@FinanceYF5: Kimi K3 went viral in two days. Deedy says there's a whole set of compute economics behind it. 1/ K3 has only been online for two days, but it has already reached #10 on OpenRouter, processing 140 billion tokens daily. However, the servers can't handle the load: throughput dropped from 30 Token/s to 13, latency spiked to 72 seconds, and time to first token exceeded 20 seconds...

2/ Even using the quantized version of K3, eight B300s alone cost over $500,000.

The recommended GB300 NVL72 full rack costs $4 million.

His assessment: Moonshot likely doesn’t have enough compute to meet this demand, and US inference providers may not be able to scale up easily either — 2.8 trillion parameters is no joke.

3/ GPU vendors now commonly sign 3- to 5-year contracts, requiring a 30% down payment, with prices skyrocketing.

Two major labs, hyperscale cloud providers, Grok, and Meta have already locked in compute; the rest are fighting for scraps.

Meta is the biggest wildcard: it holds roughly 7 GW of compute by year-end but hasn’t yet committed to any specific large model.

4/ Frontier model prices haven’t dropped much in three years: GPT-4 and o1 both cost $60 per million tokens, Opus 4 at $75, GPT-5 at $10, Fable at $50; now Sol is $30 and K3 is $15 — a decline of only 4x to 5x.

Over the same period, frontier demand has risen by at least three orders of magnitude, and task processing capacity (by METR’s measure) has grown at least 32x.

5/ His conclusion: demand approaches infinite growth with continuous performance improvements; even a slight price drop means the total value of the frontier still skyrockets.

The counter-assumption: if performance peaks, most tasks don’t need top-tier models, or frontier intelligence suddenly becomes achievable at low cost, then the logic weakens; money flowing to the frontier doesn’t automatically mean it’s cost-effective — high costs could also lead to losses.

6/ Coincidentally, Deedy says he wrote this analysis himself and scheduled it to post — and it happened to go live just before Kimi’s official announcement of compute constraints: pure coincidental validation.

Read the original:

That’s all.

If you like this topic:

1. Follow me (@FinanceYF5)
2. Like + repost the first tweet below

This weekend in San Francisco has been a bit crazy.

Two straight days of the AGI Summit, squeezed in between watching a World Cup final — switching from an AI keynote to the roar of fans in real time. Only a San Francisco summer can give you that contrast.

Too many new memories packed into 24 hours. Sharing them slowly.