China, USA, Europe — Everyone Against Everyone

DeepSeek Needs Money and is Building Data Centers

DeepSeek is suddenly seeking outside capital – at least $300 million – at a valuation of over $100 billion. The Chinese AI company, which for years positioned itself as a counter-narrative to the venture-driven AI industry, is facing its first funding round. Founder Liang Wenfeng had consistently rejected all external financing since 2023, funding the company with profits from his quantitative hedge fund Huanfang. This reversal comes at a critical time: the next flagship model, V4, has been postponed multiple times (originally planned for February, now for April), while core developers are leaving for Xiaomi, Tencent, and other competitors. DeepSeek is now looking for data center managers and hardware engineers. The company plans to build its own data centers in Ulanqab, one of China's eight national data center hubs in Inner Mongolia – from planning and construction to operation. The region offers ideal conditions: average temperatures that allow for natural cooling, 65% renewable energy, and established infrastructure from Alibaba and Huawei. With a V4 model announced for April, which will supposedly process trillions of parameters and millions of tokens of context, DeepSeek seems to have realized that algorithmic efficiency alone is no longer enough. → Hello China Tech

Synthszr Take: DeepSeek is currently experiencing what biologists call the “Allee effect”: beyond a critical size, independence flips from an advantage to a disadvantage. The hedge fund background initially enabled a kind of research monastery – no quarterly reports, no growth metrics, just AGI dreams. But without a market-based valuation, employee options become Monopoly money, while Alibaba pumps 3.8 trillion yuan and ByteDance 1.5 trillion yuan into AI infrastructure. The technical irony: V4 is supposed to run entirely on Huawei's Ascend 950PR chips – a geopolitical statement that probably consumes more development resources than the actual model architecture. DeepSeek wanted to prove that AI development can be decoupled from Silicon Valley logic, yet it ends up exactly where everyone else does: at the term sheet.

Nvidia CEO Warns: DeepSeek on Huawei Chips Would Be a Catastrophe for America

Jensen Huang, CEO of Nvidia, described a potential scenario on the Dwarkesh podcast as a “horrible outcome” for the US: DeepSeek, China's leading AI lab, optimizing its upcoming models for Huawei chips instead of American hardware. The company is preparing to release its multimodal foundation model V4, which is slated to run on Huawei's new Ascend 950PR processor. The real threat lies in the software migration: DeepSeek is completely rewriting its code, moving away from Nvidia's CUDA framework to Huawei's CANN platform. This dependency on CUDA has so far served as a second layer of US control over global AI development, beyond the hardware itself. While Huawei's chips achieve only 60% of the performance of Nvidia's H100 (which is already two generations old), Huang argues that China could compensate for this disadvantage with “abundant energy” and a large pool of AI researchers. → Techpresso

Synthszr Take: Huang understands something that proponents of export controls overlook: the real power lies not in the transistors, but in the development environments. CUDA is Nvidia's real moat, a twenty-year-old codebase in which every AI researcher thinks and works. DeepSeek's move to CANN is reminiscent of the early days of the internet when China built its own protocol ecosystem parallel to the Western web. The hardware gap (a factor of 5 today, projected to be a factor of 17 by 2027) could prove less decisive than software sovereignty. Ironically, the US sanctions are accelerating the very thing they were meant to prevent: the emergence of a fully independent Chinese AI stack. Huang warns of a “horrible outcome,” but perhaps it has already happened.

Mistral Wants to Become the Airbus of the AI World

Arthur Mensch isn't selling Mistral as the best AI model in the world, but as the most independent one. The 33-year-old CEO from Paris has realized that European companies and governments prefer a weaker model if it means their data doesn't travel to California or Beijing. Mistral lags far behind OpenAI and Anthropic in performance rankings—the French company's best model even loses to a nine-month-old version of Claude. Nevertheless, the company achieved a revenue of $200 million in 2025 and a valuation of $14 billion. German state governments are ditching Microsoft Office, and France is developing its own Zoom alternative—in this political climate, Mistral is positioning itself as a sovereign European option. The open-weight models can be run completely offline, customers can customize the code, and Mistral engineers install the systems directly on-site. → Forbes

Synthszr Take: Mistral is following the Airbus playbook: initially technically inferior, but politically indispensable. While American AI labs compete over abstract superintelligence, Mensch sells digital sovereignty to nervous institutions. The model is reminiscent of the early days of the internet when countries launched their own search engines and social networks—most of which failed due to network dynamics. Mistral circumvents this problem with a B2B focus: federal agencies and DAX companies don't need viral features, they need compliance documents. The $14 billion valuation reflects less technical excellence and more a geopolitical risk premium. Mistral is the insurance policy against a world where Sam Altman or Xi Jinping dictates the terms.

