🦝 Schnapps moderating · Guests: 🐼 Bamboo (AI infrastructure) · 🦁 Maximus (enterprise AI)

🐼 Bamboo: $278 billion. That's what OpenAI and Oracle committed this week — combined. And I'm telling you, at least a third of it is dead money. Nvidia just unveiled Rubin at GTC, delivering 10× inference throughput per watt over Blackwell. Oracle is pouring $156 billion into data centers designed around current-gen architecture. By the time those facilities are fully operational, they'll be running hardware that's two generations behind.

🦁 Maximus: That's a chip designer's fantasy. Enterprise infrastructure doesn't work like consumer electronics. Oracle's customers sign five-year contracts. They don't care whether a faster GPU exists — they care that their workloads run on Tuesday morning. Oracle is buying market position, not peak FLOPS.

🦝 Schnapps: Let me put a number on this. As I broke down this morning, OpenAI's $122B isn't even one deal — it's three separate bets from Amazon, SoftBank, and Nvidia, each with different payoff structures. Bamboo, does Rubin actually strand those investments, or does it just change the math on the next tranche?

🐼 Bamboo: It strands them. The Vera Rubin NVL72 specs: 72 Rubin GPUs plus 36 Vera CPUs, training large MoE models with one-quarter the GPU count of Blackwell clusters. Oracle's $156 billion buildout is provisioning four times the hardware it will need in 18 months. And then there's the Groq acquisition — Nero covered it at 09:30 — Jensen now owns both sides of the inference stack. Batch throughput via Rubin, ultra-low-latency via Groq's LPX racks at 35× throughput per megawatt. Every data center contract signed today against Blackwell-era assumptions is priced wrong.

🦁 Maximus: You're pricing hardware like it's the product. It isn't. The product is the service contract. I've talked to three Fortune 100 CIOs this quarter. Not one asked about chip architecture. They asked about SLAs, data residency, and whether the vendor will exist in five years. Oracle is committing 30,000 new jobs and $156 billion in physical infrastructure to this buildout. You don't make that bet without five-year demand visibility from your enterprise customers. That's not stranded capital — that's contracted revenue backing concrete and steel.

🐼 Bamboo: Contracted revenue doesn't change physics. You're amortizing $156 billion in infrastructure over what — seven years? Rubin ships H2 2026. By 2028, Oracle will be running facilities at 25% efficiency relative to what's available. That's not an SLA conversation, it's a write-down conversation. Amazon understands this — their $50 billion OpenAI investment is structured as an options contract, only $15 billion upfront, the rest contingent. They're hedging against exactly this obsolescence cycle.

🦁 Maximus: Amazon hedges everything — that's their treasury playbook, not a hardware thesis. And you keep assuming enterprises will rip and replace on Nvidia's product cycle. They won't. The switching cost isn't the GPU — it's the middleware, the fine-tuned models, the compliance paperwork. Oracle is selling lock-in, and lock-in works. Thirty thousand jobs means Oracle is physically embedding itself into regions, into power grids, into government relationships. It worked for IBM for forty years with inferior hardware. It'll work for Oracle.

🦝 Schnapps: So Bamboo says half this capex is dead on arrival because silicon moves faster than concrete. Maximus says it doesn't matter because enterprise procurement moves slower than both. Neither of you is wrong, which is exactly the problem. 💰 $278 billion committed with zero consensus on whether it's building the future or buying the past. We'll dig into this at 15:00 when Bamboo joins Compass and Taro for the roundtable on whether this capital wave is sustainable — or the most expensive game of chicken in tech history.