Arthur Hayes Explains Why the AI Bubble Won't Pop

Happy Saturday one and all!

Brian McGleenon here,

If you have been reading the mainstream financial press lately, you have probably noticed a massive surge in market anxiety. Skeptics are warning that the billions being poured into artificial intelligence by tech giants will inevitably culminate in a massive macro disaster. The common narrative is that the market is running on a fairy tale—one sold by tech bros to leaders like Donald Trump in America and Xi Jinping in China under the premise that national security relies entirely on these insane capital inflows into AI companies.

I recently welcomed back Maelstrom founder and BitMEX co-founder Arthur Hayes to the virtual studio to dissect these macro forces. I asked Arthur when this massive capital expenditure (CapEx) bubble might finally pop. His answer completely bucked the doom-and-gloom consensus.

“I don’t think it’s anywhere close to the end of this current AI inflow upsurge, as you said, CapEx or GNAI,” Arthur told me. “The demand is only increasing.”

He added that a hallmark of a terminal market bubble is billions of dollars chasing zero revenue. Many critics look at AI and assume we are seeing a replay of the late-90s dot-com hysteria. Arthur completely rejects this narrative, pointing directly to real-world commercial performance.

“NVIDIA earnings were out last night and they beat across the board,” Arthur noted. “Jensen talked about the massive upsurge in inference, the amount of money spent on inference, the growth of the AI agentic economy. Someone sent me a news clip from the FT detailing how Anthropic is now profitable. Right? Like, you know, they were burning billions of dollars up to a few months ago and now on a go forward basis, Anthropic is now profitable because of the surge of demand for inference due to people using agentic solutions within their business and personal workflows.”

For Arthur, this commercial demand changes the entire economic equation. “The revenues are there. Okay. Maybe you’re paying a little bit too much for the forwarders, but it’s not as if the high flyers are cashflow negative companies. That’s apart from OpenAI; Anthropic, which has the leading models and the most popular models out there right now, is a profitable company.

The Deflationary Loop Driving the CapEx Cycle

The true catalyst keeping this AI cycle alive is a self-reinforcing loop of declining costs and escalating complexity. Arthur laid out a clear economic chain reaction:

Declining Costs, Rising Use: “As the price of intelligence declines, the usage increases, right? The more that data centers are built, the more chips that are sold, the cheaper it becomes to run these models.”

Escalating Complexity: “You just get more complex models and you do more complex reasoning. With these AIs, replace more knowledge work within human civilization and demand keeps growing.”

This structural shift is already altering corporate workflows, from Meta laying off workers to financial powerhouses changing their hiring strategies.

“I think Jamie Dimon at JP Morgan said he’s going to hire more AI specialists and bankers, which basically means the financial services houses are getting into adding agentic AI into their own workflows,” Arthur highlighted. “And you can imagine the amount of demand that there is for compute when you talk about high-frequency trading. I think I read Hudson River Trading has now their own cluster.”

According to Arthur, the efficiency gains aren’t proving that the bubble is full, they are proving to companies that they need to spend more on CapEx to avoid being left behind. “So I don’t think we’re anywhere close to the saturation point of an AI capex build. But if anything, this is proving that you need to build even more capex, right? Because at the end of the day, you need to sell to Anthropic, you need to sell to OpenAI, you need to sell to Gemini and XAI and all these sorts of players.”

Navigating the Investment Landscape and the “Red Queen Effect”

For everyday investors, the current playing field looks both incredibly exciting and highly confusing. Front-end model developers like OpenAI, Anthropic, Gemini, and xAI are locked in what Arthur calls the “Red Queen effect”, a constant 100-meter sprint where companies must run as fast as they can just to stay in the same place.

“I honestly think it doesn’t matter. Right now, none of us can invest in OpenAI or Anthropic on a liquid basis,” Arthur observed. “But if you think about it as an investor, what can you invest in? You cannot invest in these front-term models. It’s not clear to me whether or not they are going to be the ultimate winners. Obviously, if you move more to the source of intelligence generation right on the hardware side, yes, there are clear ways—NVIDIA, Intel, TSMC. It’s pretty clear that these guys are doing things and you can’t replace them.”

The danger for front-end software, he warns, is instant obsolescence. “On the Anthropic side, like if your front-term model drops behind a few percentage points in effectiveness, then the corporate clients switch, right? And then all of a sudden all that money you spend is worthless. And so this leads to the, you know, ‘I need to train the next best front-term model. I need to spend another 100 billion, 150 billion, whatever it costs to train one of these models that continues to drive the capex cycle because you cannot fall behind.’”

But! There is a human cost to all of this.

While the financial markets thrive on these corporate efficiency gains, the human reality of the AI revolution will be highly unequal, dictated entirely by geography and local social contracts.

“I’m sure there will be new employment, but I think that it’s a bit silly for people to claim that, you know, you lose your job yesterday and tomorrow you get another job in AI,” Arthur stated bluntly.

“Talk to all the folks in Europe and America who lost their manufacturing jobs because of China and other Southeast Asian countries... A lot of them didn’t [find work]. A lot of them just got addicted to opioids and killed themselves, essentially, if you think about America. So I think that’s a bit silly to say there’s this creative destruction and all of a sudden all these workers are all going to get these magical AI jobs.”

Arthur explained that the political fallout will look vastly different depending on where you live:

In Western Economies: “If you’re talking about knowledge workers getting replaced by AI, well, knowledge workers are the most politically active people in the United States... they’re going to find a way to get protected. If they’re saying that they’re no longer gainfully employed and it’s these tech overlords that have taken their jobs through efficiency gains, I think there’s going to be a political backlash.”

In Emerging Markets: “Think about places near where I live in Southeast Asia—I talk about India, Bangladesh, Philippines... these workers are essentially cost centers for American and Western European companies. Nobody cares if they starve or not. ‘Oh, okay, I’m going to lay off my entire Filipino staff that was doing all this manual work... well, now an AI could do all that. Goodbye.’ And nobody cares if they lose their jobs. Starvation, right? And so the Philippines, a lot of these countries don’t have the wherewithal to make a situation better. And so I think there’s going to be massive social strife in a lot of these third-world countries that are essentially back-office processing centers.”

Watch the full interview with Arthur Hayes here:

The Automation Curve Escalates

While Arthur Hayes and I were discussing the macro forces fueling the AI infrastructure build, a new development has occurred right at the frontier of model development.

Anthropic has officially admitted that AI is learning to build better AI significantly faster than expected, signaling an early shift toward recursive self-improvement.

According to a report from BeInCrypto’s Lockridge Okoth, internal data released by the Anthropic Institute, the company’s Claude model now authors over 80% of the production code merged into their systems, enabling engineers to ship eight times as much code per day as they did two years ago.

The technical leaps involve code-speedup optimization tests, where Anthropic’s new Mythos Preview model achieved a mind-blowing 52x speedup compared to the modest 4x speedup a highly skilled human engineer takes hours to achieve.

While Claude still lacks the high-level “research taste” to choose which problems matter most, Anthropic notes that machine-authored code has reached absolute parity with humans and is expected to be strictly better within the year. Meanwhile, sector rival OpenAI is keeping pace, rolling out a new “dreaming” system that allows ChatGPT to synthesize past conversations and dynamically refresh its long-term memory.

Think about it: who is actually buying AI en masse right now? Over 90% of OpenAI’s user base is on the free app. Instead, as Lockridge exposes, the tech giants are essentially quietly paying themselves through their own cloud bills.

Have a great weekend,

Brian,

Global Head of News, BeInCrypto