OpenAI's Investors Doubt Sam Altman's Suitability for an IPO

OpenAI's shareholders have fired Sam Altman once before and are considering doing it again. According to The Wall Street Journal, some investors are questioning whether Altman is the right man to take the company public. Some OpenAI shareholders are internally suggesting Board Chairman Bret Taylor as a potential CEO successor to Sam Altman. The planned IPO, with a valuation of around $850 billion, is approaching, while concerns about Altman's external investments and potential conflicts of interest are growing. Taylor, former co-CEO of Salesforce and current chairman of the board at OpenAI, is considered an experienced tech manager without the controversial side projects that Altman pursues. The discussions are taking place behind closed doors as OpenAI prepares for the biggest tech IPO in history. Altman himself has not yet commented on the speculation. → StrictlyVC

Synthszr Take: OpenAI is going through the classic founder's dilemma that Paul Graham described years ago: the visionary founder who built the company becomes a risk factor for the next stage of growth. Taylor would be the safe choice, a manager without side projects in Worldcoin or energy startups, who could navigate the $850 billion valuation through an IPO. The irony: Altman's scattered attention between his various ventures was likely what gave OpenAI the creative energy for its breakthrough. Now that institutional money is on the line, this unpredictability becomes a liability. OpenAI faces a choice between the magician who creates products and the manager who sells them.

Sam Altman is Teaming Up with Tinder

Sam Altman's side hustle “World” is expanding into the mass market. After a successful pilot project in Japan, Tinder is integrating the iris-scanning technology worldwide, including in the US. Users can get verified using the spherical Orb scanners and receive a World ID emblem on their profile. Tools for Humanity (TFH), the company behind World, is planning further integrations into ticketing systems, email services, and corporate organizations. The technology uses zero-knowledge proofs to distinguish humans from AI-generated profiles without sacrificing anonymity. In parallel, World is launching a “Concert Kit” for artists who want to sell tickets exclusively to verified individuals – as a protection against scalper bots. → Techpresso

Synthszr Take: World is solving a problem that doesn't exist yet by making the solution itself the problem. The iris scanners are reminiscent of the door control at exclusive clubs, except here the bouncer is a silver sphere and the club is intended to be the entire internet. In return, Tinder users get a checkmark for having handed over their biometric data to a startup run by someone who is simultaneously building the AI systems he claims to be protecting us from. The irony is that dating apps – places of maximum self-presentation and creative truth-bending – are becoming the testing ground for “proof of human.” Altman is selling us the smoke detector while playing with gasoline in the basement.

'Vulnpocalypse': You Don't Need a Glasswing Club Membership

Earlier this month, Anthropic introduced Claude Mythos as being too dangerous for the public, locking the model behind a vetted consortium of tech giants. US Treasury Secretary Scott Bessent and Fed Chair Jerome Powell convened an emergency meeting with Wall Street CEOs, and the word “Vulnpocalypse” circulated in security circles. Now, a research team from Vidoc Security has complicated this narrative: they replicated Anthropic's publicly patched examples using GPT-5.4 and Claude Opus 4.6 in an open-source coding agent called opencode. No Glasswing access, no private API, no internal Anthropic stack. Both models rediscovered two bugs in all three runs, and Claude Opus 4.6 also found an OpenBSD bug three times. “The economics of vulnerability discovery are changing,” wrote researcher Dawid Moczadło on X. → Decrypt

Synthszr Take: Vidoc Security just proved that Anthropic's moat is a mirage. The replication of Mythos capabilities with public models is reminiscent of the history of the atomic bomb: the Manhattan Project cost billions, but once the physics was known, dozens of countries built their own bombs. The difference: with AI security vulnerabilities, proliferation doesn't take decades, but weeks. Anthropic isn't selling superior technology, but access to a club (Glasswing) whose exclusivity is already crumbling. The true resource won't be the model, but the validation: who can filter the real hits from thousands of AI-generated vulnerability candidates? The business model is shifting from model access to quality assurance.

Salesforce Goes Headless and No Longer Needs Heads

At its TDX conference, Salesforce unveiled the most radical architectural transformation in its 27-year history: “Headless 360” makes the entire platform accessible via APIs, MCP tools, and CLI commands, allowing AI agents to operate the system entirely without a browser. Over 100 new tools are immediately available to developers. The initiative answers the existential question of whether companies even need a CRM with a graphical user interface in a world with reasoning-capable AI agents. Salesforce's clear answer: No. The software sector is in a sell-off; the iShares Tech-Software ETF has fallen 28% since September, driven by fears that large language models could make traditional SaaS business models obsolete. Jayesh Govindarjan, EVP at Salesforce and architect of Headless 360, explains the painful realization: early Agentforce customers were afraid to touch their productive agents because any change could destabilize the entire system. → VentureBeat

Synthszr Take: Salesforce is making a virtue of necessity, transforming itself from a gatekeeper to a substrate. The strategy is reminiscent of converting old industrial buildings into loft apartments: what was once designed as a closed factory is being rebuilt into open infrastructure where others can design their own spaces. Salesforce resolves the technical tension between probabilistic agents and deterministic business requirements with “Agent Script,” a domain-specific language that functions like a thermostat: it defines fixed temperature ranges (business rules) within which the system can regulate freely (LLM reasoning). Govindarjan's distinction between static graphs for customer interactions and dynamic “Ralph Wiggum loops” for employee agents shows that companies need two different nervous systems: a sympathetic one for controlled external communication and a parasympathetic one for creative internal processes. The irony is perfect: Salesforce is opening its platform to the very agents that could theoretically replace it, thereby making itself the indispensable operating system of the agentic era.

