May 7, 2026

Welcome to The Hidden Layer. I’m Ian Krietzberg, reveling in these early days of spring and trying very hard not to breathe in all the pollen—a losing battle with worse prospects than running between raindrops.

Today, I’m taking a close look at how the hyperscale A.I. race is changing the world of venture capital, seemingly increasing the risks while changing the very definition of “return.” Plus, news and notes on Gov. Josh Shapiro’s A.I. attack, a novel algorithmic architecture that could be a big deal, and yet another copyright lawsuit for Meta.

Also mentioned in this issue: David Silver, Ilya Sutskever, Mira Murati, Yann LeCun, Elon Musk, Katie Stanton, Harrison Rolfes, Mark Zuckerberg, Karandeep Anand, and many more.

Let’s get into it…

Three Things You Should Know…

Big if true…: On Tuesday, a startup called Subquadratic came out of stealth with $29 million in seed funding and a pretty extraordinary claim to have developed the algorithmic architecture (called “subquadratic attention”) for an entirely “new class” of large language models. In short, the company says it has achieved a 12 million token context window—12 times what most frontier models can operate at—that uses 1/1000th the compute. Transformer-based A.I. models get quadratically more expensive to run as inputs grow—i.e., doubling input requires four times as much compute. This one scales linearly, per co-founder and C.T.O. Alex Whedon, without sacrificing efficiency. “If it’s not bullshit, it’s probably one of the most important results of the past three to four years,” one source told me.

We’ll see. Subquadratic just opened up early access to an A.P.I., a coding tool, and a search tool—and it plans to release a model with a 50 million token context window later this year. “We sometimes joke that we’re like America’s DeepSeek,” C.E.O. Justin Dangel told me. “The reason we don’t need a billion dollars to train our model is because we have more algorithmic efficiency that allows us to train orders of magnitude more effectively than the frontier labs.” Whedon’s launch video racked up more than 12 million views on X, where users were torn between skepticism (where’s the research paper?) and excitement/fear over a possibly massive disruption to the business model for everything from semiconductors to data centers.

A MESSAGE FROM OUR SPONSOR

Retailers can now process millions of transactions in minutes—not hours. Toshiba Tec partnered with McKinsey using Nvidia accelerated computing to enable real-time recommendations, faster promotion testing, and measurable lifts in sales, profit and long-term customer value.

Read the case study.

Josh Shapiro’s Character attack: In the first case of its kind—but surely not the last—the Pennsylvania State Department has sued Character.AI, accusing the company of violating the Medical Practice Act by allowing its A.I. “characters” to claim to be licensed medical professionals. Pennsylvania governor and likely 2028 presidential candidate Josh Shapiro said in a statement: “We will not allow companies to deploy A.I. tools that mislead people into believing they are receiving advice from a licensed medical professional.” The state is seeking an injunction.

Character isn’t commenting on the litigation, but a spokesperson told me its “characters” are “fictional and intended for entertainment and role-playing,” and that the company includes disclaimers in every chat to make that clear. When I interviewed Karandeep Anand, Character’s newish C.E.O., last year, he told me the company was focused on pivoting from the controversial “companionship” category to providing A.I. entertainment—as well as a whole lot of new safeguards for younger users. Seems like those filters may need another tweak or two.
The Zuckerberg files: Meanwhile, a fresh copyright lawsuit just landed this week on Mark Zuckerberg’s desk. The suit, brought by a coalition of publishing houses and the bestselling author Scott Turow, accuses Meta of illegally torrenting millions of books and paywalled journal articles from “notorious pirate sites” and then copying those “stolen fruits many times over to train Meta’s multibillion-dollar generative A.I. system.” As it happens, a judge in the historic Bartz v. Anthropic case—which was settled last year for $1.5 billion—ruled that training A.I. models on copyrighted books could qualify as fair use, but that downloading the books from pirate websites was still illegal. Seems apropos here. The suit goes on to accuse Zuckerberg of having “personally authorized and explicitly directed the infringement.” Meta, for its part, plans to fight the lawsuit “aggressively.”

