Software: Does "Token Shock" hurt AI adoption?

Research evidence excerpt

Software: Does "Token Shock" hurt AI adoption?

tangible economic activity rather

than remaining largely hype-driven. For instance, Fireworks AI Inc.,

a US inference platform, serves roughly 30 trillion tokens a day for ENTERPRISE AI ADOPTION: ECONOMICS ARE BECOMING

more than 10,000 customers — including Cursor, Uber, Samsung, THE MAIN CONSTRAINTS

Notion and Shopify — and its annualized revenue, an estimated

Our conversations with corporate users show that the primary$305m at the end of 2025, had reportedly reached $800m by May

challenge facing enterprise AI adoption is no longer whether(link).

the technology works (even though those issues have not gone

While some could argue that this is because the usage is away), but whether it can be deployed and scaled economically.

predominantly in coding tools rather than broadly with agents, but Unlike traditional software, where costs are largely tied to licenses,

we heard a similar issue last year from early adopters of Salesforce’s subscriptions, or user seats, AI is now shifting to a consumption-

Agentforce (which an Agentic platform and not a coding tool) that based model in which costs depend on the cost per token, token

periodically costs got out of hand. The underlying issue was that usage, inference activity, API calls, and increasingly autonomous

usage did not match to value being created. agents. The complexity of modeling the cost can be seen by simply

looking at all the components of just the token cost is composed of

Regarding the issue of model degradation, we talked to two (Exhibit 4) and that does not include all the other cost factors (e.g.

large European insurers about their own domain-specific models non-GPU costs). As a result, AI spending is far less predictable and