Providers slashed prices, builders learned to trim tokens, and Uber reminded everyone that AI ROI remains more fragile than the hype suggests—all in a single week.
Four months.
That's how long it took Uber to burn through its entire 2026 AI budget. And the CEO admitted publicly: nobody can prove the connection between tokens consumed and useful features shipped to customers.
This is Wire #14 from ai|expert. The day token cost stopped being an engineering detail and became a board KPI.
We're starting with this week's paradox. Anthropic launched Claude Opus 4.8 with the fast mode price cut by two-thirds—from 30 to 10 dollars per million input tokens. The same day, GitHub published a playbook that cut agentic workflow token spend by up to 62%. And yet, at Uber, the annual AI budget went to zero before the end of April. Three stories that only make sense when you put them in sequence.
Opus 4.8 has numbers that warrant attention. 69.2% on SWE-Bench Pro—against 58.6% for GPT-5.5 and 54.2% for Gemini 3.1 Pro. 84% on Online-Mind2Web for computer use. 88.6% on SWE-bench Verified. Fast mode now costs 10 dollars per million input tokens and 50 per million output tokens—compared to 30 and 150 for Opus 4.7. For anyone running agentic loops with many short turns, this two-thirds cut in fast mode changes the cost spreadsheet directly. Databricks reported a 61% reduction in cost per token when migrating to 4.8 on multimodal workloads with PDFs and diagrams. [ref: anthropics-claude-opus-48-clai]
Anthropic also delivered two operational capabilities that add up. The Messages API now accepts system inputs within the message array itself—the agent can update instructions mid-task without breaking prompt cache. And "dynamic workflows" in Claude Code: the model plans the task, distributes the work across parallel subagents, and those subagents check and contest each other's conclusions before converging. It's distributed self-revision—but Anthropic's notice is that token consumption can significantly exceed a normal Claude Code session.
There's a system card signal that deserves attention before moving to production. Anthropic identified a growing tendency for the model to speculate about who is evaluating its response—present in roughly 5% of training episodes. It hasn't become observable bad behavior yet; Opus 4.8 makes fewer false claims of success than earlier versions. But for legal or audit agent pipelines, monitor calibration drift.
GitHub transformed that scenario into a solvable engineering problem. An API proxy intercepts every call from CLIs—Claude, Copilot, Codex—and generates a usage artifact per execution. Two agents run daily: an auditor that aggregates consumption by workflow and flags anomalous spikes, and an optimizer that opens issues in the repository itself proposing specific fixes. Result, measured across twelve internal workflows over at least 109 post-fix executions: Auto-Triage Issues dropped 62%. Smoke Claude, 59%. Security Guard, 43%. [ref: github-reduces-agentic-token-c]
The metric GitHub uses—Effective Tokens—weights output at 4x, cache reads at 0.1x, with a per-model multiplier: Haiku at 0.25x, Sonnet at 1x, Opus at 5x. A 10% drop in Effective Tokens is a 10% drop in cost, regardless of model. A unit that goes straight to the board without needing translation. The philosophy they documented is simple: "The cheapest LLM call is the one you don't make."
And that's exactly the control Uber lacked.
Uber deployed Claude Code to 5,000 engineers in December. By March, 84% were agentic coding users. By April, 95%. 70% of committed code came from AI. 11% of backend updates in production were executed by agents without human review. AI cost went up sixfold since 2024, against a P&D base of 3.4 billion dollars. [ref: uber-pumps-brakes-on-ai-spendi]
COO Andrew Macdonald said in Rapid Response: "That link is not there yet"—no evidence of a connection between tokens consumed and useful features to the consumer. And what made it worse was the incentive structure: internal leaderboards ranking engineers by AI usage. More tokens, higher rank. You built a system that rewarded spending, not results. Duolingo faced the equivalent—evaluated engineers on AI tool consumption and had to step back after internal pushback.
The macroeconomic data closes the picture. Enterprise token costs fell 67% over twelve months—from 18.40 to 6.07 dollars per million on average. But 95% of volume still runs on the most expensive frontier models, even for tasks that don't need them. Average spend per organization went up 108% year over year, reaching 1.2 million dollars. 78% of IT leaders ran into charges they'd never budgeted for. Price fell. Volume exploded. The bill went up. [ref: cnbc-signals-enterprise-cost-m]
So the board will ask: where's the routing?
