for specialized skills, apprenticeships for craft mastery. Robots have nothing.

This is a multi-trillion dollar market waiting to be built. As humanoids and general-purpose robots proliferate, they will need places to learn. Not just simulation—actual physical facilities where robots can practice manipulation, develop skills, and receive human supervision and correction.

Imagine the tiers: Robot preschools teaching general manipulation—picking, placing, grasping objects of varying shapes and materials. Robot high schools covering household and commercial tasks. Specialized vocational centers for electricians, plumbers, welders, warehouse operators. Industry-specific training facilities for oil rigs, semiconductor fabs, surgical theaters.

Each tier requires different physical infrastructure, different curricula, and different feedback mechanisms. The founders who build these institutions will own critical chokepoints in the robotics value chain—every humanoid company will need to send their hardware through your schools.

We're looking for founders who want to build the physical and digital infrastructure for robot education. This is institution-building at a civilizational scale.

Space is an appealing frontier for compute: abundant solar power, fewer terrestrial constraints, and the possibility of entirely new infrastructure layers beyond Earth. But there is a hard physical bottleneck—heat. In orbit, you cannot rely on air or water cooling. If you want to run meaningful compute in space, the heat must be rejected through radiators, and today’s systems are often too heavy, too fragile, or too inefficient to make that practical at scale.

We’re interested in startups building the thermal infrastructure that makes in-space compute possible. That could include advanced radiator architectures, lightweight deployable systems, better heat transport loops, new materials and coatings, micrometeoroid-resilient designs, or fully integrated thermal-compute platforms. The goal is not just a better component, but a system that makes persistent, high-density compute in space economically and operationally feasible.

What excites us is the hidden bottleneck. Most people think the future of compute is constrained by chips or power. But in many extreme environments, it is equally constrained by where heat can go. If in-space compute becomes real, it will likely be because someone solved thermal rejection well enough to unlock the entire category.

If you’re building in spacecraft thermal management, in-space datacenters, orbital infrastructure, or adjacent enabling hardware, we’d love to hear from you.

The scientific method has a throughput problem. A materials scientist can run tens of experiments per week. A chemistry lab can test hundreds of formulations per year. Progress in physical science is gated by how fast humans can set up, execute, observe, and iterate on physical experiments.

Robots do not have this constraint. A robotic experimentation platform can run thousands of physical tests per day. Systematically varying parameters, measuring outcomes, and adapting protocols based on results. Not just automating a fixed procedure, but exploring a physical search space with a breadth and patience that human researchers cannot match.

This is already happening in narrow domains. High-throughput screening in pharmaceutical research has produced drug candidates that manual processes would have missed. Automated materials discovery platforms have identified novel alloys and compounds by searching combinatorial spaces too large for human-directed exploration. But these remain bespoke, expensive systems built for specific applications.

The general opportunity is robotic experimentation as infrastructure: platforms that let researchers define a physical question and let robots answer it through systematic, high-throughput interaction with the real world. Testing material properties under thousands of conditions. Discovering manufacturing process parameters that optimize for outcomes no human would have intuited. Probing physical phenomena at scales and speeds that reveal structure invisible to manual investigation.

The convergence is clear. Robot manipulation is capable enough to handle laboratory tasks. Sensing and measurement are precise enough to capture subtle physical outcomes. And machine learning can direct the search, choosing what to test next based on what has already been learned: closing the loop between experimentation and hypothesis generation.

We are interested in teams building robotic experimentation platforms that accelerate physical discovery - systems that turn physical questions into answers at throughput levels that fundamentally change the pace of scientific and industrial progress.

Robotics has two testing problems, and they compound each other.

The first is simulation fidelity. Modern robot simulations are photorealistic and physically wrong. They work for navigation and gross motion planning. They break down wherever physical fidelity matters: contact, deformation, soft bodies, friction, cable and cloth behavior, multi-body interaction. Teams routinely build policies that work perfectly in sim and fail on the real robot because the sim's physics were not predictive.

The second is scenario coverage. Most testing relies on hand-designed cases, replayed logs, or random perturbation. These approaches consistently miss the failures that matter most: compound interactions where two or three ordinary conditions combine in ways no engineer anticipated. A warehouse robot collides when unusual pick sequences coincide. A manipulation system fails when surface texture and grip angle combine in ways never seen in training. The industry tests for what it can imagine and ships blind to the rest.

These problems need to be solved together. Better simulation without better scenario generation means running the wrong tests more accurately. Better scenario generation without predictive simulation means testing plausible situations in a world that does not behave like reality.

Both are now feasible at a useful scale. GPU-accelerated solvers make high-fidelity physics tractable at near-real-time speeds. Data-driven approaches can learn residual physics: the gap between simplified models and real behavior from interaction data. And generative models can synthesize physically coherent scenario variations, not just noise, from real-world fleet data. The open-source ecosystem around MuJoCo, Isaac Lab, and related platforms has matured enough to support serious engineering.

We are interested in teams building testing infrastructure for robotics that combines predictive simulation with systematic scenario generation. Therefore producing pre-deployment evidence that a system will survive the real world, not just the test suite.

Starlink and LEO satellite constellations have quietly solved a 50-year problem: you can now send high-bandwidth data to any edge, anywhere on earth.

This coupled with increasingly powerful frontier models changes everything. Systems that previously required proximity to infrastructure can now operate in complete isolation, with maximum intelligence. A weather observation station on a remote mountain that can keep itself alive using an onboard generator and utilize leading AI problem solving to conduct tasks.

 Autonomous agricultural equipment operating in fields beyond cell coverage. Ocean sensors, pipeline monitors, remote mining operations—all permanently connected to cloud intelligence.

The unlocked use cases are enormous. Environmental monitoring at unprecedented scale. Precision agriculture in developing regions. Industrial operations in extreme environments. Scientific observation in previously impossible locations. Disaster response infrastructure that works when everything else fails.

We want founders building hardware and software for this new paradigm—devices that assume permanent connectivity to intelligence, can operate autonomously for extended periods, and can maintain themselves in environments where no human will visit for months or years.