Mᴀᴇꜱᴛʀᴏ

Zuck Jensen TPU/GPU Game theory convo ZUCK: “Google is giving me early TPU v7/v8. I’m moving big Llama runs there. I want B200 pricing to reflect that.” JENSEN: “Then take the TPUs. If you believe they’ll hit schedule, go all-in.” ZUCK: “I’m serious. I can cut my NVIDIA footprint by a third.” JENSEN: “Good. That frees up CoWoS for customers who want guaranteed delivery. Your allocation drops starting next quarter.” ZUCK: “…Wait. We still need a baseline of H100/B200 for the mainline models.” JENSEN: “And you’ll get some allocation. But not priority. You split your demand — someone else gets the front of the line.” ZUCK: “Jensen, TPU v7/v8 will close the gap. We’ll be fine.” JENSEN: “That’s what Anthropic said before their compiler delays. That’s what OpenAI said before they came back for more H100s. Everyone thinks they can dual-stack. Few can.” ZUCK: “We have the engineering talent.” JENSEN: “You don’t have the calendar. Training windows slip. Debug cycles multiply. And Google will always give Gemini first rights on peak TPU capacity. You’re betting your entire roadmap on your biggest competitor’s generosity.” ZUCK: “We’ll make it work.” JENSEN: “Maybe. But you just ceded two quarters of training speed to labs staying unified on CUDA. They’ll ship next-gen models before you finish porting.” ZUCK: “…We still want 40k B200s.” JENSEN: “They’re sold out. You’re welcome to join the waitlist.” Zuck realizes too late: He played the TPU card too hard. Jensen didn’t blink. Meta is stuck with half their training fleet on an ecosystem still maturing, and their NVIDIA allocation halved. And the labs staying pure CUDA now sprint ahead.