This week, Nvidia entered into an agreement with Groq—a 10-year-old company that makes chips for fast A.I. inference—to acquire its I.P. and engineering talent for roughly $20 billion in cash. It’s a significant premium given that Groq raised $750 million at a $6.9 billion valuation as recently as September—and, a few months before that, slashed its 2025 revenue projections by around 75 percent due to capacity constraints.

Groq confirmed the “nonexclusive” licensing deal in a Christmas Eve blog post. Meanwhile, founder Jonathan Ross, president Sunny Madra, and other employees will join Nvidia. While it’s not technically an acquisition, it’s still Nvidia’s largest deal to date—and the majority of Groq employees are reportedly being paid cash for their vested shares. “We haven’t acquired Groq,” an Nvidia spokesperson told me. “We’ve taken a nonexclusive license to Groq’s I.P. and have hired engineering talent from Groq’s team to join us in our mission to provide world-leading accelerated computing technology.”

Runners-up: Japanese conglomerate SoftBank announced on Monday that it’s acquiring A.I. infrastructure investment firm DigitalBridge for $4 billion in cash, a step toward its mission to “realize Artificial Super Intelligence for the advancement of humanity.” (Thanks, guys!) SoftBank is one of OpenAI’s leading investors, and recently sold its entire stake in Nvidia to fund its commitments in OpenAI. Then, a day later, Meta said it was buying Manus, a young startup devoted to producing A.I. agents. The FOMO must have gotten to Zuck.

And now for the main event…

In many ways, the A.I. race came into its own this year. Model upgrades, updates, and integrations continued at a pace so breakneck it was nearly impossible to keep up. The firehose of announcements, blog posts, and releases included the long-awaited (and widely dismissed) launch of OpenAI’s GPT-5 alongside major releases from Google DeepMind, Anthropic, and a slew of Chinese developers. And while those inside Silicon Valley rejoiced at the leapfrog-esque benchmark improvements, this year was about much more than rote model capability.

Yes, yes, there was the obvious stuff: the politics of A.I., the enormous capex fueling hundreds of new data centers, and the technical orchestration required to make these systems work reliably. But the year’s dominant narrative was one of an ambitious, profligate-spending upstart and a slightly caught-off-guard hegemon—a classic battle of Little Tech vs. Big Tech, and one long in the making: OpenAI vs. Google.

OpenAI was founded in 2015, after all, as a nonprofit research lab devoted to realizing artificial general intelligence before Google—a mission that always seemed quixotic, even if the path forward was nebulous for both parties. A couple of years later, Google introduced the Transformer, and OpenAI took that architecture and began scaling it up through more chips and data. “OpenAI did not set out with a thesis that compute was the path to progress,” co-founder Greg Brockman said in a preamble to the company’s latest infrastructure announcement. “It’s that we tried everything else, and the thing that worked was compute.” Indeed it did: The surprise release of OpenAI’s scaled-up L.L.M. in 2022 effectively began the A.I. race—a fuzzy competition with no clear finish line. For a long time, OpenAI was handily winning. Shortly after the launch of ChatGPT, Google issued a widely reported “code red,” an acknowledgement that its relevance was slipping in an era it had inadvertently invented.

Now, though, the shoe appears to be on the other foot. For the past three years, Google has pushed its internal tech forward, built out an almost incomprehensible amount of data center infrastructure, and reinforced its moat through massive investments across the A.I. startup ecosystem—all in pursuit of catching up to OpenAI and its 800 million users. And this year, while OpenAI busied itself with productizing ChatGPT and finally converting into a for-profit public benefit corporation to close on its multibillion-dollar V.C. tranches, Google finally caught up—and then surged ahead.

Even Sam Altman has acknowledged the reversal. Earlier this month, OpenAI declared its own “code red,” a response to the surging popularity of Google’s latest Gemini model. “Google was the biggest winner of the year. It turns out that having a total vertical advantage on everything from research to chips to data acquisition to distribution creates a tough-to-outdo competitor,” Rak Garg, a partner at Bain Capital Ventures, told me. “They started the year with mediocre models and no real winning A.I. product other than ephemeral wonders like NotebookLM—which is excellent! Today, they’ve outdone most on the cost-performance curve with Gemini 3 Flash, tightly integrated models into their core product surface areas that already serve a billion active users, and Gemini has quickly become the A.I. tool of choice for many of my use cases.”

Meanwhile, Google’s access to one of the largest, highest-quality troves of data available—known to non-techies as YouTube—combined with its control over the design and economics of its specialized T.P.U. chips has helped the company “retake its old throne,” Garg said. On a recent podcast appearance, Altman defended OpenAI’s code red as a “low-stakes thing” that the company does from time to time. “I think it’s good to be paranoid and act quickly when a potential threat emerges,” he said, while also acknowledging that Google is “a huge threat.” Altman continued: “If Google had really decided to take us seriously in 2023, I think they would’ve been able to smash us. Google has probably the greatest business model in the whole tech industry. And I think they will be slow to give that up.”

And yet, some industry sources I consulted believe that Google’s level of dominance is even greater than what Altman laid out. Manos Koukoumidis, a software engineer and the C.E.O. and co-founder of the open A.I. platform Oumi, expects that the concerns animating OpenAI’s code red will soon become an issue for plenty of other companies, too. “I’m surprised that Anthropic is not starting to do something similar,” he told me.

