Is AI a Bubble? I Asked the People Building It | WTF is AI?
Executive Summary
Is AI a bubble? Nikhil Kamath puts the number on the table - about $700 billion spent or committed on AI this year against roughly $100 billion of AI enterprise revenue - and asks four people actually building it: Notion co-founder Akshay Kothari, AlphaChip creators Azalia Mirhoseini and Anna Goldie (Anthropic employees 25 and 26, now co-founders of frontier lab Ricursive Intelligence), and data-training founder Vijay. Their answer is layered, literally: they map the AI stack from chips and data centres through frontier labs, neolabs, data trainers, inference providers and applications to services, then show what is happening inside each layer - AI that designs its own chips, training data that expires every nine months, recursive self-improvement measured by the cost of intelligence, and an internet where agents already generate a large share of clicks. The panel's common ground: demand looks infinite, the winners will win on intelligence per dollar, and every engineer is becoming a manager of agents. FULL COVERAGE: the captions appear to cover the entire episode. All figures and claims are the speakers' own, not independently verified. Sponsor: none stated in the episode. Nothing here is investment advice.
Key Takeaways
$700 billion in, $100 billion out
Nikhil's framing: about $700 billion is being spent or committed on AI this year while AI enterprise revenue is about $100 billion - the panel's counter is that software is only 1% of GDP while knowledge work is roughly 50%, so the bet is on disrupting human labour directly, where even 10% of services justifies the spend.
The AI stack has seven layers
Chips and data centres at the bottom, then frontier labs (OpenAI, Anthropic, Google, Meta), neolabs (Ilya's SSI, Ricursive, Thinking Machines), data trainers (Turing, Mercor, Surge, Scale), inference providers (Fireworks, Together, Baseten), applications and agents, and services on top - and every layer is now competing with every other layer.
Ricursive wants AI to design its own chips
Azalia and Anna's frontier lab brings scaling laws to chip design: their platform designs custom accelerators that adapt to AI workloads, attacking a deadlock they call the hardware lottery - chips are designed for today's models while models are designed for today's chips, and a tape-out takes about two years.
In five years we need 100,000x more compute
Azalia's math: AI inference demand is growing 10 to 100x, so a conservative five-year projection needs ten to the power of five more compute than today - which cannot come from just more data centres, so every wafer must become dramatically more efficient.
Training data expires every nine months
Vijay explains the data-training business: experts create problems that are beyond current models and grounded in realism (like debugging realistic codebases), and whatever the industry produces today is obsolete in nine months because the models learn it - coding is far ahead, law and investment banking lag, and robotics has no known recipe yet.
Recursive self-improvement is already starting
Azalia defines RSI as AI improving itself - curating its own data, designing its next architecture and its own chips - turning progress into a function of compute; the metric she proposes is the cost of intelligence, pushing intelligence per dollar at every level.
Agents are becoming the majority of users
Akshay shares that agents interacting with Notion went from about 1% of users at the start of the year to 30% today, with some estimating half of all internet clicks are already agentic on the way to 90% - and in the Valley, nobody writes code anymore: every engineer has become a manager of agents.
The Conversation Breakdown
Meet the people building AI
⏱ Untimed topicAkshay Kothari (Notion co-founder, ex-LinkedIn India), Azalia Mirhoseini and Anna Goldie (Google Brain's first residency cohort, AlphaChip, Anthropic employees 25 and 26, now Ricursive Intelligence), and data-training founder Vijay introduce themselves.
Mapping the AI stack
⏱ Untimed topicNikhil lays out the ecosystem: chips and data centres, frontier labs, neolabs, data trainers, inference providers, applications and agents, and services - and asks where the value will accrue.
The hardware lottery and AI-designed chips
⏱ Untimed topicRicursive's thesis: chips are designed for today's models and models for today's chips, a two-year tape-out deadlock; AI that designs chips that adapt to workloads can unlock orders-of-magnitude efficiency and even new neural architectures.
How human expertise becomes training data
⏱ Untimed topicVijay's process: five million knowledge workers across 70-plus countries create tasks and evaluation datasets that sit beyond the frontier models' reach - and every recipe goes obsolete within nine months as models catch up.
Recursive self-improvement, explained
⏱ Untimed topicRSI means AI curating its own training data, designing its next architecture and even its own chips; progress becomes a function of compute, and the metric to chase is the cost of intelligence.
The agentic internet
⏱ Untimed topicNotion's agent users went from 1% to 30% in under a year; agentic clicks may already be half the internet; engineers have become managers of agents - and Notion is positioning as the neutral playground across models.
So is it a bubble?
⏱ Untimed topic$700 billion of spend against $100 billion of revenue, VCs trading secondaries like potato chips, talent diffused by easy money - and the counter-case that knowledge work is half of GDP and demand looks infinite. Plus: what the panel would do with $100 million.
Core Frameworks & Ideas
The Hardware Lottery
Chips are designed for the models of today, and models are designed for the chips of today - while a chip tapes out about two years late. Ricursive's answer is AI that designs chips adapted to the workloads, breaking the deadlock and enabling architectures that current hardware makes impractical.
Cost of Intelligence
The metric Azalia proposes for the recursive self-improvement era: intelligence is the success rate at tasks, and the race is to maximise intelligence per dollar - reduce the cost at every level of intelligence, and raise the intelligence achievable at every cost.
Efficiency Branch vs Ambition Branch
Two ways companies adopt AI: the efficiency branch does yesterday's work with 30% of the people, the ambition branch does 5x the work with the same people. The median software company is stuck in the first - its growth actually fell from 11% to 10%.
Actionable Next Steps
- ✓Plan in nine-month cycles: assume whatever advantage your AI workflow has today will be absorbed by the models within nine months, and budget for continuous reinvention.
- ✓Compete on intelligence per dollar, not on spend - find your multiplier in the model, data, software or chip layer rather than trying to outspend the frontier labs.
- ✓Treat agents as first-class users: open your APIs and make your product usable across models, because agentic traffic is compounding faster than human traffic.
- ✓Pick the ambition branch over the efficiency branch: use AI to attack more scope with the same team instead of only cutting headcount.
Notable Quotes
“whatever it is we are doing today will be obsolete nine months from today because all models have learnt it.”
“every individual engineer has now become a manager of agents.”
“Maybe you shouldn't start a company then.”
The One Thing to Remember
Whatever the AI industry is doing today will be obsolete nine months from today, because all the models will have learnt it - so the only durable edge is intelligence per dollar.