Overview

In a new episode of the WTF podcast recorded in San Francisco, host Nikhil Kamath brings together four AI‑industry founders—Akshay Kothari (Co‑founder, Notion), Azalia Mirhoseini (Co‑founder, Ricursive Intelligence), Anna Goldie (Co‑founder & CEO, Ricursive Intelligence) and Vijay Krishnan (Co‑founder, Turing)—to debate whether the current AI boom is a genuine productivity revolution or a speculative bubble.

Core Economic Estimates

The panel puts total AI‑related spending for the year at roughly USD 700 billion, while enterprise AI revenue is estimated at about USD 100 billion, highlighting a ten‑fold gap between investment and realized revenue. One speaker notes that software accounts for only 1 % of U.S. GDP, suggesting that the real upside lies in AI‑enabled services and knowledge work rather than pure software licensing.

Productivity Claims

Vijay Krishnan asserts that software teams can now operate at 50 % to 60 % of their former headcount, describing this as a “giant productivity improvement” that is widely accepted in the Bay Area. Anna Goldie adds that AI‑coding tools allow firms to staff only a small core of deep experts, a model that would have been infeasible a few years ago.

Stack Mapping and Diffusion

Kamath outlines the AI stack as: chips & data centres → frontier labs → neo labs → data trainers → inference providers → applications → services. Kothari emphasizes diffusion, citing a study where the top 1 % of firms spend ten times more on AI than the top 10 % and noting that most companies outside Silicon Valley still treat AI as “a better search tool.”

Compute Demand and Custom Chips

Azalia Mirhoseini warns that global compute demand could increase 100,000‑fold over the next five years, a scale that cannot be met by simply expanding existing data centres. Ricursive Intelligence’s answer is AI‑designed custom chips that break the “hardware lottery.” Their AlphaChip system, originally built at Google Brain, has been deployed across five generations of TPUs, as well as in Google’s Axion, Pixel phones, Waymo chips, and has been adopted by MediaTek.

Data‑Driven Model Improvement

Krishnan explains that high‑quality, targeted data can dramatically improve model performance, citing a 2022 OpenAI paper. Turing now leverages more than 5 million knowledge workers in 70 plus countries—including software engineers, PhDs, lawyers and doctors—to supply such data, but warns that today’s work will become obsolete within nine months as models rapidly learn.

Application Layer Shifts

Kothari reports that Notion’s agent‑based usage rose from 1 % of users at the start of the year to 30 % today. Users interacting with Notion via agents are 2‑3 times more active, reflecting a shift from “token maxing” to “token economics,” where companies select cheaper models for routine tasks. He also notes that engineers in the Valley have moved from writing code to managing AI agents within a year.

Competitive Landscape and Risks

Goldie recounts being present at Google Brain when the Transformer breakthrough occurred and observing Google’s early chatbot being withheld over reputational concerns. She argues that Nvidia’s moat is no longer CUDA but the scale of its hardware ecosystem, as models can now generate assembly code directly. Krishnan likens autonomous‑weapon logic to nuclear‑age game theory, while Mirhoseini and Goldie—both former Anthropic employees—state that Anthropic’s safety concerns are genuine.

Implications for India

The guests converge on a simple recommendation for India: build massive compute capacity. They note that Indian IT firms have not yet become “the world’s Turings,” despite a large pool of talent. Kothari points out that securing a single angel investment in India can take three months of paperwork and notarisation. Goldie mentions a Stanford collaborator who received thousands of GPUs from the Indian government to develop a speech model for Indian languages. Kamath cites Skyroot, the Indian space startup attracting U.S. investors, as an example of India delivering world‑class hardware at a fraction of global cost.

Investment Allocation Question

When asked where he would allocate USD 100 million in the AI stack, Mirhoseini suggests focusing on compute efficiency, arguing that a 10× efficiency gain is equivalent to a billion‑dollar spend. Goldie advises founders to ask whether only they can solve a problem; if not, she recommends not starting a company.

Closing Thoughts

The episode ends with suggestions to create a university for product design in India, develop AI tools that accelerate child learning tenfold, and a personal view from Kamath that the next wave of social media will encourage users to “get off your device.” The full conversation is available on YouTube.