The Hybrid AI Trade

In 1954, engineers started bolting digital computers to analog machines. It made sense at the time. Analog was absurdly fast at continuous math but terrible at memory and reprogramming. Digital was precise and programmable, and painfully slow at the exact calculations analog ate for breakfast. So they wired them together with converters. The analog machine simulated the missile’s flight in real time while the digital machine changed the settings mid-flight and set up the next run. By 1958, serious people had concluded that hybrids were the future of computing. They had the present right and the duration wrong.

The Future Lasted About a Decade

The hybrid era didn’t end with a press release. It ended because digital computers kept getting faster on a curve analog couldn’t match. Once digital was fast enough, the analog half stopped earning its place, and you simulated it in software instead. Simulation languages in the 1960s recreated analog patch panels on a screen so engineers could keep thinking the same way while the hardware disappeared underneath them. Vendors kept selling hybrids into the 1970s to a shrinking set of customers, then stopped. The advantage analog offered got absorbed into software, and the architecture built to deliver it went away.

History Rhymes

Now look at how we’re building AI in 2026. The pattern is a probabilistic core that’s powerful but unreliable, wrapped in a deterministic layer that’s reliable but dumb. Guardrails, rules engines, orchestration frameworks, and structured-output enforcers all live in that outer layer. The model handles judgment and the deterministic half handles guarantees. Everyone calls it the pragmatic middle ground, the best of both worlds until the models grow up.

That describes 2026 accurately, and it described 1958 accurately too. The diagram is fine. Where it goes wrong is the assumption that the arrangement holds still, because the two halves of a hybrid never improve at the same speed. A deterministic layer is bounded by what you can specify in advance. Rules cover what someone already anticipated, and AI-assisted coding makes that layer cheaper to build without making it any smarter. The probabilistic core expands what nobody has to specify at all, generation by generation, funded by one of the largest capital buildouts this industry has ever run. When one side of a hybrid keeps getting better and the other only gets cheaper, the balance shifts every time a new model ships.

The deterministic layer exists because the model can’t do those jobs yet, and every one of those jobs sits on a frontier lab’s roadmap. Each item the labs cross off is a feature that used to be billable.

Pricing the Transition

Which brings me to the question I ask in every diligence meeting, which is what the bet actually underwrites. If a company’s whole value is the deterministic layer around someone else’s model, the answer is uncomfortable. You’re betting that models won’t get good enough to make your layer unnecessary before you exit. That’s a short position on model progress, and most founders holding it don’t know they’ve taken it. I’ve seen decks that cite model unreliability as a tailwind on slide three, then project ten years of growing revenue as if slide three stays true the whole way. Only one of those can be right. Either models stay broken and the revenue holds, or models improve and the product expires along with slide three.

None of this makes it a bad business. Hybrid computer makers made real money for two decades while digital caught up, and there’s nothing wrong with collecting cash while a gap exists. The price has to match what you’re buying, though. Transition revenue demands fast payback, because you’re putting on a trade rather than buying a business you intend to hold. The trouble starts when a company built on temporary workarounds raises at a valuation that treats the revenue as permanent. A high price today only pays off if the revenue is still there in year seven or eight, which means the buyer is betting the gap stays open for most of a decade. The same deck argues that models are improving fast, because rapid improvement is why anyone needs that layer in the first place. The valuation and the pitch are betting on opposite outcomes.

The exit assumption inherits the same problem. Who buys these companies? The platforms do, and they can read their own roadmaps. They won’t pay permanent-revenue multiples for a product their next release replaces. A transition trade and a durable business are both fine things to underwrite. Buying one at the price of the other is the mistake, and right now a lot of these companies are priced as if the window never closes.

What Could Go Wrong

Model progress could plateau. If the next few generations ship and enterprises in 2029 are still buying the same guardrails, the deterministic layer keeps its value and the 1958 comparison breaks, because digital never got fast enough this time. That’s a real possibility. But paying growth prices today requires believing that labs deploying historic capital will collectively stall soon and stay stalled, which is an enormous macro call to make implicitly through a seed check.

The more interesting break happens when determinism gets bought for reasons that have nothing to do with model weakness. A regulator wants a reproducible explanation for a denied loan. An auditor wants a decision trail in a clinical workflow. That layer serves an institutional demand for accountability, and the demand survives any amount of model progress, the same way analog’s advantage survived as software after the hardware went away. The value migrates from patching a bug to serving a need that was never about the technology in the first place. In healthcare and financial infrastructure, that demand is durable, and it’s where I’d rather be.

So the diligence question breaks down into one test. Would this layer still get bought if the model were perfect? If yes, you might have infrastructure. If no, you have revenue with an expiration date, and the job becomes estimating how long the window stays open and what you’re paying to enter it.