For the second half of 2026, look to the data layer. That’s the segment of the software stack that’s above storage and below applications, where data is cleaned, formatted and organized so that it can be used by artificial intelligence. And it’s where top analyst Heath Terry, head of AI investment research at Citigroup, thinks the next crop of winning companies will emerge in the AI race. “The companies that are winners are the ones that are exposed to consumption. AI is driving massive growth in consumption, particularly at the data layer. The Snowflake s, the MongoDB s, the Datadog s, even companies like Elastic that have exposure to consumption-driven business models – that’s where you want to be,” Terry said on CNBC’s ” Squawk on the Street ” Tuesday. At the level of AI applications themselves, frontier models like Anthropic and OpenAI are jockeying for position ahead of expected public offerings, and both are up against a host of open source models that have much the same overall capability at a fraction of the cost. But all of these models need access to increasing volumes of maneuverable data, and software companies that can deliver on that demand will be a crucial cog in the system. Investor confidence in that demand is creating a more nuanced and positive view of the software sector as a whole after its across-the-board devaluation earlier this year, informally termed Saas-mageddon. “We’re getting to a better place in software where we’re back to picking winners and losers,” Terry said. Some companies are already showing signs of coming success. Snowflake soared at the end of May after reporting 33% annual revenue growth. Shares trade at 112 times earnings for the next 12 months, and the company has an enterprise value of more than 14 times sales, according to FactSet data. SNOW YTD mountain Snowflake year-to-date. Datadog made a similar move at the beginning of May after reporting 25% annual revenue growth. The company has a forward price-to-earnings multiple of 98 and EV-to-sales multiple of about 19. The current “agentic” phase of AI development – which means lots of little pieces of software, or “agents,” working together – has contributed to chip system designs at the hardware level that emphasize coordination between different processing units. The term for this is orchestration . Orchestration has an analogue at the application level – a process called model routing that coordinates different end-user models for different tasks. Citigroup’s Heath Terry also sees this as a sector where value is being created. “The reason that routing layer is important is because, as a company, optimizing for which model is most efficient for the specific workload that you have is going to be an important part of how companies actually adopt and implement AI,” he said. Model routers like Not Diamond, Martian and OpenRouter are private companies focused on delivering cost savings for enterprise AI customers by dividing up tasks between models. Cost has become an increasing concern during the AI buildout as some companies pull back on their use of AI computing power, which is typically measured in units called tokens.
