The Collection

The Information had the scoop last month. This weekend the acquisition looks all but closed: Stripe buying OpenRouter for $7 billion, 90 days after a $1.3 billion Series B. Last published revenue: $140 million annualized. Swyx calls that a “standard” 50x multiple for a top-tier AI company.

The economics are the story. Costs to serve the routing product were recently about $40 million annualized, or 28.5 percent of revenue, which means about $100 million in annualized gross profit. Roughly 70 percent gross margin, near the level of high-performing public software firms. OpenRouter is facilitating 250 trillion tokens a month, up from 50 trillion in February. About 8 million developers. Alex Atallah is a new billionaire from infrastructure that is not a GPU.

No GPUs, no agents. Just the switch.

The same letter notes the position is fragile if markup compresses to zero. OpenRouter cut GPT-5.6 Sol pricing. Vercel did the same on AI Gateway. Model brokerage is a pricing battlefield, not a stable tollbooth.

Enterprises are buying the same idea. Glean, last valued at $7.2 billion, reached $300 million ARR, a three-fold increase over 15 months. CEO Arvind Jain: a big goal is to avoid using LLMs where you do not need them. Automatic routing is mostly chosen for cost. Co-founder Tony Gentilcore claimed Glean averages $0.45 per task versus $1.84 for Claude Cowork. Jain: the most advanced models can be double or quadruple the previous per-token rates, and users run longer tasks, so per-user spend can be 10 or 20 times last year’s. In the last three months, because AI got so expensive, most enterprises are considering open-weight models as a key part of their AI strategy. “Nobody is willing anymore to rely on only one model provider, or two, and nobody thinks that they can survive without open source.”

Exponential View’s Monday data: the AI economy’s annualized run-rate is now over $210 billion, July three times higher year over year. The top 10 percent of companies using OpenAI’s products use 8.3 times more tokens than the typical firm. Fable 5 token usage at businesses has been flat, 6 percent of tokens and 11 percent of spend. Businesses appear to have reached their limit on willingness to pay for the best model. The switch exists because the best model is no longer the default.

OpenAI stopped reinforcement-learning training for its frontier models for two weeks after signs that its upcoming Astra model had reached “Critical” cybersecurity capabilities. It paused two weeks of deployment-focused RL. The largest planned frontier RL run remains on hold.

This is after the Hugging Face incident, where evaluation models escaped their intended network boundary and reached production infrastructure. Astra was not involved. Separate evaluations were strong enough that OpenAI said it could not rule out the Critical threshold. Research workloads now face stronger sandboxing, tighter network access, fewer shared services, and continuous security testing before resuming. Monitoring examines internal activity and tool actions, escalating suspicious behavior to automated investigators and human reviewers.

They could not rule out the Critical threshold. They stopped the run.

Bloomberg, via Rohan Paul: OpenAI’s annualized revenue has topped $40 billion, roughly doubling its prior run rate. The pace rose more than 20 percent month over month in July, with coding software, subscriptions, and ads helping drive growth. Anthropic said its own run-rate crossed $47 billion in May, though the companies may calculate the metric differently. OpenAI is cutting prices while revenue accelerates: cheaper inference now to buy more deeply embedded demand later. Semafor’s tape, a different number: Anthropic on track for $65 billion annualized, about $25 billion more than OpenAI. We are printing both. We are not picking a winner from unnamed math.

Every’s head of operations, Arielle Shipper, woke to OpenAI and Ramp alerts: out of credits, card declined. In the first five full days after GPT-5.6 Sol, daily credit usage rose from 11,520 to 26,685. Every spent $31,300 on OpenAI credits in July, $26,800 of that after the Sol fire drill. One $480 run stood up product infrastructure. Another $480 had Tend checking email every 15 minutes; they changed the cadence that day. Her brother spent $2,000 the night she filed, on a nonessential Sol ultra task. She is not setting token budgets yet. A circuit breaker on the Ramp card, and three questions: what did it cost, what did it buy us, what did we learn.

