Hi Dotikers!

Here's a headline that would have seemed unthinkable two years ago: OpenAI is officially turning up its nose at Nvidia chips. Not the ones that train models, where the green giant still reigns supreme, but the ones that run them day-to-day. According to Reuters, the company behind ChatGPT finds its GPUs respond too slowly, particularly for coding tasks with Codex where every millisecond counts.

The issue is technical and rather elegant to understand. Nvidia GPUs use external memory, creating back-and-forth trips that slow down processing. Specialized startups instead embed SRAM memory directly on the chip. The result: responses up to 15 times faster according to Cerebras, which just signed a $10 billion contract with OpenAI for 750 megawatts of capacity through 2028.

OpenAI was also eyeing Groq, but Nvidia whipped out the checkbook on Christmas Eve: $20 billion to lock up the startup and grab its key engineers, including the creator of Google's TPU. A "non-exclusive licensing agreement" that looks suspiciously like a disguised acquisition to dodge regulators.

This battle reveals a major shift in the industry. Training models? That was yesterday. Tomorrow's war is fought on inference, where the money actually comes in. Nvidia has figured that out and is playing defense, while OpenAI diversifies its supply chain. We've gone from a perfect marriage to a couple consulting a lawyer, each keeping one eye on the exit door.

Alexis