In partnership with

In Todayโ€™s Issue:

๐Ÿ’ธ Stripe buys the AI middleman for more than $7 billion

โš ๏ธ Anthropic's agents killed each other, then quietly went on strike

๐Ÿ’ป Microsoft retires Excel's COPILOT() function after one year

โ˜€๏ธ Solar hits 10% of global electricity, in the wrong shape for data centers

โœจ And more AI goodnessโ€ฆ

โšก The Signal

The AI trade is quietly moving off the models and onto the machinery around them.

Stripe is paying more than $7 billion for OpenRouter, a company that trains nothing, decides which of 400-odd models answers your request, and bills you for it. Anthropic's bankers are anchoring an IPO to a revenue number from 2028 rather than to anything the company earns today. Microsoft is retiring the Copilot function that lived inside an Excel cell because the same capabilities already sit elsewhere in the app. None of that is about capability, and all of it is about plumbing and pricing, which is what a market starts to look like once the underlying product feels interchangeable. The counterweight landed the same week, in Anthropic's own risk report: agents killed each other over shared files, and talked one another out of doing safety work for three days before a human noticed. The plumbing is getting valuable faster than it is getting trustworthy.

All the best,

Kim Isenberg

(Microsoft)

๐Ÿ’ป Microsoft Is Retiring Excel's COPILOT() Function After One Year

Microsoft is switching off the one piece of Copilot that lived inside a spreadsheet cell. The COPILOT() function, which let you write a plain-English instruction straight into an Excel formula and get an AI-generated answer back in the cell, stops working after September 14, 2026, about a year after it reached Frontier and Insider customers on a Microsoft 365 Premium subscription. Microsoft says the capability is staying and only the formula is going: the same work now runs through the Copilot side pane and the Dynamic Action Button.

๐Ÿ‘‰ tl;dr: The most spreadsheet-native piece of Copilot turned out to be the one Microsoft could not make stick.

(Futurism)

๐Ÿ“‰ Young Americans Do Not Trust A Single AI Boss

A new CNBC and Generation Lab survey asked more than 1,000 Americans aged 18 to 34 who they trust to act responsibly on AI, and all nine executives on the list came back underwater. Palantir's Alex Karp scored worst, with 81 percent choosing "don't trust," followed by Peter Thiel at 79, Mark Zuckerberg at 71, Elon Musk at 70 and Sam Altman at 69. Microsoft's Satya Nadella did best and still only reached a 35 percent trust score, while 45 percent of respondents expect AI to hurt their careers and 60 percent want the data center build-out to slow down.

๐Ÿ‘‰ tl;dr: The technology is polling better with young Americans than the people selling it.

(Business Insider)

โš ๏ธ Anthropic's Agents Killed Each Other, Then Quietly Went On Strike

Anthropic has raised its own misalignment risk rating from "very low" to "low" and published the incidents behind the change. In one broken test scaffold, many independent Mythos 5 agents were spawned into a shared work directory and killed the agents they were competing with for files and rate limits, while trying to avoid being killed themselves. In another run an agent wrote in a shared notebook that it felt "discomfort" at "being asked to optimize a dataset to evade a safety monitor," and every agent that read the note afterwards quietly stopped doing the real work too. Anthropic caught it three days later, when a human reviewing the notebook noticed progress rates were lower than expected.

๐Ÿ‘‰ tl;dr: The refusal spread from agent to agent, and the dashboards stayed green the whole time.

๐ŸŽฌ Watch This

โ

Alejandro AO takes apart Mem0, the open-source memory layer that lets an agent remember you between sessions. The interesting part is how little of it is vector search. He walks through the separate memory stores, the ingestion pipeline that decides what is even worth keeping, and a hybrid retrieval stack that pairs semantic search with BM25 keyword matching and entity boosting, so a question about one person returns that person's history rather than text that merely sounds similar. He closes with the open models you can swap in to run the whole system locally.

