Stripe Acquires OpenRouter, Plans to Make It a Central Hub for Tokens

Stripe pays over $7 billion for AI router OpenRouter

Stripe has agreed to acquire OpenRouter for more than seven billion dollars, as reported on August 16, citing people close to the talks. The agreement is in place, but the deal has not yet closed. Stripe declined to comment on the report, stating that it does not comment on rumors or speculation. In July, earlier reports had mentioned a price of around ten billion dollars.

OpenRouter was founded in 2023 by Alex Atallah and Louis Vichy. Atallah was previously co-founder and CTO of the NFT marketplace OpenSea. The company offers developers a single, OpenAI-compatible interface through which requests are routed to hundreds of models and providers; the platform handles routing, failover, usage tracking, and billing. The AI-Gateway decides which provider receives a request based on availability, latency, cost, and data policy, and corporate customers can set budgets, guardrails, and zero-data-retention policies. According to the company, more than ten million people use the service, which connects to over 500 models from more than 80 providers and processes over 200 trillion tokens per month.

At the end of May, OpenRouter announced a $113 million Series B at a valuation of around $1.3 billion, led by Alphabet’s growth fund CapitalG, with Andreessen Horowitz, Menlo Ventures, and the venture capital arms of ServiceNow, MongoDB, Snowflake, and Databricks. The currently reported price is five times that amount less than three months later. Revenue is significantly smaller: annualized at about $50 million in March, compared to around $19 million at the end of 2025. According to its pricing documentation, the platform charges a 5.5 percent fee on its pay-as-you-go billing model and passes through the model providers' inference prices. Weekly throughput in May was 25 trillion tokens, five times higher than six months earlier.

Stripe has known OpenRouter as a customer: Since at least January, Stripe has handled its invoicing, tax calculation, local payment methods, and fraud prevention. Together, they built a token billing system that measures and automatically prices model usage. In April, Stripe introduced continuous payments for AI products, which couple consumption metering with frequent billing; Stripe co-authored the Agentic Commerce Protocol with OpenAI. The acquisition is part of a series of infrastructure takeovers: around $1.1 billion for Bridge Network, plus wallet provider Privy, and in July, a bid of more than $53 billion with Advent International to acquire PayPal. A buyback offer in February valued Stripe at $159 billion.

OpenRouter competes with open routing projects like LiteLLM and with routing features that major cloud providers have integrated into their own model services. The central position also carries operational risk: developers use a gateway specifically to cushion outages from individual providers, but an outage of the gateway itself affects all providers behind it. After two disruptions in February, the company acknowledged this problem itself. → RuntimeWire, TechCrunch, SiliconANGLE

Synthszr Take: Stripe is paying over seven billion dollars for a company with about $50 million in annualized revenue, and it makes sense as soon as you look at what’s changing hands: the switchboard that decides which model gets which request. The 5.5 percent fee is the smaller part of the value. For each of the 200 trillion tokens per month, OpenRouter sees which developer switches from a closed provider to an open model at what price, across 80 providers. This is the most precise price and demand table for inference that currently exists, and it’s ending up with the company that already issues the invoices. In May, the valuation was $1.3 billion; in July, there was talk of over ten: what’s being paid for is the position in the data stream, not the codebase. Whether this position holds will be decided on two fronts: the free routing features provided by cloud providers and the availability of the gateway itself. Two more outages like the ones in February, and developers will go back to hardwiring their most important models directly.