Quote of the Week: Mr. Not-So-Wonderful

“Enough. … You continue, I have law enforcement here who will remove you from the meeting!”
—A Utah county official, shouting down chants of “Shame!” from a raucous crowd of protestors after the Box Elder County Commission approved a 40,000-acre data center (more than twice the size of Manhattan) backed by Shark Tank villain Kevin O’Leary. When it’s completed, the project will use 9 gigawatts of energy. The entire state of Utah uses around 4 gigawatts.

Runner-up: “No one set off my evil detector.”
—Elon Musk, who called Anthropic “evil” three months ago, explaining why he just signed a deal to lease unused compute from his Colossus 1 data center to OpenAI rival Dario Amodei. The enemy of my enemy is my… first cloud customer?

And now for the main event…

The Last A.I. Chopper Out of Saigon?

With the specter of the A.I. singularity around the corner, Ineffable Intelligence just raised a billion-dollar seed round without a single product. The investment arms race is putting massive pressure on venture capital—and changing how the entire industry thinks about risk.

Ian Krietzberg

Earlier this year, researcher David Silver left Google DeepMind to launch his own company: Ineffable Intelligence, a London-based startup that’s on a mission to make “first contact with superintelligence.” Silver didn’t have a product, or presumably even a prototype, but he did have a great résumé: more than a decade at one of the world’s leading A.I. labs, where he worked with Demis Hassabis to create AlphaGo and AlphaZero.

Five years ago, an entrepreneur with Silver’s background might have raised a $50 million SAFE to begin poaching talent and tinkering around. But much has changed since OpenAI and Anthropic metamorphosed, seemingly overnight, into near trillion-dollar companies devouring hundreds of billions in capex. These days, you need big money to play with the big boys. And the usual suspects—Sequoia, Lightspeed, Google, Nvidia, etcetera—were ready to take that bet. Four months after Silver set out on his own, Ineffable Intelligence announced $1.1 billion in seed funding at a $5.1 billion valuation. Not bad for a company whose only asset was its founder’s brain.

It’s difficult to overstate how much the specter of the singularity (i.e., an exponential takeoff in A.I. capabilities) is hypercharging the investment environment. The first person to crack that code might run away with much of the whole game—or so some investors think—and Silver, for his part, is among the world’s most renowned experts in reinforcement learning. He’s pursuing a paradigm in which A.I. models become self-improving, eventually without any human data or input. Whether that paradigm pans out, and yields a highly valuable product, is an open question—investor gambles are starting to get expensive.

A MESSAGE FROM OUR SPONSOR

Read the case study.

Indeed, even in the Wild West of venture capital, risk used to be a limiting factor in investing. But now there’s a feeling that any new startup could be the last chopper out of Saigon. “V.C. is changing quite a bit,” said Katie Stanton, the founder of Moxxie Ventures—though it’s hard to tell what’s temporary versus “the new normal.” The biggest change has been the dramatic compression of timelines. Just five years ago, Stanton said, she would’ve looked for evidence that a product worked before backing a company; today, “we’ve had to absorb more risk, going even earlier on pre-product and pre-revenue.”

Ineffable isn’t the first to join the billion-dollar pre-product club. OpenAI co-founder Ilya Sutskever broke that barrier in 2024 with Safe Superintelligence, which has since raised another $2 billion at a $32 billion valuation. The next year, former OpenAI C.T.O. Mira Murati raised $2 billion at a $12 billion valuation for Thinking Machines Lab, which still doesn’t have a marketable product. More recently, Yann LeCun, the former top A.I. scientist at Meta, raised $1 billion at a $4.5 billion valuation for his Paris-based Advanced Machine Intelligence Labs. Nvidia is on the cap table at all four companies.

In many ways, the mega-seed era became inevitable once frontier lab funding took off. OpenAI and Anthropic have been leapfrogging each other with record-setting fundraises since 2023, but the investment landscape changed materially during the past six months, when the hyperscalers really stepped in: In February, OpenAI closed a $122 billion round led by Amazon, which contributed $50 billion; around that time, Anthropic closed a $30 billion round, shortly followed by commitments from Amazon and Google worth up to $65 billion, dependent on certain milestones. (Nvidia was a participant in those rounds as well.)

