Open Source (AI) Strategy For The Physical Economy

There’s currently a debate raging in AI about open source models and their “inherent danger”, closed source models and “safe” release procedures, and a seemingly orthogonal but related debate about geopolitics between China and the US.

We won’t be talking about any of that in this piece – instead we think the debate about open vs. closed in AI is the wrong debate entirely when it comes to value generated with AI, and the next phase of the market and economy.

The better question is: given the continuous and obvious edge to commoditizing your complement in Joel Spolsky’s famous rendition of the phrase, which layers of the market should (and will) become commons, and which layers should companies own?

Bill Gurley’s recent P3 Institute piece, From Open Source Software to Open Source Strategy, has been useful for us in thinking about this, because it moves open source out of the ideology bucket and into the strategy bucket. His core point is simple and important: open source is no longer only a software development model. Used correctly, it is a deliberate corporate and ecosystem strategy.

Companies use open source to neutralize stronger incumbents, reduce supplier power, align industries around shared standards, prevent closed tollbooths, and shift value toward the operators who can deploy, distribute, and monetize on top, and we think that is the lens physical AI needs now.

Open source matters in the physical economy because it can turn shared chokepoints into neutral infrastructure: maps, identifiers, methodologies, evals, protocols, datasets, demos, and coordination standards.

The physical economy does not need open source as charity. It needs open-source strategy.

Android gave the mobile ecosystem an open alternative to a world where Apple could control the handset interface — today ~70% of devices run Android. Open Compute Project helped Meta and other hyperscalers commoditize data-center hardware supplier power. Kubernetes turned cloud orchestration into a neutral standard, reducing AWS lock-in and making workloads more portable, leading to a much more competitive cloud computing market beyond ~80% AWS marketshare. RISC-V created an open alternative to proprietary instruction-set architectures, and is now the dominant semiconductor architecture. Overture Maps created a shared geospatial reference layer against Google’s structural map advantage.

NVidia, Microsoft, Google, OpenAI, and many open source companies recently publicized and signed an open letter imploring policymakers not to restrict open weight models and allow geopolitical fear mongering to destroy innovation in the space. This is also because every one of these signatories understands that wholesale pricing transfer power is centralizing around the foundational AI labs, and the global economy (especially the Global 1000 corporates) cannot afford it.

To prevent this, the playbook is very clear. The recurring pattern is not that “open beats closed” in some abstract moral sense, because it’s the “good choice”, the pattern is sharper:

1. A layer becomes strategically critical.

2. One incumbent or supplier has too much power over that layer.

3. The rest of the ecosystem cannot afford to let that control compound.

4. A coalition turns the layer into a neutral commons.

5. Value moves from owning the bottleneck to operating, deploying, and building trusted workflows on top.

This is why foundation governance matters. Android is the cautionary example. It was open enough to rally the ecosystem, but weak neutral governance let Google recapture much of the control through adjacent services. Kubernetes, Open Compute Project, RISC-V, and Overture Maps are cleaner examples because neutral foundations make it harder for one sponsor to take the commons back.

As AI is just beginning to reach into the physical world with geospatial embeddings, physics-based frontier models, time series language models, world models, and robotics, physical AI will need this same discipline from market players.

AI is moving from internet-native work into physical-world systems, and the numbers are huge. At Cerulean, we’ve been looking across energy markets, food and ag, supply chain, oceans and maritime, and industrial manufacturing and optimization for ~5 years now — these markets all have in common the fact that they heavily resisted digital transformation, but are now beginning to yield their data trails to AI.

The early AI economy was mostly about text, code, images, search, chat, and knowledge work, but the next frontier is different. These energy systems, supply chains, farms, factories, ports, satellites, oceans, sensors, inspections, logistics networks, chemistry, biology, weather, materials, and industrial operations require notions of time, conservation of mass, and other parameters language models struggle mightily with. Tectonic market shifts like this are once in a generation and create the conditions for a systemic change that aligned investors can create and capture value from.

Google DeepMind is the obvious reference point. AlphaFold, WeatherNext, AlphaEarth, Gemini, Gemma, robotics, and AI for science all point in the same direction: frontier AI is pushing into scientific and physical-world domains where the ground truth is not sitting neatly on the public internet, and that is a strategic tension. The labs can build powerful general models, they can produce new representation layers, they can create developer surfaces, agent frameworks, and cloud-scale deployment paths.

But no single lab can own every farm boundary, supplier relationship, parcel record, grid constraint, facility graph, carbon methodology, ocean sensor stream, and field workflow. The physical economy is too fragmented, regulated and tied to tacit knowledge.

That does not make AI less important, or a failure in delivery, but it does make open-source strategy more important.

If every physical AI system depends on a few closed chokepoints, the market recreates the same problem Gurley describes across mobile, cloud, hardware, maps, telecom, and silicon. One company controls the reference layer, everyone else pays tax, and innovation bends around the tollbooth.

We believe the better path is to identify the shared layers early, build coalitions, and make them neutral.