His reasoning was simple: The business model these startups adopted confers every advantage to the company with the most capital. And Google, with a roughly $4 trillion market cap, has effectively unlimited resources. “[How] are you going to be the one winning when the other guy has already caught up, and you’re playing on the strategy that favors them and not you?” Koukoumidis said. In other words, if scale is all you need to keep producing better, more effective A.I. systems, then the contest boils down to the ability to spend. Google’s revenue is approaching $400 billion this year; OpenAI, meanwhile, is years away from turning a profit, according to internal projections reported by The Information, and has stacked up around $1.4 trillion in infrastructure-related spending commitments.

All of America’s biggest A.I. players—with their closed, proprietary offerings—are now also facing competition from the surging popularity of open-weight models coming out of China, which are often cheaper, more customizable, and increasingly competitive on performance. Perhaps most concerning, Koukoumidis told me, is that China is reaching parity with the U.S. despite a scarcity of compute—belying the brute-force strategy adopted by much of Silicon Valley.

The big, expensive hyperscaler approach may yet pay off. The trouble is that it remains entirely theoretical whether real cognition can be genuinely mimicked through brute force. The human mind, after all, is an enormously complex, algorithmically elegant, and computationally efficient thing. Engineers have been telling me for years that scale alone won’t unlock A.G.I., and that the next breakthrough is more likely to come out of left field—perhaps a small, under-the-radar startup working on complex, more efficient architectures. (Yann LeCun, Meta’s former top A.I. researcher, who left to start his own company last month, has famously declared that L.L.M.s are a dead end.) But OpenAI isn’t hedging its bets with different approaches: They’re fighting fire with fire against an opponent who can’t run out of fuel.

What all this portends for OpenAI depends entirely on whom you ask. Altman, for his part, doesn’t expect code red to last for much longer—as he said, it’s basically strategic paranoia. Sure, the company’s losses keep piling up despite steady gains in revenue, but OpenAI hasn’t struggled to raise new capital. To that end, it’s reportedly in talks to secure a $100 billion funding round at a $750 billion valuation—a remarkable 50 percent jump from its reported $500 billion valuation just two months ago.

However, DataRobot C.E.O. Debanjan Saha believes OpenAI will be stuck in this code red whether it publicly acknowledges it or not. “OpenAI is way out over its skis on infrastructure commitments and capital intensity, and the competition is exploding,” he said. “Prices will come down and differentiation will get harder. When that happens, the economics start to look fragile. Code red is real.” The big question, he said, is whether Microsoft will buy OpenAI outright, or whether OpenAI will become “the Netscape of the A.I. era—the company that defined the moment but couldn’t survive what came next.”

The year 2026, Saha argued, will be “critical” for OpenAI—not just because the better-funded competition has caught up in terms of tech prowess and mindshare, but because the ferocious investment cycle is also beginning to shift. For instance, concerns about rising debt levels caused Blue Owl Capital to pull out of a $10 billion data center project with Oracle—a major pivot, and a telling signal to the industry given Blue Owl’s multibillion-dollar role in bringing data centers online. Meanwhile, Oracle’s stock has fallen by about a third from its all-time high this fall as investor euphoria collides with the realities of debt financing.

There’s also the issue of chip infrastructure outpacing power supply. In other words, you can build as many data centers as you want—but there’s little point if you can’t plug them in. “By 2026, OpenAI must show it can convert massive capital investment into platform leverage, otherwise hyperscalers like Microsoft and Google will set the terms,” Saha said. He doesn’t expect Microsoft to acquire OpenAI next year, mainly because it would be a difficult merger from a regulatory standpoint. But he thinks it’s possible that Microsoft—which currently owns a little more than a quarter of the company—will continue tightening its controls of the economics and distribution. “If OpenAI can’t bend its cost curve or own more of the platform layer,” Saha warned, “it risks becoming a premium model provider inside someone else’s ecosystem—rather than the defining A.I. platform.”

Sharon Goldman chronicled the battle—on the ground—between small, rural towns in Arizona and Silicon Valley–funded land developers eager to drop a few more multibillion-dollar data centers into the A.I. boom. [Fortune]

The Democratic Party has a bit of a problem when it comes to A.I. Its constituents are basically unequivocally opposed to the industry’s unrestrained advancement, which presents both an easy political opportunity and an enormous challenge. How should it proceed? [Politico Magazine]

Earlier this month, the Science Fiction and Fantasy Writers Association posted rules that would’ve allowed some form of A.I.-assisted writing to be eligible for the Nebula Award. Then came the backlash, swift and intense, and the organization took it all back. It reraised a debate that sci-fi author Chuck Wendig explored in this post, which compares generative A.I. to a garlic press. [Terribleminds]

Speaking of A.I. and writing: A writer from a small town in Indiana, after losing freelance work ostensibly to the growing prominence of generative A.I., chronicled his fraught career shift into tree trimming. “In towns like mine, outsourcing and automation consumed jobs,” he writes. “Then purpose. Then people.” [N.Y. Times]

That’s all for today. Happy New Year, and I’ll see you next week.

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