Upside-down World: AI Startups Demand Regulation, Software Monopolists Lobby for Open Source

Open Source: OpenAI and Anthropic Against the Rest of the World

Nvidia CEO Jensen Huang has published an open letter describing open AI models as necessary for the US's technological leadership. According to The Business Engineer, 25 of the industry's biggest names have signed, including Meta, Microsoft, Nvidia, Palantir, IBM, Hugging Face, and Mistral. The arguments are familiar: portability, customer choice, no single points of failure, and genuine competition instead of concentration at the model level.

The analysis sorts the signatories into groups with different motives. Eight open-weight natives like Meta, Hugging Face, Mozilla, Mistral, and the Linux Foundation are defending their own product or founding purpose. Nine companies like Microsoft, Palantir, IBM, ServiceNow, and Perplexity sell above the model layer and benefit from more competition below it. Nvidia signs alone as a compute representative because model diversity equals compute diversity. Also notable are the absentees: OpenAI, Anthropic, Google/DeepMind, AWS, Databricks, Snowflake, Salesforce, and Apple are missing.

The New York Times frames the dispute as an open conflict that escalated this week. On one side are Anthropic and OpenAI, who argue that some models are too dangerous for open development and must be controlled. On the other side is the rest of the industry, which considers open models indispensable for a healthy AI landscape. Huang posted on X for the first time ever, and nine minutes later, Microsoft CEO Satya Nadella followed suit.

In parallel, according to five people familiar with the talks, OpenAI and Anthropic have lobbied regulators in Washington about Chinese open-source models. Treasury Secretary Scott Bessent and science advisor Michael Kratsios are involved. The dispute was triggered by China's pace: startups Z.ai and Moonshot AI have recently released models that rival those of American labs. Anthropic and OpenAI claim that the Chinese companies illegally scraped data from their systems to do so. → The Business Engineer, nytimes

Synthszr Take: Look at who wrote this letter: Mozilla, the Linux Foundation, Hugging Face, Mistral. This is the DNA of a movement that has believed since the nineties that open code is a public good and control is the real risk. Anthropic and OpenAI, on the other hand, were founded on the opposite premise that the technology is too dangerous to be released and is therefore safest in their hands. Between these camps lie two completely different founding myths, and that's precisely what makes the dispute so bitter. Nine minutes separated Huang's and Nadella's X posts, cleanly orchestrated by the old guard. The punchline: The young labs, posturing as security guardians, need regulation from Washington to maintain their lead over DeepSeek and Moonshot. Yesterday's idealists sell openness, today's security apostles sell fear, and now Bessent and Kratsios are supposed to decide which of the two narratives becomes American policy.

Nvidia Announces $500 Billion Partnership with SK Group for AI Memory

Nvidia CEO Jensen Huang announced a partnership with SK Group worth around $500 billion during a meeting with South Korean President Lee Jae Myung in San Francisco. According to Yonhap and ChosunBiz, it is the largest commitment in AI infrastructure to date. Neither company has yet released details. The focus is on SK Hynix, one of two suppliers that dominate the market for High-Bandwidth Memory (HBM)—the stacked DRAM memory that AI accelerators need for fast data processing. According to TrendForce, SK Hynix has postponed the launch of its HBM4 products and is focusing on margins in standard DRAM, while Samsung is already reporting over a billion dollars in revenue from HBM4. Huang also announced an investment in Naver, a collaboration with Hyundai on a self-driving Genesis, the relocation of several AI researchers from California to South Korea, and a joint Korean LLM with KAIST. → Techpresso

Synthszr Take: $500 billion isn't a number you just invest; that's roughly the gross domestic product of Norway, bundled into a bilateral supply chain commitment. And Lee met with Nvidia, OpenAI, Anthropic, and Broadcom in a single day. Seoul is selling access to capital and in return gets researchers, its own language model via KAIST, and a position that no amount of geopolitical bargaining power can outweigh: Whoever controls HBM controls the pace of the entire AI build-out. The direction of the talent flow is interesting. Previously, the brightest minds gravitated toward the Bay Area; now Nvidia is sending researchers from California to South Korea, to where the memory is manufactured. That's the real story behind the large sum. Capital no longer follows the cheapest location; it follows the scarcest resource, and right now, the scarcest resource is sitting in two Korean factories. For Europe, this raises the uncomfortable question that no one wants to answer on stage: What scarce resource do we actually have that someone would bid for with half a trillion dollars?