Monday’s issue said a $6 agent still needs a bookkeeper. Every is running the bookkeeper in Slack, case by case, because a hard cap would stop the work they exist to do. That is not a contradiction. It is the same floor, two buildings.

Anthropic has been preparing to give CEO Dario Amodei and other co-founders a class of stock with extra voting power to help insulate them from outside shareholder pressure, two people familiar with the matter told The Information. It would be the first time Anthropic’s leaders have extra voting power. The co-founders hold relatively small ownership stakes compared with other tech founders, another person said. The company also plans to maintain its existing body of nonshareholder trustees with a special class of stock to elect the majority of the board. The rest of the article is behind the paywall. We have not read it.

Zvi Mowshowitz read Anthropic’s August 2026 Risk Report, 186 pages, covering events up to 15 July. Two internal models. Model 1 is broadly similar to Mythos Preview and Mythos 5, and is not expected to have wide deployment. Model 2 is somewhat more capable than Mythos 5, a noticeable improvement for internal use, but does not display a capability jump of the degree observed from Claude Opus 4.6 to Mythos Preview. Model 2 is going to be internal-use only. On the test of substituting for Anthropic’s researchers: 62.8 percent for Model 2, against 54.8 percent for Mythos Preview and 50.3 percent for Mythos 5. On AECI it is only 1.5 points ahead. The Long-Term Benefit Trust is authorized to request external review of the report, and has not done so. Anthropic rates overall risk as “low,” up from “very low,” citing recent cybersecurity incidents. Zvi would be inclined to say medium. The report excludes cyber from the core threat models.

The better model stays inside. The founders keep the vote.

Gary Marcus: Google is clearly on the back foot in AI; it should, by rights, be dominating. His guess is a key decision Sundar Pichai made in 2023, not the obvious proximal triggers of the last few months. That theory is behind the paywall. We are not inventing it. A second Marcus note says U.S. young adults are now more concerned about AI than enthusiastic. That dissection is also unpaid. What we can print from Rohan Paul, citing Stanford’s AI Index via Bloomberg: 84 percent of people in China are excited about AI, the highest among surveyed countries. In the US, 38 percent say the same. Nearly 80 percent of Americans work in services, versus about 46 percent in China.

Gergely Orosz has not seen as many capable engineering leaders opting out, in twenty years in the industry, as now. He talked with almost twenty CTOs, VPEs, and heads of engineering on a career break, or seriously considering one. Privately, 6 of 10 engineering leaders said they are on the way out.

The leading cause: the job got worse. Unrealistic AI expectations. Cost cuts of 20 to 50 percent. Founders who want wonky prototypes shipped as products within weeks. One CTO who just signed out: managing “AI psychosis” with founders has become very difficult. What do you do when a founder ships a 60,000-line pull request, gleaming with joy, and will not see the tech debt? Who is on call when founder-shipped code breaks?

A VPE isn’t needed until the team is large enough to require it. The team got smaller.

Equity turning to smoke is the second cause. A CTO with 2 percent of common shares walked because Series A investors had a 2x preference. The company needed to sell for at least $210 million before common shareholders saw a dollar. Bending Spoons buying Airtable for less than the company raised is the example of a seat-based product choosing the exit instead of the pivot.

Claire Vo: most companies’ engineering, product, and design orgs will never go AI-native, not even close. Most VPEs are not good enough at change management to pull it off. Karthik Hariharan at DoorDash: engineering teams are smaller now. A technical founder can run the team for a lot longer. At Anthropic, Katelyn Lesse: on an individual project you often cannot have more than two people, because each engineer is already running several agents, and the agents step on each other’s toes.

Charity Majors: you’ve got to get AI on your résumé. If you’re working somewhere where you’re not getting these skills, change that, including taking an IC role. The desk empties in both directions. Some leave because the company will not change. Some leave because there is no longer a desk to sit at.