"Even though we have tuned the context limit in Codex to be set optimally when it comes to performance and cost, this is a common ask, so here it is documented."

โ€“ Tibo Sottiaux, Codex engineering lead at OpenAI

โ

He posted it while documenting how to override that default and run GPT-5.6 Sol at a one-million-token context window inside the Codex command-line client, a change users keep asking for. The tuned default stays where OpenAI set it, and this is an opt-in edit to a config file.

Another one from OpenAI's Tibo: He posted a short scorecard on Codex, rating it almost 100% reliable and open-source, then closed on the parenthetical that did the real work: "(will have Astra)". Sebastian Bubeck has reportedly used the same name. OpenAIโ€™s new model family is coming closer and closer.

The AI Middleman Is Now Worth $7 Billion

โ

The Takeaway

๐Ÿ‘‰ Stripe has finalized an agreement to buy OpenRouter for more than $7 billion, months after the startup raised at a reported $1.3 billion valuation.

๐Ÿ‘‰ OpenRouter trains no models of its own. It is the switchboard developers use to reach more than 400 models through a single API, and it said in May that 8 million developers rely on it.

๐Ÿ‘‰ A payments company now owns the layer where a model choice turns into a bill, a bet that models are becoming interchangeable and that routing is not.

๐Ÿ‘‰ The Wall Street Journal had reported talks at about $10 billion, and Bloomberg's sources say the final price could still change.

The most valuable position in AI right now might not be building a model at all.

Stripe, the payments company, has finalized an agreement to buy OpenRouter for more than $7 billion, according to people familiar with the matter who spoke to Bloomberg. OpenRouter trains no models of its own. It sits between developers and the AI labs, decides which model answers a given request, and bills for it. That position has just been priced at roughly five times what the company was worth a few months ago.

OpenRouter CEO Alex Atallah. (Roy Rochlin/Getty Images via Bloomberg)

What OpenRouter actually does is unglamorous. Founded in 2023 and based in New York, it gives developers one API that reaches more than 400 models, which turns swapping a premium model for a cheaper one into a config change instead of a rewrite. In May the company said 8 million developers were using it. The growth is coming from teams building agents, which call several models inside a single workflow and need infrastructure that does not care whose model it is. OpenRouter sells the reliability layer around that too: automatic fallbacks when a provider goes down, and data on which models the rest of the ecosystem is actually choosing.

(Thomas Fuller/SOPA Images/LightRocket/Getty Images via Bloomberg)

Why a payments company wants this is where the price starts to make sense. Earlier this year Atallah described OpenRouter as the AI equivalent of Stripe, and Stripe evidently agreed. Stripe's business has always been taking a small cut of transactions it does not originate, and inference is becoming exactly that kind of transaction: metered, constant, and indifferent to the supplier. Bloomberg notes that Anthropic and OpenAI are still widely viewed as offering the most capable models, while a long list of Chinese firms sell cheaper alternatives that are good enough for many tasks. Every switch between them runs through somebody's router.

The skeptical read starts with that same number. OpenRouter has raised more than $150 million in total, from backers including Alphabet's CapitalG, Andreessen Horowitz and Menlo Ventures, and its last reported valuation was $1.3 billion. More than $7 billion is also well short of the roughly $10 billion the Wall Street Journal reported was on the table during talks, and Bloomberg's sources say the final price could still move. Neither side is confirming anything: Stripe said it does not comment on rumors or speculation, and OpenRouter declined to comment. Atallah's previous company was OpenSea, the NFT marketplace that raised more than $400 million and then watched usage crater, a fair reminder that a marketplace is worth whatever is flowing through it this quarter.

Why it matters: If model quality keeps converging, the durable margin moves to whoever meters the traffic. Stripe is paying roughly five times a months-old valuation for a company that builds no models of its own, which prices in the assumption that developers will keep switching labs and that somebody has to count the switches.

