Token Prices Are Falling. So Why Are AI Bills Going Up? | ServiceNow, TransUnion, WEKA

Recorded live at NeoCloud Summit 2026, Hotel Kabuki, San Francisco. Token prices keep falling, yet enterprise AI bills keep climbing. As companies move from simple chat interactions to agentic workflows that reason, retry, and call tools, token consumption is growing quickly, and many organizations are only starting to understand what they're actually paying for. In this panel, Vyoma Gajjar of ServiceNow moderates a conversation with Arzoo Mittal, a technical product manager at TransUnion, and Val Bercovici, Chief AI Officer at WEKA. They look at enterprise tokenomics from the other side of the invoice. Arzoo explains why agents are moving enterprise cost curves. A chat answer may take one or two model calls, while an agentic workflow can take ten or fifteen. She also shares how product teams in regulated industries decide when to use a fully autonomous agent and when a human-in-the-loop workflow makes more economic sense. Val compares today's AI bills to the early days of cloud, when invoices ran to hundreds of pages nobody could decipher. He makes the case for bringing FinOps maturity to AI by using premium tokens for planning and orchestration and lower-cost tokens for most execution. TOPICS COVERED • Why falling token prices haven't lowered AI bills • How agentic workflows multiply model calls and token usage • Applying FinOps discipline to AI spend • Premium, mid-tier, and low-cost classes of tokens • Open weight alternatives and dedicated token contracts • Continuous evals to measure model quality • Choosing between autonomous, human-assisted, and human workflows • Why memory, not compute, is AI's real bottleneck • Prompt caching and the mismatch with long-running agents • KV cache offloading to storage • Cost per successful outcome vs. cost per token • Benchmarks and evaluations for business outcomes • Reserved infrastructure vs. pay-per-token pricing • Model routing across multiple providers • Visibility and observability into token spend Val explains why he believes memory is the real limit on AI. Agents that run for hours or around the clock don't fit the short cache windows that pricing models were originally built around, which puts enormous pressure on HBM and DRAM and is pushing the industry toward storage offload. Arzoo argues that the metric that matters most is cost per successful outcome. If an agent can complete a workflow for a fraction of what a human would cost, the raw token count matters less. To close, both panelists share what enterprises need to succeed: model routing across multiple providers, rigorous evaluations, and far more transparency into where tokens are actually being spent. Watch the full panel for a practical look at the economics behind enterprise AI and token consumption. :mega: Join us at AI INFRA SUMMIT 6 on December 9, 2026, in San Francisco. Register now → luma.com/aiinfra6