In partnership with
In Todayโs Issue:
๐ฐ Alphabet's record $205B AI bill spooks the market
๐ซ 200 startups fight a ban on China's open models
๐ Anthropic opens its AI-economy data to everyone
๐ค NVIDIA puts an open world model on the robot
๐งฌ The first AI-designed drug reaches Phase III
โจ And more AI goodnessโฆ
โก The Signal
AI's sharpest fights right now are about control: who gets to spend, who gets to restrict, and who gets to measure the technology.
Today's stories all circle that theme. Alphabet is pouring up to $205B into AI infrastructure and getting punished for it; nearly 200 U.S. startups are begging Washington not to wall them off from China's best open models; and Anthropic just turned its map of who-actually-uses-AI into a public tool. Underneath the politics, the science keeps compounding: an AI-designed drug entered late-stage trials this month, and DeepMind's Demis Hassabis is still betting AI helps cure disease within a decade. The frontier labs solved capability faster than anyone solved distribution, trust, and policy. That gap is now the story.
All the best,
Kim Isenberg
Whoโs actually reading this?
Weโre planning next yearโs coverage and building it around you.
Three taps, no typing.
Why it matters: we need to know who our readers are; itโs the first question any serious sponsor asks and we canโt answer it right now. It also gives us a real click in every issue, which helps our numbers.
(Alphabet)
๐ฐ Alphabet's AI Bill Balloons to $205 Billion
Alphabet beat on almost every line, and investors sold it off anyway. Q2 revenue rose 24% to $119.8B, Google Cloud jumped 82% to $24.77B, and the Gemini app hit 950M monthly users. But the stock slid about 5% after hours once the company lifted 2026 capital spending to $195B to $205B (from $180B to $190B), a record AI-infrastructure bet the CFO tied to "an acceleration in the delivery of capacity to meet growing demand."
๐ tl;dr: Google's AI business is booming, but the market now cares more about the capex than the earnings beat.
(Anthropic)
๐ Anthropic Turns Its AI-Economy Data Into a Chatbot
Anthropic released a free connector that lets anyone query the Anthropic Economic Index in plain language inside Claude. The Index tracks how people actually use Claude across the economy, and it still skews heavily technical: "computer and mathematical" work makes up 37.2% of Claude conversations versus just 3.4% of U.S. workers. Anthropic cautions the data reflects Claude usage, not the whole labor market.
๐ tl;dr: You can now ask Claude which jobs lean on AI most and get Anthropic's real usage numbers back.
(NVIDIA)
๐ค NVIDIA Shrinks a "World Model" Onto the Robot Itself
NVIDIA released Cosmos 3 Edge, a 4-billion-parameter open "world model" (an AI that predicts how a physical scene will unfold) small enough to run on the robot instead of in a data center. On NVIDIA's Jetson Thor chip it reasons about its surroundings and outputs control actions in real time, 15 times per second and 32 actions per inference, and it ranks #1 among 4B-parameter models on the VANTAGE-Bench vision benchmark. The weights are open on Hugging Face, aimed at factories, warehouses and hospitals.
๐ tl;dr: Physical AI is getting small and open enough to run onboard the robot itself.
๐ฌ Watch This
In her latest episode, Dr. Rhonda Patrick sits down with immunologist Dr. Derya Unutmaz to make the case that the next decade could be the most consequential longevity window in human history. His forecast is bold: longevity escape velocity within 8 to 10 years (the point where medicine adds more than a year of life expectancy for every year that passes), and the possibility of real age reversal within fifteen. A grounded, mechanism-heavy conversation on where AI-driven biology is actually pushing human healthspan.
โ Demis Hassabis, CEO of Google DeepMind, on CBS's 60 Minutes
๐ OpenAI's Tibo Sottiaux is teasing that something big and "codexy" is about to drop. The rumour mill says it points to GPT-5.6 Sol going live on Cerebras at roughly 750 tokens per second, fast enough to make agentic coding feel close to real-time.
(Cerebras)
Silicon Valley to Trump: Banning China's Open Models Would Hurt America Most
The Takeaway
๐ Nearly 200 U.S. startups, including Replit, Proton and Y Combinator, signed letters urging the Trump administration not to cut off access to Chinese open-weight AI models.
๐ A new trade group, the Little Tech Association, sent the letters to President Trump and Commerce Secretary Howard Lutnick on Wednesday.
๐ The trigger: reports the White House may restrict Chinese models after Moonshot's Kimi K3, a 2.8-trillion-parameter open-weight model, launched on July 16.
๐ Founders argue a ban is unenforceable, since the weights are already downloaded worldwide, and would mostly kill U.S. companies built on them.
