Today's News
Anthropic hits $1B quarterly profit on $60B ARR run rate, vastly outpacing OpenAI's money-losing model
SemiAnalysis
Anthropic achieved 3Q26 GAAP EBIT of $1B (6% margin) with mid-60% gross margins, powered by 75-85% API revenue mix versus OpenAI's consumer-heavy subscription model. The company reached $60B ARR run rate—a dramatic acceleration from $9B at end-2025—with Net Dollar Retention hitting 500% as existing customers massively increased token consumption, driven by Claude Code dominance in B2B coding workflows. With $60M ARR per MW of compute (up from $16M nine months prior), Anthropic can reinvest 60% of compute spend into training new models while maintaining profitability, versus OpenAI's negative EBIT margins and $0.70/month cost to serve 900M free users. This margin gap compounds as better models drive higher token prices and throughput, positioning Anthropic as the clear clubhouse leader in capturing the B2B market and creating a structural reinvestment advantage that could extend its lead over a money-burning competitor.
Benedict Evans Newsletter via Benedict Evans
Export controls on Anthropic's Claude Fable 5 and Mythos 5 models, imposed June 12 after Amazon researchers discovered a jailbreak technique, were lifted June 30. Fable 5 is now available globally on Claude Platform, Claude.ai, Claude Code, and Claude Cowork starting July 1, while Mythos 5 access was restored for US organizations. Anthropic deployed an improved safety classifier blocking the reported bypass in over 99% of cases and worked with the US Department of Commerce's Center for AI Standards and Innovation to validate the strengthened safeguards. The company also coordinated with Amazon, Microsoft, Google, and other Glasswing partners to develop a shared industry framework for assessing and fixing AI model jailbreaks, signaling deeper government collaboration on pre-release testing and research.
Aadit Patel via Techmeme
Chinese AI models have captured 30-46% of token usage by US companies via OpenRouter since February, up from 11% annually, driven by cost advantages of 60-90% versus Anthropic and OpenAI models. Z.ai's GLM 5.2 achieved the fastest adoption of any model tracked by Vercel in 2026, with daily token volume growing 27x and customer count growing 80x in its first full week; GLM 5.2 landed within a percentage point of Anthropic's Opus 4.8 on agentic benchmarks at roughly one-fifth the cost. Meanwhile, major platforms are pursuing vertical integration: Microsoft is replacing OpenAI and Anthropic models with its own MAI models in Excel and Outlook to reduce costs, Amazon plans a $25B+ bond sale for AI infrastructure, and DeepSeek is developing its own inference chip to reduce reliance on Nvidia and Huawei. These moves reflect intensifying cost pressure and structural shifts toward internalized model economics across the AI stack.
Benedict Evans Newsletter via Benedict Evans
The two companies announced a major chip hub investment in response to AI-driven memory demand that has forced Apple to raise Mac/iPad prices 20-25% and driven overall memory price increases of 5-10x. This represents a 5-10% annual capex growth rate and signals the industry's commitment to addressing the memory bottleneck constraining AI infrastructure buildout.
Lilian Weng via AINews
Weng's comprehensive research covering 35 papers reframes recursive self-improvement (RSI) around harness design—the system surrounding a base model that orchestrates execution, tool use, planning, and evaluation—rather than direct weight modification. The analysis establishes that goal/context specification will remain necessary even as harness improvements get internalized into core models, signaling the field's consensus shift toward agent/harness engineering as a primary scaling axis alongside parameter count. Key patterns include workflow automation with goal-oriented loops, file systems as persistent memory for long-horizon tasks, and sub-agent parallelism, with the prediction that harness improvements will eventually be internalized into core model behavior but external context and tool interfaces should remain.
The Next Platform
Meta is formalizing plans to sell cloud services under the Meta Compute initiative, positioning itself as a neocloud provider to manage massive AI infrastructure investments and reduce reliance on external suppliers. The effort, led by infrastructure head Santosh Janardhan alongside Daniel Gross from Superintelligence Labs, reflects a broader competitive dynamic where hyperscalers are vertically integrating compute to control costs and supply chain risk. Meta has already committed to $21B in capacity from CoreWeave and $10B from Google Cloud, but now plans to offer API access to its Muse Spark model and basic infrastructure-as-a-service to customers, leveraging its homegrown MTIA accelerators and GPU clusters to monetize excess capacity and offset enormous AI infrastructure spending.
The Next Platform
Datacenter Ethernet switch sales grew 61% in Q1 2026 to just over $10B, significantly outpacing overall server market growth of 30.4%, signaling that networking infrastructure is becoming strategically as important as chip selection for AI infrastructure builders. Sales of 800 Gb/sec switches surged 10.3x to $3.58B in Q1, now representing 35.8% of datacenter Ethernet sales, with cost per bit dropping more than half year-over-year as competitive pressure and technological advancement accelerate. Nvidia surpassed both Cisco and Arista in datacenter Ethernet switch revenues, with Q1 datacenter Ethernet sales of $2.1B (up 2.9x year-over-year), reflecting the broadening of GenAI from model training to production inference and the resulting unprecedented datacenter bandwidth demands.