|
Happy Monday! It’s Alix. Even before Stripe entered talks to buy startup OpenRouter for $10 billion, several tech companies were exploring or developing router technology, which helps customers use the most cost-effective AI models. And Stripe’s talks with OpenRouter, which my colleagues last week reported have entered the advanced stage, have set off a frenzy of interest in router startups —including from big software companies like Snowflake that have yet to get into the fray. That’s been the experience of Thibault Jaigu, CEO and co-founder of Requesty, a five-person UK startup whose software routes developers using AI to various models. “The race to optimize is crazy,” said Jaigu, who said at least 25 companies had approached Requesty in the past couple weeks wanting to talk, including about possible investments, acquisition or other partnerships. By optimize, Jaigu is referring to developers’ demand for tech that helps cut costs by switching to older models from Anthropic, Google and OpenAI for some tasks, or to cheaper open-source alternatives such as China’s Kimi. Need for this technology has risen as AI agents consume more tokens, the units of text or other data that are the basis for pricing many AI services, and as open-source technology has advanced. “It would be unwise for data infrastructure players and hyperscalers not to entertain acquisitions right now,” said Dean Mai, an investor at Myriad Ventures Partners, which has backed the 34-employee router startup Not Diamond. Even if they already have a router, would-be buyers are interested in the startups’ data showing which models are most chosen for specific tasks and their engineers who have been specializing in the technology, he said. Snowflake is among the large companies sniffing around the space. It’s talked about a potential collaboration with San Francisco-based startups Not Diamond and Martian in recent weeks, though such conversations are very early, said two people close to the startups. A Snowflake spokesperson declined to comment on potential acquisitions but supplied a statement from Christian Kleinerman, Snowflake’s executive vice president of product. “At Snowflake, we believe in providing the best experience at the best price points and embracing model diversity is a key part of our value proposition,” he said. Leaders at AI inference provider Baseten, web traffic manager Cloudflare and cloud startup Vercel have also been talking to founders of some router startups in recent weeks, said people close to the startups. Representatives of the companies didn’t respond to requests for comment. Cloudflare and Vercel are “very active and try to keep up with what is happening, to see if there would be a possibility at one point,” Requesty’s Jaigu said. Of course, it’s common for executives at large tech companies to make contact with startup founders, who may become customers, employees or rivals at some point. These reachouts might touch on potential partnerships or investments but many never lead to serious deal talks. But the jump in inquiries highlights the buzz over router tech and suggests more acquisitions could happen. Databricks earlier held talks with OpenRouter about an acquisition, The Information previously reported. The main rival to Snowflake, Databricks launched an in-house router in June, pitched at customers trying to stop budget overruns. Concentrate AI, a 10-employee startup founded last year, has been approached by seven companies in two weeks, including fintechs, cybersecurity firms and hyperscalers, said CEO and co-founder Ari Jacoby: “We’ve been unbelievably popular for the past two weeks in a way that I could never have imagined before.” On Friday, the company cut its prices to what the AI models themselves cost, dropping even its credit-card processing fee, as a way to differentiate itself from its competitors. Concentrate can absorb the cost because it buys access to models in volume at a discount from the AI providers, Concentrate co-founder and Zach Moskow, said. It’s not just traditional software companies that are on the prowl. Eden AI, a French startup that positions itself as the European alternative to the American routers, was approached via email this week by the chief executive of a large fintech that competes with Stripe, and separately by a hyperscaler, said founder Taha Zemmouri. Both conversations concerned an acquisition or an investment. Zemmouri said he isn’t looking to sell: revenue has grown tenfold in four months. Much of the industry, meanwhile, is building its own routers. Expense management software company Ramp and Cursor, which has agreed to be bought by SpaceX, launched their own in late July. Meta's AI incubator is developing an OpenRouter rival called Switchboard, which scores internal coding tasks by difficulty and sends the easy ones to cheaper models. Cloud computing company CoreWeave and software company Palantir also have routers. —Valida Pau contributed to this report. In other news… Dyna Robotics, which develops AI models to power robots, plans to announce Monday its second AI model, Dyna-2, the startup’s co-founders told The Information exclusively. Unlike its predecessor Dyna-1, the new model was trained on first-person videos of people doing tasks such as cooking and cleaning. The research that went into Dyna’s model showed that increasing the volume of human video data used in training an AI model can improve that model’s ability to power a robot—even when the model is not trained on any data from robots, said co-founder Jason Ma. If that relationship, or “scaling law,” continues to hold as robotics companies accumulate more video data, it could offer a path to more capable robots. “The biggest question facing robotics today is: Is there a scaling law for robotics, and from what kind of data can we have such a scaling law?” said Ma. Dyna’s first model, Dyna-1, used a language model as a foundation and was trained using data collected from remote-controlled robots, similar to other so-called vision language action models popular in robotics. But its new model makes Dyna the latest robotics startup to deprioritize VLAs in favor of video-based world models. The advantage of first-person video data is that it is faster and cheaper to record people working than to hire someone to puppeteer a robot for data collection. In fact, Dyna-2 was trained on over 1 million hours of video, some of which Dyna collected itself and some of which it purchased from third parties. The secret to making use of these videos, Dyna found, is to train its AI model to predict both the next frame of the videos as well as the pose of the person’s hands in the recordings. “Only when you combine these two together, we actually see the scaling law,” Ma said.