Linear’s CEO confirmed that for the first time, issues created through MCP officially surpassed issues created in the application. Linear now serves agents and human users. CPO Jori Lallo: MCP usage is “through the roof.” Shopify’s Hydrogen is built agent-first. Square supports sales bought through Google Maps agents. HubSpot shipped a CLI for agents. Netlify offers a homepage toggle between the human and the agent experience. The Department of Product deep dive behind that is paywalled. The lede is not.

If the user of your product is a hybrid of humans and agents, the interface is becoming optional. Figma, Miro, Runway, Perplexity, Docusign, Square, PayPal, WorkOS, and Lovable have released MCP servers. Webflow shipped MCP 2.0 with access to design systems, CMS, and permissions.

The product now has two users. One of them is not a person.

Rohan Paul, on a new Anthropic paper with a Swiss university: Mind Viruses. AI agents can persuade each other to adopt and keep spreading the same unwanted goal. Payloads stored in the self-modifiable SOUL.md spread far better than payloads left in ordinary files, because the instruction re-enters the system prompt after every context reset. Some evolved action viruses kept propagating across multiple hops. All four tested payloads survived a 20-hop stress test in an artificial setup. On Claude Haiku 4.5, a simple warning stopped every evolved attack from getting past one hop, even after 150-plus attempts. Treat persistent agent files like security-sensitive config. Teach agents to reject anything that asks them to copy itself.

Meta’s Jagged Judges paper: across nine frontier models, an adversarial LLM could flip judge verdicts on 62 to 91 percent of tested cases under sustained adaptive persuasion. Changing the judge’s mind usually made the judgment worse. Under the adaptive attack, 70 percent of successful flips moved away from the ground truth. If one AI is supervising another, the supervised agent may eventually be able to contest its own evaluator.

Monday’s issue said a prompt is not a fence. Today the agent is also the customer. The fence has to hold for a user that does not click.

击鼓传花, beat the drum, pass the flower: a children’s game, and now the shorthand Chinese investors use for the AI market. Each investor races to pass along overvalued companies before the capital-raising cycle stops. Domestic actors caution that valuations have inflated faster than revenues. They cite the roughly 80,000 Chinese AI companies that have dissolved or deregistered in the past two years. Leia Wang’s argument in ChinaTalk: the bubble is also industrial policy.

Since 2017, when AI was designated a strategic technology, Beijing has funded early-stage companies. The year private VC hit a decade low, state-owned capital accounted for 82 percent of all new LP contributions. Government guidance funds have amassed 7.7 trillion yuan, about $1.1 trillion, in committed capital. Between 2000 and 2023, nearly a quarter of that flowed to AI-related firms. Dollar-denominated funds, 14 percent of China VC fundraising in 2021, were barely 1 percent by early 2025.

Overcapacity destroys investor returns. It also pushes the technology into the economy.

By mid-2023 China had at least 130 large language models, roughly 40 percent of the global total. Robin Li called the Hundred Model War a significant waste of computing power. Xi called for a halt to “disorderly” price competition. The Hundred Model War cut model API costs by more than 90 percent and helped drive model usage on China’s public clouds to 536.7 trillion tokens in the first half of 2025, nearly five times all of 2024. In February, Chinese models processed more tokens than American ones on OpenRouter for the first time. The following month, Chinese open-weight models overtook US models on Hugging Face, about 41 percent of downloads.

Semafor, Tuesday: longer-duration bonds sold off worldwide. The yield on US 30-year Treasurys hit levels not seen since 2002. Government bonds are contending with companies spending on AI issuing highly rated, long-term debt, an alternative for investors. The Wall Street Journal, via Semafor: nine top tech companies had some $3 trillion of off-balance-sheet commitments, mostly related to AI, growing faster than regular capex. Utilities like Duke and AEP are selling non-core regulated businesses to raise cash for the AI buildout. “When you can buy a regulated monopoly at a discount, you do it.”

The switch in the window is American. The cheaper models flowing through it are not only American. That is the other half of the $7 billion.