InvestAnswers: AI agent swarms could spark 1,000x compute demand

AI agents are turning into a 24/7 "workforce" that multiplies compute demand

The host frames the episode around the idea that the world is underestimating future compute needs, possibly "by could be a factor of a thousand." He describes a shift from single chatbot queries to swarms of AI agents that "burn everything around the clock" — GPUs, CPUs, memory, electricity — comparing the dismissed fears over Bitcoin mining energy use to a coming, larger version of that debate for AI. He cites an MIT view that a solo founder running agent swarms could in theory produce the output of a 100,000-person company, and references Immad's reported use of 400 agents building a video game in 20 hours.

He points to Meta's agent product Muse as evidence this is already mainstream, saying it went from zero to 2.2 million daily active users within weeks and could reach "500 million, could be a billion users" given Meta's reach across Instagram, Facebook and WhatsApp. The host stresses this isn't a 2030 or 2040 story: "this will be over the next 12 months if not sooner."

"One person theoretically... could be running a team of agent swarms could produce the output of a 100,000 person company." — the host

Tokenized stocks on Solana and high-frequency AI trading show markets colliding with AI

The host discusses an unverified claim, sourced from "multiple sources," that Jane Street earns about $200 million per megawatt using Cerebras chips for inference-driven high-frequency trading. By his rough math, a Cerebras chip costs about $3 million and draws 20 kW, so a megawatt (~50 chips) costs roughly $150 million and could be paid off in nine months at that revenue rate — a claim he flags as unconfirmed but notes could draw other market makers into similar hardware.

He also cites Solana Foundation's Lily Liu arguing that "internet capital markets" using stablecoins like USDC to trade tokenized assets globally could become the largest capital market in the world, powered by Solana. As evidence, he points to Securitize — backed by BlackRock, Morgan Stanley and Cantor Fitzgerald, working with Jump Crypto — launching 12 tokenized stocks including Apple, Nvidia, Tesla, SpaceX, Palantir and MicroStrategy on Solana rails.

"The biggest tech platforms over the last 30 years serve distinct languages... within this future... everybody uses internet money... to trade tokenized stocks, anything, anywhere, anytime." — Lily Liu, as described by the host

Compute bottlenecks, not model quality, are where the host says investors should focus

The host argues AI compute is growing 10x a year versus Moore's Law's roughly 6x over five years, implying a gap he frames as "100,000x in 5 years" against available chip supply — part of why, he says, Elon Musk is pursuing space-based data centers ("Terra Bay"/orbital compute) since terrestrial chip supply "can never scale for what the world needs." He cites a Gadget Review report that Samsung has locked up AI chip supply for five years, and notes Finland reportedly ordered Google to pause clearing for two planned data centers, which he frames as a power/permitting bottleneck pushing more compute toward space.

His named picks for "bottlenecks" include Nvidia and Broadcom for compute, Micron and SK Hynix (and Samsung) for memory, Marvell and "ALAB" for networking, and TSMC and Intel for foundry capacity. He also raises, with explicit uncertainty, a claim attributed to Intel's CEO about an alternative EUV light-source method that could eventually challenge ASML's lithography dominance, though he stresses this "could be the wild card" and is not confirmed. He lists risks including China locking up GPU supply, agent adoption failing (which he dismisses as unlikely given current growth), decentralized GPUs working for inference but not frontier training, and power/permitting constraints.

"Own the bottlenecks, not the chat bots. They're race to the bottom." — the host