CoreWeave, which built its business on renting out Nvidia G.P.U.s—mainly to infrastructure-constrained hyperscalers—reported earnings last night that sent the stock plummeting some 16 percent. Though third-quarter revenue more than doubled year over year to $1.36 billion and net losses narrowed to $110 million (practically infinitesimal in A.I. terms…), the company’s operating income fell by more than half, to $51 million, and its operating margin plummeted from 20 percent to 4 percent. Meanwhile, its debt spiked to the tune of billions of dollars.

In the days leading up to the report, CoreWeave C.E.O. Michael Intrator, S.V.P. of engineering Chen Goldberg, and chief strategy officer Brian Venturo dumped tens of thousands of shares of the stock, according to S.E.C. filings. Just the sort of leadership you want to see from the management team!

Runner-up: SoftBank, meanwhile, reported record net income for the quarter, a substantial portion of which came from unrealized “fair value” gains from the company’s investment in OpenAI—not the first time we’ve seen “fair value” factor into a corporate balance sheet. At the same time, SoftBank sold its entire position in Nvidia for $5.8 billion, partially sold its T-Mobile position for $9.2 billion, and secured an $8.5 billion bridge loan to help with its investment in OpenAI.

And now for the main event…

Over the past few years, the astronomical, market-bending reallocation of capital into A.I. has been predicated on a simple thesis: It will take trillions of dollars, thousands of data centers, and hundreds of new power plants to unleash a transformative superintelligence. “It’s hard to even imagine today what we will have discovered by 2035,” Sam Altman wrote in June. “Maybe we will go from solving high-energy physics one year to beginning space colonization the next year, or from a major materials science breakthrough one year to true high-bandwidth brain-computer interfaces the next year.” Heavy cake.

The next month, Elon Musk said that while Grok 4 had not “yet invented new technologies or discovered new physics,” it was “just a matter of time.” Dario Amodei, the C.E.O. of Anthropic, has echoed that language, asserting (baselessly) that A.I. will soon surpass the smartest Ph.D.s, if it hasn’t already. Indeed, the entire sales pitch of the frontier model companies—OpenAI, Anthropic, xAI, Meta, and Google DeepMind—rests on the notion that bigger is better.

But while the Big Five pour money into the infrastructure required for a theoretical A.I. breakthrough and their valuations swell to unprecedented levels, a quieter revolution is taking place among smaller companies constructing intricate systems of lighter-weight models to solve highly specific, tangible, real-world problems—rather than devoting all their attention, time, and money to the intangible dream of one model to rule them all.

One realm in need of exactly this kind of practical advancement is materials science—think better battery technology, more powerful solar panels, more efficient engines, faster planes, better semiconductors, etcetera. Progress in the field has been sluggish given the exorbitant costs and time required for breakthroughs. To speed up the process, Radical AI, which was founded in 2024, is working to create what C.E.O. Joseph Krause described to me as the “materials flywheel,” which uses different A.I. models, automation, and robotics technologies to predict and autonomously test new materials. The prediction process takes advantage of a sizable system of diffusion models, neural networks, and generative models, in addition to advanced atomistic models, which essentially simulate materials at an atomic level. Large language models—basically, fine-tuned versions of some of the offerings from the major labs—are also part of the mix, but they’re primarily an interface layer between the more specialized systems and the end user.

Radical, which secured a million-dollar contract with the U.S. Air Force in August, is currently focused on developing high-entropy alloys to help the military advance its hypersonic flight capabilities. Their A.I. system is designed to predict new structures for the alloy, which it then sends to its “self-driving lab” to test. By the end of this year, Krause hopes to conduct 100 experiments each day. For comparison, he told me that he ran around 50 similar experiments a year as a Ph.D. student studying materials science, and added that materials experiments tend to have a nearly 90 percent failure rate. With Radical’s autonomous lab, each failure will feed data back into the A.I. system, hopefully refining its predictions.

In July, the company secured a $55 million seed round with participation from Nvidia, among others. A significant portion of that money, Krause said, is going into the autonomous lab. “We envision a world where now, in one brain, you have discovery all the way through to light manufacturability in like, a week, versus the 15 years it takes today,” he added. For Radical, the business opportunity is selling valuable materials; A.I. is just how it’s getting there.

Meanwhile, a systems-driven approach to simulation also sits at the core of Geminus, which operates a platform designed to make industrial operations more efficient—for example, by helping oil companies optimize production processes while reducing their environmental impact. Often, the data associated with industrial applications is very noisy, while the decisions derived from it are high-impact. That makes precise predictions—with quantifiable confidence scores—incredibly valuable. Geminus achieves this by fusing A.I. model architectures with more classical, physics-based models and plenty of guardrails and constraints.

Dr. Karthik Duraisamy, founder and chief scientist at Geminus, and a professor at the University of Michigan, doesn’t believe we’ll reach a point where a single model will be able to robustly function in scientific or engineering environments on its own. He explained to me how, in a paper over the summer, he’d argued that the best frameworks involve a “big reasoning model at the top, [which then] needs to be talking to smaller, expert foundation models that know only molecules, or biology, or weather,” etcetera. Large models, he added, play a big role, but “on their own, they’re very deficient.”

This multi-model philosophy has also been adopted by xCures, a platform designed to process and organize unstructured data from medical records to provide clinical insights. Mika Newton, the company’s C.E.O., also argued that instituting a “high degree of constraints” around the platform’s central large model using smaller, narrowly trained, expert models leads to more reliable and robust performance. “When we try to charge people for our software, they go, Well, I can just throw my medical records on ChatGPT. We’re like, Go for it—and the reason is, those models are not fit for purpose,” Newton said. “When you’re using a general model, you get the hallucination issue. The Minority Report of healthcare is not coming soon. … I think you’d have way too many degrees of freedom and potential error.”

While these smaller players might uncover more cost-efficient and creative ways to achieve what many of the hyperscalers are ostensibly aiming for, it’s unlikely that the relentless tide of data center expansion will ebb anytime soon. The market forces are overwhelming, the upside remains tantalizing, and the stakeholders are still innumerable. But some of these smaller companies are starting to develop workarounds that challenge the narrative that comic book–level expansion is essential for the industry’s continued growth.

For example, SiMa.ai is focused on powering A.I. at the edge, which basically refers to hardware—medical devices, robotics, vehicles, etcetera—that has been imbued with A.I. tech. Krishna Rangasayee, SiMa’s C.E.O., told me that the hardware side of things has largely been overlooked in deference to the cloud, and that the company has been building specially designed silicon chips and software to meet what he described as a growing need. So far, the company has raised a total of $355 million in venture capital.

When we spoke, Rangasayee was quick to note that, when it comes to efficiency, SiMa consistently outperforms Nvidia by a healthy margin. “We have gotten comfortable with megawatts and gigawatts [from] nuclear reactors now going [to] data centers and such. We just don’t have the capacity, and I also think there’s no need for it,” he said. “We are a proof point that there’s no need to burn power like we are. I’ve joked, If you want to burn more power and spend more money, please use Nvidia. If you think there is an alternative, we are not too shabby.”

That’s all for today. I’ll see you on Thursday.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006