About $700 billion is committed to AI this year against roughly $100 billion of AI enterprise revenue, by the count of Vijay Krishnan, founder and CTO of Turing, on WTF is with Nikhil Kamath.
The usual case for AI spending treats that gap as a timing problem. Vijay and Akshay Kothari, co-founder and COO of Notion, said the bigger problem is adoption: outside coding, most companies and most people have barely changed how they work. Azalia Mirhoseini and Anna Goldie, co-founders of Ricursive Intelligence, made the opposite case on capability, saying AI is already starting to improve itself.
"I think we're definitely in a bubble here in terms of how we use AI. Most people around the world that I met... are essentially using it as a better search tool." — Akshay Kothari
The four build different layers of the industry. Akshay runs operations at a software company where agents now make up 30% of users. Vijay's company produces the training data frontier labs buy. Azalia, formerly of Google Brain and Stanford's faculty, and Anna, who worked on Google's Gemini team, both worked at Anthropic before starting their own lab.
The full episode is covered here so you can skip it. 2 hours 29 minutes of audio, 15 minutes of reading.
Here are the 14 arguments that matter.
WTF is with Nikhil Kamath · Oct 23, 2026 · 2h 29m · 134m saved
With Akshay Kothari, Co-Founder and COO of Notion · Vijay Krishnan, Founder and CTO of Turing · Azalia Mirhoseini, Co-Founder and CTO of Ricursive Intelligence · Anna Goldie, Co-Founder and CEO of Ricursive Intelligence
Key Takeaways
Inference companies are valued on the assumption that global AI revenue passes $1 trillion by 2030
AI is a bubble in how it is used, not in what it can do: most people treat it as a better search tool
$700B of AI spending against about $100B of enterprise revenue, and no non-tech Fortune 500 company Vijay can name that has gained from it
The bull case rests on AI taking a share of human labor, not of software budgets
Software teams run at 50–60% of their old size for the same output, yet median software-company growth slipped from 11% to 10%
Agents went from 1% of Notion's users to 30% within 2026
About $100 a month now pays for an AI that checks a person's whole context every 30 minutes and looks for work to do
Even a small amount of friction left in an automated process limits the whole system, which is why no AI optimist Vijay knows forecasts 10%-plus US GDP growth
1. Scarce GPUs Lift Inference
Nikhil Kamath, co-founder of Zerodha, opened by asking the panel to lay out the AI stack from the bottom up. Chips and data centers sit at the base, then the frontier labs (OpenAI, Anthropic, Google, Meta), then newer labs such as Ricursive and Ilya Sutskever's SSI. Above them sit data and training firms such as Turing, Mercor, Surge and Scale, then inference companies such as Fireworks, Together and Baseten, then applications and agents.
Nikhil asked why inference companies, which he described as repackaging compute and open-source models, command such high valuations. Azalia and Akshay pointed to scarce GPU capacity: chip-hours on older-generation hardware now cost far more than a year ago.
Companies stopped maximizing AI use and started pricing it within six months
"We've gone from token maxing in the beginning of the year — everybody use AI as much as you can — to within six months... token economics. It's like, wait, where is all the money going?" — Akshay Kothari
A token is the unit an AI model reads and writes, and the unit providers bill by. Akshay's point is that companies now ask which tasks need a frontier model and which can run on a cheaper one. Akshay and Azalia gave that shift, along with scarce chips, as reasons inference companies are valued so highly.
Inference valuations assume AI revenue has no near-term ceiling
"A lot of the valuations are predicated on the assumption that global AI revenues will exceed a trillion dollars by 2030... if you're one of the leading players helping make tokens of open-source models accessible, and the market does not expect any ceiling to be hit anytime soon, you may be valued very highly." — Vijay Krishnan
Vijay added that hundreds of billions of that revenue are expected to run through serving open-weight models specifically. An open-weight model is one whose trained parameters are published, so anyone can run it on their own hardware.
