In partnership with
In Today’s Issue:
🌶️ OpenAI's first custom chip, built with Broadcom
🧠 Google loses two more AI stars to Anthropic
🌍 The EU and allies join the US "Pax Silica" pact
🫁 Stripe, Anthropic and OpenAI fund a war on colds
✨ And more AI goodness…
⚡ The Signal
OpenAI just stopped being only a software company. On Tuesday it unveiled Jalapeño, its first custom silicon: an inference chip co-designed with Broadcom and taken from design to tape-out in about nine months. The reveal lands in the same week that Google watched two more Gemini researchers leave for Anthropic, and that the US widened its "Pax Silica" pact to wire allied chip supply chains around China. Read together, the day points at one thing: control of the AI stack. Compute, talent, and supply lines are the three levers that decide who leads, and right now every major player would rather own them than rent them.
All the best,
Kim Isenberg
(Tasos Katopodis/Getty Images via FT)
🌍 The US Builds an Allied Chip Bloc to Box Out China
The US-led "Pax Silica" pact just grew to 24 countries as the EU, Germany, the Netherlands and Greece signed on to pull AI supply chains away from China. Architect Jacob Helberg, the State Department's under-secretary for economic affairs, told the FT the group will span chips, critical minerals and energy, with Argentina, Chile, Costa Rica, Kazakhstan and Panama joining this week. He pitched it as a counter to both China's Belt and Road and the UN's "digital sovereignty" push, warning that fragmented rules risk "synchronised mediocrity." Washington also plans a Stanford manufacturing-curriculum deal and a new economic-security zone in Kazakhstan.
👉 tl;dr: The AI cold war now has a formal alliance, and the West wants its own end-to-end chip pipeline.
(Damian Lemanski/Bloomberg)
🧠 Two More Google AI Stars Defect to Anthropic
Google is set to lose two more key Gemini researchers, Jonas Adler and Alexander Pritzel, to Anthropic, Bloomberg reports. The pair worked on Google's AI coding push and model training, and will join Nobel laureate John Jumper, who is also leaving DeepMind for Anthropic; star researcher Noam Shazeer recently jumped to OpenAI. The drain is fueled by pre-IPO equity: Anthropic, freshly valued at $965 billion and weighing a public listing as soon as this fall, can dangle a rare payday. Per a 2025 SignalFire analysis, DeepMind staff are nearly 11 times more likely to leave for Anthropic than the reverse.
👉 tl;dr: The model race has become a talent race, and Anthropic's IPO upside is bleeding Google's bench.
(MIT Technology Review)
🫁 Stripe, Anthropic and OpenAI Declare War on the Common Cold
A new $500 million nonprofit called Intercept wants to wipe out respiratory infections, from the common cold to the flu. Funded by Stripe and backed by Anthropic, the OpenAI Foundation, Bill Gates, Flu Lab and Jane Street traders, it pairs next-generation vaccines with large-scale air cleaning for public spaces. The AI angle runs through the Arc Institute, co-funded by the Collison brothers, whose AI models for biology are speeding up vaccine and protein design.
👉 tl;dr: AI's biology tools are now aimed at a problem everyone has: never catching a cold again.
Use AI to pressure-test a big decision before you commit to it.
Why it helps: This week's news is full of expensive bets, custom chips, poached talent, new alliances. The same logic scales down. Before a hire, a tool purchase, or a strategy pivot, have a model argue the other side so you see the failure modes first.
Try this: "Act as a skeptical board member. Here is a decision I am about to make: [describe it in 3 to 4 sentences]. Give me the three strongest reasons it fails, the second-order effects I am not seeing, and one cheaper experiment that would tell me whether I am wrong before I spend real money."
🎬 Watch This
Bloomberg's Emily Chang sits down with Nobel laureate Jennifer Doudna, the Berkeley biochemist who co-invented CRISPR, for the latest episode of The Circuit. They trace how gene editing went from a lab curiosity to a medical platform, and where Doudna thinks Silicon Valley keeps misreading biology, including the limits of treating living cells like code. It is a sharp, accessible counterpoint to this week's AI-meets-biology headlines, from the Intercept vaccine push to AI-driven longevity.
Inference is thinking. Thinking is way harder than reading.
Claude Fable 5 appears to be re-emerging in AWS Bedrock documentation. AWS model cards now list the lifecycle as Active, with Bedrock IDs including anthropic.claude-fable-5 and global.anthropic.claude-fable-5 . The data-retention docs also show it as status: "available" when provider_data_share is enabled. This does not yet confirm that Fable 5 is fully back online, but it is the clearest sign so far that access may be returning.
(Anthropic)
Jalapeño: OpenAI Builds Its Own AI Chip
The Takeaway
👉 OpenAI unveiled Jalapeño, its first custom inference chip (the hardware that runs a trained model for users), co-designed with Broadcom.
👉 Early engineering samples already run real workloads in the lab, including OpenAI's GPT-5.3-Codex-Spark, at production frequency and power.
