Quick Answer: Is the AI boom real? Partly. Capability is compounding fast, but the money case is unproven: one panellist put committed AI spending this year near USD 700 billion against roughly USD 100 billion in enterprise AI revenue. Returns may show up first in services and knowledge work, and they're arriving unevenly across companies.
Companies six months old are valued at a billion dollars. Investors trade secondary shares "like they're selling potato chips." And the productivity gains everyone promises still haven't shown up in GDP. That's the picture Nikhil Kamath painted in the latest episode of his WTF podcast, recorded in San Francisco and distributed through NewsVoir and MENAFN. He brought a stock market investor's skepticism to four builders. Their answers are more useful than the usual hype, and more divided than you'd expect.
The Money Math Is the Weakest Link
The central debate is arithmetic. One panellist put committed AI spending this year at around USD 700 billion, against roughly USD 100 billion in enterprise AI revenue. That's a gap of about seven to one, and it's the number skeptics keep quoting.
Anna Goldie, co-founder and CEO of Ricursive Intelligence, questions the denominator itself. What should count as AI revenue at all? If a coding tool lets a small team do work that once needed twice the headcount, the value shows up in cost lines and margins, not in an AI revenue line.
Here's the counter-intuitive part. The panel argued that software accounts for only about 1% of US GDP, so even a big software win barely moves national accounts. The real bet is on services and knowledge work, which together are far larger. That's why the productivity signal looks missing: it's hiding in categories nobody tags as AI.
Vijay Krishnan, co-founder of Turing, offered a sharp test. Name one Fortune 500 company outside the AI business, "a Pepsi or a Coke," that has meaningfully grown revenue or cut costs. The challenge is still open in the episode. Yet Krishnan also said software teams can now safely run at half to 60% of their former size, which he called "a giant productivity improvement." Both can be true. The cost savings are real. The revenue story isn't proven yet.
Diffusion Explains the Gap Better Than Hype Does
Akshay Kothari, co-founder of Notion, offered the most useful explanation for the mismatch. Capability is growing exponentially, but most people outside the Valley still use AI "as a better search tool." He cited one study finding that the top 1% of companies spend ten times more on AI than the top 10%.
Kothari's historical analogy helps. The first printing presses copied handwriting, and the first films were recordings of stage plays. New technologies take time to absorb, and early uses are always timid. He then split AI use into two branches:
- Efficiency branch: doing yesterday's work with 30% of the people. Easy to measure and quick to justify, but it caps your upside at the work you already do.
- Ambition branch: a team of a hundred doing five times as much. Harder to measure and slower to pay back, but it's where new revenue comes from.
Most companies fund the first and call it transformation. Kothari's summary is blunt: "the biggest barrier is just imagination." One caveat: the 1% versus 10% figure is a spending ratio, not proof that the heavier spenders earn more. It tells you who is investing, not who is winning.
Compute Demand Is Outrunning the Data Centers
Azalia Mirhoseini, co-founder of Ricursive Intelligence, starts at the bottom of the stack. Inference demand is growing 10 to 100 times. On what she called a conservative estimate, 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.
Her company's answer is AI that designs custom chips for specific workloads. That's a way out of what she calls the hardware lottery: chips get built for today's models, and models get built for today's chips. When chip design cycles run for years, the workload you target may be obsolete before tape-out.
Ricursive's credibility rests on prior work. Goldie said AlphaChip, the chip-layout system she and Mirhoseini built at Google, has been used across five generations of TPUs, as well as Axion, Pixel phones and Waymo chips, and has been adopted by MediaTek. Treat the 100,000x projection as directional. It depends on assumptions the episode doesn't fully unpack.
Goldie also argued that Nvidia's moat is no longer CUDA. Since models can now write assembly directly, the advantage lies in its racks, ecosystem and scale. That's a claim worth testing against your own infrastructure costs before you act on it.
Expert Data Has a Shelf Life of About Nine Months
Krishnan explained the less visible work of making models smarter. A 2022 OpenAI paper, as the panel described it, 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 can't yet solve.
Turing draws on more than 5 million knowledge workers across 70-plus countries, including software engineers, PhDs in physics and chemistry, lawyers and doctors. The catch is the half-life. As Krishnan put it, "whatever it is we are doing today will be obsolete nine months from today because all models have learned it."
This is a treadmill, not a warehouse. The value lies in finding where models still fail, not in producing a fixed dataset. The episode doesn't address how much of that gap closes as models generate their own training data, and that's the question I'd want answered before betting a business on expert data for frontier AI models.
