- Apple cut daily AI spending allowances for some engineers from $300 to $200 in October, a 33% reduction, as “tokenmaxxing” drives aggregate bills higher even as per-token prices fall.
- OpenAI faces projected inference costs of $14.1 billion in 2026, with its compute already contributing $10 billion to Microsoft’s Q1 FY2027 revenues.
- AT&T cut advanced AI task costs by up to 56% by dynamically routing queries to cheaper models, targeting a 60-70% open-source share, up from 40%.
Apple cut daily AI spending allowances for some engineers from $300 to $200 in October, joining Microsoft, Meta and Accenture in clamping down on what the industry now calls “tokenmaxxing.” The structural problem: as per-token prices fall, consumption rises faster, and aggregate AI bills climb regardless of unit cost reductions.
Big Tech Pulls Back on AI Spend
Employees encouraged last year to experiment freely with AI tools have driven consumption-based billing to levels that do not reliably map to business output. Daniel dos Santos, VP of research at Forescout, has noted that high token consumption does not translate reliably into business value. As frontier model prices fluctuate, aggregate spend has become difficult to forecast, prompting tighter fiscal controls across large technology companies.
What Tokenmaxxing Actually Costs
Most enterprises lack governance over how employees select and use AI models. Agentic workloads compound the problem: employees building small agents frequently reach for expensive frontier models for routine tasks, and some of those agents, once deployed, run unattended, accumulating costs with no visibility. Conal Gallagher, CISO and CIO at Flexera, has flagged this as a hidden drain on resources.
Inference is now the second-largest line item in enterprise AI budgets. A rebound effect drives much of that growth: declining unit costs stimulate greater overall usage, pushing total spend higher even as per-token prices fall.
OpenAI’s Azure Dependency
At $8.67 billion in inference spend through the first nine months of 2025 alone, and with full-year 2026 projections reaching $14.1 billion for 2026, OpenAI’s compute costs contribute materially to its operating losses and require continuous external capital. The majority of that compute runs on Microsoft Azure. OpenAI accounts for roughly 45% of Microsoft’s $625 billion cloud backlog and has committed up to $250 billion in Azure spend over time, with its compute contributing $10 billion to Microsoft’s Q1 FY2027 revenues. The dependency runs in both directions: Microsoft’s financial performance is now materially tied to OpenAI’s consumption.
Efficiency Tools Beyond Budget Caps
Emergent’s RTK CLI proxy intercepts shell commands and optimises context sent to AI coding assistants such as GitHub Copilot and Cursor. Microsoft Research’s LLMLingua takes a different approach: prompt compression at inference time, reporting up to 20x reduction with minimal performance loss on reasoning and conversation benchmarks, according to Microsoft Research. Provider-side caching, which bills repeated context at a lower rate, and retrieval filtering, which limits what gets included in a prompt, are the other main levers. The practical standard for evaluating these tools is shifting from raw token reduction to completed-task quality relative to total cost.
AI FinOps Takes Shape
Model tiering is the highest-impact lever: routing tasks dynamically to models matched to their cost and performance requirements, rather than defaulting every query to a frontier model. Prompt caching, batch APIs and AI gateway guardrails follow. Shadow AI auditing, identifying and rationalising unsanctioned model usage, closes the loop on spend visibility. Logging and observability infrastructure underpins all of it; without per-task attribution, cost governance stays theoretical. As inference optimisation strategies mature, the gap between managed and unmanaged deployments widens considerably.
The Provider Paradox
Anthropic released Claude Haiku 5.5 on October 7 at $0.10 per million input tokens and $0.50 per million output for prompts under 100,000 tokens, roughly a tenth of the previous Haiku’s cost, according to Anthropic. The Stanford AI Index 2025 tracked the longer arc: querying a model at GPT-3.5-level performance cost roughly $20 per million tokens in late 2022 and around $0.07 by late 2024. That price compression is precisely what drives the rebound effect, cheaper tokens invite more of them.
Routing, Not Restricting
AT&T cut the cost of coding and other advanced AI tasks by up to 56% by dynamically routing employee queries to less expensive models, with a reported 2% drop in quality. The company targets a 60-70% share of queries handled by open-source models, up from 40% currently. That trajectory points to where enterprise AI architecture is heading: hybrid deployments combining frontier models for high-complexity tasks with self-hosted open-source models for high-volume, stable workloads, and on-device AI for scenarios requiring zero marginal inference cost.
Gartner advised CFOs in November 2026 to treat AI as a tool for workforce transformation rather than headcount reduction, with active governance needed to redirect AI-generated capacity toward measurable outcomes, time saved, first-contact resolution rates and opportunity cost, rather than infrastructure spend alone. The 56% cost reduction AT&T achieved by routing rather than restricting is the cleaner model: optimise what gets sent to which model, and the budget problem becomes more tractable than any blanket cap.