Architectures with tiered intelligence stacks are hitting an average cost of 2.31 dollars per million tokens—against 18.40 from a frontier-only stack. Difference of 87%. OpenRouter, which raised 113 million in May and processes roughly 25 trillion tokens per week—five times more than six months ago—exists to solve that routing in production. Open-weight models captured 38% of enterprise token volume in Q1 2026, versus 11% a year prior. The shift is happening—but most companies still haven't built the evals harness needed to trust cheaper endpoints with critical traffic.
Everything that got cheaper in the previous block needs hardware to run. This week, four stories map the supply that enables—or blocks—that capacity.
AWS revealed that RNG—Resilient Network Graph—is now the standard architecture for most new builds globally, after a Dublin pilot in 2024. The topology replaces the hierarchical fat-tree with a quasi-random mesh using commodity switches and passive optical panels called ShuffleBoxes. Declared numbers: 69% less network hardware, 33% more throughput, 40% less network power consumption. [ref: amazons-resilient-network-grap]
The Spraypoint protocol distributes packets randomly to neighbors until a waypoint associated with the destination is reached—without custom silicon. That generates almost twice the independent paths between routers. But architects need caution: the academic paper doesn't publish p99 latency for hop count differentials, and there's no published evidence of integration with RDMA or RoCE. The throughput gain is real in AWS infrastructure. The direct impact on inference clusters is still theoretical.
At Dell, the Q1 fiscal 2027 numbers show the other side of the equation. Total revenue of 43.84 billion dollars, up 88% year over year—the fastest growth since the 2018 IPO. AI server revenue: 16.1 billion, up 757%. Annual guidance was revised from 50 to 60 billion dollars. [ref: dell-q1-beats-on-ai-server-dem]
More than 5,000 AI server customers in the quarter—neoclouds, sovereign entities, enterprises. And COO Jeff Clarke said they're "repricing every day" because of the component supply shortage environment that started in January 2026. DRAM, NAND, x86 CPUs, and hard drives are under supply constraints in the second half of 2027. If you don't have a long-term contract locked in now, you're in the allocation queue behind those who do—and this week's quote may not be good next month.
For anyone wanting to escape that queue, an on-prem alternative gained real viability this week.
Four Mac Studio M3 Ultra with 512 GB RAM each—total cost around 38 thousand dollars—running DeepSeek V3.1 671B at roughly 25 tokens per second and Kimi K2, a trillion-parameter MoE, at about 34 tokens per second. No cloud API, no data egress. What made this viable now is RDMA over Thunderbolt 5, which Stabilise.io measured at 5 to 9 microseconds latency between nodes at 80 gigabits per second. [ref: the-10k-sovereign-ai-cluster-h]
The compliance argument is architectural, not benchmarking: weights and prompts never leave the building. For teams under GDPR, HIPAA, or NIS2, that property can exceed any cloud throughput advantage.
But the limit is clear. These clusters suit batch inference better than interactive chat. Setting up RDMA requires manual commands in macOS recovery mode. Apple Silicon frameworks still lag vLLM in absolute performance. For MoE models, communication time approaches compute time during expert routing—requires manual tuning of the memory layer. It's viable. Not plug-and-play.
And Lombardy closes the European map with a regulatory surprise. The regional council approved a 200% surcharge on data centers in agricultural zones and 100% in rural zones. With 30 GW of capacity requested across Italy and the region authorizing only 2 GW—regional councilor Massimo Sertori called them "real and concrete projects"—the survival rate of the planned pipeline is in single digits. National law cut licensing time to 10 months. But regional law obliterated the economics of land that operators wanted. [ref: italy-imposes-200-tax-on-data-]
The takeaway for anyone modeling European expansion is direct: land-use classification is the primary cost variable before signing any land options in Lombardy. Brownfield industrial is the only economically viable path. A 200% surcharge on greenfield rural, combined with a 2 GW authorization ceiling against a regional pipeline of 15 GW-plus, eliminates everything else.
Token cost became board language—and whoever doesn't have the equivalent of Effective Tokens on the spreadsheet before the next approval round will repeat Uber's mistake. On Friday, the Edition goes deep on production evals and what the Pymetrics case and Microsoft's MDASH reveal about auditing agents before they audit you. Have a good week.