Google rents GPUs from SpaceX 💰, smart TVs scrape the web 🧑‍💻, your AI vendor is your rival 💼

Google has signed a multi-year deal to buy compute from SpaceX for $920 million a month, starting in October and running to mid-2029. The agreement covers 110,000 Nvidia GPUs plus the CPUs, memory, and supporting hardware to run them. SpaceX's projected annual data-centre revenue from arrangements like this is now set to outstrip its combined income from Starlink, launches, and its other lines.

The Financial Times reports that the NSA is using Anthropic's Claude Mythos model for offensive cyber operations, with about half a dozen Anthropic engineers embedded inside the agency as forward-deployed staff. The arrangement is awkward: the same government designated Anthropic a supply-chain risk and the Pentagon blacklisted it earlier this year. It is unclear whether the engineers are helping with live hacking or only customising the model.

OpenAI has added Lockdown Mode, an optional setting that hardens ChatGPT against prompt-injection attacks by stripping away the features attackers tend to abuse. Turning it on limits some capabilities in exchange for tighter protection of sensitive data, a trade modelled loosely on Apple's phone-level Lockdown Mode. It signals that prompt injection has moved from a researcher's curiosity to a default threat OpenAI now builds protection against.

Nvidia has open-sourced Nemotron 3 Ultra, a 550-billion-parameter mixture-of-experts model with 55 billion active parameters, built for orchestrating long-running agents. The pitch is a system of models: a frontier reasoner plans while cheaper models handle high-volume tool calls and validation, holding down the token costs and goal drift that multi-agent workflows accumulate over many turns. Open weights mean you can wire it into your own stack.

A researcher used Anthropic's Opus 4.8 to surface a flaw in Zcash's Orchard shielded pool that had gone undetected by the world's best cryptographers since 2022. The bug could have been exploited to mint unlimited counterfeit ZEC without anyone noticing. After Shielded Labs disclosed and patched it, the token fell more than 40% in a day, shedding $3.8 billion in market value.

New York's legislature has passed a one-year moratorium on new large data centres, which would be the first statewide ban of its kind once Governor Hochul signs it. The bill targets facilities drawing 20 megawatts or more and adds mandatory public hearings. It lands as local opposition to AI data-centre buildouts hardens into actual legislation, giving anyone siting compute a real regulatory chokepoint to plan around.

A research team including Huawei says it used the company's Ascend 910C chips to complete post-training for DeepSeek-V4-Pro, a step China's chipmakers had struggled to reach. Inference on domestic silicon was already working; training and refining a model's behaviour is the harder problem the sanctions were meant to block. Clearing it moves China closer to building frontier models without Nvidia hardware in the loop.

Film and TV actors voted to ratify a four-year contract with the studios, with more than 90% in favour. The headline protection: a synthetic AI performer must bring significant additional value over a live actor before a studio can use one, a bar that narrows where AI doubles are commercially viable. Like the writers' contract in April, it runs four years instead of three for added stability.

Include Security documents how Bright Data turns ordinary consumer devices into a residential proxy network of 400 million home IPs that AI companies route scraping traffic through. The supply comes from an SDK embedded in apps which, with consent buried in the terms, turns a phone or smart TV into an exit node. It is a clear look at the plumbing behind where AI training data quietly comes from.

The operator of an on-chain analytics platform pointed Claude Code at a Postgres database of 1.3 billion Polymarket trades using the community Postgres MCP, then asked questions in plain English. Claude wrote the aggregates, knew to hit the pre-aggregated stats table instead of scanning every row, and handled the data quirks. The headline result: only about one in five wallets ever turns a profit.

Simon Willison has built micropython-wasm, an alpha package that runs untrusted Python inside a WebAssembly sandbox, plus a Datasette plugin on top of it. Plugin systems are his favourite way to extend software cheaply, but they normally run plugin code with full privileges. Compiling MicroPython to WASM gives him a way to execute agent-written or third-party code with the isolation that full-fat Python plugins never had.

