Anima Anandkumar Turned Down Bezos Billions to Build Physics AI Instead

Caltech professor Anima Anandkumar and MIT-trained engineer Benedikt Jenik turned down a board seat, a $2 million salary, and a stake in Jeff Bezos-backed Project Prometheus to build Accelerated Understanding, an AI system based on neural operators instead of Transformers. The company claims its model processed 5 trillion data points in a single prompt, roughly 5 million times larger than what flagship models from Anthropic and Google typically handle. Project Prometheus, which raised $12 billion in its Series B in June, is focused on automating manufacturing of complex physical systems.

Two AI researchers turned down a board seat. They also said no to a $2 million salary and a stake in Jeff Bezos-backed Project Prometheus - and chose to build a physics AI instead. Accelerated Understanding says its new model, built on neural operators rather than Transformers, processed 5 trillion data points in a single prompt. A Caltech professor and an MIT-trained infrastructure engineer turned it down anyway: a board seat, a $2 million salary, a stake in a $12 billion Jeff Bezos-backed company. That's not a small ask to refuse. They bet instead that the next leap in AI won't come from making chatbots smarter, but from teaching machines to understand physics. If you've spent the past three years assuming every serious AI bet runs through a bigger Transformer, this one doesn't fit that story. Anima Anandkumar and Benedikt Jenik just unveiled Accelerated Understanding, an AI system that skips the Transformer architecture powering ChatGPT, Gemini and Claude entirely. In its place sits a neural operator. It's a mathematical approach Anandkumar helped pioneer during her years at Nvidia, built to learn how physical systems evolve across space and time rather than how words follow other words. The claim that got people talking: in testing, the model ingested 5 trillion data points in a single prompt, according to the company, a figure it says is roughly 5 million times larger than what flagship models from Anthropic and Google typically handle. That's not a language benchmark. It's an attempt to simulate reality itself, at a scale no text-based model was ever built for. Late in 2024, Jeff Bezos and biotech entrepreneur Vik Bajaj came calling with an offer few founders would refuse. As Reuters reported in an exclusive account, picked up by outlets including the Japan Times and Techstartups, Bezos and Bajaj wanted Anandkumar as the public face and scientific leader of what became Project Prometheus. Jenik would sit as a board observer. The terms were staggering: a combined 35% equity stake, a salary starting at $1 million a year and rising to $2 million after three months, and a Series B round of more than $2 billion with Bezos contributing his own capital. Instinct's AI Assistant Grabbed User Data Rights, Then Raised $250 Million https://startupfortune.com/instincts-ai-assistant-grabbed-user-data-rights-then-raised-250-million/ Instinct, the viral AI life assistant from 23-year-old founder Noah Shinn's Spear Street Technology, granted itself a "perpetual and irrevocable" license to user data before revising its terms under pressure. Two days after TechCrunch exposed the backlash, the company still raised $250 million at a $2.5 billion valuation. - ai assistant data rights controversy explained https://startupfortune.com/instincts-ai-assistant-grabbed-user-data-rights-then-raised-250-million/ - how startups use user data for fundraising https://startupfortune.com/instincts-ai-assistant-grabbed-user-data-rights-then-raised-250-million/ They said no. Project Prometheus went on to raise $12 billion in its Series B this past June, according to Reuters, and it's now focused on building AI that can automate the manufacturing of complex physical systems. Anandkumar and Jenik kept building Accelerated Understanding on their own instead, without Bezos's money or his name attached. Why turn down that kind of financial security? Anandkumar spent five years at Nvidia, from 2018 to 2023, working under Jensen Huang on the neural-operator research that now sits at the center of her startup's technology. She clearly believes the architecture is worth more outside someone else's org chart than inside it, even one funded by Bezos. Betting against the Transformer Every major AI lab right now is racing to scale the same basic idea: predict the next token, then the next, then the next. Accelerated Understanding is making the opposite bet. Its model works in four dimensions, three of space plus time. It's aimed squarely at problems where language models fall apart: chip design, robotics, extreme-weather forecasting and reading geological data for energy companies. Anandkumar has declined to say how much funding Accelerated Understanding has raised or which hardware providers it relies on. That's a notable silence for a company making claims this large. It's worth reading with some skepticism until independent benchmarks or paying customers confirm what the 5-trillion-data-point figure actually buys a chip designer or a utility company in practice. Still, the signal matters more than the number. Nvidia gave Anandkumar a platform for neural-operator work years before the current physical AI rush. Turning down a $2 billion war chest to chase a non-Transformer architecture instead is a vote of confidence from inside the industry, not a marketing claim from outside it. That's the tell. Frankly, the AI industry has spent three years treating scale-the-Transformer as the only serious strategy worth funding. Not everyone bought it. Accelerated Understanding shows that some of the people who helped build that consensus don't fully believe it anymore. US Startups Are Quietly Replacing OpenAI and Anthropic With Chinese AI https://startupfortune.com/us-startups-are-quietly-replacing-openai-and-anthropic-with-chinese-ai/ Chinese open-source AI models now carry more than 60% of OpenRouter's traffic, up from roughly 30% a year ago, as Airbnb, Perplexity and Nvidia turn to models like Alibaba's Qwen and DeepSeek to cut inference costs by 60% to 90%. Andreessen Horowitz says the shift is reshaping how startups build, even as Congress investigates the data risk. - US startups replacing OpenAI with Chinese AI models https://startupfortune.com/us-startups-are-quietly-replacing-openai-and-anthropic-with-chinese-ai/ - why American companies switching to Alibaba Qwen instead https://startupfortune.com/us-startups-are-quietly-replacing-openai-and-anthropic-with-chinese-ai/ Whether neural operators actually out-predict the specialized physics models built over decades for narrower jobs remains unproven outside the company's own tests. Nobody knows yet. But the founders who walked away from Bezos's money now have every incentive to prove it fast. Also read: Claude Code's Usage Limit Boost Ends September 14 With a Real Cut https://startupfortune.com/claude-codes-usage-limit-boost-ends-september-14-with-a-real-cut/ • Rillet turned an unsolicited board update into a $1 billion accounting startup in 48 hours https://startupfortune.com/rillet-turned-an-unsolicited-board-update-into-a-1-billion-accounting-startup-in-48-hours/ • ServiceNow patches three maximum severity flaws inside its AI agent platform https://startupfortune.com/servicenow-patches-three-maximum-severity-flaws-inside-its-ai-agent-platform/