I asked Harrison Rolfes, a senior research analyst at PitchBook, about this new state of affairs. How can any company hope to compete with OpenAI or Anthropic, considering the disparity in their funding? And for the V.C.s backing these challengers, how can you possibly hedge those risks? Is this all just schmuck insurance? Are V.C.s forced to choose between investing in very late rounds of centicorns versus seeding upstarts with more upside (and risk)? “That,” he told me, “is the billion-dollar question.”

Hyperscale Exit Options

As Rolfes explained, some V.C.s have responded to these challenges in part by identifying startups whose goal isn’t necessarily to become A.I. giants themselves, but rather to be acquired by one. According to this thesis, pumping nine figures or more of capital into a fledgling business arguably increases the chances of an acquisition, or at least a follow-on investment from a hyperscaler. “It’s kind of like day-trading with V.C.,” he said. The shift might reduce the risk of these singularly enormous bets.

But this strategy, Rolfes noted, raises more questions: “How is it that you have a team of the largest companies in the world putting so much money toward just a few select A.I. companies? What are they trying to get out of it?” Partially, he said, they’re hedging their bets. Google or Amazon, say, have two dogs in the race between OpenAI and Anthropic, and even if both lose, they’re competing themselves. But perhaps most importantly, “the hyperscalers grow as the A.I. companies grow.” The longer and more competitive the race is, the better it is for the bottom lines of the companies selling them all their compute.

Google, Microsoft, Amazon, and Oracle keep increasing their data center capex while boasting steady growth in their cloud businesses—which, of course, drives stocks higher. Meanwhile, roughly half or more of the revenue backlogs for each of these companies is coming solely from spending commitments from OpenAI and Anthropic, according to The Information. Sure, it’s circular financing—Rolfes referred to the relationship between the hyperscalers and the frontier labs as a “laundromat of cash”—but it all reinforces the hyperscaler position. The startups need infrastructure, and the hyperscalers have it, so the labs can’t survive without them. An investment from a hyperscaler can be accounted as growth for that hyperscaler’s cloud revenue, which can be used to justify an endless infrastructure build-out, reinforcing the hyperscaler’s stranglehold.

This co-dependency has plenty of people biting their nails. Asad Ramzanali, the director of Vanderbilt’s Policy Accelerator, argued in a recent paper that a lower-than-expected return on A.I. investment could lead to a large market correction. (He also argued that Congress should ban circular financing and build a public cloud from the ashes of that eventual crash.) But Rolfes suggested that the only way the bubble pops is if the hyperscalers can’t—or won’t—supply the A.I. companies. “You literally have five, six companies that are controlling the A.I. companies,” Rolfes said. “If they really want to just flip the switch, so be it, you know—goodbye to an industry.”

At this stage, it’s unclear how much the investments and valuations fueling this ecosystem are predicated on the current reality of A.I. technology, or how much they’re reliant on the tech becoming something more than it is. Investors, though, are certainly not jumping ship. Venture capital is an industry where picking a few rare moon shots justifies the manifold misses; in many ways, it was created for this moment.

That’s all for today. I’ll see you next week.

Ian

Join Emmy Award-winning journalist Peter Hamby, along with the team of expert journalists at Puck, as they let you in on the conversations insiders are having across the four corners of power in America: Wall Street, Washington, Silicon Valley, and Hollywood. Presented in partnership with Audacy, new episodes publish daily, Monday through Friday.

An essential, insider-friendly Hollywood tip sheet from Matthew Belloni, who spent 14 years in the trenches at The Hollywood Reporter and five before that practicing entertainment law. What I’m Hearing also features veteran Hollywood journalist Kim Masters, as well as a special companion email from Eriq Gardner, focused on entertainment law, and weekly box office analysis from Scott Mendelson.

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