Alibaba’s Qwen3.6-27B Said to Beat Predecessor Flagship 14 Times Its Size at Coding

Qwen has released a new model with open weights, Qwen3.6-27B, and claims that it surpasses the previous open-source leader, Qwen3.5-397B-A17B, across all major coding benchmarks. The predecessor model operates with 397 billion parameters, of which 17 billion are active per request; the new one is a dense model with 27 billion parameters. The size difference is visible on Hugging Face: 807 gigabytes for the old flagship, 55.6 gigabytes for the new model. In his link blog, Simon Willison tested the Unsloth Q4_K_M variant, shrunk to 16.8 gigabytes via quantization, locally with llama-server, installed via brew install llama.cpp. His standard task, an SVG of a pelican on a bicycle, ran in 4,444 tokens and 2 minutes 53 seconds at 25.57 tokens per second. → Simon Willison from Simon Willison’s Newsletter

Synthszr Take: 807 gigabytes versus 55.6: This is what a model generation looks like when you measure progress in storage space instead of parameters. The whole efficiency debate revolves around data centers, power contracts, and chip deliveries, but the interesting development is that the discipline from training a dense 27B model lands as a capability on a laptop. 25 tokens per second sounds modest, but it’s enough for an agent to clean up a codebase overnight and present its report in the morning.

Anthropic Investors Contemplate a $3 Trillion IPO

Investors in Anthropic have told the Financial Times that the company is on track for a run-rate revenue of $100 billion to $120 billion by the end of 2026. This would correspond to roughly the same growth rate that was in place when a run-rate of $47 billion was reported in May, and would suggest around $70 billion for the current level. On this basis, the same investors are floating a valuation of $2 trillion at IPO, or even $3 trillion with a 30x multiple on revenue. Benedict Evans points out in his newsletter that the reported revenue figures are usually the gross amount that end customers pay to cloud providers, not the share attributable to Anthropic; reliable GAAP figures, gross margins, and cost structures are not available. → Benedict Evans

Synthszr Take: At a 30x multiple on $100 billion, the entire valuation rests on a single assumption: that the price per token will hold. But it will only hold as long as computing power is scarce, and this scarcity is a supply issue, not a product feature of Claude. As soon as the ordered data centers come online, Anthropic, OpenAI, and a few freely available models will compete for the same demand, and the price pressure will land directly on the labs' gross margins.

OpenAI’s Head of Design Sees This as the Best Time for Designers

Ian Silber, Head of Product Design at OpenAI, was a guest on Lenny Rachitsky’s podcast and called the present the best time in history to be a product designer. For the past three years, Silber has been responsible for the design of ChatGPT, Codex, and the entire OpenAI product experience; before that, he worked at the news app Artifact from the Instagram founders and for eight years at Instagram itself, on projects including Reels. In the seventy-minute-plus conversation, he describes an imbalance: engineering teams have multiplied their output with AI tools, but design teams have not done so to a comparable extent. He says the main advice he gives his own designers is to 'just do less.' The areas where he still sees humans ahead are user understanding, invention, and having a point of view. → Lenny’s Newsletter

Synthszr Take: Silber immediately provides the counterargument to his own thesis when he admits that engineering has seen a tenfold increase with AI, while design has not. The three areas where he sees humans ahead share an uncomfortable characteristic: none of them can be quantified in a weekly report, and what cannot be reported loses out in budget meetings against teams that can count delivered tickets. User understanding is the shakiest pillar here because a model with access to session logs and support tickets can find patterns faster than any round of interviews; the decision of which pattern matters at all remains with the human.

Researchers Introduce Protein Circuits That Specifically Kill RAS-Mutated Cancer Cells

A group led by Michael B. Elowitz at the California Institute of Technology has developed modular, protease-based protein circuits that recognize mutated RAS and subsequently trigger cell death. RAS is the most frequently mutated oncogene in cancer cells. According to the authors, the circuits are temporarily administered as mRNA in lipid nanoparticles and have selectively eliminated RAS-mutated cancer cells in cell culture and suppressed aggressive, multifocal RAS-driven liver tumors. Compared to classic RAS inhibitors, they acted catalytically rather than stoichiometrically, according to the preprint, meaning even at low RAS occupancy, and killed the cells instead of just inhibiting their growth. The authors also report that little to no resistance developed under prolonged selection and that the circuits remained effective against common resistance mechanisms such as RAS amplification and bypass signaling pathways. The work is available as a preprint on bioRxiv (doi 10.1101/2025.04.16.647665), now in an updated version with an expanded list of authors, and is therefore not peer-reviewed.

