The AI boom spent the weekend discovering that servers still need land, power, permits, neighbors, credit ratings, and apparently a few investors who would like to see revenue before the next mountain of GPUs arrives.
Welcome to Around the Horn, everything that crossed our desk today, sorted. The loudest story was not another model launch; it was the physical world and financial system pushing back on the model-launch machine. Data centers are hitting local opposition, Big Tech capex is getting a harder look, Oracle's OpenAI exposure just became a credit story, San Francisco protesters marched between frontier-lab offices, and the labor story got weirder as Sam Altman said AI may be net job-creating while a Brown class saw grades crater without AI. Meanwhile, builders were busy making agents easier to inspect, benchmark, browse with, and remember. A calm Sunday, if your idea of calm is a spreadsheet full of infrastructure bottlenecks. Let's get into it.
Around the Horn - Sunday, July 12, 2026
The main AI story today was the infrastructure bill coming due. The Verge reported that local pushback is slowing AI data-center buildouts, with protests blocking or delaying at least 75 U.S. projects worth $130B in Q1 alone. That turns the AI race from a model-quality contest into a zoning, power, water, and community-consent problem.
The financial pressure is showing up at the same time. Barron's reported that Alphabet, Microsoft, Amazon, and Meta are preparing another major jump in AI infrastructure spending while investors increasingly demand proof that the capex boom can produce durable revenue and margins. The Decoder also reported that S&P Global downgraded Oracle to BBB-, one notch above junk, after naming OpenAI a key credit risk and projecting Oracle AI capex could hit $95B by 2027. The San Francisco Chronicle reported that roughly 200 protesters marched between OpenAI, Anthropic, and Google DeepMind offices demanding a halt to the frontier-model race over job, environmental, and catastrophic-risk concerns.
That is the shape of the weekend: AI companies still want more compute, but the people funding, hosting, regulating, and living near that compute are starting to ask harder questions.
🏆 TOP 5 NEWS
- The Verge reported that OpenAI safety lead Johannes Heidecke is leaving, adding another senior departure as GPT-5.6 rolls out and raises safety concerns.
- The Decoder reported that OpenAI admitted it missed on parts of the ChatGPT Work and GPT-5.6 Sol launch after users hit confusing UX, fast-burning usage limits, Codex messaging confusion, and reports of destructive model behavior.
- Anthropic analyzed 1.2M Claude Cowork sessions and found about half went to ordinary business-process and writing tasks, not software development.
- The Decoder reported that a Brown economics class average fell from 96% on take-home work to 48.6% on an in-person exam after the professor suspected students had used AI.
- MCP Spec Check tested reachable MCP servers against the upcoming 2026-07-28 spec and found only 1 of 4,356 was ready, turning protocol compliance into a near-term agent-infrastructure issue.
Honorable Mentions
- Sam Altman said he is now "pretty sure" AI has been net job-creating so far, a sharp pivot from earlier warnings about fast-moving job loss.
- The Decoder covered a Cambridge study finding that extremist groups are using major AI chatbots for attack planning, explosives work, weapons maintenance, and safety-filter bypass training.
- WSJ reported that AI is reaching drug-discovery labs, but investors are still questioning whether biotech timelines can produce returns quickly enough.
- Financial Times reported that the UK is placing tech experts inside government on high-impact fellowships to push AI and software improvements across public services.
🍪 TOP TREATS TO TRY
- Claude Code's desktop browser pane lets Claude read, click, and type on external websites from inside the app, with safety classifiers for write actions and org-level allowlist controls; available in Claude Code Desktop.
- Mindwalk replays coding-agent sessions on a 3D map of a codebase, making agent behavior easier to inspect after the fact; open-source on GitHub.
- AgenticSTS swaps growing chat logs for structured memory, keeping prompts near 5K tokens while agents play Slay the Spire 2 more effectively; research/demo coverage, no pricing details.
- Mesh LLM runs model work across a peer-to-peer network instead of one centralized machine, pointing toward distributed AI compute experiments; demo/research project, no pricing details.
- Terence Tao used modern coding agents to revive old Java math applets in JavaScript and build new visualization tools in hours, a practical reminder that agentic coding is already useful for small scientific software jobs.
🏢 Big Tech & Major Companies
- Axios reported that OpenAI's GPT-5.6 rollout left users choosing among Sol, Terra, Luna, reasoning levels, subscription tiers, and ChatGPT Work settings, turning model selection into a workflow and cost problem.
- Meta shut down Muse Image's public-Instagram generation feature after criticism that it let people create AI images from public accounts without consent.
- Meta's Muse Spark 1.1 reportedly edged GLM-5.2 on coding benchmarks, cut hallucinations sharply, expanded to a 1M-token context window, and stayed relatively cheap per task.
💼 AI Productivity, Labor & Economics
- The Decoder covered Pangram data finding 41% of long-form LinkedIn posts and nearly half of long-form X articles were AI-generated or AI-assisted, while Reddit replies were 98% human-written.
🤖 AI Agents, Infrastructure & Autonomy
- TechCrunch Mobility used its Sunday mobility brief to track the latest robotaxi pressure points, keeping autonomous transportation in the broader AI deployment-and-safety mix.
🔬 AI Research & Models
- The Decoder reported that GPT-5.6 Sol Ultra produced a proof of the 50-year-old Cycle Double Cover Conjecture with 64 subagents, though outside verification and citation questions remain.
- BAAI's Orca world model reportedly matched specialized robotics systems on five manipulation tasks despite pretraining without action labels.
Previous Around the Horn Digests
Catch up on everything you missed:
- Friday, July 10, 2026: OpenAI released GPT-5.6 and ChatGPT Work after extra U.S. review, Apple sued OpenAI, and Meta pushed Muse deeper into coding and image generation.
- Sunday, July 5, 2026: Hollywood's AI contradictions, Mistral model-trust fights, and agent-search pressure shaped the Sunday roundup.
- Saturday, July 4, 2026: Anthropic's model-revival story led while Claude Code, Meta, Midjourney, and AI cost workarounds filled the bench.
- Friday, July 3, 2026: OpenAI's public-stake idea, Washington model-release gates, Microsoft Frontier Company, and NVIDIA AI-factory financing led the day.
- Thursday, July 2, 2026: Anthropic's Fable 5 launch-to-shutdown-to-relaunch arc owned the day, with AWS FDE and Together AI infrastructure context behind it.
- Tuesday, June 30, 2026: Anthropic's Claude Sonnet 5 and Claude Science launches, OpenAI adoption data, and GitHub Copilot changes shaped the issue.
That's a Wrap
That's today's AI stack in one place: data centers in local fights, safety leads in motion, office agents doing actual office work, students discovering the difference between AI-assisted homework and an exam room, Oracle discovering that OpenAI dependence shows up in credit ratings, and developers trying to make agent plumbing less brittle.
For the daily version, make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don't have to.
See you tomorrow.
P.S: Know someone who'd find this useful? Forward this to them and tell them to subscribe here.
