Anthropic eyes Microsoft Maia chips; Cursor hits $3B

Nvidia reported $81.62 billion in Q1 revenue with data centre revenue nearly doubling, beating estimates and adding an $80 billion buyback. CEO Jensen Huang said the AI factory buildout is accelerating at extraordinary speed and agentic AI has arrived. The stock still slid for the fourth straight post-earnings session, with investors flagging the Iran conflict and rising competition from custom silicon.

SpaceX filed its S-1 prospectus with the SEC, kicking off what could be the largest IPO in history under ticker SPCX. The filing shows $18.67 billion in 2025 revenue, mostly from Starlink, against a $4.9 billion loss. SpaceX has pumped 60% of capex into its AI division and is pitching orbital data centres as a structural growth story for investors.

An internal OpenAI model has disproved a conjecture Paul Erdős posed in 1946 about how many point pairs in the plane can sit exactly one unit apart. Mathematicians had assumed square grids were optimal. The model produced an infinite family of counterexamples yielding a polynomial improvement, with the proof checked and explained in a companion paper by external mathematicians.

Anthropic is in talks with Microsoft to run Claude on Microsoft's Maia AI chips, following Microsoft's $5 billion investment and Anthropic's $30 billion Azure compute commitment. Anthropic CEO Dario Amodei recently flagged difficulties with compute, and the move would diversify it beyond AWS Trainium and Google TPU. If signed, Maia would mark the third silicon family powering Claude.

Cohere released Command A+ under an Apache 2.0 license: a 218B-parameter mixture-of-experts model with 25B active params, 128K context, multimodal input, and 48 language support. It is built for sovereign agentic workflows on-prem and runs on a single Nvidia B200 or two H100s with W4A4 quantisation. Weights are on Hugging Face with managed inference in Cohere's Model Vault.

The US Commerce Department signed preliminary deals to take $2 billion in equity stakes across nine quantum computing companies, including IBM, IonQ, Rigetti, and PsiQuantum. The package, funded under the Biden-era CHIPS Act, gives Washington direct ownership in firms it views as strategic. Some beneficiaries are linked to Trump-aligned investors, drawing criticism over selection criteria. Quantum stocks jumped sharply on the news.

Cursor's annualised revenue hit $3 billion in late April, up from $2 billion in February, with more than 3,000 customers paying at least $100,000 each. The figures appear in SpaceX's IPO filing, which gives the rocket company a 30-day window to acquire Cursor for $60 billion after its June 12 listing. Cursor's new Composer 2.5 model partly trained on SpaceX data centres.

A single season of Netflix can generate 2,000 hours of raw footage, more than 216 million frames. Netflix built a three-layer pipeline that runs character, scene, and dialogue models, then normalises outputs into one-second temporal buckets. Hybrid text-plus-vector queries against Elasticsearch let production teams find any moment in seconds. The bucketing strategy is the architectural trick most teams skip.

Spotify ships only 12 percent of A/B tests as positive wins, with another 64 percent producing valid learning. The team now treats LLM-as-judge evals as an upstream funnel that filters bad ideas before they hit users, not a replacement for online experiments. Evals get calibrated against shipped outcomes, raising the hit rate while keeping experimentation honest.

A long essay arguing that the AI coding industry has forgotten Michael Polanyi's 1966 claim that we can know more than we can tell. Senior engineers carry tacit knowledge that never makes it into AGENTS.md, retros, or prompts. When AI assistants generate code without this context, codebases accumulate what the author calls tacit bankruptcy, and only the humans who held the knowledge can repay it.

Tom Tunguz dissects SpaceX's S-1 and finds three companies under one roof: Starlink ($11.4B revenue, 39% operating margin), Space ($4.1B, negative 16% margin), and AI ($3.2B, negative 199% margin after the xAI merger). The AI segment consumed $12.7B in capex, more than triple Starlink's. Tunguz frames SpaceX as an AI conglomerate funded by orbital broadband.

Ed Zitron picks apart the WSJ story claiming Anthropic will hit operating profitability in Q2. He argues the math relies on a temporary discounted compute arrangement with SpaceX for May and June, and points to inconsistent CFO statements about whether the company is profitable on training or just inference. The framing treats compute discounting as durable margin.

Epoch AI estimates OpenAI uses just 10 to 15 percent of global operational AI compute and the top frontier labs combined still account for under half. The rest is spread across consumer apps, enterprise inference, and smaller labs. The piece reframes 'lab vs. everyone' competition: frontier labs are not yet where most AI work happens, but their compute share is growing fast.

An analyst joins LLM-extracted themes from call transcripts to a customer table. NULLs get filled with zero or 'no issue mentioned' and the pipeline silently converts 'did not call support' into 'did not experience billing frustration'. The essay names four ways generated variables smuggle bias into downstream models: selection, generation, truncation, and frame imposition. Builders treating LLM outputs as observations are building on quicksand.

Every CEO Dan Shipper argues AI does not eliminate expert work, it expands it. Once models commoditise a skill, output sameness drives demand for human framing and taste. Every still employs nearly 30 people despite aggressive AI adoption. Shipper's in-house Senior Engineer benchmark jumped from 32 to 62 with GPT-5.5, but only because the prompt smuggled in human framing.

Tracea is an open-source observability platform for AI agents, capturing every LLM call, tool execution, cost spike, and error path. It runs on your own infrastructure with no per-event pricing, exposes real-time dashboards, and uses LLM-powered root cause analysis to surface why an agent failed. Pitched as a self-hosted alternative to LangSmith and similar SaaS observability tools.

Rmux is a Rust rewrite of tmux that exposes a Playwright-style typed SDK for driving terminal sessions from code. It supports persistent sessions, structured snapshots, and a Ratatui widget for TUI integration. The pitch is that coding agents and humans can script long-lived shell workflows with the same ergonomics they use for browser automation.

Weaviate now bundles an MCP server into the database, hosted at /v1/mcp on the same port as the REST API. Cursor and Claude Code can connect with one env var and run hybrid search over your code and docs. The tutorial walks through schema design for source files, AST-based chunking, and how to wire retrieval into agent workflows.