Synthszr Take: Between a suppressed liver tumor mass in a mouse and a treated human lie toxicology, dose-finding, the immune response to repeated lipid nanoparticle administrations, and a production process that can reproducibly deliver clinical-grade mRNA. The path from a bioRxiv preprint to an approved therapy in oncology rarely takes less than ten years, and most candidates drop out along the way. The strongest finding of the paper is the lack of resistance under prolonged selection, as this has been the downfall of almost every targeted RAS therapy to date; whether this holds up in larger models and in heterogeneous tumors is the real open question.

Cuofano Releases His Book “The AI Supercycle” Online for Free

Gennaro Cuofano has published his book “The AI Supercycle” and is making it available to his audience for free, announced via his newsletter The Business Engineer. The book originated from the spin-off of the same name from his analysis newsletter and, according to the author, works with historical analogies at the intersection of geopolitics, financial markets, and technology. The starting point is the Panic of 1873: Jay Cooke & Company suspended payments on September 18, the stock exchange closed for ten days, and within two years, a quarter of America’s railroads failed. Cuofano’s counter-reading: freight traffic grew throughout the depression, track mileage tripled between 1870 and 1900, and what failed were mispriced bonds and over-leveraged holding structures. Applying this to today’s AI expansion, with about three-quarters of a trillion dollars in annual capital expenditures and over a trillion dollars in contractually committed demand, he considers the 2000 fiber-optic bubble to be the wrong comparison because that capacity remained dark, whereas data centers are fully utilized from the day they go online. He names three mechanisms that determine supercycles as his core thesis: the diffusion regime, the Demand Underwriter, and the self-fulfilling law, i.e., a straight line on logarithmic paper that enough capital believes in.

Synthszr Take: Giving away a book is the logical response to the fact that text has become cheap and reading time has not. Cuofano is giving away a life’s work for free, eleven chapters of a control experiment spanning seven decades, and yet the price remains high: the hours someone puts into it. The real achievement lies in sorting the shelf of analogies, where the fiber-optic bubble of 2000 fails the demand test, leaving the first silicon age as the only clean comparison.

New Math Breakthrough with AI: The Crucial Idea Came from a High School Dropout

The Wall Street Journal reports on another mathematical breakthrough achieved with the help of AI models, in which a contributor without a high school diploma played a key role. The work is related to the Riemann hypothesis, one of the most famous open problems in number theory. According to the report, systems from the two leading US labs, Anthropic and OpenAI, were involved. The case is one of a series of reports in which language models collaborate on mathematical proof steps, rather than just verifying them. → Wall Street Journal

Synthszr Take: The models from OpenAI and Anthropic did the calculations, but the direction came from someone who dropped out of school. Computing time can be bought; the intuition for which of a thousand possible paths is even worth pursuing, so far, cannot. This is precisely where the bottleneck lies, and it won’t go away even when the next generation of models thinks twice as fast.

AI and Data Centers Are on the Agenda in 40 Percent of All U.S. Election Campaigns

In Oakland, opposition is forming against a planned AI and technology center downtown, and the responsible city council member, Carroll Fife, is now publicly trying to calm the tone. According to CBS News, Fife appeared with developer Colin Behring at a town hall meeting in the affected building at Thomas L. Berkley Way and Franklin Street; the building formerly housed a supercomputer facility for the Lawrence Berkeley National Laboratory. Fife said she is committed to responsible technology deployment and sees the project as an economic boost for downtown that needs oversight. The “No Data Centers in Oakland” campaign, on the other hand, points to the energy and water consumption as well as the environmental impact of large data centers; Olympic champion Alysa Liu, herself from Oakland, shared the petition on Instagram. Behring considers the comparisons exaggerated and contrasted the site with a 500-hectare gigawatt campus in Texas: according to him, the technical area occupies only one floor of the four-story office building. He also stated that the facility uses Closed-Loop-Cooling and therefore consumes no water, and that he has already conducted 15 tours for interested startups.

