Overview

In a San Francisco‑recorded episode of the WTF podcast, investor Nikhil Kamath brings a skeptical lens to the AI boom, interviewing four founders across the AI stack: Akshay Kothari, co‑founder of Notion; Azalia Mirhoseini and Anna Goldie, co‑founders of Ricursive Intelligence; and Vijay Krishnan, co‑founder of Turing.

Core Assertions on AI Spending and Revenue

The panel puts committed AI spending this year at roughly USD 700 billion, while enterprise AI revenue is estimated at about USD 100 billion. The argument is that the upside is not in software—software accounts for only about 1 % of U.S. GDP—but in services and knowledge work, which together represent a far larger share of economic activity.

Productivity Claims

Krishnan notes that software teams can now operate at 50 % to 60 % of their former headcount, a productivity improvement that he says is undisputed in the Bay Area. Goldie adds that AI‑coding tools enable firms to staff only a small core of deep experts, a model that would have been infeasible a few years ago.

Diffusion and Adoption

Kothari attributes the gap between spending and revenue to diffusion: while capability is expanding exponentially, most firms outside Silicon Valley still treat AI as “a better search tool.” He cites a study showing the top 1 % of companies spend ten times more on AI than the top 10 %. He also distinguishes two AI use‑cases—efficiency (doing yesterday’s work with 30 % of the people) and ambition (a hundred‑person team achieving five‑fold output).

Compute Demand Explosion

Mirhoseini warns that AI inference demand could grow 10‑ to 100‑fold, and conservatively the world will need 100,000 times more compute in five years than today. She argues that scaling data centres alone will not meet this need.

Custom Chip Design as a Solution

Ricursive Intelligence’s answer is AI‑designed custom chips. Their system AlphaChip, originally built at Google for chip layout, has been deployed across five generations of TPUs, as well as in Axion, Pixel phones, Waymo chips, and adopted by MediaTek. This approach aims to escape the “hardware lottery” where chips are built for today’s models and models for today’s chips.

Human‑Generated Training Data

Krishnan explains that high‑quality, targeted data can dramatically improve models, citing a 2022 OpenAI paper. Turing leverages more than 5 million knowledge workers in over 70 countries, ranging from software engineers to doctors, to supply such data. He cautions that today’s work will become obsolete within nine months as models learn it.

Application Layer Shifts at Notion

Kothari reports that agent‑based usage at Notion rose from about 1 % of users at the start of the year to 30 % today. Users interacting with Notion through agents, without opening the app, are two to three times more active. Companies have moved from “token maxing” to “token economics,” opting for cheaper models when frontier performance is unnecessary. Engineers in the Valley have transitioned from writing code to managing agents in under a year.

Industry Perspectives on Moats

Goldie argues that Nvidia’s moat is no longer its CUDA software platform—models can now generate assembly directly—but rather its racks, ecosystem, and scale. She also recounts being present at Google Brain when the Transformer breakthrough occurred and witnessing Google’s early chatbot being withheld over PR concerns.

Societal and Risk Considerations

Krishnan likens autonomous‑weapon logic to nuclear‑age game theory, while Mirhoseini and Goldie (both former Anthropic employees) affirm that Anthropic’s leadership takes safety concerns seriously. Krishnan worries that children lacking basic skills may struggle to add value beyond models.

Implications for India

The panel’s advice for India is straightforward: “Build compute.” They note that AI‑driven services will erode traditional IT outsourcing, creating a missed opportunity for Indian firms to become the world’s “Turings.” Kothari highlights the difficulty of launching a startup in India, where a single angel investment can take three months of paperwork and notarisation. Goldie mentions a Stanford collaborator selected by the Indian government and provided thousands of GPUs to develop a speech model for Indian languages. Kamath points to Skyroot, the Indian space company now attracting U.S. investors, as an example of India excelling by delivering global‑level capability at a fraction of the cost.

Investment Allocation Question

When asked where he would deploy USD 100 million in the AI stack, Mirhoseini suggests focusing on compute efficiency—spending $100 million on a model that is ten‑times more efficient is equivalent to a $1 billion spend. Goldie advises founders to ask whether only they can solve a problem; if not, they should perhaps not start a company. Other suggestions include establishing a university for product design in India and creating an AI tool that enables a child to learn ten times faster.

Closing Thoughts

Kamath predicts that the next wave of social media will encourage users to “get off your device,” reversing current trends. He also notes that his own advice would now sit behind a paywall, as the “principle of the idea with the anonymised author” is what travels, not the author’s identity.

---