The guests are Akshay Kothari, Co-founder of Notion, the collaboration platform now built for humans and agents alike; Azalia Mirhoseini, Co-founder of Ricursive Intelligence; Anna Goldie, Co-founder and CEO of Ricursive Intelligence, the frontier lab using AI to design AI chips; and Vijay Krishnan, Co-founder of Turing, which supplies the expert data used to train frontier models.
Kamath opens by mapping the AI stack as he sees it: chips and data centres, frontier labs, neo labs, data trainers, inference providers, applications and, as the panel adds, services. The conversation then moves through the economics of the boom, the slow spread of AI beyond Silicon Valley and what countries such as India should build next.
The central debate is whether the money adds up. One panellist puts committed AI spending this year at around USD 700 billion, against roughly USD 100 billion in enterprise AI revenue. The bet, the argument goes, is not on software, which accounts for about 1% of US GDP, but on services and knowledge work, which together account for far more.
Krishnan adds a challenge of his own: name one Fortune 500 company outside the AI business, “a Pepsi or a Coke”, that has meaningfully grown revenue or cut costs. Yet he is equally clear that software teams can now safely run at half to 60% of their former sizes, calling it “a giant productivity improvement” that nobody in the Bay Area disputes.
Goldie questions what should count as AI revenue at all. At Ricursive, she says, AI coding tools have made it possible to hire only a small team of deep experts, a strategy that would not have worked a few years ago.
Kothari's explanation for the gap is diffusion. Capability is growing exponentially, but most people outside the Valley still use AI “as a better search tool”. One study he cites finds that the top 1% of companies spend ten times more on AI than the top 10%.
New technologies have always taken time to absorb, he says. The first printing presses copied handwriting, and the first films were recordings of stage plays. He separates two ways of using AI: an efficiency branch, doing yesterday's work with 30% of the people, and an ambition branch, where a team of a hundred does five times as much. Today, “the biggest barrier is just imagination.”
Mirhoseini starts at the bottom of the stack. AI inference demand is growing 10 to 100 times, and on a conservative estimate, she says, the world will need 100,000 times more compute in five years than it has today.
“We are not going to get that simply by just scaling our data centers,” she says.
Ricursive's answer is AI that designs custom chips for specific workloads, a way out of the “hardware lottery” in which chips are built for today's models and models are built for today's chips.
The two founders have done this before. AlphaChip, the AI system they built at Google for chip layout, has now been used across five generations of TPUs, Goldie says, as well as Axion, Pixel phones and Waymo chips. The system has also been adopted by MediaTek.
Krishnan explains the less visible work of making models smarter. A 2022 OpenAI paper showed that a small amount of high-quality, targeted data could improve a model substantially. That made it worth hiring experts to write problems today's models cannot yet solve.
Turing draws on more than 5 million knowledge workers across 70-plus countries, from software engineers and PhDs in physics and chemistry to lawyers and doctors. The catch, he says, is that “whatever it is we are doing today will be obsolete nine months from today because all models have learned it.”
Kothari brings the view from the application layer. At the start of the year, about 1% of Notion's users were agents; today, that figure is 30%. People who use Notion through other agents, without ever opening it, are two to three times more active because users want a neutral place for their context as they switch between models.
Companies, he notes, have moved within six months from “token maxing” to “token economics”, choosing cheaper models wherever the frontier isn't needed. In software, he says, engineers in the Valley have gone in under a year from writing code to managing agents.
Goldie offers an insider's view of the companies that shaped the field. She was at Google Brain the night the Transformer team got its breakthrough result and watched Google hold back an early chatbot over fears of bad press before its creators left to start Character.AI.
Nvidia's moat, she argues, is no longer CUDA, its software platform, since models can now write assembly directly. It is the racks, the ecosystem and the scale.
Kamath brings ideas of his own to the table. If AI agents can move money, he asks, will humans become a business model, a “human friction layer” that sits between AI and execution, the way a nuclear launch still needs a person to turn the key?
Krishnan points to a recent Stanford Business School talk arguing much the same: even a little friction becomes the rate limiter, which may be why no AI optimist is forecasting double-digit US growth.
Kamath expects sovereign AI to follow the path of social media, with governments wary of foreign companies holding their citizens' attention and pushing for domestic alternatives. He says anything he writes now would sit behind a paywall rather than be absorbed by a model, because what travels is “the principle of my idea with the anonymized author”, not the author.