The fight over Chinese AI has turned inward, and its loudest new combatants are American startups afraid of being cut off from the best open models. The Little Tech Association, a new Silicon Valley trade group, sent letters on Wednesday to President Trump, Commerce Secretary Howard Lutnick and other officials, warning against any move to block U.S. developers from downloading Chinese open-weight models. Nearly 200 companies signed on, among them Replit, Proton and startups backed by Y Combinator. Their core line: "American leadership requires two things: world-leading American open-weight models and continued access for U.S. builders to open models already available worldwide."
(via TechStartups)
The pressure answers reports that the administration is weighing new curbs after Moonshot AI released Kimi K3, a 2.8-trillion-parameter system it bills as the largest open-weight model yet, on July 16, with full weights due July 27. Chinese families like DeepSeek, Qwen, Kimi and GLM have already overtaken every other country in Hugging Face downloads, and plenty of U.S. products quietly run on them because they are cheap, capable and easy to modify.
(UPI)
The founders' warning is blunt. "There'll be hundreds of companies that instantly die," said Suhail Doshi, founder of AI-infrastructure startup Particle and a member of the group. Their case rests on enforceability: once open weights are public, a U.S. download ban does little to slow China but a lot to hobble American builders who already wired those models into their products.
The administration's rationale is security: officials worry Chinese models could carry hidden vulnerabilities or deepen U.S. reliance on a strategic rival. Critics counter that open weights can be inspected and run locally, which is arguably more auditable than a closed American API, and that walling off the ecosystem hands the cheap layer of AI to China by default.
Why it matters: A ban aimed at China would mostly bite the U.S. startups the White House says it wants to protect, and it cannot pull weights that are already public back off the internet.
Sources:
๐ https://www.politico.com/news/2026/07/22/startup-founders-urge-trump-not-to-shut-off-chinese-open-weight-ai-01008992
๐ ChinaTechNews: Nearly 200 startups urge Trump not to ban Chinese AI models
๐ Tom's Hardware: Trump reviving push to ban Chinese AI models after Kimi K3
The New Blueprint for AI Powered Support
Most support teams are experimenting with AI. Few are transforming because of it. The gap is where the most consequential decisions in support are being made, and where most teams get stuck. Hear from industry leaders on moving from pilot to production.
(Sherwood / FactSet)
The chart: Sherwood's standardized quarterly capital expenditure for the four U.S. hyperscalers, Amazon, Alphabet, Microsoft and Meta, from 2015 through Q4 2025. Combined spending has climbed from roughly $5B a quarter a decade ago to about $120B, with Alphabet the second-largest slice.
The lesson: The AI build-out now dominates Big Tech's balance sheets. Alphabet just raised its 2026 capex guidance to $195B to $205B (from $180B to $190B) and its stock still fell about 5%, a sign investors are shifting from cheering the scale of the spend to asking when it pays off.
The caveat: These are combined quarterly figures through Q4 2025 and exclude finance leases; Alphabet's new 2026 guidance would push its share higher still. Capex is money committed up front, and none of these bars show the returns it eventually earns.
๐งฌ The First AI-Designed Drug Reaches Late-Stage Trials
โก Bottom line: Insilico Medicine just moved Rentosertib, a drug its AI both discovered and designed, into a 320-patient Phase III trial.
๐ก Why it matters: It is the strongest sign yet that generative AI can design real drugs, moving past the protein-structure predictions it became known for.
๐ What it means: If it holds up, AI could compress the slowest, most expensive stretch of drug discovery for diseases of aging.
Idiopathic pulmonary fibrosis (IPF) is a brutal, age-related disease in which the lungs slowly harden into scar tissue until breathing fails. There is no cure and few good drugs. Earlier this month, Hong Kong-based Insilico Medicine began a Phase III trial for Rentosertib, an oral IPF drug, and what makes it a landmark is where the molecule came from: Insilico's AI both found the biological target and designed the compound to hit it.
(Insilico Medicine)
The target is a protein called TNIK, which Insilico's "Pharma.AI" platform surfaced by mining the biology of aging, before generative chemistry drew a molecule to block it. The Phase III study is a randomized, double-blind, placebo-controlled trial in 320 IPF patients over 40, run across 52 weeks. In earlier Phase IIa testing, patients on a 60 mg daily dose showed improved lung function versus placebo.
The bigger bet is Insilico's thesis that aging biology can be a discovery engine: TNIK sits on pathways tied to inflammation, fibrosis and cellular aging, so a drug that works here could point toward other age-related conditions. It is early, most Phase III drugs still fail, and one trial does not validate "AI drug discovery." But it is among the first medicines with an AI-discovered target and an AI-designed molecule to reach the last stage before approval.
8 AI positions. +85% average gains.
Over 19 months: Sandisk +196%, Micron +150%, Western Digital +129%, Nvidia +69% โ average +85% across 8 positions. The same founder now writes MavSource's daily AI brief: newsletters, podcasts, labs, earnings, investor letters โ synthesized in 5 minutes. Free.
Past performance does not indicate future results. MavSource is informational only, not investment advice.