MPAI is quietly building the missing layer under the AI economy: standards

The most consequential AI standards body you have never heard of is MPAI, the Moving Picture, Audio and Data Coding by Artificial Intelligence organization. It is an international, unaffiliated, non-profit group developing standards for AI-based data coding with explicit intellectual property licensing frameworks. Its name is a mouthful. Its ambition is not small: interoperable, multi-source AI components that can be swapped, licensed, and governed like the codecs that standardized video and audio decades ago.

MPAI’s current published work includes a draft “MPAI as a Service” V1.0 standard and an “MPAI Metaverse Model – Technologies” V2.2. It has also published Version 1.1 of “Connected Autonomous Vehicle – Technologies (CAV-TEC)”. These are not flashy model releases. They are plumbing. But the plumbing is the point. The AI economy today runs on proprietary, vertically integrated stacks: one lab owns the model, the inference, the tooling, and the distribution. MPAI’s bet is that a different architecture is possible, one where AI components are standardized, independently developed, and licensed under clear terms.

That bet is easy to dismiss. Standards bodies move slowly. MPAI is a small organization with minimal associate membership fees and weekly videoconference meetings. It does not have the gravity of the International Organization for Standardization or the Institute of Electrical and Electronics Engineers. It does not have the marketing budget of OpenAI or Anthropic. What it has is a specific, coherent theory of how AI should be built: as a market of components, not a monopoly of monoliths.

The theory deserves attention because the current moment is defined by its absence. Consider what happens when a developer wants to swap one vision model for another inside a deployed agent. There is no common interface. There is no shared format for describing what a model does, what it costs, what its license permits. There is no neutral registry of components. The developer rewrites integration code, renegotiates terms, and prays the new model behaves like the old one. MPAI’s standards are an attempt to make that process routine.

The organization’s framing is explicit: “Interoperable multi-source AI components enable advanced and explainable AI solutions.” That sentence carries two claims. The first is about interoperability, which is uncontroversial in principle. The second is about explainability, which is more interesting. MPAI is not merely standardizing interfaces. It is standardizing the idea that AI systems should be assembled from parts whose behavior can be understood and audited. That is a governance position dressed as a technical specification.

MPAI’s application areas span human-machine communication, enhanced audio, financial data, video coding, online gaming, connected autonomous cars, and mixed-reality collaborative spaces. The breadth is striking. This is not a niche standards body for video compression. It is trying to cover the entire surface of AI application. Each area gets its own standard: AI components for human-machine dialogue, AI components for improved audio experience, AI interpreting complex data structures, AI adding more compression to video codecs, AI-based end-to-end video coding, AI-based standards for connected autonomous vehicles, AI standards for interoperable metaverses, and standards for avatar creation and animation.

The avatar work is worth pausing on. MPAI is developing standards for “portable avatars for real and virtual world experiences.” In a world where Meta, Apple, and a dozen startups are building walled-garden virtual worlds, the idea of a portable avatar is quietly radical. It says your digital representation should not be trapped inside one platform’s ecosystem. It should move. That is the same logic that drove the original Moving Picture Experts Group (MPEG) standards: your video file should play on any device, from any manufacturer, without asking permission.

That historical parallel is the strongest argument for MPAI’s relevance. MPEG standards did not just make video files portable. They created an entire economy of encoders, decoders, players, and content distributors built on a shared substrate. The licensing frameworks were imperfect, sometimes litigious, but they worked. MPAI is explicitly modeling itself on that legacy. Its name is a deliberate echo. Its mission statement says it develops “standards for AI-based data coding with clear Intellectual Property Rights licensing frameworks.” The emphasis on IP licensing is not an afterthought. It is the core.

The problem is that AI does not map cleanly onto the codec model. A video codec is a deterministic function. Feed it a bitstream, get a picture. An AI model is a stochastic, learned system. Its behavior is not fully specified by its weights and its license. Its outputs depend on training data, prompt context, and sampling parameters. Standardizing the interface is tractable. Standardizing the behavior is not. MPAI’s response has been to focus on components rather than end-to-end systems, which is a reasonable hedge. But it means the standards will encode interfaces, not guarantees.

There is also the question of who actually adopts these standards. The largest AI labs have no incentive to join. Interoperability is a threat to their moats. OpenAI benefits from a world where Claude Code and Codex sessions are not interchangeable. MPAI’s membership model, with minimal fees for associates, is designed for small players and researchers, not for the frontier labs. That is a feature and a limitation. It keeps the organization independent. It also keeps it marginal.

The most telling detail in MPAI’s public materials is the governance language. The organization says it is “governing the MPAI ecosystem” and “delivering on MPAI promises through a governed ecosystem.” Standards bodies do not usually talk about governance. They talk about specifications. MPAI is explicit that standards are not enough; there must be a system that enforces them. That is a step beyond MPEG, which relied on market forces and patent pools rather than an internal governance structure.

Interoperability is a threat to the largest labs’ moats. MPAI’s bet is that a governed market of components beats a monopoly of monoliths.

What does this mean for AI builders? The practical answer is: not much, yet. MPAI’s standards are drafts and early versions. No major platform has adopted them. No major model vendor has announced compliance. But the direction of travel matters. As AI moves from single-model demos to multi-agent, multi-vendor production systems, the cost of non-interoperability will rise. Teams already spend a large share of engineering effort on glue code: connecting model outputs to tools, tools to data, data to evaluation. MPAI is trying to make that glue a commodity.

The deeper implication is for the AI economy. If MPAI’s component market materializes, it changes who captures value. Today, value accrues to the model owners and the platform operators. A component market would shift some value to the specialists who build narrow, well-defined AI functions: a better speech recognizer, a better financial-data parser, a better video compressor. Those specialists would not need to compete with frontier labs on general capability. They would compete on specific, standardized, licensable components. That is a different game.

None of this is guaranteed. Standards bodies fail more often than they succeed, and the ones that succeed usually have the backing of powerful incumbents. MPAI has neither. It has a coherent theory, a patient structure, and a historical analogy. That may be enough to matter, or it may be enough to be ignored. The industry will decide by its actions: whether builders adopt the standards, whether vendors implement them, whether the frontier labs eventually find it cheaper to join than to fight.

For now, the notable fact is that MPAI exists at all. In an industry defined by speed, scale, and consolidation, a small non-profit is building the slow infrastructure of an alternative future. Its published standards are drafts. Its governance is untested. Its licensing frameworks are unproven. But the question it is asking is the right one: what does a competitive, explainable, interoperable AI economy look like, and who builds the rules for it? MPAI is betting that the answer is a standards body with a clear IP framework, and it is working on the answer one weekly videoconference at a time.