Op-ed: Salesforce just revealed the next battleground in AI — and it’s not the models

Analysts and pundits in finance have an endless supply of clever, simplistic explanations for complex market movements. Over the years, culprits have ranged from El Niño and climate change to the World Trade Organization, rock ‘n’ roll music and the fluoridation of water. Many of these explanations proved wrong, yet commentators continue to reach for overly simple narratives to explain market moves. The latest example is the sell-off in software stocks.

For the past year, markets have vaporized roughly $2 trillion in software value by asking the wrong question about AI. The question behind the so-called “SaaSpocalypse” — can an AI agent do what this software does? — is misguided because it treats a software company as nothing more than a bundle of tasks waiting to be automated away.

The right question, and the clearest test we know of for separating AI’s winners from its roadkill, is different: Does this company own something AI agents cannot operate without?

Salesforce’s recent blowout earnings show why the balance of power may be shifting away from frontier models and back toward software companies that investors had prematurely left for dead.

For months, analysts have assumed that frontier large language models will accrue an ever-greater share of the value created by AI, with the valuations of those companies soaring commensurately. But intelligence is precisely the thing being commoditized before our eyes: Frontier models now leapfrog one another every few months, growing ever more capable but also increasingly interchangeable.

AI may also be approaching RSI — recursive self-improvement — which could accelerate those improvement cycles even further, while the price of raw intelligence collapses amid heated competition from well-funded rivals and increasingly capable Chinese open-source models.

So in this new world, if LLMs are the commodity, where is the moat? The answer, we believe, is data. As analysts at Wells Fargo put it, “Lower cost of intelligence increases value of incumbent data.” That is Economics 101: When an input becomes abundant and cheap, value migrates to its scarce complements. And for AI agents, the scarce complement is trusted proprietary data.

An agent closing a sale still needs somewhere to research the customer, log the interaction, store the contract and customize the terms. Without trusted proprietary data, AI agents are far less useful. As the old adage goes, “Junk in, junk out.”

Salesforce is one of the largest repositories of enterprise customer data in the world, and far from being the beneficiary of a lucky earnings beat or one-time aberration, it may be positioned as a long-term structural winner from the commoditization of AI. Agents cannot work without data, so the data keeps flooding in: Salesforce’s Data 360 ingested a staggering 104 trillion customer records this quarter, up 355% year over year. Meanwhile, AI agents themselves generate still more data, all of which has to land somewhere trusted. Salesforce delivered 3.2 billion units of agentic work this quarter, nearly double the prior quarter.

That flywheel — agents generate work, work generates data, data deepens the moat and makes each successive generation of AI agents more valuable — is why Salesforce’s combined AI and data annual recurring revenue has reached $3.9 billion, more than tripling in a single year, as the accompanying chart shows. Agent force alone rocketed from $100 million to over $1.5 billion in ARR within 18 months of launch, up 240% year over year.

As Salesforce CEO Marc Benioff said on CNBC with Jim Cramer following the earnings report Wednesday, amid the stock soaring 20% and more, “The ‘SaaSpocalypse’ narrative has been such nonsense. … Here you can see, net new AOV growth is the strongest in four years. Skeptics said seats would decline, and Agentforce sales and service and Slack all grew seats year over year. Skeptics said customers are going to leave, and attrition is near its lowest level ever, … bookings more than doubled quarter over quarter and contract-length terms improved across all segments. … Agents are using more Salesforce than ever.”

Indeed, the numbers show why his argument deserves attention. Pricing power was supposed to evaporate and margins compress; instead, non-GAAP operating margins hit 34.1%, adjusted earnings of $5.90 per share nearly doubled and demolished consensus of roughly $3.27, revenue grew 11% to $11.35 billion, bookings grew even faster with current remaining performance obligation up 14%, and management raised full-year guidance to as much as $46.4 billion.

But the most decisive evidence came from the customers best positioned to know. If AI agents could truly run without Salesforce, the frontier AI labs would be disintermediating it. Instead, they are paying for it — 9 of the top 10 AI companies now run on Salesforce and Slack, with their combined spending up 435% year over year — and partnering with it, exemplified by the debut of Claudeforce: “The No. 1 AI in the world, Anthropic, and the No. 1 CRM, Salesforce, coming together,” in Benioff’s words.

The companies building the supposedly all-conquering models have concluded that those models, as Benioff put it, “depend on CRM … They do not replace them.” That is what he means when he says “Salesforce is first and foremost in the data business” — and why he is betting on that thesis with a $25 billion buyback, the largest in company history.

The test generalizes far beyond Salesforce, as we have argued. Not all software firms will flourish in this new world. Firms with proprietary data that is valuable to AI agents should be better positioned to prosper, while software companies that offer little beyond functionality — and have few other sources of customer stickiness — could struggle. Balance sheets will matter too: Companies with strong free cash flow and little leverage will have more room to reinvest in the AI transition, while heavily indebted firms may be forced to devote scarce cash to servicing debt instead.

Nevertheless, the direction of power in the AI economy is becoming clearer: away from companies whose primary advantage is manufacturing intelligence, which grows more abundant by the month, and toward companies that own the scarce assets that intelligence cannot function without. Increasingly, that means data.

Jeffrey Sonnenfeld is Lester Crown Professor in Leadership Practice at Yale School of Management and president and founder of the Yale Chief Executive Leadership Institute. Steven Tian is research director of the Yale Chief Executive Leadership Institute and former quantitative analyst at Rockefeller Capital Management. Stephen Henriques is senior research fellow at the Yale Chief Executive Leadership Institute and a former McKinsey consultant.