👉 OpenAI says performance per watt is "substantially better" than today's best chips, the number that sets the cost of every answer.
👉 Design to tape-out (a chip's final pre-manufacturing milestone) took about nine months, with first deployments targeted for end of 2026.
OpenAI has started building the chips it used to only buy. On Tuesday the company and Broadcom unveiled Jalapeño, OpenAI's first in-house chip and the opening piece of a multi-generation hardware platform the two will build together. Unlike the GPUs that dominate AI today, Jalapeño is an inference chip, purpose-built to run finished models for users rather than to train new ones, the workload that now eats most of OpenAI's compute bill.
(OpenAI / Broadcom)
The pitch is efficiency. OpenAI says early Jalapeño samples are already running machine-learning workloads in the lab, including its GPT-5.3-Codex-Spark model, at the frequency and power it expects in production, and that the chip's performance per watt is "substantially better" than the current state of the art. President Greg Brockman said the team built it around jobs general-purpose hardware handles poorly: "We have a deep understanding of the workload. We've really been looking for specific workloads that are underserved." OpenAI even used its own AI models to help design the chip and compress the schedule.
The other headline is speed. The companies took Jalapeño from first design to tape-out in roughly nine months, which they call one of the fastest cycles ever for a high-performance chip. First silicon is slated to deploy by the end of 2026 and scale from there, part of a plan to stand up gigawatt-scale data centers with Microsoft and other partners.
(Stargate data-center sites; satellite imagery via Epoch AI)
Why it matters: Custom inference silicon is how OpenAI attacks its single biggest cost and loosens its dependence on NVIDIA. If Jalapeño's efficiency claims hold, OpenAI can serve more users per watt and per dollar, a structural edge in a market where the price of a token, not the size of a model, increasingly decides who wins.
Global HR shouldn't require five tools per country
Your company going global shouldn’t mean endless headaches. Deel’s free guide shows you how to unify payroll, onboarding, and compliance across every country you operate in. No more juggling separate systems for the US, Europe, and APAC. No more Slack messages filling gaps. Just one consolidated approach that scales.
The chart: Artificial Analysis plots its AA-Briefcase Elo (an agentic-task quality score, higher is better) against wall-clock time per task in minutes (answer plus reasoning plus tool use, lower is better). Claude Opus 4.8 (max) tops the quality axis but is the slowest at 22.6 minutes per task; GPT-5.5 (xhigh) sits near the efficient frontier at 10.8 minutes, while lightweight gpt-oss-20B finishes in 2.1 minutes at far lower Elo.
The lesson: Raw scores are only half the picture now; the real axis is quality per minute. Top reasoning models buy their extra capability with time, so the "best" model now depends on how long you can afford to wait for each answer.
The caveat: Elo is a relative ranking on one benchmark (Briefcase), and time-per-task swings with the test harness, tool latency, and each model's effort setting (max, high, xhigh), so the absolute minutes will not match your own deployment.
The Race to Build an AI That Understands Aging
⚡ Bottom line: AI labs and longevity startups are racing to build "foundation models" for human aging, not just chatbots for text.
💡 Why it matters: If models can read the body's aging signals, they could flag disease years before symptoms and test anti-aging drugs faster.
🔎 What it means: Longevity is turning into a data problem, and whoever models aging best may shape the next decade of medicine.
For years, "AI and longevity" meant wellness apps guessing your "biological age." In 2026 the ambition is bigger. In May, drug-discovery company Insilico Medicine and Human Longevity announced a multi-million-dollar effort to build what they call the first large-scale AI foundation model for longevity science, trained to spot age-related disease early, predict health risk, and surface new anti-aging drug targets. It is the same recipe behind chatbots, big models learning patterns from huge datasets, pointed at the biology of getting old.
(Insilico Medicine)
Why now? Because aging looks increasingly urgent and measurable. A Washington University study in Nature Medicine (June 22) found younger generations are aging faster biologically: US adults born in the 1990s showed systemic aging scores far above those born in the 1960s, tracking with an 8% higher risk of early-onset cancer. Those "aging clocks," algorithms that read blood and molecular markers to estimate true biological age, are exactly the tools AI is now supercharging.
The money is following. Retro Biosciences, backed by OpenAI's Sam Altman, says AI helped make cellular reprogramming 50 times more efficient and has pushed its first candidate into human trials, reportedly while raising at a $5 billion valuation. The shared bet: model aging well enough, and you can intervene before disease starts.
(Retro Biosciences)
Sources:
HubSpot AEO
Picture this. A buyer opens ChatGPT and asks for a recommendation in your category. Your competitor's name comes up. Yours doesn't. And that buyer never makes it to your website.
That's happening right now in markets everywhere. And most teams don't know it's happening because it never shows up in their analytics.
HubSpot AEO shows you exactly where your brand stands in AI search, where competitors are getting recommended instead of you, and tells you specifically what to fix. No expertise needed.
Try it free for 28 days. Just $50 a month after.