Microsoft Replaces OpenAI Models with In-House Models

Microsoft AI has brought two new in-house models into public preview: MAI-Image-2.5-Pro as a high-resolution image generator and MAI-Voice-2-Flash for high-volume voice applications in enterprise use. At the same time, the company is presenting production data intended to prove that it can run its own products without OpenAI's frontier models. According to Microsoft, Bing Image Creator now runs entirely on MAI-Image-2.5. In PowerPoint, the model reduces GPU costs by up to 84 percent compared to OpenAI's GPT-Image-2. For the Dynamics 365 Contact Center, used by T-Mobile and EasyJet among others, Microsoft reports GPU cost reductions of up to 89 percent. MAI-Voice-2-Flash operates twice as fast as its predecessor at $15 per million characters, while also costing 32 percent less than the previous model. In healthcare, Dragon Copilot now runs in 58 languages on MAI-Transcribe-1.5; according to the company, the application is used by 170,000 medical providers and was responsible for 28 million patient contacts in the last quarter. → MyClaw Newsletter

Synthszr Take: An 89 percent drop in GPU costs in the contact center is a number that gets every purchasing manager out of their seat, but it's not useful for your own negotiations. Microsoft is reducing internal transfer prices here across the entire vertical stack, from its own model to Azure to the final product. This isn't a rate you can negotiate as a customer. Still, it means something for procurement: If even Microsoft finds it worthwhile to ditch GPT-Image-2 from PowerPoint, then the frontier premium you're paying today for standard workloads like transcription or voice agents has become negotiable. The useful leverage lies in segmenting your own use cases: a call center with millions of calls needs latency and price per call, not a maximally expressive premium model. Separate the tasks that truly require frontier quality from the high-volume cases, and you'll have two separate budget lines with very different price pressures tomorrow. From now on, vendor lock-in with a single provider will measurably cost you more than it saves in convenience.

Hermes Goes Viral: Open Agent with a More Complete Architecture than OpenClaw

The MyClaw Newsletter reports that the open agent Hermes is currently going viral because users consider it more than just an open alternative to OpenClaw. According to the newsletter, the underlying architecture is more complete: memory, skills, workflows, model routing, local deployment, and autonomous execution are designed as a single system. This allows users to switch providers without having to rebuild their setup. The provider argues that this foundation delivers more reliability, more control over one's own data, and more customizability for serious agent use. → MyClaw Newsletter

Synthszr Take: What's interesting is the layer that's creating lock-in for people with Hermes: the fact that memory, permissions, and workflows are in ONE place that the user owns. That's the exact lock-in point. When a father builds a JARVIS for his six-year-old and can swap Gemini for another model at will without redoing the whole setup, then he holds the customer relationship in his own hands, and no provider can trap him in a closed system. That's the real shift: The agent that orchestrates the user daily owns the relationship; the model behind it becomes a utility.

Runway Launches Media Router That Automatically Selects the Right Media Model

Runway has released the Runway Media Router, a tool that automatically selects the appropriate image, video, or audio model for a request—depending on whether developers are optimizing for quality, speed, or cost. According to TechCrunch, it is the first router of its kind built specifically for generative media; such routing is already common for large language models. The router runs via Runway Dev, the developer platform launched earlier this month that provides API access to Runway's own models as well as a growing number of third-party models. Customers like Adobe, Cloudflare, ElevenLabs, Expedia, Shutterstock, and Quora are using it to build generation directly into their own products instead of directing their users to Runway's app. Head of Product Anthony Maggio cites token prices and quality as the main purchasing criteria, but also points to the option of favoring US providers while the Trump administration considers sanctions against Chinese open models. Runway's last in-house frontier video model, Gen 4.5, was from December, and its own text-to-video models are no longer at the top, according to Artificial Analysis. → AI Secret

