Token Bill Dwarfs CPU Bill | Ben Pouladian on Meta's Muse, Anthropic S1, and Why All Roads Still Lead to NVIDIA

September 30, 2026

AI Summary

5 min read

When Meta released Muse, an AI assistant that connects to Instagram, books appointments, and manages soccer practice schedules, Ben Pouladian finally had an app he could hand to his wife without being called over to troubleshoot. "She's like, wow, look, it connects to my Instagram. It's helping me book appointments. This is what AI really is?" Muse is not a frontier model, Pouladian says, but it does 90% of what ordinary people want — and that distinction matters for understanding where the real costs and investment opportunities in AI are heading.

The Token Bill Dwarfs the CPU Bill

The central insight Pouladian presses is that agentic AI — where a model runs a task multiple times, iterating until it gets it right — consumes enormous numbers of tokens. The CPU cost of orchestrating those loops is real but not the dominant expense. "The token bill dwarfs the CPU bill," he says, estimating the ratio could be 20x, 40x, or even 84x. Many investors have been bidding up CPU stocks like AMD and Intel on the thesis that agentic workloads will drive CPU demand. Pouladian does not deny that demand will rise, but he argues the real cost center is token production, which is a GPU problem.

Continue reading the full summary in the app — free to try.

Read Full Summary →

Free • No credit card required

Listen to Audio Summary Open in App

Never miss an episode of Monetary Matters with Jack Farley

Get every new episode summarized in your inbox — free, ~5 minutes to read.

No spam. Unsubscribe anytime.

What you'll learn

1 Timestamped Navigation Outline
2 (00:00) **Introduction: Ben Pouladian Returns** - BEP Research founder joins Jack Farley to discuss AI's latest developments, Meta's Muse, and the competitive landscape
3 (00:47) **Meta's Muse: The First Consumer AI That Works** - Ben explains why Muse is a breakthrough consumer application with 3 million monthly active users
4 (03:41) **The Token Bill Dwarfs the CPU Bill** - Ben breaks down the cost structure of serving AI agents to consumers
5 (05:55) **Why All Roads Lead to NVIDIA** - The investment implications of token-dominant cost structures
6 (08:33) **Ben's Open Letter to AMD's Lisa Su** - Why Ben challenged AMD's marketing claims and what he's skeptical about
7 (13:40) **CPU Competition: AMD, Intel, and ARM** - How traditional CPUs may not be suited for the agentic AI era

+ Full timestamped outline available in the app

Show Notes

Ben Pouladian of BEP Research returns to Monetary Matters to break down who's winning and losing in AI right now, starting with Meta's Muse, the consumer agent he calls the "McDonald's version" of AI, and why he thinks its revenue will come mostly from ads rather than subscriptions, especially across WhatsApp-first markets like India and Brazil. Ben explains why the token bill for agentic AI dwarfs the CPU bill, why the key metric is becoming cost per token per megawatt, and why he believes "all roads and rockets lead to NVIDIA." He makes the case against AMD, including his open letter to Lisa Su and the scaling challenges of the Helios rack, and argues that NVIDIA GPUs are the "Rodeo Drive" of digital real estate because they are fungible and financeable in a way custom ASICs are not. Jack pushes back with TrendForce data showing every part of the AI supply chain is tight except GPUs, and Ben responds that commodity bottlenecks like memory and hard drives get crushed once supply catches up, while the true bottleneck is powered, permitted land. They discuss the growing data center backlash and whether the base case should be 25 gigawatts rather than 43, Anthropic's S-1 and its roughly half-trillion-dollar compute commitments, and whether AI lab revenue can double again next year. The conversation closes with scaling laws, why models will segment like grades of gasoline, AI's move into biology and drug discovery, and stock ideas including Meta, Lam Research, and Lattice Semiconductor. Recorded September 29, 2026.

“Meta's Muse Rally Bought the Wrong Layer. We modeled a billion agent users. The investment case turns on activity, token costs and trust”: https://newsletter.bepresearch.com/p/metas-muse-rally-bought-the-wrong?utm_source=profile&utm_medium=reader2

Jack Farley on X https://x.com/JackFarley96

Ben Pouladian on X https://x.com/benitoz

BEP Research: https://www.bepresearch.com/

Follow Monetary Matters on:

Apple Podcasts https://rb.gy/s5qfyh

Spotify https://rb.gy/x56dx5

YouTube https://rb.gy/dpwxez