Chamath’s comment came during a discussion of AI startups raising $100M rounds, much of which was being spent on GPU compute. He said that by "next week" (i.e., roughly late February 2024) "when all these announcements are done ... you're about to see a one‑tenth of the compute cost," framing it as going from spending $70–80M of a $100M round on compute to spending $7–8M for the same throughput. (podscripts.co)

In the very next episode (E167 on Feb 23, 2024), one of the main segments is explicitly titled "Groq's big week, training vs. inference, LPUs vs. GPUs," indicating that the "announcements" he was previewing were about new inference hardware like Groq’s Language Processing Unit (LPU). (metacast.app) Just before and during that week, Groq’s LPU was publicly benchmarked and heavily covered: a Feb 13, 2024 press release and independent benchmarks showed its LPU inference engine delivering dramatically higher throughput (hundreds of tokens/sec) than existing GPU-based providers on Llama 2‑70B, marketed as a 10x step-change in speed and efficiency. (groq.com)

On February 22, 2024—within the "next week" window—Beebom and other outlets summarized Groq’s own claims that its LPU clusters could perform LLM inference 10× faster and at roughly 1/10th the cost per token compared to Nvidia H100 GPU clusters, explicitly using the “10x speed” and “1/10th the cost” language Chamath had alluded to. (beebom.com) Follow‑up technical coverage (e.g., The Next Platform) clarified that the “1/10th cost” is best understood in terms of per‑token/time/energy cost for inference rather than literal sticker price of full systems, but it still represents an order‑of‑magnitude reduction in effective compute cost for those workloads relative to the prevailing GPU setups. (nextplatform.com)

By contrast, there was no immediate, industry‑wide 10× drop in GPU prices or in the overall cost of AI compute in that same week—H100s remained extremely expensive, and Nvidia’s next major efficiency jump (Blackwell, announced March 18, 2024) was future‑dated and did not instantly cut users’ bills by 10×. (investors.com) So if you interpret his remark as “the whole market’s compute costs will suddenly be 10× lower next week,” it would be wrong. But taken in context—previewing upcoming announcements about new inference hardware that offers ~10× better cost‑per‑throughput versus the then‑standard Nvidia GPU stacks—the prediction that such announcements were imminent was essentially borne out by Groq’s LPU launch coverage and benchmarks in that exact time window.

Given the format of your normalized prediction (“announcements will be made that reduce the cost of AI compute (for comparable throughput) by roughly 10x versus the then‑prevailing cost structure”), and the fact that those Groq announcements with 10x cost‑and‑speed claims did occur in the following week, the most reasonable classification is that the prediction was right, albeit narrowly (it applied to specific inference offerings, not to all AI compute).