Friday 1 May 2026
The industry’s center of gravity is shifting from training to inference. Noam Brown calls inference compute “a strategic resource, currently undervalued”; Sam Altman says OpenAI must become “an AI inference company now.” Jensen Huang put a number on it: computing demand up 10,000× in two years. This week’s moves — Google selling TPUs directly, OpenAI abandoning Stargate data centers for flexible leases, Nvidia acquiring Groq, Amazon’s custom chips hitting $20B — all point to the same conclusion: the inference inflection is here, and the infrastructure race is the new frontier.
🤖 Models & Infrastructure
Google to Sell TPU Chips to ‘Select’ Customers in Latest Shot at Nvidia — Alphabet plans to sell its custom TPUs directly to customers for on-premise installation. Already in deals with Anthropic and Meta, Google is escalating its competition with Nvidia for AI silicon dominance.
Mistral Medium 3.5 Powers Remote Vibe Agents — Mistral’s 128B dense model powers Vibe, a new platform for long asynchronous coding tasks running in the cloud. Scores highly on SWE-Bench Verified and enables Le Chat’s new Work mode for complex multi-step tasks.
Granite 4.1: IBM’s 8B Model Matching 32B MoE — IBM’s new Granite 4.1 family (3B/8B/30B) trained on 15T tokens with a five-phase pre-training approach. The 8B model matches the previous 32B MoE’s performance through multi-stage RL focused on data quality.
OpenAI Has Effectively Abandoned First-Party Stargate Data Centers — The original Stargate plan for 20 data centers collapsed over control disputes. OpenAI is now leasing compute instead. Analysts estimate the company could run out of cash by mid-2027.
White House Opposes Anthropic’s Plan to Expand Access to Mythos Model — US officials cite national security concerns over Anthropic’s plans to widen access to its Mythos model, adding another data point to the growing tension between AI safety policy and commercial deployment.
🛠️ Agents & Tools
Build Programmatic Agents with the Cursor SDK — Cursor’s new SDK exposes the same runtime, harness, and models powering the IDE for CI/CD, automations, and embedded agents. A clear platform play that shifts Cursor from seat-based IDE to programmable agent infrastructure.
Lessons on Building MCP Servers — Practical guide to making MCP toolchains work: models don’t plan — they scan tool lists and grab whatever looks most probable. Effective chains make the next call “blindingly obvious” at every step.
Agents Can Now Create Cloudflare Accounts, Buy Domains, and Deploy — Cloudflare is making it possible for AI agents to autonomously provision infrastructure end-to-end: create accounts, purchase domains, and deploy — with Stripe integration for payments.
DataPRM: Reliable Data Analysis Agents — New environment-aware process reward model for data analysis tasks, addressing the reliability gap in agent-driven data workflows.
🔬 Research & Engineering
AI Evals Are Becoming the New Compute Bottleneck — Evaluation costs have escalated to rival or exceed training costs, with some runs costing tens of thousands of dollars. The field needs standardized documentation and data reuse to maintain equal access and external validation.
Apple’s LaDiR: Latent Diffusion Enhances LLMs for Text Reasoning — Apple’s novel reasoning framework unifies continuous latent representations with iterative refinement from latent diffusion models, enabling parallel generation of diverse reasoning trajectories.
Introducing AutoSP for Sequence Parallel Training — AutoSP automates converting standard transformer training code into sequence-parallel code for long-context LLM training, integrated with DeepSpeed. Eliminates complex manual code changes.
Microsoft World-R1 for 3D-Consistent Video Generation — Reinforcement learning framework that improves 3D consistency in video generation by leveraging feedback from 3D and vision-language models without modifying the base architecture.
Agentic Harness Engineering Beats Human-Designed Baselines — Research shows harness evolution via revertible components improves Terminal-Bench 2 pass@1 from 69.7% to 77.0%, beating a human-designed Codex-CLI baseline. The throughline: model quality alone is insufficient — the harness around the model determines production performance.
