Open App →
Back to Blog
Original BAB News analysis November 22, 2026 By Steve

The IBM Test

The worst deal in tech history looked rational the day it was signed. The AI buildout is full of deals like it, and the only way to read them is to ask what each side gave away because it looked like a commodity.

The most rational mistake in tech history

In August 1980, IBM’s board approved a personal computer project that broke every rule the company had. The team’s plan was open architecture, non-IBM technology, non-IBM software, non-IBM sales and service. IBM was late to a market that Apple, Commodore and Radio Shack were already shaping, and the only resource it couldn’t buy was time.

So it went outside for the operating system. IBM first approached Digital Research about CP/M, and those talks went nowhere. Microsoft, already supplying IBM’s BASIC, stepped in. When Microsoft signed its IBM contract in November 1980, it didn’t have an operating system to deliver. It licensed one from a small Seattle hardware company, Seattle Computer Products, for a flat $25,000, then bought the full rights for another $75,000 shortly before the PC shipped in August 1981.

The clause that mattered was small. IBM’s license wasn’t exclusive. Microsoft kept the right to sell its own version, MS-DOS, to anyone else who built a compatible machine.

From IBM’s seat, that was fine. The value was in the hardware. The operating system was a component, the same as the chips and the disk drives IBM was also buying off the shelf. Nobody pays a premium for a component.

Then the clones arrived. Compaq and dozens of others built IBM-compatible machines, and every one of them needed the same operating system. Hardware became the commodity. The operating system became the standard every machine had to run. IBM had signed away the one layer that turned scarce.

There’s a second loser in the story who rarely gets mentioned. Seattle Computer Products sold the operating system outright for about $100,000 in total, without knowing IBM was the customer. It made the same mistake one rung lower: it valued the layer at what it was worth that day.

Why it wasn’t a mistake on the day

It’s easy to call this a blunder in hindsight. It wasn’t one at the table.

IBM optimized for the constraint it actually had in 1980, which was speed. Licensing a non-exclusive operating system got a product to market in about a year. The deal solved IBM’s problem. The trouble is that it solved the problem of 1980, and the industry’s constraint didn’t stay in 1980.

The deal turned bad on the day the scarce layer moved, from building the machine to running the software every machine needed. Nobody at the table could see that day coming. That’s the uncomfortable part of the lesson: a deal’s quality isn’t fixed at signing. It depends on which layer becomes scarce later.

The test

That gives a simple question to ask of any deal: what did each side give away because it looked like a commodity on the day they signed?

It isn’t “who won the negotiation.” IBM arguably won that one. It got a fast, cheap operating system on its own schedule. The better question is who kept the rights to the layer that looked boring.

Three things to check:

The AI buildout is a hard place to run this test, because its bottleneck has moved more in five years than most industries move in twenty. Chips, then advanced packaging, memory, power, permits, community consent. Every move rewrites which side of a deal was holding the valuable piece.

Who seems to be running the test

Google is keeping layers others rent. Its TPU program is the most mature custom AI silicon effort among the hyperscalers, which means its compute isn’t only a purchased input. In December 2025 it went a step further on power. Alphabet agreed to acquire Intersect Power for $4.75 billion, the first time a data center giant bought an entire power company rather than signing supply contracts. Chips and power are two of the scarcest rungs in the buildout, and Google is moving toward owning both.

Google is also playing Microsoft’s 1980 role. In January 2026, Apple announced that its next-generation foundation models will be based on Google’s Gemini models and cloud technology. Bloomberg reported the price at about $1 billion a year. The model maker is supplying the layer, and the device maker is building on top of it. That’s the 1980 arrangement with the names changed.

The hyperscalers building chips are keeping optionality, not cutting ties. Every major hyperscaler now designs its own AI silicon, yet Nvidia still holds about 70% of the AI chip market. Microsoft says it will keep buying Nvidia and AMD chips alongside its own Maia accelerator. That’s the opposite of IBM’s move. Rather than pick one supplier for speed, they’re paying for the right to switch.

Microsoft and OpenAI rewrote the IBM clause, from both sides. This one is worth reading closely. Under an April 2026 amendment, Microsoft’s license to OpenAI’s models and products runs through 2032 but is no longer exclusive, and OpenAI can now serve customers on any cloud provider. In 1980 terms, the model maker won the right to sell to the clones. One source familiar with Microsoft’s position described the trade as giving up exclusivity in return for certainty.

On the surface, Microsoft looks like IBM here. The difference is the part IBM never had: Microsoft holds roughly 27% of OpenAI, and OpenAI keeps paying it a revenue share through 2030, capped at a reported $38 billion. IBM owned none of Microsoft. If the layer Microsoft let go of becomes the standard, Microsoft still owns a large piece of whoever controls it. That’s the hedge IBM didn’t think it needed.

Who may be making IBM’s bet

These aren’t verdicts. Each of these deals has a reasonable case behind it, the same way IBM’s did. The test just shows which bet each side is making.

Apple is renting the model layer. The bull case is strong. Apple runs the Gemini-based models on its own servers through Private Cloud Compute, uses them as a base for its own Apple Foundation Models, and keeps the customer relationship with more than two billion devices. If the device and the customer are the scarce layer, Apple kept the right piece. The IBM case: if the model becomes the platform people actually choose, Apple is the hardware maker that licensed the operating system from someone else. Which reading wins depends on where the scarcity settles, and that isn’t knowable yet.

AMD paid in equity for a seat. In October 2025, AMD agreed to supply OpenAI with up to 6 gigawatts of GPUs and issued OpenAI warrants for up to 160 million shares, about 10% of the company, vesting as deployments and AMD’s share price hit milestones. AMD gets a marquee customer and a place in the largest buildout in the industry. OpenAI gets cheap upside in its own supplier plus a second source alongside Nvidia. Run the test and the question is which side gave away more of the scarce layer. AMD handed over equity to win volume. OpenAI kept optionality across suppliers.

Utilities signing long-term supply deals with hyperscalers. These look like clean wins today: a creditworthy customer, decades of demand, guaranteed load. Google’s Intersect purchase is the warning sign. A hyperscaler that learns to build and run its own generation is a customer turning into a competitor, the same way the clone makers came to compete with IBM using IBM’s own standard. Some analysts already expect the utility relationship to shift from supplying power toward providing transmission and interconnection. That’s a narrower role than the one utilities are pricing today.

The catch

The test doesn’t tell you who wins. Only hindsight does that, and hindsight is cheap. IBM’s PC created the industry that made Microsoft, and IBM went on to build large businesses elsewhere. A deal can lose the war for one layer and still be the right decision for the company that signed it.

What the test does is make the bet visible. Every deal in this buildout is a wager on where the scarcity goes next. The party optimizing for speed against today’s bottleneck is betting it stays put. The party holding onto the layer that looks boring is betting it moves.

History mostly favors the second bet. A constraint that gets solved doesn’t vanish. It moves to the next rung, and whoever owns that rung sets the terms. You don’t remove a bottleneck. You relocate it. The best deals are written for where it’s going.

The question worth asking about any deal on your own desk is the one IBM didn’t ask in 1980: what are we signing away today because it looks like a commodity?