There is a question sitting underneath a lot of AI pricing conversations right now, and it has nothing to do with the model. The pricing pencils out. The demo lands. The ninety-day outcome is defensible. What is harder to answer cleanly is the quieter question underneath: if the customer actually uses the product the way it is being sold, can the seller sustain the throughput across a year of real load without the experience degrading? Not the launch experience. The Tuesday-in-March experience.

That is the question that just got more interesting.

Last week, Apollo and Blackstone partnered with Broadcom on a $35 billion debt-backed AI infrastructure platform tied to Anthropic's compute, designed to enable more than 20 gigawatts of capacity through 2028. Days later, KKR launched a $10 billion AI infrastructure company with Nvidia and Vistra, pairing committed capital with the two named bottlenecks of the build: chips and electricity. Read these as finance stories and they are interesting. Read them as marketing stories and something more useful comes into focus.

Compute is becoming financeable like an asset. Which means, for the first time, capacity itself can be contracted.

The unit of customer value is shifting

For most of the last two years, AI go-to-market has been a potential business. You sold what the model could do. You sold the demo. You sold the roadmap and the velocity of the underlying labs. Customers bought because the upside was obvious and the downside was, in practice, somebody else's problem.

That era is closing, quietly, because the thing being sold is starting to acquire a balance sheet. When a $35 billion loan sits behind a compute platform, the capacity at the other end of that loan has to clear cash flows. It has to be allocated, metered, contracted, and defended. PitchBook noted that holding the hardware in a separate vehicle keeps it off Anthropic's balance sheet ahead of a possible listing, but the more important effect is downstream. Compute that is financed as infrastructure starts to behave like infrastructure. It can be promised in a way that demos cannot.

My read is that the unit of customer value is moving from model capability to capacity-backed reliability. The promise is no longer "look what this can do." The promise is "this will keep doing it, at this latency, at this throughput, on Tuesday in March, when your support volume triples."

That is a different sale. And it requires a different kind of marketing.

Capacity-backed promises as a new marketing primitive

If I am thinking about this correctly, the new primitive in AI-native go-to-market is not the feature or the use case. It is the capacity-backed promise. A specific, contractable claim about throughput, latency, continuity, and degradation behavior that is underwritten by financed compute the seller actually controls.

The interesting thing about a capacity-backed promise is that it cannot be generated by a content team. It is produced where finance, product, and operations meet the customer. Marketing's job becomes translating the underlying capacity into a credible commercial offer, and refusing to sell past it.

A few moves get easier once you accept this frame.

You can build pricing around sustained service levels instead of seat counts, because you actually know what your compute envelope is. You can design enterprise offers with explicit continuity clauses, because your supplier just did. You can compete against vendors who are still selling demo energy by quietly out-promising them on the boring parts: response time at peak, recovery behavior on degraded routes, year-two cost trajectory. The shift sounds operational. It is actually creative. It opens a different language for talking to customers, one that respects how seriously they are starting to take AI inside their own businesses.

This is the part that quietly disqualifies a lot of current marketing output. If your story is still pitched at the level of "we use AI," you are speaking last year's language to a buyer who is starting to ask underwriter questions.

The zombie problem

There is a less flattering side to this. Not every AI company is going to be on the right side of the financing line.

Many venture-funded AI startups raised on narrative and now need to grow into valuations that assumed a different cost of capital. They sold potential because potential was what they had. As compute increasingly gets financed by parties who expect to be repaid against measurable capacity, the gap between the well-backed platforms and the narrative-only players becomes visible in the place that matters most: what they can credibly promise a customer for the next twenty-four months.

This is not a moral problem. It is a structural one. A company with no claim on durable capacity cannot underwrite a serious enterprise commitment. It can sell pilots. It can sell into curiosity budgets. It cannot sell into the customer's actual operating plan, because its own plan is too contingent.

Buyers will figure this out faster than sellers want them to. They will start asking, in different words, the question sitting underneath every serious AI pricing conversation: can you sustain this. The companies that have rehearsed the answer, and can show the financed capacity behind it, will win deals that look on the surface like feature competitions but are actually credibility competitions.

The advantage conversion problem

There is a second risk on the buyer side, and it is the one I keep coming back to. Having access to frontier capability is not the same as converting it into durable customer advantage. The lag between the two is where most of the disappointment in enterprise AI is going to live.

Financed capacity makes speed available. It does not make speed yours. You still have to redesign the workflow, instrument the outcome, and shape the offer in a way customers can sign. As I argued in Token Cost Is the Real Infrastructure, the unit economics of the execution loop decide whether capability becomes margin. Capacity financing makes the question sharper, not softer. You now have to know which of your promises your loop can sustain at what cost, and which ones are still pilot-grade.

The failure mode is recognizable. A leadership team sees the capability, announces the offer, and discovers, two quarters in, that the operating reality cannot carry the marketing. The brand pays for that gap. So does the pipeline. The fix is not louder marketing. It is the discipline to promise only what the financed loop can hold.

This is the same mechanism I sketched in Announcements Don't Move Ships. Underwriters Do. Announcements move attention. Throughput moves only when a third party will stand behind the operational claim. The AI version of that third party is now visible. It is the capital stack.

What this means for the next twelve months

A short list, written as honestly as I can write it.

The AI-native marketing function is going to look more like an underwriting desk than a content studio. Its job is to know what the company can credibly promise, and to refuse claims the system cannot hold. That is a skill set most marketing organizations have not staffed for.

Competitive advantage will increasingly run through the offer, not the message. Two companies with similar models will diverge on the precision of what they will commit to, in writing, for a year. The one with financed capacity behind the commitment will win the buyer's serious budget.

And the gap between AI theater and AI business will start to be legible in contracts. If you cannot find the capacity-backed promise in the contract, you are looking at a pilot dressed up as a platform.

The thing I would watch is not the next model release. It is the next round of enterprise AI contracts that quietly include continuity and throughput language. That is where the new primitive shows up first. Not in the press release. In the redlines.

If your marketing is still selling potential, you are late. If you are selling a capacity-backed experience your financed loop can actually carry, you are doing the only kind of AI marketing that scales without becoming a credibility problem. The speed is real. The discipline that turns it into advantage is what you promise, and what you refuse to.

Sources