GameStop is trying to buy eBay, and almost everyone covering the story is asking the wrong question. The real question isn't whether Ryan Cohen can afford it — it's whether the underlying business logic actually works. According to a former Andreessen Horowitz general partner who dissected eBay's 10-K line by line, the answer is yes. And the reason is hiding in a number so absurd it almost reads like a typo.
Why Is GameStop Trying to Buy eBay?
When Ryan Cohen went on CNBC to talk about GameStop's interest in acquiring eBay, the finance world lit up with skepticism. Most analysts focused on the capital structure: 50% cash, 50% stock, and the question of whether a $10 billion meme stock company can swallow a $48 billion marketplace. But that framing misses the point entirely.
The real thesis, as laid out in this conversation, is straightforward: Cohen isn't buying eBay because he thinks he's smarter than eBay's product team. He's buying it because he can see roughly $2 billion in fat that Wall Street has been pricing as fixed cost. Cut that, park the savings in treasuries, and the interest on any acquisition debt essentially pays for itself. That's the floor of the deal. The ceiling is something much more interesting.
GameStop, which most people have written off as a dying retail chain, happens to own approximately 1,600 physical store locations. In a world where AI agents are about to do most of our shopping, those stores may be worth more than at any point in the company's history.
What Is eBay Spending $2.4 Billion on Marketing For?
Here's the number that stops the conversation cold. In fiscal year 2025, eBay spent $2.4 billion on marketing. That's not revenue — that's the marketing budget alone. So how many net new users did that $2.4 billion buy?
One million. eBay's user base went from 134 million to 135 million active users.
That works out to roughly $2,400 of marketing spend per new user — on a platform that virtually every American already knows exists. This isn't growth marketing. This is a company spending billions of dollars just to tread water, presumably re-acquiring lapsed users or fighting churn it shouldn't have in the first place.
For an investor with operational instincts, this isn't a red flag — it's a flashing green light. It means the business has an enormous, identifiable cost structure that isn't tied to product or infrastructure. It's pure overhead waiting to be rationalized. Strip it out, and the economics of the business look dramatically different.
Can GameStop Actually Beat Amazon in Collectibles?
This is where the bull case gets genuinely compelling. Amazon has been trying to crack the used goods and collectibles market for six years. They've attempted renewed programs, collectibles verticals, and trade-in schemes. None of them have worked. Amazon's used and collectibles business has been essentially flat across that entire period.
Why? Because you cannot put a 1962 Mickey Mantle card through the same warehouse as a phone charger. The category is structurally different. Collectibles — rare cards, vintage sneakers, antique pens, rare keyboards — require authentication, condition grading, and physical inspection. Amazon's warehouse model, optimized for fungible goods at scale, simply doesn't work for one-of-a-kind items where fraud is the primary risk.
eBay already has the marketplace and the buyer trust for collectibles. GameStop brings something eBay has never had: a national network of physical locations where items can be brought in, inspected, and verified by a human being. That combination — digital marketplace plus physical verification infrastructure — is something Amazon cannot easily replicate without building it from scratch.
Why AI Agents Need Physical Stores to Buy Rare Goods
This is the part of the thesis that most people aren't talking about yet, but it may be the most important piece of all. We are entering an era of agentic commerce — where AI assistants like Claude or future agents do your shopping for you, searching listings, comparing prices, and completing purchases autonomously.
For commodity goods, this works fine. An agent can verify a phone charger by checking the SKU. But for rare or used items — a vintage Montblanc pen, a signed rookie card, a first-edition watch — the agent hits a wall. There's no way to verify authenticity without a trusted physical check.
The insight here comes from direct experience. When Discord briefly became an e-commerce platform during the NFT boom of 2021, it worked because blockchain provided a trustless verification layer — you didn't need to physically verify ownership because the chain handled it. But when the team looked at expanding into rare sneakers, rare keyboards, and other physical collectibles, they ran into the same problem every time: physical verification was out of scope, and without it, the marketplace couldn't scale safely.
GameStop's 1,600 stores solve that problem. Imagine an AI agent that finds a rare pen for you, then says: "Verified — the item was brought into a GameStop location, inspected, and received a physical authentication stamp." That's a new primitive for agentic commerce, and it's one that would be extraordinarily difficult for any pure digital player to replicate quickly.
What Is AMP PBC and Why Is It Structured as a Public Benefit Corporation?
Beyond the eBay thesis, the conversation covers a new fund called AMP PBC, launched eight weeks prior to this recording by a former Andreessen Horowitz general partner. In that time, the firm has secured more than $1.3 billion in capital commitments — a solo GP operating with a five-person team.
AMP operates on a simple macro thesis: we are roughly in 1885 industrial England. The steam engine — in this case, AI — has been invented. Everyone knows what products it can enable. But the key input, compute, is being hoarded and massively underutilized. Elon Musk's Memphis cluster of 500,000 H100s is reportedly running at around 11% MFU. That's $12 billion in compute being wasted.
AMP's infrastructure business buys compute on long-term leases, aggregates it onto a shared grid, drives utilization up, and passes it at cost to portfolio companies. The venture capital arm, called the AMP Foundry, co-creates new labs one at a time — the current flagship being Periodic Labs, focused on AI-driven discovery of high-temperature superconductors.
The PBC structure is both substantive and practical. Substantively, both venture capital and infrastructure generate positive externalities when done well — funding innovation and enabling focused research teams to outperform much larger organizations. Practically, it protects the firm legally when it gives away compute at cost in ways that might otherwise look like shareholder value destruction in the short term.
The analogy offered: think less robber baron, more self-regulated utility. AMP is positioning itself as the independent system operator of the AI compute grid — the same role that grid operators play in electricity markets, ensuring access isn't monopolized by the largest players.
What Is Periodic Labs and Can AI Find New Superconductors?
Periodic Labs is the current flagship investment of the AMP Foundry. The team is led by a co-creator of ChatGPT and a former quantum physics lead from DeepMind. The firm operates a 30,000 square foot facility in Menlo Park running a closed-loop materials discovery process:
- AI models predict candidate materials with superconducting properties
- Robots synthesize the predicted materials in the lab
- X-ray diffraction machines test whether the synthesized material matches the AI's predictions
- Verification results feed back into the training run to improve the next round of predictions
In the 90 days prior to this conversation, the lab reported more material verifications than the field had seen in the prior decade. The approach mirrors what made AI successful in protein folding — applying the bitter lesson (scale and compute beat hand-crafted heuristics) to a physical science domain where the verification loop is expensive but clearly defined.
The broader point is that the most interesting investments right now aren't in software. They're in domains where the bottleneck is execution, not imagination — where we know the verification loop works, where the science is sound, and where the missing ingredient is the right team, the right compute, and a partner willing to sit in daily standups to make sure things actually ship.
Is the AI Boom Actually a Bubble?
The closing question cuts to something most people in the industry quietly wonder. The answer here is direct: most of the world still has no idea what AI is. Flying to cities where AI diffusion was expected to be well underway, the reality is that people are barely using ChatGPT. The technology remains alien to the majority of the global population.
That context matters. If everyone in the AI ecosystem disappeared tomorrow, very little would change for most people on Earth. We are still extraordinarily early. The infrastructure being built now — compute grids, physical verification networks, materials science labs — is the equivalent of laying railroad tracks before anyone has figured out what to ship on them. The boom is real. The bubble framing misses how much of the world hasn't even shown up to the station yet.