The closest analogy we’ve found for physical AI is Overture Maps.

The strategic problem is obvious right off the bat — Google has spent years building one of the most valuable proprietary maps in the world. That map is not only a consumer product, it is a reference layer for search, logistics, mobility, local commerce, advertising, autonomous systems, and AI agents that need to reason about real places.

If every AI system grounds itself in one company’s proprietary map, that company holds quite a lever over the entire ecosystem.

Overture Maps was the open-source strategy response: a Linux Foundation project backed by companies that each needed a reliable, interoperable, production-grade base map, but did not want to depend on Google’s map moat. Its Global Entity Reference System is especially important because it gives physical places stable identifiers that let organizations join their proprietary data to a shared reference layer.

This could be a template for the rest of the physical economy. Physical AI will not work without stable references. A model cannot reason reliably about a supply chain if farms, mills, warehouses, products, claims, and counterparties are inconsistently identified. A grid agent cannot coordinate decisions if assets, constraints, interconnection queues, load profiles, and market signals cannot be compared. A carbon or nature market cannot price risk if projects, parcels, methods, measurements, and claims live in incompatible ledgers.

The world needs more GERS-like infrastructure, and it needs it for all kinds of things, like assets, facilities, suppliers, parcels, claims, energy nodes, nature projects, industrial processes, materials, and verification events.

The physical economy needs open commons where one company’s control would slow the whole market.

Take maps and identifiers. AI systems need stable references for places, assets, farms, facilities, projects, suppliers, products, vessels, grid nodes, and claims. Without shared identity layers, every company pays the same matching tax. The market spends money reconciling records instead of improving decisions.

While we’re at it, how about evals?

Physical AI needs domain-specific evals for geospatial classification, grid forecasting, supply-chain extraction, industrial anomaly detection, MRV, nature monitoring, logistics routing, and field-data quality. Generic benchmarks will not tell a utility, insurer, regulator, retailer, or project developer whether an AI system is reliable enough to touch real decisions.

Then methodologies and rubrics.

Carbon, biodiversity, supply-chain claims, nature finance, and externality accounting cannot be trusted if every actor grades itself with a private rubric. The market needs transparent methods that can be inspected, debated, improved, and implemented across many systems.

Then public datasets and demo layers.

Founders need enough open data to prototype. Labs need enough public structure to test model behavior. Customers need enough demos to understand what is possible. The physical economy cannot rely only on PDFs, gated data rooms, proprietary dashboards, and one-off integrations.

Then interoperability protocols.

Supply chains, energy systems, carbon markets, industrial workflows, and geospatial applications all cross company boundaries. If every integration is bespoke, AI adoption remains consulting. Shared protocols turn fragmented workflows into markets.

These are the layers where we believe open-source strategy is strongest.

They reduce supplier power. They prevent closed tollbooths. They make fragmented industries legible. They give startups, labs, enterprises, regulators, and researchers a common substrate.

The answer is not to open everything, necessarily.

In the physical economy, the most valuable company moats should and will stay proprietary.

Workflow access should be owned. The company that earns the right to sit inside a utility’s planning process, a retailer’s supplier workflow, a farmer’s field interface, a port operator’s operating system, or an industrial customer’s maintenance loop has earned a moat by winning the customer in the market.

Private data rights should be owned. The best physical-world data is usually not public. It comes from repeated workflow participation: grid telemetry, supplier records, field observations, parcel evidence, sensor streams, lab data, machine logs, claims history, and human corrections.

Customer trust should be owned. A trusted system in the physical economy is hard to displace because it touches regulated, operational, financial, and reputational risk. Trust compounds through accuracy, uptime, auditability, security, and judgment.

Deployment loops should be owned. The company that can convert machine intelligence into real operational decisions has the advantage. Prediction is not enough. The question is whether the system improves interconnection, dispatch, procurement, underwriting, verification, compliance, routing, maintenance, or capital allocation.

This is the core distinction:

Open the shared layers. Own the operating loops.

Open-source strategy is a market design tool.

It tells an ecosystem where to avoid tollbooths, where to create neutral infrastructure, and where founders should build durable companies.

That distinction matters most when the underlying market is fragmented. The physical economy is exactly that kind of market. Energy, supply chains, agriculture, geospatial systems, carbon, nature, industrial operations, and oceans all depend on many parties coordinating around shared facts while still protecting private relationships, workflows, and data rights.

The strategic mistake would be letting the reference layers become closed before the market has a chance to form.

From our perspective, maps, identifiers, methodologies, evals, protocols, and demo layers could (and will) become commons. Workflow access, private data rights, customer trust, deployment, and operating feedback should become company moats.

That is the first principle.

In the next piece, I want to make it practical.

How does a builder decide which layer to open and which layer to protect? What does this look like inside a real operating harness? Where do companies like CommonShare, LGND, and others show the boundary between shared infrastructure and earned company advantage? What should founders, labs, and foundations actually build?

Part one is the strategy: the physical economy needs a commons layer.

Part two is the operating question: how do we build on top of it without giving away the moat?