Ask a room full of technology executives whether artificial intelligence is producing measurable results and most hands go up. Ask whether those results are strong enough to justify interrupting the chief executive's summer holiday, and the room nearly empties. That gap, described by British tech entrepreneur Azeem Azhar, is a compact illustration of where enterprise AI actually stands — real progress, but progress that has not yet become unignorable.
Azhar, founder of the research group Exponential View, told the story during a podcast conversation with Nicholas Thompson of The Atlantic. The AI bubble debate, he suggested, ultimately narrows to a single question: is revenue at AI labs growing fast enough to pay for the enormous data center buildout now underway? The anecdote from Las Vegas does not answer that question, but it frames the smaller ones that hang off it.
A Show of Hands in Las Vegas
Azhar found himself in front of roughly 160 IT vice presidents gathered in Las Vegas. He put a straightforward question to them: who could point to measurable AI results inside their organizations?
About two-thirds indicated that they could. Azhar acknowledged afterward that the number was higher than he had expected — a signal that AI deployments are no longer purely experimental in many large companies. Pilots have become production systems, and some of those systems produce numbers that executives can quote.
Then came the follow-up. Who among them had results good enough to interrupt the CEO's summer vacation? According to Azhar's account, only about eight people remained. From roughly a hundred affirmative answers to a handful, the drop-off is stark.
Reading the Gap
The two questions measure different things. The first asks whether AI is doing something measurable. The second asks whether it is doing something so valuable that a senior leader would drop everything to hear about it. Most enterprises appear to have cleared the first bar and stalled well short of the second.
Azhar's own conclusion is that companies are making progress, but slowly. Whether that pace is sufficient to justify current investment levels remains unresolved. He notes that some executives describe boards becoming more ambitious after early wins, and that even in slower markets such as Italy, chief executives report trust building and budgets rising despite missteps along the way.
Why the Anecdote Sits Inside a Bigger Argument
The reason a show of hands in a conference hall matters is that the industry is arguing about money. The central dispute is whether AI labs are generating revenue fast enough to fund the data center capacity being built to serve them. Every smaller question feeds into that one, and the most consequential of them may be how much real value companies extract from the AI they already buy — because that determines whether they keep buying more, ideally at higher prices.
At the economy-wide level, that return on investment still cannot be documented. It shows up instead in anecdotes, and Azhar's is one of them. His characterization of the overall situation is that it remains "finely balanced" — not obviously a bubble, not obviously a boom.
Chip Lifespans and the Cost Equation
One underappreciated variable is how long purchased AI chips remain useful in production. Whether that productive life is four, six, or eight years changes the math considerably, because depreciation schedules determine how much value each accelerator must generate before it is retired. A shorter useful life raises the bar for every workload running on it; a longer one eases the pressure. That single assumption can swing the profitability picture for a data center operator without any change in demand.
Open-Weight Models and the Bear Case
A second thread in the debate concerns which models companies actually run. Many organizations are shifting away from expensive frontier systems toward open-weight alternatives. From a user's perspective, that is a rational cost decision. From an infrastructure investor's perspective, it is more troubling: adoption could keep climbing while the revenue flowing back to fund the buildout does not keep pace.
Azhar labels this the bear version of the story. In that scenario, AI usage expands across the economy, organizations report genuine productivity gains, and the bubble still bursts — because the money concentrating at the application layer never reaches the capital-intensive layer beneath it. Growth in usage and growth in funding are not the same thing, and the gap between them is where the risk sits.
The Incentive to Sound Optimistic
Another complication is that self-reported success is not neutral evidence. Azhar points to a Boston Consulting Group survey finding that roughly 70 percent of chief executives worldwide say AI success matters for how they are perceived in their roles. When a leader's reputation is tied to a technology program, there is a built-in incentive to describe that program more favorably than the underlying numbers might warrant.
That does not make the Las Vegas tally meaningless — if anything, the executives who could not claim vacation-worthy results were volunteering an unflattering answer. But it does mean that broad surveys of AI satisfaction should be read alongside harder evidence such as revenue, retention, and the persistence of deployments once the initial novelty fades.
Signals Worth Watching
Azhar is explicit that he does not have a clean answer to the bubble question. Nobody does yet. The practical approach is to track the indicators that would eventually settle it:
- Whether AI lab revenue growth continues to outpace the cost of building and operating new data center capacity.
- How long enterprises assume their AI chips will stay productive, and whether those assumptions shorten.
- How quickly spending shifts from frontier models to open-weight alternatives, and what that does to vendor margins.
- Whether AI budgets keep rising in slower markets, where optimism is harder to sustain on hype alone.
- Whether the current crop of measurable results translates into decisions a CEO would interrupt a vacation to make.
The honest reading of the moment is that AI has moved past the demonstration stage in many companies without yet reaching the stage where it dictates strategy. Two-thirds of a room can point to results. Roughly eight can point to results that command attention at the very top. The distance between those two numbers — and how fast it closes — is what the next phase of the investment cycle will be judged on.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com
