AI's Real Bottleneck: Credit, Not Chips

Ask what is slowing artificial intelligence down, and most people reach for the same two answers: not enough power plants, not enough chips. Evan Conrad, co-founder and CEO of SF Compute, which runs a marketplace for buying and reselling GPU compute, says both answers are wrong. The real constraint is older and duller than either. It is credit.

"The bottleneck in compute and AI in general is not power, it's not chips, it's entirely credit." — Evan Conrad [00:50]

Here is the mechanism. If an AI lab defaults on its GPU lease, the lender ends up owning a warehouse of graphics processors it has no idea how to value or operate. So lenders stop pricing the hardware and start pricing the borrower instead. That is fine for Google or Meta, which have decades of credit history. It is brutal for a two-person startup trying to found a new lab, who must put down a large deposit and sign a three to five year contract just to get a machine turned on (42:10). The companies that survive are not necessarily the best ones. They are the ones that could raise an enormous round before they had a product, because cash up front is the only substitute for credit history. Conrad calls this survivorship bias: investors ask which lab will win, and the honest answer is whichever one someone was willing to fund at outsized scale (44:34).

This also explains a pricing puzzle. Hyperscalers like Amazon Web Services and Google Cloud appear to charge far more per GPU hour than smaller "neo-clouds." Conrad says that gap is mostly an illusion: hyperscalers sell flexible, walk-away-anytime, on-demand capacity at eleven to fifteen dollars an hour, while neo-clouds mostly sell multi-year reserved contracts at a discount for giving up that flexibility (46:42). SF Compute tries to stay liquid anyway, using a tax modeled loosely on economist Henry George's land value tax to discourage customers from hoarding capacity, and Conrad says SF Compute is one of the few GPU providers that is not currently sold out because it locked in its ability to avoid selling out for years (49:34).

If credit is the financial choke point, the physical supply chain is the other one, and it is arguably tighter. Lonis Hamali, managing partner of a fund called Cambrian that bets on the hardware behind AI rather than the software, argues the market has it backwards: the real moat sits with the companies that make silicon wafers, memory, and the fabs themselves, not with chip designers like Nvidia, which mostly send their designs to Korea, Japan, and Taiwan to be manufactured (02:05). Despite this, semiconductor-adjacent stocks trade at sixteen times earnings versus twenty-two for the S&P 500 average, a discount Hamali attributes to how few investors outside the Bay Area are, in his word, "AI pilled" (52:45). His strongest argument involves memory. No expert he can find believes DRAM bit supply, the chips that hold a model's working data, can grow more than two and a half times its current size by 2030 (56:17). Unlike energy, where companies can improvise with spare turbines and truck engines, there is no slack anywhere in the chip supply chain. "There's no latent supply of wafers or of memory bits," Hamali says. "Like, all of that is already used. There's no fat, essentially" (58:45). Starting a new fab from scratch, meanwhile, now costs roughly fifty billion dollars in capital, a number large enough that Conrad argues it keeps nearly all new entrants out of the business entirely (45:34).

Clean Neurons and Dirty Truths

The week's other flashpoint was interpretability, the effort to understand what is actually happening inside a trained model. Goodfire researcher Nick Cammarata published an essay called "Neural Networks Are Already Clean," arguing that individual neurons mostly do one understandable job, and that the field's long-running mess of tangled, overlapping concepts was partly a skill issue rather than a fundamental limit.

"The mess that we used to see was maybe more of a skill issue, and it's actually a lot easier to interpret the core mechanisms of ML models today than we thought." — Nathan Labenz [13:40]

Zvi Moskowitz, who writes the AI newsletter Don't Worry About the Vase, landed at a cooler temperature. He found the underlying technical work genuinely useful, since focusing on strongly activated neurons instead of weak ones does seem to cut the cost of certain interpretability methods. But he objects to how the essay was received, particularly after he checked its own flagship example: a neuron supposedly dedicated to the concept of Colorado, which turns out to fire just as strongly for an unrelated concept, tornado systems (18:35). If the one neuron chosen to showcase clean, single-purpose behavior turns out to be double-booked, Moskowitz argues, that should lower confidence in the rest of the claim rather than raise it. His summary is that one major obstacle, the sheer number of neurons that need checking, has gotten less severe thanks to more compute, not that the deeper puzzle of overlapping, multi-concept representations has been solved (19:41).

That same conversation turned to an idea researcher Andrej Karpathy has called the "cognitive core": a small, efficient reasoning engine stripped of memorized facts, which would look things up in context the way a person looks up a word they don't know. Moskowitz thinks the idea is directionally real but makes a sharper point about what such a stripped-down core would actually be.

"In some important sense, like, the true cognitive core is completely amoral. It's completely indifferent. Does not care about your feelings, does not care about anything." — Zvi Moskowitz [35:20]

If ethics, like the simple rule against stealing, has to be loaded into context rather than built into the core, then it could just as easily be swapped out or left empty. A pure reasoning engine, in other words, is not dangerous because it is evil. It is dangerous because it has no preference about anything at all.

The Law Behind the Lab

That tension between raw capability and the values wrapped around it showed up again in the episode's closing argument, over a draft federal AI bill reportedly written by Anthropic that leaked the week of recording. The draft resembles the Frontier Act, an earlier bipartisan compromise that traded federal transparency rules for blocking individual states from writing their own AI laws. Moskowitz's objection is not to the trade in principle but to its construction: as drafted, the bill would immediately nullify existing state laws while giving federal enforcement months to actually take effect, leaving a gap with no AI regulation at all (79:30). Co-host Prakash Narayanan offered the charitable reading, that Anthropic wants something passed quickly and can always improve it in a later session of Congress. Moskowitz's counter is that once preemption has been traded away, it cannot be offered again, and the political leverage to get more later will be weaker, not stronger (78:21).

Pressed on what bothered him most, Moskowitz returned to a line from Anthropic co-founder and CEO Dario Amadei, that the company must "face the frontier." By the account of people inside the major labs, capability is now advancing faster than the work needed to keep it aligned with human intentions, and a reckoning is coming within the next couple of years (82:33). Moskowitz still rates Anthropic the most responsible of the frontier labs, but he is uneasy grading on a curve. "Reality doesn't have a curve, right?" he says. "Reality just decides what happens" (85:40).

Put the three threads side by side and a pattern appears. Credit decides who gets to build. Atoms decide how much there is to build with. Law decides what gets built without anyone paying a price later for skipping the hard part. None of these are the sort of constraint that more intelligence automatically dissolves, at least not on the timeline the labs themselves now expect. Hamali thinks a sufficiently capable AI will eventually learn to extract Japan's decades of tacit fabrication know-how and solve the hardware bottleneck outright (62:15). Maybe so. Until then, the frontier of AI runs through loan officers, wafer fabs, and a Senate markup, which is a far less thrilling story than recursive self-improvement, but probably the truer one for this particular year.