A hackathon field report on running each agent in a simulated market on a different lab's small model: gpt-oss-20b, MiniCPM3-4B, Nemotron-Mini-4B, and a fine-tuned Qwen 0.5B. You play a shadow financier, lending, shorting, and planting tips while a magistrate hunts you. The engineering lesson is that the friction sat almost entirely at the serving layer, in CUDA images and per-model config, not in the modelling.

Tired of copy-pasting files into a chat, a developer set out to give the AI direct access to a project folder, then found the existing MCP implementations all wanted FastAPI, LangChain, or the official SDK. Reading the spec, they realised MCP is just JSON-RPC over stdio or HTTP, which Python's standard library already covers. The result is a server with no dependencies, both transports, fifty tests, and no pip installs.

Prompted by the Mythos export-control debate, a researcher benchmarked open-weight models against Opus on genuine vulnerability research, not a toy app. The result: open models trail on the harder artefacts, but a good harness closes most of the gap, and post-training matters more than architecture. GLM-5.1 stood out, matching Opus across the board. The policy point: open weights set the floor, not the ceiling, of what attackers can reach.

Within a single fortnight, Anthropic, OpenAI, and Google all moved into consulting. Anthropic formed a $1.5 billion joint venture with Blackstone and Goldman to embed engineers inside mid-sized companies; OpenAI stood up a $4 billion Deployment Company and bought consultancy Tomoro for 150 forward-deployed engineers; Google committed $750 million to Accenture, Deloitte, and McKinsey. The platforms you build on now compete for the same services revenue you do.

You can write a clean RL environment with solid rubrics and still train a model nobody wants, if the tasks don't map to work people actually pay for. The argument: ground task design in Bureau of Labor Statistics data, and target the $200-an-hour part of a job rather than the $15 part. A model that handles the high-value judgment is worth far more than one that automates the cheap busywork.

A practical guide to outcome-based pricing for AI products, built around twelve design decisions: defining the outcome, proving it happened, deciding what does not count, and making the invoice auditable. The sharp line it draws is between real outcome pricing and usage billing dressed up with better marketing. Sierra is the example, charging for resolved conversations and saved cancellations, with no charge when the agent fails to resolve.

Before a demo gets booked or a form gets filled, a buyer opens ChatGPT or Perplexity and gets a ready-made shortlist of tools in your category. New data finds nearly 90% of the AI citations shaping that shortlist come from off-site sources, not your own pages, and 71% of buyers now lean on AI chatbots for software research. The selection happens before anyone reaches your homepage.

Microsoft has open-sourced pg_durable, which brings durable execution into Postgres so long-running, fault-tolerant workflows live next to the data they touch. Define the workflow in SQL, let it checkpoint each step, and resume cleanly after crashes, restarts, or failed steps, instead of stitching together cron jobs, workers, queues, and status tables. It runs inside Azure HorizonDB, Microsoft's new Postgres cloud service, and targets backend engineers, DBAs, and SREs.

Coral turns any API, database, or file system into a read-only SQL schema, so you can JOIN PagerDuty incidents against GitHub pull requests, or Slack messages against Linear issues, in a single query with no custom integrations or ETL. It handles auth, pagination, rate limits, and schema mapping. Point it at your sources and Claude Code pulls data at, the makers claim, higher accuracy and lower cost than hand-written glue.

office-open-xml-viewer renders DOCX, XLSX, and PPTX files faithfully in the browser, with parsers written in Rust and compiled to WebAssembly and renderers using the Canvas 2D API. Each format also exposes a headless engine, so you can build your own scroll views, thumbnail grids, or master-detail panes instead of being locked into the bundled viewer. The whole codebase was written by Claude through iterative prompting, with no human-authored application code.

Kyushu is an open-source CLI that lets you write a JavaScript or TypeScript fetch handler, compile it to a self-contained WebAssembly binary, and run it anywhere with a single command. The pitch is Cloudflare Workers without Cloudflare: a familiar handler API and a sandbox isolated from the host, with no Node, Bun, or Docker required. It is an early experiment that expects breaking changes.