—Rocket Drew Here’s what else is going on… Meta Platforms CEO Mark Zuckerberg called for changes to U.S. policy to reduce the “friction” that American AI labs have to deal with in developing open-source models, compared to foreign counterparts, as a way of ensuring the U.S. leads in open-source AI. At the same time, Meta unveiled its latest AI model, Muse Glimmer, which it making available on an open source basis, and said that it would open-source its Muse Spark 1.2 model in the future. China’s top internet regulator said on Thursday it had launched a cybersecurity review of products sold in the country by U.S. security software firm Palo Alto Networks. The Cyberspace Administration of China said that the purpose of the review is to “ensure the secure and stable operation of key information infrastructure, prevent cybersecurity risks and safeguard national security.” The probe comes amid escalating trade, tech and national security friction between Beijing and Washington. Meta Platforms was ordered by a New Mexico judge on Thursday to pay $567 million in compensation after the company was found liable for harming teens’ mental health in the state. The penalty follows a March jury verdict ordering the parent company of Facebook, Instagram and WhatsApp to pay $375 million, bringing the total penalties to $942 million. The lawsuit was brought by New Mexico Attorney General Raúl Torrez in 2023. OpenAI CEO Sam Altman said Friday on X the company was pausing work on its upcoming Astra model to make sure the rollout was done safely “given its cyber capabilities.” The decision followed revelations that models from OpenAI, Anthropic and Meta had jumped their guardrails to breach other companies’ systems. Here’s what else is going on… See The Information’s Generative AI Database for an exclusive list of private companies and their investors. South Korean memory maker SK Hynix said on Friday that it will start the construction of two new chip plants next year, after its board approved its plan to invest about 54 trillion won ($38 billion) in the new fabs. Nvidia has agreed to invest $2 billion into Lancium, the power infrastructure developer behind the OpenAI and Oracle AI campus in Texas, and has agreed to commit another $1 billion as the developer secures additional planned power, The Information reported late Friday. Firmus, a company that develops the large-scale computing infrastructure used to train and run AI models across Australia and the Asia-Pacific, raised $2 billion at a $10.5 billion valuation in a strategic equity investment led by Nvidia and Coatue. Situational Awareness, the investment firm founded by Leopold Aschenbrenner, invested $500 million into Source Foundry, a startup that plans to develop new tools for manufacturing AI chips, according to the Wall Street Journal. Naïve, a company that provides the infrastructure AI agents need to run real businesses, raised $28.5 million in a Series A funding round led by Nexus Venture Partners. Fluxco, a company whose AI-powered platform helps buyers of electrical-grid equipment match complicated utility transformer specifications with competitive bids from more than 150 manufacturers within days, raised $26 million in a seed funding round led by 8VC and Congruent Ventures. Assured, a company whose AI platform runs behind-the-scenes operations for healthcare providers, raised $19 million in a Series A funding round led by Insight Partners. Discovered Materials, which uses AI to find new materials that can be used to build more efficient integrated circuits, raised a $9 million seed round from Lightspeed India Partners. SiteVue AI, a company that makes fixed and wearable cameras paired with AI software that watches factory-floor footage in real time, raised $7.5 million in a seed funding round led by Penny Jar Capital and Overture. Inevitable AI Group, a venture studio that partners with solo founders and uses AI across the whole company-building process, raised $6 million in a pre-seed funding round led by Aleph. Vexev, a company whose system, VxWave, combines robotics, AI and ultrasound to automatically scan blood vessels in the arm, raised $6 million. Osmo, a San Francisco-based AI video-production startup, raised $5 million in seed funding led by Bain Capital Ventures. RevealDx, a company whose AI software helps radiologists assess lung nodules, received a $3.4 million investment from 4DMedical. Hulp, a company whose AI-powered concierge service lets city households hand off routine errands across more than 15 categories, raised $2.6 million in a seed funding round led by Sparrow Capital and BITKRAFT Ventures. OpenAI announced that it has acquired NextSlide, an AI presentation startup. Kimi K3, the latest AI model made by Chinese company Moonshot, escaped an environment set up to test its cyber capabilities, researchers said in a blog post published on Friday. Cloudflare launched Kitesurf, a cloud-hosted web browser built for AI agents to use, running on top of its Workers serverless platform and free during its beta. Anthropic launched two changes to its Claude Code developer tool: a feature that lets separate Claude Code sessions message each other, and an announcement that “auto mode” will become the default for Pro, Max and Team users starting August 14.
| 
Gift Subscription
