🎯 Top 3 Things to Know
1. Anthropic told investors it expects its first quarterly operating profit in Q2 2026, on $10.9 billion of revenue. Revenue is up 130% over Q1 and operating income lands around $559 million. The shift comes from two places at once: customers spending more than $1 million a year roughly doubled between February and April, and compute cost as a share of revenue is projected to fall from 71 cents per dollar in Q1 to 56 cents in Q2. Frontier AI companies have been treated as structurally loss-making, so an operationally profitable Anthropic, in the middle of training new frontier models, recasts how investors will read the coming AI IPO cycle. Watch the Q2 actuals against the projection, and watch whether OpenAI's confidential S-1 includes a comparable profitability path. PYMNTS
2. SpaceX's IPO prospectus disclosed that Anthropic is paying SpaceX $1.25 billion a month for compute through May 2029. The contract totals about $45 billion, covering Colossus 1 and an expansion onto Colossus 2's NVIDIA GB200 capacity in June. Either side can exit with 90 days' notice. The bill is roughly the size of SpaceX's standalone 2025 revenue, which sets a new floor for what serious frontier training now costs and explains the urgency behind Anthropic's other deals with Amazon, Google, Microsoft, Broadcom, NVIDIA, and Fluidstack. Watch how the contract is amortized in the Q2 cost-of-revenue line, and whether other labs follow with similar long-term hyperscaler commitments. Axios
3. CNBC argued cheap near-frontier models could compress the pricing assumptions behind the OpenAI and Anthropic IPOs. The same ten-evaluation workload costs $4,811 on Claude, $3,357 on ChatGPT, $1,071 on DeepSeek, $948 on Kimi, and $544 on Zhipu's GLM, per Artificial Analysis. Chinese models' share of usage on the OpenRouter developer marketplace rose from about 1% in 2024 to over 60% this month. Databricks CEO Ali Ghodsi described an "advisor model" pattern in which enterprises route most calls to a cheap default and reserve frontier models for the hardest tasks. Anthropic itself wrote in a May policy paper that US models are only "several months ahead" and that Beijing leads on cost-driven adoption. Watch whether the OpenAI S-1 quantifies pricing power as a separate risk factor. CNBC
🚀 Frontier Models & Features
- OpenAI is preparing a confidential S-1 with the SEC as early as today, targeting a September listing. Goldman Sachs and Morgan Stanley are on the deal, against a private-market valuation near $852 billion and roughly $25 billion of annualized revenue. CNBC
- Anthropic confirmed Colossus 2 GB200 capacity ramps through June, scaling Claude Code and API rate limits already lifted earlier this month. TechCrunch
- Google DeepMind connected Genie 3 to Street View at I/O. The world model can now be seeded with real-street imagery to generate interactive 720p environments at 24 fps. TechCrunch
🔬 Research Worth Reading
δ-mem: Efficient Online Memory for Large Language Models (Lei, Zhang, Li, Wang et al.). arXiv
- TL;DR: Augments a frozen full-attention LLM with a small online associative-memory state, updated by a delta rule, that injects low-rank corrections into attention at decode time instead of widening the context window.
- Stat: With only an 8×8 memory state, δ-mem reaches 1.31× the score of the frozen backbone on MemoryAgentBench and 1.20× on LoCoMo, while preserving general benchmarks.
- Apply it: Before pushing context windows past 200K to fix a long-document RAG bug, try bolting a small associative-memory state onto the existing model and measure delta on a memory-stress eval. The fixed-size state caps the cost and is easier to reason about than ever-larger prompts.
MemReread: Enhancing Agentic Long-Context Reasoning via Memory-Guided Rereading (Ji, Weng, Li, Tang, Lou & Zhang / Soochow University and Peking University). arXiv
- TL;DR: Skips intermediate retrieval and instead has the agent decompose its question and reread targeted chunks of the source when its working memory turns out to be insufficient, so retrieval cost is paid only when needed.
- Stat: Avoids the quadratic-attention penalty of standard long-context decoding by deferring rereads to the moments the agent actually fails to answer, rather than processing everything up front.
- Apply it: In an agent that already loads big documents into its context, add a check that detects insufficient-evidence answers and triggers a scoped reread of the original source, instead of reissuing the full query against a vector store.
🏢 Enterprise in the Wild
- AppliedAI and McKinsey deployed an agentic onboarding system at a European chemicals manufacturer that cut a two-week vendor onboarding process to under five minutes of active processing, with a reported 99%-plus reduction in manual effort. PR Newswire
- SAP Sapphire 2026 announced new SAP-on-Azure capabilities for fine-tuning models inside SAP's Business Data Cloud, aimed at finance, supply chain, and HR workflows. Microsoft Azure blog
- Japan's Ministry of Finance said the government and major banks will gain access to Anthropic's Claude Mythos within two weeks, after a meeting between the finance minister and US Treasury Secretary Scott Bessent. BuildFastWithAI summary
🛠️ Tooling & Ecosystem
- Anthropic added Compliance API integrations with third-party security and compliance tools, giving IT a unified governance surface across Claude products. Anthropic news
- Managed Agents now support self-hosted sandboxes in public beta and MCP tunnels in research preview, so tool execution can run inside customer infrastructure while the agent loop stays on Anthropic's side. Anthropic news
- OpenAI Codex shipped Appshots on macOS, a hotkey that attaches an app window plus available text to a Codex thread without prompt setup. OpenAI release notes
⚖️ Policy & Regulation
- The White House AI executive order, which would have set up a voluntary 90-day pre-launch review framework for frontier models, was postponed again on May 21. Reporting points to internal disagreement over whether participation should be genuinely voluntary or carry NSA-backed classified testing. No new signing date has been announced. CNN via BuildFastWithAI summary
- Anthropic's May policy paper states US models are only "several months ahead" of Chinese counterparts and that China leads on cost-driven global adoption, framing its argument for stronger US infrastructure and export-control support. The acknowledgment is notable given the company's IPO timeline. BuildFastWithAI summary
📌 Watch List
- AI IPO sequence and how SpaceX, OpenAI, and Anthropic price each other.
- Long-context memory mechanisms competing with context-window expansion.
- Self-hosted sandboxes and tunnel-based MCP execution inside enterprise boundaries.
- Cost-driven model routing and the "advisor model" pattern in enterprise stacks.