📅 Tuesday, July 14th 2026 · ⏱️ Read time: 5 min · 🔗 Issue No. 17
Long time, no loop. It's been roughly a year since our last issue, and a lot has changed — so we're back to doing what we do best: cutting the AI noise down to what actually matters. Glad you're still with us.
You're in the loop — five of the biggest names in enterprise software — Google, Microsoft, Salesforce, Snowflake and ServiceNow — just lined up behind a shared agent protocol built to blunt Anthropic and OpenAI's grip on how AI connects to your data. Meanwhile, a courtroom spat went personal: two days after Apple sued OpenAI for allegedly poaching its hardware secrets, Elon Musk and Sam Altman spent the weekend trading shots on X, turning a legal filing into the industry's loudest feud.
Today: the Big Tech protocol bloc, Apple's suit-turned-brawl, and Wall Street's pick of Chinese models — plus why the physical world is catching up with the AI buildout, how to cut your API bill by routing models, five fresh tools, and a prompt that pressure-tests your plans.
🔁 The Loop
Big Tech draws battle lines over AI's plumbing
Editorial illustration — rival software giants wiring their agents into one shared standard. Original art, no logos.
Google and Microsoft rally a rival to Anthropic's agent standard. Google, Microsoft, Salesforce, Snowflake and ServiceNow have agreed to back a shared protocol for how AI agents plug into enterprise data and tools, The Information reports — a direct shot at Anthropic's Model Context Protocol, the de facto standard for the past 18 months, and at OpenAI. The twist: all four labs also sit inside the Linux Foundation's new Agentic AI Foundation, so they're cooperating on open standards and knife-fighting over them at once. Watch whether a five-company committee can ship something coherent before MCP's head start hardens into a moat. See who signed on.
Apple sues OpenAI — and Musk piles on. Apple hauled OpenAI into federal court, alleging a coordinated campaign to lift its hardware trade secrets "at every level," from staff engineers to its chief hardware officer, tied to the 400-plus former Apple employees now at OpenAI. Within 48 hours the filing became a spectacle: Musk amplified it and needled OpenAI's hiring, Altman fired back, and the two spent the weekend brawling on X. OpenAI says it has "no interest in other companies' trade secrets." Read the complaint.
Goldman tells clients which Chinese models to buy. Goldman Sachs named DeepSeek and ByteDance its top picks among China's AI labs, per CNBC — a milestone in Chinese open-weight models going mainstream in Western stacks. The math is blunt: open Chinese models run 60–90% cheaper than OpenAI and Anthropic's flagships, and Goldman projects Chinese model API and subscription revenue leaping from roughly 35 billion yuan this year to 879 billion by 2030. See the picks.
Unlocked reads: how Chinese models quietly took 30–46% of US enterprise tokens, and Brookings on why the data-center fight is really a fight over power.
🌊 Deep Current
The AI buildout is running into the physical world
Editorial illustration — an AI's ambitions meet power lines, water, and the neighbors. Original art, no logos.
The map fills with red. At least 75 data-center projects worth roughly $130 billion were blocked or delayed in the US in the first quarter of 2026 — about as many as in all of 2025 — and the number of local groups organizing against them has climbed past 430 across 40-plus states. This is no longer NIMBY noise; it's the first hard constraint on a buildout that assumed land, power and permits would keep flowing.
It's a pocketbook story now. The country's 4,000-plus data centers already draw about 6% of US electricity, up from 4% two years ago, and in dense markets like Virginia residential power prices have jumped 267% in five years. A Gallup poll found 71% of voters — 75% of Democrats and 63% of Republicans — don't want one built nearby. When a fight crosses party lines and lands on the electric bill, it stops being ignorable.
Model quality improves every month — but transmission lines take a decade, and community consent can't be A/B tested.
The other ledger. The same boom is distorting the world beyond the substation. Reporting from the Bay Area describes homes selling millions over asking, some paid via stock swaps by AI employees, even as communities elsewhere push back on water use and grid strain. The industry's reflex — build bigger, faster — collides with a resource it can't provision on a GPU roadmap: local goodwill.
