Why the value concentrates here

The About page says why I am building WunderOS. This page says why the market pays for it, and why it pays at the point where WunderOS sits rather than upstream or down.

The shape of the AI market

Right now the AI market has three regions. On one side are the inputs to intelligence: compute, models, data, energy, capital. These are abundant and getting cheaper, and each new frontier model turns the ones before it into a commodity. So far, so ordinary.

On the other side are the outcomes that an enterprise actually pays for: a trial initiated, a component manufactured, a claim adjudicated, a mission completed. These do not commoditize because each is bound to the particular systems, rules, and liabilities of its industry.

Between the two sits a narrow region where abundant intelligence is turned into a specific outcome someone is willing to stand behind, that is, pay for. Ontologies, agent orchestration, verification, regulatory approval, execution.

The shape has a name in venture circles, the bow tie. The framing as I found it is 8VC’s; and the shape of WunderOS is mine.

Value concentrates in the narrow region. Not because the inputs do not matter but because they are abundant, and abundance does not command a price. The outcomes command a price, but that price is captured industry by industry, by the firms that already own the trial, the factory, the claim.

What is both scarce and general is the conversion in the middle. It is scarce because it is hard, and it is general because every industry needs the same thing done: intelligence made into action that holds up.

The constraint in the middle is permission

The scarce thing in the middle is safety and permission, not intelligence. An enterprise cannot let a cheap, abundant, non-deterministic process act on expensive real-world systems without a layer that makes each action authorized rather than merely produced: governed, bounded, deterministic, audited, replayable.

The largest outcomes are exactly the ones a company is least willing to let an unbounded process touch. Weapons, factories, clinical development, money, customer relationships. The more expensive the system the action reaches into, the more the action has to be one the enterprise can defend afterward, since the cost of a wrong action there is not a bad token but a recall, a fine, a casualty, a lawsuit, a real-world harm.

So the difficulty is not producing a plausible action. Frontier models produce plausible actions cheaply, and will produce them more cheaply next year. The difficulty is producing an action a regulated enterprise can authorize, and proving after the fact that it was authorized. And the trick to the difficulty is to channel machine intelligence safely, such that the enterprise isn’t sued or fined, without domesticating machine intelligence to the extent that it’s ineffective.

That is the whole of the middle, and it is where the money is, because it is the step no abundant input performs on its own and no industry incumbent supplies to the others.

I’ve described the difficulty precisely elsewhere. The regulated enterprise faces a Fourfold Problematic: agents are unreliable, slow, expensive, and hard to audit and control.

Every one of those is a failure to convert intelligence into authorized action. An agent that does not complete without a human, that cannot be replayed, that runs up an unbounded bill, or that cannot show what it did, has produced intelligence but not permission.

How WunderOS fits in the bow tie

WunderOS is an autonomic agentic operating system for the regulated enterprise, and each of its architectural commitments is one edge of the bow tie’s middle.

An LLM is a plant in the control-theory sense, a source of non-determinism. You do not make a plant deterministic, since that’s the ineffective domestication path; you put a controller around it. You leash it.

WunderOS is the controller, and the controller is what turns an abundant, non-deterministic input into a bounded, replayable, auditable action. That is determinism doing the work the middle requires: a deterministic plan can be replayed, audited, and budgeted, and a deterministic failure is a repairable failure, while a non-deterministic one scatters and plateaus.

The named components of the bow tie’s middle are the named properties of WunderOS:

• Agent orchestration is what it does on the OTP BEAM, hosting a family of agents at high concurrency, continuously, recovered and replayed after a crash.
• Ontologies are its temporal knowledge substrate, where every assertion carries provenance and every fact carries the two times of when it was learned and when it became true.
• Verification and regulatory approval are its audit-grade non-repudiability, where an account of what the system did is available after the fact and provable, not reconstructed.

WunderOS is not adjacent to the middle of the bow tie. It is built as the middle.

WunderOS is a bridge

The most valuable node in the narrow region is the one many industries have to cross. Regulated-enterprise automation is horizontal, not one vertical: defense, clinical development, financial services, and industrial operations differ in their outcomes and share their constraint, which is that an agent’s action has to be authorized and provable under audit.

Testing, evaluation, and verification is a bridge every one of them crosses, and a substrate that provides it serves them at once. Remove that node and several clusters come apart. That is the definition of a bridge, and a bridge is the thing whose value does not depend on any single industry’s cycle.

This is why WunderOS is not a bet on defense, or on biotech, or on agentic back-office services. It is a bet on the layer all of them route through. The first place I prove it is the enterprise agentic data enclave, a sovereign governed space where a company hosts someone else’s agents on its own data.

The enclave is one crossing of the bridge. The bridge is the asset.

Why now

The timing is not incidental. The left side of the bow tie is commoditizing fastest exactly now: models that were frontier a year ago are cheap today, and the marginal intelligence available to an enterprise is climbing while its price falls.

That moves the scarcity, and therefore the value, off the inputs and into the conversion. At the same time the work is passing through a phase change. The first wave of agent tooling answered how to make one clever agent do one thing well.

The question now is how an organization runs a fleet of families of agents doing a wide range of things, reliably and acceptably, under the audit, compliance, and risk requirements that are uniquely its own. That is the phase where the demo is in production, the spend shows up on a bill someone has to defend, an auditor asks what happened, and the pager goes off at three in the morning.

It is the phase in which the middle stops being a convenience and becomes critical to the enterprise’s mission.

Why it compounds

Value in the middle does not merely sit; it compounds, and determinism is why.

Because WunderOS wraps the model in a deterministic shell, every agent action, classifier output, and plan execution writes back into the substrate, and those writes are folds that compress operational experience into the surface later cycles read as context. The substrate improves with use.

On top of that, WunderOS includes CARL, the continuous autoresearch loop, our version of recursive self-improvement. CARL sharpens the deterministic harness around a frozen model from accumulated traces, and some other tricks I’m not ready to talk about just yet, so the operator-to-agent ratio improves and the fraction of expensive frontier calls shrinks the more often an agent runs.

Reliability and cost improve on their own. And because every improvement is an audited, replayable, human-gated edit rather than silent drift, the loop that makes the system better is itself inside the audited middle.

The moat I’m building isn’t a single clever component. It’s that the layer gets deterministically better with use while remaining provable, which is the one thing an abundant input cannot do and an industry incumbent will not build.

This is the WunderOS wager in short. Raw intelligence is going to be everywhere and nearly free. The scarce, general, compounding asset is the layer that turns it into action a regulated enterprise can authorize and prove, and that is what WunderOS is built to be.