Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
FULL INTERVIEW: Anjney Midha on Fixing AI’s Biggest Bottleneck
TL;DR
Midha’s eBay/GameStop thesis is really a cost-cutting and verification thesis, not a product thesis — he argues eBay spent $2.4 billion on marketing in fiscal 2025 to add just 1 million net users, so Ryan Cohen may be seeing roughly $2 billion of removable “fat” plus a way to pair eBay’s marketplace with GameStop’s 1,600 stores for physical verification.
Physical verification is the missing layer for agentic commerce — Midha says AI agents can help find rare goods like vintage pens or a 1962 Mickey Mantel card, but they can’t reliably confirm authenticity without an in-person verification system, which is why he thinks marketplaces for collectibles are structurally defensible against Amazon.
His conviction comes from Discord’s NFT boom, where he accidentally ran into the same problem at scale — after selling his startup to Discord, he watched more than $10 billion of GMV flow through buy/sell channels during the NFT era and concluded that once commerce moves from on-chain assets to sneakers, keyboards, or collectibles, liquidity breaks without physical verification.
AMP’s core bet is that compute is today’s coal, and most of the industry is wasting it — Midha says the AI ecosystem looks like “1885 industrial England,” with companies hoarding scarce compute and running their own underutilized generators, citing xAI’s Memphis cluster of 500,000 GB200s at 11% MFU and under 60% node allocation as an example of billions in wasted infrastructure.
AMP is structured like a self-regulated AI utility, not a normal VC fund — the new PBC has over $1.3 billion in commitments after eight weeks, plans to bring several billion dollars of compute online this year, gives compute to portfolio companies at cost, and reinvests fees and carry into acquiring more capacity.
Midha’s investing style is hands-on, lab-building venture rather than passive check-writing — he describes AMP Foundry as creating “new labs one at a time,” including Periodic Labs in Menlo Park, where AI, robots, and X-ray diffraction are used in a tight verification loop to search for new high-temperature superconductors, producing more material verifications in 90 days than he says the field saw in the prior decade.
The Breakdown
He opens with a hot take on the GameStop-eBay chatter
Instead of leading with his new fund, Midha wants to talk eBay. His claim is that nearly everyone reading the potential GameStop/eBay deal is asking the wrong question: not whether GameStop can afford eBay, but whether eBay’s business becomes much better if you cut the right costs and rethink what it’s for.
The wildest number in the segment: $2.4 billion in marketing for 1 million users
Midha says eBay’s February fiscal 2025 10-K shows $2.4 billion in marketing spend for net user growth from 134 million to 135 million. That works out to roughly $2,400 per new user on a site every American already knows, which is why he thinks Ryan Cohen may be underwriting a simple move: slash around $2 billion in spend Wall Street treats as fixed, and let that alone support the deal economics.
Why collectibles beat Amazon — and why AI makes that truer
The bull case, in his telling, is not generic e-commerce. Amazon’s used and collectibles efforts have been flat for six years, he says, because “you cannot put a 1962 Mickey Mantel card through the same warehouse as a phone charger.” In an agent-driven shopping world, he argues, AI can find the item but still needs a trusted physical verification stamp — which is where GameStop’s 1,600 stores could matter.
Discord taught him that liquidity dies without verification
Midha grounds the argument in his own story: after selling Ubiquity6 to Discord in 2020, he joined as head of platform and unexpectedly found himself around an e-commerce engine during the NFT boom. Discord had more than $10 billion of GMV flowing through buy/sell channels, and the lesson was clear — NFTs worked because ownership was on-chain, but rare sneakers, keyboards, and other physical goods hit a wall because Discord couldn’t build the verification layer.
AMP’s real mission: fix the compute grid
Asked how he’d rebuild eBay with agents, Midha basically says he’s too busy fighting a more urgent war: compute. He frames AMP as a public benefit corporation acting like an “independent system operator of the compute grid,” comparing the current AI market to 1885 England, with every factory running a half-empty generator in the backyard instead of sharing power efficiently.
His pitch is that billions of dollars of AI infrastructure are sitting idle
Midha says underutilization is the hidden scandal of the AI boom, pointing to xAI’s Memphis cluster: 500,000 GB200s, 11% MFU, and less than 60% node allocation. AMP’s answer is to secure long-term compute leases, pool clusters, coordinate capacity, and pass that compute to portfolio companies at cost; he says the firm already has more than $1.3 billion in commitments in just eight weeks and expects several billion dollars of compute online by year-end.
He wants to build labs, not just make bets
The most vivid part of the interview is how operational his venture model is. Through AMP Foundry, he says they help create frontier labs directly, highlighting Periodic Labs — co-founded with Liam, a co-creator of ChatGPT, and Doge from DeepMind’s quantum physics teams — where AI predicts materials, robots synthesize them, X-ray diffraction tests them, and the results feed back into training.
Why AMP is a PBC: positive externalities, not robber-baron vibes
Midha gives both the economist answer and the vibes answer for choosing a public benefit corporation. Substantively, he thinks venture capital and compute infrastructure both create positive externalities and should be run like a self-regulated utility; emotionally, he says he doesn’t want to get sued by shareholders for “giving away billions of dollars of compute at cost” when he believes that’s exactly how to create long-term value and avoid repeating the uglier parts of the industrial buildout.
His closing note: people still underestimate how early this is
The interview ends on a sharp question: is the world prepared for AI not to be a bubble? Midha’s answer is basically no — inertia is too powerful, most people still barely use Claude or ChatGPT, and if capabilities froze today, much of the world wouldn’t notice. His parting point lands cleanly: for all the hype, it’s still unbelievably early.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