The bottom line. If you're betting on cheap, abundant AI, the constraint to watch isn't chips — it's kilowatts and consent. The labs that learn to be good neighbors, buying power without spiking residential rates and paying for water infrastructure, will quietly out-build the ones that treat counties as obstacles. Expect "community strategy" to become as core to an AI roadmap as model training by year's end.
🛠️ The Workbench
Cut your AI bill by routing models to the task
Frontier models are overkill for most calls. Here's how to send the grunt work to cheap models and save the expensive ones for what matters — the exact arbitrage Goldman just blessed.
List your AI tasks and tag each as high-stakes (customer-facing, legal, code that ships) or commodity (summaries, tagging, first drafts, extraction).
Sign up for a model router that exposes many models behind one API, so you can switch per call without rebuilding integrations.
Point commodity tasks at a cheap open-weight tier (a DeepSeek or Qwen model at cents per million tokens) and keep a frontier model on the high-stakes lane.
Run 20–50 real examples through both tiers and eyeball the cheap output — accept it only where quality actually holds.
Add a fallback rule: if a cheap call fails a confidence or format check, auto-escalate that one request to the frontier model.
Sample Prompt: "You are a routing assistant. Given this task description and a sample input, tell me the cheapest model tier likely to handle it at acceptable quality, and name one failure mode I should test before switching."
🗣️ Overheard
What the timeline's buzzing about
🇨🇳 Companion curfew: China's new anthropomorphic-AI rules take effect July 15, and ByteDance's Doubao and Alibaba's Qwen are already disabling user-made companion agents ahead of the deadline.
🏛️ Fed goes to the Valley: The Federal Reserve tapped a16z's Marc Andreessen to co-lead a new task force on AI, productivity and jobs.
🔌 Machines with wallets: Cloudflare opened a waitlist for a gateway that charges AI agents per request over the x402 protocol, settled in stablecoins.
🧮 Proof, not vibes: Mistral open-sourced Leanstral 1.5, a model that writes Lean 4 proofs for its own code — topping Claude Sonnet on a verification benchmark at a fraction of the cost.
🎬 Fix it in one frame: A workflow pairing GPT Image 2 with Runway's Aleph 2.0 lets editors change a single frame and propagate the edit across a whole clip.
🔎 Fresh Finds
Five tools worth a look
🎙️ Bono AI: Record or speak once and it publishes formatted content across every platform.
🍋 LemonLime: Automate an existing workflow from a single plain-English prompt.
🛠️ Aura: An open-source IDE for supervising AI coding agents with built-in review loops.
🧩 Sim: An open-source workspace for building and running AI agents and workflows.
💬 Scarlett: An AI co-worker that lives in Slack and iMessage and runs tasks on request.
★ = sponsored placement, if any.
🧪 Prompt Lab
The "devil's advocate" plan review
Paste any plan or decision and let the model argue against it before reality does — great for a launch plan, a hiring call, or a big vendor commitment.
You are my sharpest, most skeptical advisor. Here is a plan I'm about to commit to: [PASTE PLAN]. 1) Restate what I'm actually betting on in one sentence. 2) List the three assumptions that, if wrong, break the plan — ranked by damage. 3) For each, give the strongest real-world reason it might be wrong and the earliest warning sign I'd see. 4) Name one cheaper or reversible test I could run this week before committing fully. Be blunt. I want the case against, not reassurance.
Want art to go with it? Try this named style:
A single tightrope-walking robot balancing a glowing decision-orb, modern gouache illustration, deep indigo and warm amber on an off-white background, soft paper-grain texture, generous negative space; no text, no words, no logos.
⏪ Rewind
Yesterday’s most-opened link
Readers kept clicking our breakdown of why Grok 4.5's $2/$6 pricing reset the value floor for every frontier model — and what it means for your stack. Catch up here.
Stay in the loop — the InTheLoop team