Hundreds of billions in LLMs, yet AI pricing is a black box for enterprises

For AI researchers and ML engineers, the pricing dilemma reveals a deeper flaw: current LLM architectures produce inherently stochastic outputs, making cost prediction almost impossible without rigorous guardrails. This unpredictability threatens to slow the rollout of mission-critical AI agents in enterprise settings.

The scramble to monetize artificial intelligence is hitting a wall: the economics of token-based pricing are so chaotic that even the industry's biggest players can't create stable cost models. As Microsoft, Google, and Anthropic collectively pour hundreds of billions into large language models (LLMs), the task of converting that spending into predictable revenue streams is proving fraught.

“

As a result, the same business workflow could cost $10 one day and $100 the next, depending on how the AI interprets the task and invokes sub-agents.

At the heart of the problem lies the token — the mathematical subunit that LLMs use to process prompts and generate responses. When a user asks ChatGPT to draft an email or an agentic system orchestrates a multi-step workflow, every input and output is broken into tokens. The bill for that interaction depends on how many tokens are consumed, which varies dramatically. A seemingly minor rephrase of a question can double or triple the token count; the same prompt can yield different-length answers from different models; and as noted by Saviynt’s Simon Gooch, agentic setups where multiple AI agents collaborate compound this variability unpredictably.

Goldman Sachs’ analysis highlights a paradox: while per-token costs have plummeted over recent years, total token consumption by enterprises has surged, making aggregate costs harder to forecast. For AI service providers, this means the unit economics that underpin pricing are constantly shifting. A fixed-price contract for 12 or 24 months, once standard in SaaS and cloud services, becomes a gamble. “Trying to tie someone into a cost model for the next 12 months, two years, three years, it doesn't make any sense, honestly, because we don't know,” Gooch said.

The implications ripple across the AI ecosystem. For SaaS companies building on top of LLMs, the inability to price reliably threatens margins and slows go-to-market strategies. When a startup cannot confidently quote a client a monthly fee for an AI-powered analytics tool, the sales cycle stalls. For enterprise buyers, sticker shock from variable monthly bills under consumption-based models can kill adoption. CIOs used to predictable cloud invoices are wary of AI line items that swing by thousands of dollars overnight.

This pricing fog is more than a temporary friction — it reflects fundamental architectural traits of current LLMs. The stochastic nature of generative AI means outputs aren't deterministic, and agentic frameworks, which increasingly chain model calls for reasoning and action, amplify token usage exponentially. As a result, the same business workflow could cost $10 one day and $100 the next, depending on how the AI interprets the task and invokes sub-agents. The industry has responded with attempts to set usage caps and credits, but these are blunt instruments that often frustrate power users.

The three tech giants — Microsoft, Google, and Anthropic — each pursue different monetization paths, from Copilot subscriptions to per-token APIs to enterprise deals, yet none have cracked the predictability code. The pressure is likely to accelerate consolidation or push toward end-to-end platforms that control the entire stack, from model inference to application, to better forecast costs. Meanwhile, the existence of free, high-quality AI like ChatGPT sets a consumer price anchor of zero, making it even harder for third-party developers to charge a premium.