Listen to the audio version of this article (generated by AI).

Scientific progress has historically been limited by “siloed” thinking. Even the most diligent of researchers can only read a fraction of the millions of published biomedical papers, which means an unknown amount of potential discoveries remain hidden.

So researchers at the University of Chicago brought AI into their workflow to remedy this blindspot.

By building a map of 20 million biomedical papers and the 9 million researchers who authored them, they produced a mountain of research that could find new connections that change how we treat diseases.

Their system traces where scientists concentrate their attention. Then it searches in the opposite direction: the overlooked territory between established fields, where useful connections may be waiting.

The resulting ideas were dubbed “alien hypotheses.” In other words, they’re possibilities that fall outside the research paths humans tend to follow.

That is incredible! AI gives scientists a way to search for connections (that could otherwise never be found) systematically, and bring them into the laboratory for testing.

Meanwhile, another research team is attacking a different bottleneck: time.

Global climate simulations can occupy supercomputers for weeks. That limits how many scenarios scientists can examine as they study a changing planet.

Researchers at UC San Diego and the Allen Institute for AI have developed Spherical DYffusion, a model that simulates 100 years of global climate patterns in 25 hours. By combining generative AI with physics-based data, it runs roughly 25 times faster than the methods used for comparison, without requiring massive supercomputers.

A century of simulated climate. About a day of computing.

Now put those developments next to the conversation on Wall Street.

Investors are questioning whether the AI build-out can earn back its enormous cost. That is a necessary question. But the answer depends partly on applications still moving from research labs into practical use.

These projects show why the demand story extends well beyond writing emails and answering chatbot questions.

They do not guarantee profits. A hypothesis still needs an experiment. A climate simulation still needs validation. And an AI infrastructure company still needs customers who can pay.

But the range of problems AI can tackle is expanding while investors are reassessing what that capability is worth.

That tension is central to our latest episode of Being Exponential. We examine whether the sell-off reflects weakening demand, where I see opportunity, and the concrete warning signs that would change my bullish outlook.

Watch the full conversation by clicking the image below:

As we discuss in the episode, Anthropic’s growth decelerates while OpenAI gains momentum. My interpretation is that competition explains much of that divergence: Customers have more models and coding tools to choose from.

Measure the market before declaring that demand has disappeared. One company losing momentum can coexist with an industry gaining customers.

Then there is financing.

Building data centers requires enormous upfront spending. Servers arrive before years of customer payments do. Borrowing costs, construction delays, and access to capital all affect whether a project earns an attractive return.

S&P Global Ratings projects that combined hyperscaler capital expenditures will exceed $1.3 trillion in 2027, with external financing playing an increasingly important role. These are legitimate business risks.

But different companies carry different risks.

In the episode, I explain why concerns surrounding CoreWeave (CRWV) do not automatically justify the same conclusion about Nebius Group (NBIS), IREN (IREN), or Applied Digital (APLD).

You need to examine each company’s debt, funding arrangements, customer commitments, and ability to bring capacity online. A crowded customer pipeline does not eliminate financing risk. An expensive loan does not prove customers have stopped buying.

Networking remains a major focus, including Astera Labs (ALAB) and Marvell Technology (MRVL). As AI systems expand, moving data between processors becomes increasingly important. Buying more computing power accomplishes less if those processors spend too much time waiting for information.

Memory is another area to watch, including Micron Technology (MU) and SanDisk (SNDK).

Here, investors need to interrogate the earnings forecast. A low valuation can signal opportunity, but it can also reflect expectations that profits are approaching a cyclical peak. The next reports need to show whether demand and earnings strength can endure.

So, what changes my mind?

A meaningful slowdown across the major AI developers. Suppliers missing sales expectations and lowering forecasts. And, especially, the largest cloud companies cutting their capital spending plans.

That is why the spending outlooks from Meta Platforms (META), Amazon (AMZN), Alphabet (GOOGL), and Microsoft (MSFT) deserve your attention.

Maintained or increased budgets support the infrastructure thesis. Meaningful cuts require a reassessment.

Be a disciplined investor. Follow the evidence, and stay flexible when it changes.

My bullish case rests on expanding demand and continued investment. Earnings season gives us another opportunity to test both. The full episode walks through the setups I find attractive… and the warning signs I refuse to ignore.

P.S. My colleague Eric Fry sees a different risk emerging: a widening divide between companies spending heavily to build AI and businesses using it to improve their operations. He believes the midterm elections could accelerate that “Separation,” with Nov. 4 as a potential turning point. In his new briefing, Eric explains his case and reveals one AI stock he recommends — free. If you own AI stocks, his argument is worth understanding.