When the Software Giants Buy Their Way Into AI

Share

For a couple of years, the default assumption in the industry was that AI would do to the big software companies what the internet did to the generation before them. A new capability arrives that is so general and so cheap it can route around whatever walls an incumbent has built, and a wave of smaller, faster companies rides it into the same customers. Every new model release fed that story, the sense that the giants were about to be caught flat footed, that their piles of seats and contracts would not protect them from something that changes the nature of the product itself. So it is a little strange that the biggest software companies on earth have spent this year doing the opposite of what the story predicted. They have been buying their way into AI.

The cleanest example is Salesforce, which signed a definitive agreement this week to acquire Cresta, a voice AI company that runs customer service calls end to end, for a price that press reports put at about $2.7 billion. Cresta will fold into Service Cloud, next to the agents Salesforce already sells. It is only the latest stop on a buying spree. In the weeks before, Salesforce reportedly paid around $3.6 billion for Fin, and across the year it has picked up a string of smaller AI companies, for customer messaging, for agents, for analytics. It has also put money into Genesys, the contact center giant, in a round it co led with ServiceNow. The through line is hard to miss. Salesforce has been collecting the pieces of the customer relationship, the voice, the messages, the agents that answer, and treating AI not as the thing that comes to replace it but as the thing it needs to own the exact moment a customer asks for help.

I am working from reported figures, since neither company published the Cresta price tag, and I do not want to pretend I have audited the number. But the scale is visible even in round figures. A nine year old spinoff of Stanford's AI lab that automates contact center calls is worth somewhere north of two billion dollars to a company that already owns the customer database those calls run against. The price only makes sense if you believe the value is in the conversation and the workflow, not in the software that powers it.

And that is the part worth stopping on. Salesforce is not buying models, and there is a reason. The models have become a commodity. They are cheap, interchangeable enough for a huge share of real work, and available to any buyer on the same terms. There is nothing about a frontier model that Salesforce needs to own, because owning it buys little advantage. What it cannot easily replace by writing its own code is the layer that is already wired into how a customer's business operates, the voice routing, the conversation templates, the workflows that staff and customers have grown used to, and the data attached to them. When the owner of the customer relationship is willing to spend billions on that layer, the market is giving about as direct an answer as it ever does to the question of where the durable value in AI actually lives.

This is the pattern hiding inside a pile of deal announcements, and it cuts against the story a lot of people told about this moment. The big software companies were not supposed to be able to absorb AI this easily. They were too slow, too bureaucratic, too committed to selling seats and contracts, to win with technology that changes the nature of the product. And yet they are not really competing with the frontier labs at all. They are competing to sit between the model and the customer. That is a different game, and it turns out the incumbents are remarkably well placed for it, because they already hold the relationship, the data, the distribution, and the trust. AI was supposed to be the thing that disintermediated them. It is turning out to be the thing they absorb.

To be fair, buying is not building, and a shopping spree is not a strategy on its own. Salesforce has given real evidence of how hard the integration will be. Its Agentforce agents stumbled early largely because of data problems, the agents feeding on inconsistent or siloed information and producing inconsistent results. That is a useful caution. The value these acquisitions are supposed to capture does not come free just because you now own the software. It has to be wired into the messy reality of a customer's systems, and that is exactly where acquirers tend to lose what they paid for. It would be just as possible to read the whole frenzy as a sign of weakness, a big firm buying the pieces of a game that is changing faster than it can build.

Salesforce is not alone in this, and that is what makes it feel like a shift rather than an anomaly. Around the same week, Marvell agreed to buy Lumentum in an all stock deal with an implied value near $115 billion, combining two of the biggest suppliers of the optical parts that connect AI data centers. That is consolidation at the infrastructure layer, the machine that feeds intelligence. So you have the giants buying the AI layer that touches customers on one end of the stack and the giants buying the components that feed it on the other. A great deal of the money moving through AI right now is not flowing to new challengers trying to displace the establishment. It is the establishment buying the challengers, or buying the parts it needs to avoid being challenged.

For founders, that narrows the field in a way worth noticing. If you are building an AI application aimed at enterprises, the likely ending is increasingly one of a few paths. You get bought by the incumbent who already owns the customer, because the relationship and the data, not your code, are the prize. You try to outrun a version of your product that the incumbent has integrated and now ships as a feature of its platform. Or you find a niche deep enough and unglamorous enough that the economies of a giant do not bother with it. None of these is obvious doom, and a good many of them end well. But they are a much narrower set of futures than the early AI hype promised, when the assumption was that thousands of independent AI companies would each find a corner and grow there.

This has happened before, which is both reassuring and not. When a powerful new wave washes over software, the incumbents usually find a way to buy their way across it rather than be washed away. The early cloud years and the consolidation that followed the first internet boom show the same rhythm. The new capability always seems, for a season, to belong to the young and the fast. And then the big platform companies, with their customer lists and their balance sheets, quietly take the winners into the fold. This is not an exotic outcome. It is what usually happens. What is new here is only the price, which has grown large enough to signal something real about where the value went.

There is a quiet irony at the end of it. The very things the giants were said to be too slow to build, the agility and focus of a good startup, turned out to be things they could buy. And the price is its own kind of confession. When the company with the customers spends billions on the company with the AI workflow, it is acknowledging, in the most expensive way it can, that the value in AI sits in the relationship with the buyer, not in the intelligence itself. The money is not overturning the thesis that has run through this industry for a year or two. It is confirming it.

I do not think this means independent AI companies are done for, any more than earlier rounds of consolidation wiped out everyone below the big platforms. There will always be room for products that are genuinely better, and for niches too small or too strange for an acquirer to bother with. But the shape of the next few years is getting clearer. The giants are not going to be displaced by AI. They are going to absorb it, and the competition will live in the layer between the model and the customer, where the relationship, the data, and the trust already reside. Salesforce is not paying billions for voice software. It is paying for the seat beside the customer, for the moment of the conversation, for the reason not to churn. The model was never the prize. The customer always was, and the companies who already own the customers just spent the week proving it.