Synthszr Take: Runway has essentially lost the model race, and the router is the honest response to that: no Gen 5 since December, others are leading in text-to-video. Instead of continuing to compete in a crowded field against Google and Chinese providers, Runway is turning itself into a crossroads where Adobe, Cloudflare, and ElevenLabs buy their generation services. This is clean aggregator logic: Whoever owns the routing earns a cut from every third-party model, including the better ones. The real asset here is the evaluation intelligence that Runway's creative team has built up over years: the judgment of which model best captures motion and lip-sync. That's harder to copy than a checkpoint. The risk lies in one place: A router lives on the trust that it routes neutrally, and Runway also routes to its own models. Whether Adobe will play along long-term will be decided by the first case where the router chooses the Runway model even though the third-party one was cheaper and better.

Midjourney Acquires Astrology App Co-Star with 4.3 Million Users

Midjourney has acquired the social astrology app Co-Star, according to Bloomberg. The terms of the deal were not disclosed. Co-Star has about 4.3 million monthly active users and primarily helps them create astrological birth charts that can be shared with friends. According to the app's description, the horoscopes and compatibility analyses are created from a mix of AI and human editing. The entire Co-Star team of about two dozen people will move to Midjourney. Midjourney, known for its AI-powered image and video generators, does not yet operate its own standalone app and serves its users through a Discord server. → StrictlyVC

Synthszr Take: Midjourney is buying two dozen people who know how to run a consumer app with 4.3 million active users. That's the real announcement here. A lab that still delivers its images via a Discord server and yet disrupts the market has simply never learned how to build a product where people want to live. That's exactly what Co-Star provides: daily returns, shared birth charts, a circle of friends that makes the app sticky. Pure image generation is a tool you open, use, and close again without forming a bond. Along with its medical branch and the spa, a bet is emerging: Midjourney wants to relegate the model to an ingredient and instead own the touchpoints where people return. Whether an astrology community is the right first building block for this foundation remains an open question, but the direction is right.

Alibaba Boosts Qwen's Daily Users to 58 Million with 3 Billion Yuan in Vouchers

Alibaba distributed three billion yuan in incentives during the 2026 Lunar New Year campaign, which users could redeem for food and drinks in the Qwen app. According to data from QuestMobile cited by Reuters, Qwen's daily active users rose from seven million to fifty-eight million during the promotional period. This represents more than an eightfold increase in just a few weeks. The vouchers were tied to redeeming real goods through the app. The figures come from third-party measurement by QuestMobile, not from official Alibaba data. The report does not state how many of the new users remained after the promotion ended. → 🔳 Turing Post

Synthszr Take: Fifty-eight million sounds like a blockbuster, until you do the math. Three billion yuan for about fifty-one million new daily users is just under sixty yuan per head, and people got the money for redeeming snacks. A purchased DAU says exactly one thing: someone wanted a voucher. It says nothing about retention, nothing about utility, nothing about whether anyone will still open the app on February 20th when there are no more freebies. The relevant number isn't in the report, and that's no coincidence: Those who celebrate subsidized peaks rarely show the curve that follows. Qwen will become interesting on the day its seven million core users grow organically to ten or fifteen million. Until then, it's just an expensive screenshot for the next investor deck.

US Embassy in Ottawa Supports Canada's AI Sovereignty, But With a Caveat Favoring US Tech

A representative from the US Embassy in Ottawa explained in a background briefing, according to Global News, that Washington supports Canada's pursuit of sovereign AI, as long as it doesn't come at the expense of superior American technology. The official said that promoting its own companies was not problematic for him, but pointed out that both governments should sell jointly where US providers are stronger. A completely domestic build-out across hardware, models, and software is not a rational goal for a country of Canada's size because the costs and timeframe would be too high. He named export controls and curbing Chinese influence in standardization bodies as areas for cooperation. Ottawa's own “AI for All” strategy under Premier Carney aims to raise AI adoption from 12 to 60 percent by 2034. It also aims to create over 250,000 jobs and contribute CAD 200 billion to the economy. The strategy lists international alliances with Germany, Australia, Europe, India, Qatar, Saudi Arabia, and the UK, but does not mention the US. AI Minister Evan Solomon nevertheless acknowledged the dependence, saying at the US-Canada Summit in Toronto: “Sovereignty is not solitude.” → Techpresso

