OpenAI's compute bill hits $750B 💸, a third of MCP servers flunk agents 🚧, see a model think 🧠

AMD and Anthropic signed a tens-of-billions deal for AI servers, easing Anthropic's compute crunch while strengthening AMD against Nvidia. Anthropic will buy up to 2 gigawatts of AMD's latest Instinct MI450 chips from early 2027, and AMD invests up to $5 billion in return, its first cheque into the firm, as deployment milestones are met. A second credible silicon supplier hands model builders pricing power Nvidia has mostly dictated alone.

OpenAI said it will spend $750 billion on infrastructure through 2030, 25 per cent above its earlier estimate, even as Stargate appears stalled. First up is Project Camellia, a $20 billion campus near Savannah drawing at least 3.2 gigawatts from Georgia Power, most of it new natural gas, with OpenAI covering the full cost. That committed capex sets the bar every rival and every token price now has to clear.

Treasury secretary Scott Bessent said sanctions stay on the table after a White House official accused Moonshot of improperly distilling Anthropic's Fable. Science chief Michael Kratsios alleged Moonshot ran large-scale distillation and accessed banned Nvidia GB300 servers in Thailand, though experts dispute that Kimi K3 came mainly from distilling Fable, public only since 1 July. If Washington restricts Chinese open weights, teams leaning on cheaper Kimi-class models need a fallback.

OpenAI announced Presence, a product for deploying enterprise agents that answer questions, use company systems, take approved actions and escalate to a person. Each deployment starts with one job such as billing, the agent gets only the knowledge and access that job needs, the company sets the policies, and Codex proposes updates from production sessions. The pitch is control and auditability, aimed at the reliability gap that stalls agent pilots.

Travis Kalanick's robotics company Atoms raised $1.7 billion led by Andreessen Horowitz, with Ben Horowitz joining the board. Bain Capital, Fifth Wall and Uber joined, reconnecting Kalanick with the company that forced him out in 2017; Atoms builds on Cloud Kitchens and the acquired firm Pronto toward a 'wheelbase for robots'. A16z is placing one of its largest cheques on physical-world automation rather than another software agent play.

Alibaba debuted Qwen-Image-3.0, the third generation of its foundational image model, themed 'Real'. It accepts up to 4.5k-token prompts to render dense layouts like newspapers, storyboards and exam papers, keeps 10px text and LaTeX legible, renders pores and hair strands, and handles native text across 12 languages plus nested UI mockups. It is positioned as a deployable productivity tool for design, education and e-commerce rather than a novelty generator.

An engineer who integrates MCP connectors scanned 36 servers and found a third failing agents that pass every spec check. His mcpgrade scorecard, one npx command with no API key, graded 15 servers A and 11 D or F, firecrawl bottom on 134 errors, 132 undocumented parameters. The cause is schemas generated from zod or OpenAPI with no .describe() call, so counting undescribed parameters is the fastest reliability fix.

Calvin French-Owen calls Cloudflare Durable Objects, after months building on them, a near-perfect primitive for agents. Each object pairs a V8 isolate keyed by ID with its own SQLite instance, so you model a service as stateful objects driven by requests or timer alarms rather than services sharing a database. That per-object state and addressing maps onto long-running agent sessions, though he is candid about where DOs fall short.

Sierra published the diary behind its internal MCP gateway, giving employee agents safe access to the SaaS products and systems it runs on. One-and-a-half engineers built it after the team 'grabbed the lock' on connecting agents to data, solving permissioning and auditing once rather than having each team wire up its own integrations. The write-up maps the depth in a supposedly boring, mature protocol, worth reading before you build one.

Arize lays out how to check whether an LLM judge is trustworthy, since no single number answers it. It splits alignment into three questions: how reliably humans apply criteria, whether the LLM agrees as often as humans do, and what errors it makes, scored with chance-adjusted metrics like Cohen's kappa and a confusion matrix. In LLM-as-judge evals, it is a recipe to stop trusting a judge on vibes before scaling.

Antirez argues AI coding changes how software is distributed, not only how it is written, because recipients can now modify it themselves. The stable-versus-unstable branch model loosens when a 95 per cent-ready branch lets users adapt and specialise code for their own hardware, as with his Redis sorted-set work. For maintainers, a repository grows more valuable as a template others extend than a polished product covering every case.

A markets analyst frames token economics as the variable deciding whether the trillion-dollar AI buildout can pay for itself. A token is the unit of AI output, and while its price is commoditising, what funds capex is the dollar pool of tokens, price times volume, a race volume is winning. Value drains from anyone selling undifferentiated tokens toward the toll booths below, ASML and TSMC, and the frontier layers above.

Warp's Zach Lloyd argues the startups hitting $100M ARR in a year may be in trouble. Many scale by reselling inference at very low or negative margins, passing up to $90M of $100M revenue to Anthropic, OpenAI and model clouds, so net revenue is the honest measure. As tokens commoditise and enterprises bring their own inference, a thin wrapper around resold tokens has little defence against a cheaper rival.

Anthropic economist Peter McCrory laid out why AI has not yet moved the US labour market: no material rise in joblessness. June unemployment sat at 4.2 per cent, the Fed's full-employment level, and AI acts as a skill-biased, labour-augmenting technology that amplifies human expertise rather than replacing the worker. The hedge matters: models that soon build their own successors could break the pattern, so today's stability is no forecast.

WebMCP is a standard with everything except users: a page exposes typed functions agents call instead of scraping the DOM. It has Google and Microsoft editors, a Chrome origin trial through 156 and a waiting Lighthouse audit category, yet near-zero deployment and no mainstream agent using the tools. The 21 July draft moved the API from navigator to document, so early adopters carry migration debt for something nothing calls.

serve-sim, from Expo's Evan Bacon, streams an iOS Simulator to the browser so coding agents can drive a real device. A Swift helper captures the framebuffer via simctl io, exposes a 60 FPS MJPEG stream and WebSocket control channel, then serves a React preview, no Xcode plugin or app instrumentation. Test locally first, then tunnel the URL to drive an iPhone, iPad or Watch from anywhere, Apple Silicon only.

Subtext is an interpretability tool showing the words a 4B model leans toward before it emits output. It hooks the residual stream at nine layers and decodes each token's activations into vocabulary space via a Jacobian lens isolating what the model can report, cheap enough it streams at native speed. The result: the model often settles a verdict, judging '12 + 5 = 1' wrong, while still reading the prompt.

embeddinggemma.c runs EmbeddingGemma on CPU, Apple Metal, CUDA, ROCm or Intel SYCL from one native executable, the CPU binary about 100 KiB. It packs hand-tuned kernels per accelerator plus dynamic batching, singleflight deduplication and exact-result caching, with Matryoshka outputs at 768, 512, 256 or 128 dimensions. Measured against llama.cpp on an Apple M5 Max, it returned more embeddings per second in all 54 cells, geometric mean 1.25x, up to 2.01x.