Agents Are Becoming Users, and Token Bills Are Forcing Discipline
Kothari described the application layer. At the start of the year, about 1% of Notion's users were agents. Today that figure is 30%. People who reach Notion through other agents, without ever opening it, are two to three times more active. Users want a neutral place to keep their context as they switch between models.
Cost discipline has changed fast too. Companies moved within six months from token maxing to token economics, choosing cheaper models wherever the frontier isn't needed. Kothari also said engineers in the Valley have gone in under a year from writing code to managing agents.
| Approach | What it optimizes | Main trade-off | Data point from the episode |
|---|---|---|---|
| Token maxing | Capability regardless of cost | Spend grows faster than measurable value | Companies moved away from this within six months |
| Token economics | Cost per useful outcome | Needs routing, testing and quality checks | Cheaper models used wherever the frontier isn't needed |
| Efficiency branch | Output per person | Caps growth at current work | Yesterday's work done with 30% of the people |
| Ambition branch | New output per team | Harder to measure, slower payback | A team of 100 doing five times as much |
The routing advice carries a risk. Cheaper models can quietly degrade quality on tasks that need frontier reasoning. Run a side-by-side evaluation on your real workload before you switch.
What India Should Build (and What Makes It Hard)
Krishnan gave the first answer immediately: "Build compute. That's a no-brainer." Kamath framed the stakes. As AI reshapes services, India will need something to fill the gap that IT outsourcing leaves behind.
The panel was less kind about execution. Krishnan noted that Indian IT companies haven't become the world's Turings, even though much of this expert work is already done by people in India. Kothari said it's still too hard to start a company there, with a single angel investment taking three months of paperwork and notarisation.
Goldie pointed to a Stanford collaborator selected by the Indian government and given thousands of GPUs to build a speech model for Indian languages. Kamath pointed to Skyroot, the space company US investors are now competing to back. The panel's lesson from space: India does best when it takes something the world is already doing and delivers it "at like a hundredth of the cost."
Based on the discussion, the practical sequence looks like this:
- Secure affordable compute access for local builders.
- Cut the time and paperwork needed to form a company and take angel money.
- Build domain-specific data and evaluation work, where Indian expertise is already concentrated.
Where the Human Friction Argument Gets Real
Kamath asked whether humans become a business model, a "human friction layer" between AI and execution, the way a nuclear launch still needs a person to turn the key. Krishnan pointed to a Stanford Business School talk making a similar case: even a little friction becomes the rate limiter. That may explain why no AI optimist is forecasting double-digit US growth.
Krishnan's bigger worry is children. If they never build basic skills, he said, "it is very difficult for me to see what value they can add over a model at all." That's a concern for hiring and training, not just education policy.
On risk, Krishnan compared the logic of autonomous weapons to nuclear-age game theory. Mirhoseini and Goldie, both early Anthropic employees, said they believe the safety concerns voiced by its leadership are sincere. Note the conflict: their view comes from people with close ties to the lab, so weigh it accordingly.
Frequently Asked Questions
What is the AI stack Nikhil Kamath described?
Kamath mapped it as chips and data centres, frontier labs, neo labs, data trainers, inference providers, and applications. The panel added services. The layers matter because each carries different capital needs and different revenue timing. Services is where the panel expects the largest productivity gains to land.
Is the AI boom real or just hype?
Partly. Builders agree capability is growing quickly, and some software teams can run at half to 60% of their former size. But one panellist's figures put committed AI spending near USD 700 billion against about USD 100 billion in enterprise AI revenue. The payoff isn't proven yet. Watch whether productivity shows up in services, not just software.
Why isn't AI productivity showing up in GDP yet?
Software is only about 1% of US GDP, so software gains barely move national accounts. The panel argued that returns will appear in services and knowledge work, which are much larger. Diffusion is also slow: most people still use AI as a better search tool. Gains are real inside companies but hard to see in aggregate data.
How can a company get more value from AI than a search tool?
Start with the ambition branch, not only efficiency. Pick one workflow where a team could produce several times its current output, and measure outcomes rather than headcount saved. Route simple tasks to cheaper models, but test quality first. Kothari's point is that the main barrier is imagination, not tool access.
Is the AI boom real? The strongest evidence sits inside companies that have changed how they work, not in headline valuations. Pick one workflow this week, measure output per team rather than headcount saved, and note what changes after 30 days. Pass this to a founder or operator who keeps hearing the ROI question without an answer.