Synthszr Take: Sovereignty that depends on the permission of the stronger party is not sovereignty. Canada lists seven partners in its strategy and leaves out the US, while its own minister admits at the summit that Washington remains the largest and most important partner. This gap between rhetoric and reality is the real issue. The US official has already answered the power question: You can serve niches, but in anything serious, we sell together, and the stack belongs to us. Canada could still retain real control, but not at the model level and hardly at the compute level, where the nine-figure sums from the US make any Canadian round look tiny. The leverage lies with the data, the process and domain data of Canadian companies that no one has to license out of the country. With 26 percent adoption enthusiasm (the lowest value in the Ipsos survey) and a target of 60 percent by 2034, this won't be decided in standardization bodies, but by the question of whether Ottawa defends its data sovereignty or sells it off in the name of partnership.

US Computer Science Enrollments Decline for the First Time in Nearly 20 Years

The number of computer science enrollments at US universities declined in the 2025-26 academic year for the first time in nearly two decades, as determined by Stanford economist Jacob Light, according to Business Insider. A separate survey by the National Student Clearinghouse Research Center found an 8.1 percent drop in the related field of Computer and Information Sciences for fall 2025. Light emphasizes that the decline is not necessarily a direct result of AI. The changing nature of learning due to generative AI, shifted expectations about the returns on computer science skills, and a general weakness in the tech job market could also play a role. He explicitly treats the patterns as a descriptive finding, not as causal proof. An analysis by the Washington Post suggests that some of the lost enrollments are shifting to adjacent fields like Data Analytics and Data Science. The job market for programming jobs has become more uncertain, at least in large corporations: in 2026, over 120,000 tech jobs have already been cut, according to Layoffs.fyi, a large portion at Meta and Google, which are ramping up their AI spending. A Challenger report cites AI as the reason for more than 54,000 layoffs in the previous year. → Techpresso

Synthszr Take: The interesting part will come in five years, when this cohort would typically be entering the market as juniors. Do the math: 8.1 percent fewer enrollments today means noticeably thinner graduate cohorts around 2030, precisely at the moment when the seniors being laid off today are slowly thinned out and no one is coming up behind them. Companies are currently cutting the training level where they used to shape a mediocre coder into someone with product judgment. AI generates runnable code, but the ability to decide whether a feature is ten times better or just noise comes from years of working on a real system, not from a prompt. If you streamline away those years, by 2031 you'll be missing the generation that can direct the agents meaningfully at all. The smarter students already sense this and are gravitating toward data science and domain knowledge, where intent matters, not syntax. Cutting junior training positions brings a three-year cost advantage and leaves a ten-year skills gap.

Prime Video Shows First Teaser for 'Blade Runner 2099' Starring Michelle Yeoh

Prime Video released the first teaser for the sequel series 'Blade Runner 2099' at San Diego Comic-Con, offering a first look at the continuation. The series is executive produced by 'Blade Runner' director Ridley Scott and takes place 50 years after the events of Denis Villeneuve's 'Blade Runner 2049'. In this future, the synthetic humans, Replicants, have taken control of Los Angeles. Michelle Yeoh plays Olwen, a Replicant and Blade Runner herself, who is tasked with hunting down a group of rogue Replicants while having only a few days left to live. → StrictlyVC

Synthszr Take: Ridley Scott has Michelle Yeoh playing a Replicant who hunts other Replicants, and of all things, the Voight-Kampff test is coming back. When 'Blade Runner 2049' was released in 2017, the question of how to separate human from machine was a pure thought experiment for cinephiles. Nine years later, we're arguing about watermarks for AI texts, about personality rights for synthetic voices (Scarlett Johansson vs. OpenAI was exactly this scene), and about whether a model can claim something like an identity. The series is launching into a world that is currently playing out its own premise in reality, which is why a teaser like this hits differently in 2026 than it used to: The audience is sitting in the test room itself. Science fiction was long the place where we could shelve questions for which reality was not yet ready.