He is candid about his own experience too. Last year, he recalls, Sam Altman told him to buy luxury names like Hermès and LVMH, on the theory that rising productivity would send people looking for brands that set them apart. Those stocks have since fallen.
Kamath also admits that leaning on AI has made him worse at things he once did easily. Krishnan's bigger worry is children: if they never build basic skills, “it is very difficult for me to see what value they can add over a model at all.”
The conversation also turns to risk. Krishnan compares the logic of autonomous weapons to the game theory of the nuclear age, while Mirhoseini and Goldie, both early Anthropic employees, say they believe the safety concerns voiced by its leadership are sincere.
For India, the advice is practical. The first answer comes immediately: “Build compute. That's a no-brainer.”
Kamath frames the stakes: as AI reshapes services, India will need something to fill the hole that IT outsourcing leaves behind. One guest calls it a missed opportunity that Indian IT companies have not become the world's Turings, given how much of this work is already done by people in India.
Kothari says it is still too hard to start a business there, with a single angel investment taking three months of paperwork and notarisation.
Goldie points to a Stanford collaborator who was selected by the Indian government and given thousands of GPUs to build a speech model for Indian languages. Kamath points to Skyroot, the Indian space company US investors are now competing to back.
Space, the panel agrees, shows what India does best: “take something that the world is doing and actually do it at like a hundredth of the cost.”
Throughout the episode, Kamath returns to one question: if he had USD 100 million, where in the AI stack should he build?
Mirhoseini points to compute efficiency: “If you're spending USD 100 million on a model, but you're saying, I'm 10x more efficient than my competition, then this is as if you're spending a billion dollar on your model.”
Goldie's answer is blunter. Founders should ask what has to happen that only they can do, and if they can't answer that, “maybe you shouldn't start a company.”
Other parting suggestions include a university for product design in India and, from Krishnan, an AI tool that lets a child learn at ten times the pace. Kamath's own instinct is social media, which he believes will turn the other way: “The new social media will almost be like, get off your device.”
The episode is available now on YouTube.
People by WTF is a global podcast platform hosted by Nikhil Kamath, featuring in-depth conversations with leaders across business, policy, technology, culture and academia. The show explores long-term institutional, technological and economic questions shaping global society through candid, high-signal dialogue.
Past guests include Elon Musk, Prime Minister Narendra Modi, Bill Gates, Ajay Banga, Martin Escobari, Rishi Sunak, Akshata Murty and Ranbir Kapoor.
Guest Information
Akshay Kothari is Co-founder of Notion, the collaboration platform for work. Kothari grew up in Ahmedabad and studied at Purdue and Stanford. He co-founded Pulse, a news app acquired by LinkedIn, and went on to run LinkedIn in India for three years. An early investor in Notion, he joined full-time in 2018, when the team was about eight people.
Azalia Mirhoseini is Co-founder of Ricursive Intelligence, a frontier lab founded in 2025 that uses AI to design custom chips for AI workloads. Mirhoseini did her PhD at Rice University and joined Google Brain in 2016, where she co-led AlphaChip. She was among Anthropic's first 30 employees, later worked part-time on Gemini, and is CS faculty at Stanford, where she founded the Scaling Intelligence lab.
Anna Goldie is Co-founder and CEO of Ricursive Intelligence. Goldie studied computer science and linguistics at MIT and did her PhD at Stanford on reinforcement learning for chip design and LLM post-training. She joined the first cohort of the Google Brain residency, co-developed AlphaChip and was an early employee at Anthropic.
Vijay Krishnan is Co-founder of Turing, which helps frontier AI labs improve their models with expert data and helps enterprises deploy AI. Krishnan studied at IIT Bombay and Stanford, where he worked with Chris Manning and Andrew Ng. His first company, in personalised search and content recommendations, was acquired. Turing began by building remote engineering teams and, since 2022, has worked with frontier labs, starting with OpenAI.
Key Themes Discussed
The discussion covered the AI stack, from chips and data centres to frontier labs, neo labs, data, inference, applications and services; why compute demand is outpacing supply, and how AI-designed custom chips could close the gap; how expert human data trains frontier models, and why it goes out of date within months; agents as users, and the shift from “token maxing” to “token economics”; whether AI investment is a bubble, and why productivity gains aren't yet showing up in GDP; the slow diffusion of AI beyond Silicon Valley; what Google, Nvidia and the frontier labs got right and wrong; open-weight models, distillation and sovereign AI; what India and other countries should build to share in the AI boom; and AI risk, the future of devices and interfaces, and what AI means for learning and skills.
About The Author
Comment List