Today's News
AI token spend surges 10x in six months, forcing companies to choose between cost controls and productivity gains
The Pragmatic Engineer
Across 15 companies ranging from seed-stage to 10,000+ person enterprises, token consumption has exploded in the past 2-3 months, with some developers spending $500/day on Claude Code and fintech staff engineers hitting $1,400 in a single session. The trend reflects both genuine productivity improvements and "tokenmaxxing" behavior driven by internal leaderboards and performance reviews.
Anthropic offers no volume discounts even at $5M+ annual spend, while Cursor negotiates tiered discounts starting at 5% above $1M
The Pragmatic Engineer
Custom pricing agreements exist but only at high thresholds; companies are discovering that asking vendors directly is the only way to surface available discounts, and that token cost is becoming a material line item in engineering budgets ($24-60K+ annually per heavy user).
Techmeme
OpenAI announced it has signed contracts for 10GW of compute, hitting a goal it previously aimed to reach by 2029, with 3GW+ added in the last quarter alone. This represents a major acceleration in infrastructure buildout and signals OpenAI's confidence in near-term demand, with direct implications for the competitive race for compute and power availability across the AI stack.
SemiAnalysis
Agentic workflows with high cache hit rates and input-to-output ratios compressed effective blended pricing to $0.99/MTok despite $5/$25 sticker prices, while Blackwell and software optimizations (wideEP, disagg, MTP) delivered 14-32x throughput gains over Hopper. This margin expansion from <40% to >70% represents a structural shift where AI labs now capture nearly all ecosystem value, reversing 2023-2025 dynamics where infrastructure vendors dominated.
SemiAnalysis
VR NVL72 achieves 17x higher throughput than H100 at only 70% higher TCO, yet capex-per-watt barely moved ($37.4W to $38.1W) and Nvidia's pricing framework remains anchored to prior assumptions rather than reflecting increased token economics value. With N3 utilization expected to exceed 100% and DRAM fabs running >90%, both vendors have room to raise prices materially—particularly through SOCAMM memory markup (currently ~$8/GB, potentially reaching $13/GB by end-2026)—but haven't yet.
Codex expands from coding into general computer work with 42% faster inference and dynamic UI routing
AINews
OpenAI shipped a major update positioning Codex for knowledge workers across docs, slides, spreadsheets, and planning tasks, with native integrations to Microsoft/Google/Salesforce suites and a planning UI that dynamically routes task-specific interfaces rather than static toggles. This signals OpenAI's strategic pivot from "coding agent" to general-purpose "computer-use agent" competing directly with Claude's agentic positioning.
Microsoft commits to "per-user and usage" hybrid pricing model, moving away from pure seat-based licensing
Clouded Judgement by Jamin Ball
Satya Nadella stated that Microsoft's productivity, coding, and security businesses will transition from per-seat to "per-user and usage" models, with the seat becoming a packaging mechanism for prepaid consumption beyond which customers pay per token or per action. GitHub Copilot moved to usage-based pricing effective June 1, and D365 bookings growth is already being impacted as customers pause seat renewals to evaluate the new model. This signals a fundamental shift in SaaS economics that will ripple across the industry.
Consumption-based pricing will become the dominant model for AI agent and task-based software, requiring founders to abandon seat-based assumptions early
Clouded Judgement by Jamin Ball
The shift from per-seat to consumption economics is particularly critical for companies building agent-based software, where traditional licensing breaks down—agents don't buy seats, they execute tasks and consume tokens. Companies that delay this transition risk building for a shrinking market, as metrics like ARR, NRR, and bookings become messier and less predictive under hybrid models.
Worth Reading
The Pragmatic Engineer (Substack)
I read the full post covering GitHub's infrastructure collapse under AI-driven load (reliability down to 86% this month, zero nines), contrasted with why competitors haven't faced similar breakdowns. The piece includes concrete reliability metrics, analysis of GitHub's self-inflicted vs. load-driven failures, Mitchell Hashimoto's departure and commentary on professional infrastructure standards, Anthropic's recent trust erosion (Claude Code nerfing, silent price increases, company bans), GitHub Copilot's dramatic price increases, and emerging competitive dynamics in AI coding tools including Cursor and Google's efforts. This directly addresses infrastructure resilience, competitive positioning, and economics across the AI stack.
Read moreBloomberg
I read the full Bloomberg report on OpenAI's infrastructure milestone: the company has signed contracts for 10 gigawatts of AI computing capacity, meeting a goal originally targeted for 2029 three years early. The report includes concrete breakdowns of recent capacity additions (2GW from Amazon, 1GW undisclosed in past 90 days), discusses how Stargate has evolved from a specific $500B venture into a broader data center brand, and reveals strategic pullbacks from high-cost sites in Texas, UK, and Norway. This is directly relevant to understanding OpenAI's infrastructure economics and competitive positioning.
Read moreCursor Blog
I read the full technical post on Cursor's approach to building and iterating their coding agent harness. The post details concrete engineering practices: how they've evolved from heavy context engineering and guardrails (late 2024) to dynamic context fetching as models improved; their multi-layered measurement system including "Keep Rate" (fraction of agent code that remains in codebase) and LLM-based satisfaction detection; specific tool error classification and anomaly detection for production monitoring; and deep model-specific customization (e.g., patch-based edits for OpenAI vs. string replacement for Claude, custom prompting for different providers). It also covers mid-conversation model switching challenges, subagent orchestration, and their vision for multi-agent systems where the harness becomes the coordination layer. This is directly relevant to understanding competitive product strategy and infrastructure buildout in AI coding.
Read moreAxios
I read the full Axios report on a significant reversal in Trump administration policy toward Anthropic. The piece details how the White House is drafting an executive action to circumvent the Pentagon's supply chain risk designation that previously blacklisted Anthropic from federal use, with meetings underway this week to brief companies on guidance that would walk back the OMB directive. The reporting includes concrete details on the underlying dispute: Anthropic's refusal to sign the Pentagon's "all lawful purposes" agreement due to restrictions on mass domestic surveillance and autonomous weapons, contrasted with OpenAI and Google's acceptance of similar terms. It also notes that despite the legal fight, the NSA is already using Mythos and multiple agencies are clamoring for access, suggesting real economic and strategic stakes in the resolution.
Read more