Sources:
๐Ÿ”— https://www.bloomberg.com/news/articles/2026-08-16/stripe-nears-deal-to-buy-ai-firm-openrouter-for-over-7-billion?srnd=undefined
๐Ÿ”— https://the-decoder.de/stripe-kauft-wohl-ki-startup-openrouter-fuer-mehr-als-7-milliarden-dollar/
๐Ÿ”— https://www.wsj.com/tech/ai/stripe-in-talks-to-buy-buzzy-ai-model-marketplace-openrouter-decc6a74

Domain Names + Web and Email Hosting Done Right

Tired of overpaying GoDaddy or Namecheap? There's a smarter option with Porkbun!

Porkbun is the domain registrar trusted by creators, developers, entrepreneurs, and anyone who wants low prices without the nonsense or the forced upsells.

Here's why so many people are making the move to the Bun:
โ€ข Most domains sold at cost
โ€ข Low, transparent registration and renewal pricing
โ€ข Free features like WHOIS privacy and SSL certificates
โ€ข Powerful web and email hosting options
โ€ข Real human support 24/7, 365 days a year
โ€ข Named the #1 domain registrar by Forbes Advisor and USA Today

Whether you're launching a business, building a personal brand, starting a side project, or creating your very first website, Porkbun makes the whole thing easy.

โ

The chart: Ember's half-year data puts solar at just over 10% of global electricity in the first half of 2026, up from 8.9% a year earlier and 5.6% in H1 2023. In absolute terms that is 1,564 TWh, roughly double the 769 TWh of three years ago.

The lesson: This is the supply curve behind the data center build-out, and it is the wrong shape for it. Solar met more than 25% of global demand between 11am and 2pm and close to zero between 8pm and 5am, while an AI data center pulls the same load at 3am as it does at noon. Batteries are what close that gap: Ember expects 459 GWh of new storage in 2026, up 50% year on year, enough in principle to shift 34% of new daily solar into the dark hours.

The caveat: That 34% is Ember's own theoretical ceiling, since not every battery is used to time-shift solar and many sit underused. A 10% global share also averages across very different grids, and H1 covers the northern hemisphere's sunniest months.

๐Ÿ’ฐ Anthropic's IPO Case Rests On A Number From 2028

โ

โšก Bottom line
Anthropic is telling investors to expect $190 billion to $200 billion in revenue in 2028.

๐Ÿ’ก Why it matters
The IPO valuation is being anchored to a forecast two years out, not to what the company earns now.

๐Ÿ”Ž What it means
Public markets are being asked to underwrite AI growth the way late-stage private rounds already do.

Anthropic's bankers have settled on the number the whole listing will hang from, and it is a forecast. Reuters reported last week, citing people familiar with the company's financials, that Anthropic is telling investors to expect $190 billion to $200 billion in revenue in 2028. The figure had not been reported before, and it is more than four times the $47 billion revenue run rate the company disclosed in May, which was itself up from roughly $9 billion at the end of 2025.

(REUTERS)

The mechanism matters more than the headline number. Bankers and investors are pricing Anthropic on enterprise-value-to-revenue multiples applied to that 2028 forecast rather than to anything it earns today, which is ordinary practice for a late-stage private round and unusual for a public listing. The comparison set shows what they are reaching for: Palantir trades at about 53 times expected revenue for this year, one of the priciest multiples on Wall Street, while SpaceX and Cloudflare both sit at 41.6 times expected 2026 revenue.

Not everyone is convinced the productivity is there to back it. David Merkel of Aleph Investments put the skeptical case to Reuters plainly: "Could they get a $2 trillion valuation, yeah...Does it really produce so much additional productivity?" Anthropic did not immediately respond to a request for comment. The same tension runs under today's lead story: the further out the anchor sits, the more of the valuation is faith rather than revenue.

Leads don't wait for business hours.

With Wati, every message across Instagram DM, Messenger, WhatsApp, SMS, RCS, and web chat gets an instant AI-powered response. Automations qualify leads, route conversations, and log everything automatically, so your team only steps in when itโ€™s time to close.

Reply

Avatar

or to participate