💰 Business & Capex
The $112 Billion Quarter: Hyperscalers Bet the Farm on AI — Google, Amazon, and Microsoft reported combined capex of $112B in Q1. Google is now outspending Microsoft, and the gap is widening. Bank of America forecasts hyperscaler debt issuance hitting $175B in 2026.
Apple Soars Past Wall Street Expectations as Tim Cook Prepares to Step Down — $111.2B in revenue in the first earnings report after Cook’s pending departure announcement. The transition to John Ternus begins amid exceptional financial performance.
Amazon’s Custom Chips Are Now a $20B Business — AWS’s Trainium and Inferentia chips have grown from a side project to a $20B revenue line, putting real pressure on Nvidia’s dominance in cloud AI silicon.
SpaceX Board Approves Mars Bonus for Elon Musk — 200M super-voting restricted shares tied to a $7.5T valuation and permanent Mars settlement with 1M+ residents. Additional rewards for 100 terawatts of space-based computing infrastructure.
PayPal Makes Venmo a Standalone Business Unit — As potential buyers circle, Venmo is being separated for easier tracking and possible sale. PayPal is recruiting a digital banking executive to run the new segment.
🔒 Security & Vulnerabilities
Shai-Hulud Malware Found in PyTorch Lightning AI Training Library — Malicious dependency discovered in the PyTorch Lightning library used for AI training, raising fresh concerns about supply chain attacks targeting ML infrastructure.
CopyFail Vulnerability Not Disclosed to Gentoo Developer — A security vulnerability was not properly disclosed, sparking debate about responsible disclosure practices in the open source ecosystem.
Spain’s Parliament Will Act Against LaLiga’s Massive IP Blockades — Congress moves to curb LaLiga’s aggressive IP blocking strategy, which has collateral-damaged legitimate services across Spain.
Calls Grow to Ban Palantir in Australia — After a UK MP described Palantir’s manifesto as “ramblings of a supervillain,” pressure mounts for Australia to halt new government contracts with the spy tech firm.
🔥 Hacker News
Claude Code Refuses Requests or Charges Extra If Commits Mention “OpenClaw” (462 comments) — Anthropic’s coding agent reportedly penalizes users whose commit messages reference the “OpenClaw” framework, sparking heated debate about model behavior control.
Belgium Stops Decommissioning Nuclear Power Plants (616 comments) — Belgium reverses its nuclear phase-out, a major policy shift as energy security concerns trump anti-nuclear sentiment.
Mozilla’s Opposition to Chrome’s Prompt API (203 comments) — Mozilla formally opposes Chrome’s proposed Prompt API, raising concerns about browser-based AI capabilities and standardization.
GCC 16 Has Been Released (50 comments) — The latest version of the GNU Compiler Collection ships with new features and optimizations.
Reverse Engineering With AI Unearths High-Severity GitHub Bug — AI-assisted reverse engineering discovers a high-severity vulnerability in GitHub, demonstrating the dual-use nature of AI security tooling.
🌐 World & Australia
AI Outperforms Doctors in Harvard Emergency Triage Trial — Researchers say the results mark a “profound change in technology that will reshape medicine,” with AI delivering superior diagnostic accuracy in emergency settings.
Colombia Climate Talks End with ‘Historic Breakthrough’ on Fossil Fuel Phaseout — Nearly 60 countries back voluntary roadmaps to wean the world off coal, oil, and gas at a conference prompted by frustration with UN climate summits.
Alice Springs Unrest After Arrest Over Child’s Death — Dozens gather outside hospital as police clash with rioters following the arrest of a man over the death of a five-year-old girl.
Has Gina Rinehart ‘Bought’ One Nation? — From luxury flights to meat pies, Australia’s richest person is bankrolling the rapidly expanding One Nation party. What it means for voters.
Meta Threatens to Shut Down Social Networks in New Mexico — Meta asserts in court filing that New Mexico’s proposed child safety remedies would be too onerous, threatening to pull its platforms from the state entirely.
Sources: TLDR (General + AI), AINews/Latent Space, The Guardian, Hacker News — 4 newsletters processed, ~45 stories distilled