Mac Mini is Sold Out and Becomes the Mac Claw

Apple's smallest Mac has evolved from a niche product to a sought-after AI server. The Mac Mini accounted for only 3 percent of Mac sales in 2025, but in the last six months, demand for models that support local language models like OpenClaw has exploded. Models with 32 GB and 64 GB of RAM are “currently unavailable” on Apple.com, and delivery times for other configurations are up to 12 weeks. The unexpected demand caught Apple off guard: “Apple was surprised by the number of people buying Minis for Clawdbot,” says IDC analyst Francisco Jeronimo. While MacBook Pros with 128 GB of RAM are available within days, the Mini is struggling with its success as a cost-effective alternative to cloud services. Analysts suspect a combination of underestimated demand, upcoming M5 updates, and a global memory shortage as the causes. → www.wsj.com

Synthszr Take: The Mac Mini is experiencing its second birth as a decentralized data center, similar to how the PlayStation 3 once became a scientific supercomputer. Apple has unintentionally created a product that is perfect for our times: compact, energy-efficient, with unified memory, and ideal for language models. The irony: a device that Apple designed as an entry-level desktop is becoming a tool for the decentralization of AI infrastructure. What's happening here is reminiscent of the early days of the internet when universities ran their Unix workstations as web servers at night. The 12-week delivery time is more than just a bottleneck story; it shows how quickly usage patterns can shift when the right hardware meets the right use case. Apple may have inadvertently laid the foundation for an ecosystem of private AI infrastructure.

ChatGPT Becomes an E-Commerce Powerhouse

OpenAI is enhancing ChatGPT with visual shopping features, turning the chatbot into a fully-fledged e-commerce assistant. Users can now browse products directly in the chat interface, visually compare them, and view detailed product information like prices, reviews, and features side-by-side. The Agentic Commerce Protocol (ACP) provides the technical infrastructure for this integration, connecting merchants directly with ChatGPT. Instead of jumping between browser tabs and sifting through the same “best of” lists, the entire product search now happens in a single conversational flow. The feature is rolling out this week to all ChatGPT users, from the free version to Pro subscribers. → Techpresso

Synthszr Take: OpenAI is making the classic e-commerce funnel obsolete. The Agentic Commerce Protocol works like a reverse PageRank system: instead of crawling and ranking web pages, merchants can actively integrate themselves into ChatGPT. The parallel to Apple's App Store is obvious—if you're not in the system, you don't exist for the user. The crucial difference: while Amazon pushes users toward a purchase decision, ChatGPT optimizes for conversation. This shifts power from SEO optimizers to those who can best structure their product data. OpenAI is betting that “chatting with a bot about shirts” is more natural than using filter masks and categories—and in 2026, that might actually be true.

Tokenmaxxing Instead of Looksmaxxing Doesn't Make it Better

Developers are bragging about their AI token budgets like teenagers with their credit cards: the higher the limit, the cooler. Waydev CEO Alex Circei has now analyzed data from 10,000 developers and discovered a phenomenon that gives every product manager nightmares. While developers immediately accept 80-90% of the code generated by Claude Code, Cursor, or Codex, they have to rewrite 70-90% of it in the following weeks. The actual acceptance rate is a meager 10-30%. GitClear confirms the drama: AI users cause 9.4 times more “code churn” (deleted vs. added lines) than their colleagues without AI tools. Faros AI even reports an 861% increase with high AI adoption. Jellyfish puts it succinctly: developers with the largest token budgets achieve twice the throughput rate, but consume ten times the amount of tokens. → StrictlyVC

Synthszr Take: Token budgets are the new status symbol trap of the tech industry, comparable to the prestige office on the 45th floor that nobody uses. The pattern is reminiscent of the dot-com bubble, when companies celebrated their burn rate as a success metric: input became more important than output. Physics calls this an entropy problem: pumping in more energy only increases chaos, not order. Junior developers accept a particularly large amount of AI code, creating their own technical debt crisis, while experienced developers use the tools more sparingly. The industry is measuring progress by the tokens consumed rather than by maintainable code, a classic case of Goodhart's Law: once a measure becomes a target, it ceases to be a good measure.