The debate over how companies should measure artificial intelligence progress is increasingly centered on a simple question. Should the industry keep chasing more token usage, or should it focus on whether those tokens produce real business value?
At Nebius' first Inflection forum in San Francisco, Chief Revenue Officer Marc Boroditsky argued for the second approach. He said the industry needs to move away from what he called tokenmaxxing, or treating more token consumption as a sign of success, and toward what he described as valuemaxxing, where the focus is on outcomes instead of raw usage.
But the executives on stage did not present a unified view. Conversations throughout the event showed an industry still split on whether heavy AI spending is a sign of waste, investment, or simply the cost of working through an early market.
The discussion included leaders from Databricks, Cognition, Pinecone, LangChain and other companies. The Information's Amir Efrati compared the moment to the early cloud computing era, when enterprises were stunned by unexpected bills from Amazon Web Services. His point was that early infrastructure shifts often look expensive before the longer-term economics become clear.
Databricks vice president Nikita Shamgunov said customers are now asking for stronger cost controls after what he described as the breakdown of tokenmaxxing. Even so, he said internal spending priorities at Databricks remain generous for engineering work.
Cognition chief executive Scott Wu took a more bullish view of AI consumption. He argued that GPU costs are high but still cheaper than paying people to do the same work, especially in software development. Wu said it can be far less expensive for models to generate large volumes of code than for humans to do it manually, while cautioning that useless output is just wasted spending. He said his company is increasingly interested in measuring results rather than how much is consumed.
Wu also said companies can no longer afford to delay adoption for years and expect to catch up later. In his view, AI is moving too quickly for laggards to wait out the cycle.
Another divide in the room concerned how mature enterprise AI spending really is. Wu said only a small share of enterprise AI outlays are still experimental, since the workloads that grow tend to be the ones that clearly work, with coding leading the way. DataRobot chief product officer Venky Veeraraghavan disagreed, saying most enterprise spending outside coding is still exploratory as companies search for durable use cases.
That divide reflects a broader uncertainty across the market. Some companies are already seeing clear returns from AI coding tools and other production use cases. Others are still testing potential applications without finding one that scales.
Wu also described how AI agents are taking on longer tasks than they did just a couple of years ago. He said autonomous work sessions have expanded from seconds to hours, and in some customer cases, weeks. He suggested the next stage could involve agents taking on open-ended business goals, such as reducing infrastructure costs, and figuring out their own project scope over time.
Later in the forum, Nebius vice president of ecosystem strategy Devang Sachdev offered a concrete example of how model choice affects economics. He described building a healthcare compliance agent that cost $637 per run on a closed frontier model. Replacing it with open-weight alternatives, first DeepSeek and later Nvidia's Nemotron, lowered the cost to $24 and reduced runtime from an hour to 15 minutes.
The example reinforced a theme running through the event. Open models are improving quickly, and the software layer around a model may matter as much as the model itself.
Attendees also spent time discussing the supply side of AI. Conversation in the hallways centered on Nvidia's next-generation Vera Rubin supply, with companies trying to understand not only how much compute they can secure, but how much rivals can get.
The prevailing view was that AI compute will stay constrained at least through next year, with some expecting conditions to ease in mid-2027 or even later. Tight supply is also shaping major deals. One reported agreement giving Google access to Blackwell capacity from SpaceX was seen by attendees as reflecting both steep pricing and short-term commitment in a market where leverage still sits with suppliers.
For now, the debate inside the industry is not over whether AI will be expensive. It is over what companies should be willing to pay for, and how they will know when they are getting value back.