Enterprise AI is being defined by a new, expensive reality: the token economy. Tokens are the economic unit used to measure the input and output of large language models (LLMs). As input data is tokenized and the LLM responds with output tokens, companies monetize and price their applications based on this usage.
This system has led to massive spending, with the Magnificent Seven companies alone collectively spending a trillion dollars to build infrastructure capable of supporting the massive gigawatt capacity required for more tokens.
The expense is compounded by newer LLMs, which are becoming significantly pricier—some are five to six times more expensive than their predecessors—because they are built to spend more time thinking and reasoning. This has come to be known as tokenmaxxing – akin to the upsell you’d get at a car dealer, who – after you agree on the price, you’re put in front of a salesperson who wants you to add undercoating or rim protection to the final cost. More profit for them.
Jitender Aswani, senior VP of product at data platform provider Starburst, told SD Times that he calls LLMs that do this “overzealous.”
“They will answer your question, but they might give you a very verbose response, which basically means they are tokenmaxxing. The other way they max their output tokens is they ask you, ‘I’ve answered your question, but are you interested in A, B and C?”
An outcome-based strategy
To address this crisis of spending and diminishing returns, Starburst, an intelligence platform, offers a distinct approach. Internally, the company combats wasteful token usage by rejecting the idea of unlimited and unaccountable AI use. Instead of setting quotas or competitive leaderboards based on token volume, Starburst’s strategy is purely outcome-based. They monitor the impact of AI adoption, not the quantity of prompts.
For example, an engineer who uses one billion tokens to achieve an “amazing outcome,” Aswani said, “is valued more than one who uses a trillion tokens with less impact.” The engineering metrics that matter are developer velocity and cycle time — how quickly an idea can move to stable, reliable production — not prompt volume or token consumption. This ensures that the considerable investment in AI tooling is focused on moving the business needle, not just on increasing usage.
Access to fragmented data without moving it
The platform’s core strength is its ability to access and integrate all of what Aswani calls a customer’s “ground truth” structured data, which is often fragmented across 200 or more systems, without requiring the data to be moved.
“AI is only as good as the data it can reach,” Aswani explained. “If AI doesn’t have access to all the data, it’s going to hallucinate. It’s going to make up answers, which, in the case of the enterprise, you may end up making wrong business decisions or ineffective business decisions.”
As data has exploded and different data types have emerged, this has led to fragmentation, with different data types held in different silos. Some large enterprises might have data spread over hundred of systems, powering hundreds of applications. “They have call center data, they have customer support data, they have customer experience data, they have product analytics data. A company like Bank of America or Citibank, you can imagine how many applications they have. Each application has data, and then eventually that data needs to be analyzed for us to understand, what are our customers doing, what kind of questions they’re asking, what kind of friction they’re facing. That’s the big business challenge our customers face, is that data is fragmented, but they need a system, an application or a platform like Starburst that can integrate all of that data without moving data. That’s our value proposition.+
Meanwhile, recognizing that one model doesn’t serve all of an enterprise’s needs, Starburst provides solutions that help contain token spending by maximizing the utility of the LLM ecosystem.
Starburst’s orchestration layer allows customers to “bring your own LLM”. This gives enterprises choice, letting the system determine which model is best suited for a task. For simple chats, the orchestration layer can select a less expensive model, while a different model might be chosen for summarization or multimodal input. By matching the right tool — and its associated cost — to the job, Starburst helps companies control token spending.