Jaya Gupta

On 28 May 2026, Jaya Gupta published Token Budget Wars, a thread on X that drew 230.5K views and opened with a single claim: "Enterprise AI has moved from adoption to allocation." The numbers carry the argument. Two runs of the same workflow on the same input can differ in token cost by 5-10x, a figure she casts as something a CFO has to explain to a CEO. Claude shipped in November 2025, after annual 2026 budgets were already locked, and by Q1 companies were running multiples ahead of plan. Retry tails compound the problem: cost per resolution tracks T/p, so a drop from 90% to 70% completion adds roughly 28%. Context inflation and routing everything to the frontier model do the rest.

Gupta had laid the foundation earlier. On 22 December 2025, with Ashu Garg at Foundation Capital, she argued that agents raise the bar for systems of record rather than replacing them, and that the missing layer is decision traces: the exceptions, derogations and precedents living in Slack, deal desk conversations and employee memory. Salesforce stores current state, not state at the moment of decision. Snowflake receives data by ETL, after the decision is gone.

The two pieces meet at one point. Decision traces accumulate into a context graph she calls more valuable than the cost report, and the allocation layer is the prize. Her obstacle: "the signal and the noise share the same unit."