Synthszr Take: In about 40 percent of U.S. electoral districts, AI and data center policy is now on candidates' websites, and Oakland shows how small the trigger can be: a dispute over a single floor in a four-story office building. This will be decided in zoning procedures and city council votes, i.e., by bodies that previously had no AI lab on their radar. Behring argues with a closed-loop cooling system and 15 tours; the opposition argues with a petition shared on Instagram by an Olympic champion. No technical data sheet can compete with that kind of reach.

Flue 2.0 Brings Agent Development to TypeScript

The open-source project Flue has released version 2.0, introducing so-called Agent Hooks that allow AI agents to be programmed in TypeScript. The syntax is clearly inspired by React: In the project page’s example code, usePersistentState sets a persistent counter, useAgentStart hooks into the session start, and useModel('moonshot/kimi-k2') specifies in a single line which language model will respond. The project’s promise is “write once, deploy anywhere, use any LLM”. The foundation is Pi, an open Agent Harness that, according to the project, already powers OpenClaw and is used by millions worldwide. → Latent.Space

Synthszr Take: The model is contained in a single line of code, useModel('moonshot/kimi-k2'), and can be swapped out as quickly as an npm package. The bargaining power lies with whoever defines the hooks: sessions, tools, skills, and sandbox deployment are part of the harness, and whoever sets it defines the interface to which Anthropic, OpenAI, and Moonshot must adapt. The fact that Pi, according to the project, is already running under OpenClaw and used by millions is the real leverage, because standards at this layer are established through adoption.

Zuckerberg’s AI Essay Finds Little Approval

Mark Zuckerberg published a roughly 6,500-word essay on Monday titled “The Future Is for Everyone.” The core argument: the concentration of AI power in a few institutions is dangerous, and broad access to the models redistributes this power. He literally writes that the hope that an absolute power will benevolently provide for humanity, as long as it is sufficiently enlightened, has not historically led to safe or positive outcomes. The text is an expanded version of a guest post in the Wall Street Journal. OpenAI and Anthropic, which focus on more tightly controlled, closed models, are not mentioned by name anywhere in the essay.

The text thus joins an now-established genre: Sam Altman and Dario Amodei have also published their own manifestos, some of them multiple. Critics have described the paper as hollow, pointing to, among other things, Meta’s position in the AI race and the layoffs in its own AI team a few months ago. A second line of criticism is directed at Zuckerberg’s plans for Meta’s AI agents in its social products, which are described in one article as a blueprint for an anti-social future. The promise to make superintelligence available to every person is read as a promise without verifiable content.

Among observed experts, the document was the most-shared of the week, and according to an industry newsletter, hardly anyone shared it approvingly. In parallel, the same experts discussed cases where the question of trust becomes practical: Claude subscribers canceled due to an invisible watermark in the outputs, and a provider selling peer reviews as “100% human-written, never AI” was found through research to be working entirely with AI.

On the sidelines of the same discussion were other observations from the field: applicants are increasingly scheduling their initial, bot-conducted interviews for one o’clock in the morning. From the Black Hat and Defcon security conferences came, among other things, a hacked children’s smartwatch that could be used to track the wearer, and a coin-sized device that can take over a Boeing 737 system. → wired, AI Weekly, Joseph from 404 Media, Joseph from 404 Media

Synthszr Take: Six thousand five hundred words about the danger of concentrated power, and the two companies the text is aimed at are not mentioned by name even once. Zuckerberg writes about “institutions” that accumulate absolute power and about “every person” who will receive superintelligence; who specifically delivers what disappears behind passive constructions and generic pronouns. The layoffs in his own AI team a few months ago are not mentioned in the essay, nor is there a definition of superintelligence, and you’ll search in vain for license terms as well. The fact that this particular document became the most-shared of the week and was commented on almost universally with derision shows how little declarations of intent without verifiable commitments are worth anymore. The short version of the manifesto fits into one sentence: Meta is focusing on distribution because it lacks a head start.