Allocators: $38 an hour off-peak is now the public price anchor for tokens (36Kr). Any neocloud that rents B300s at $8–10 an hour (财经) and sells tokens has to beat DeepSeek's published economics, not just its rivals. Underwrite AI infrastructure for the efficiency turn, not for more data-center square footage.

Builders: DeepInfer makes elite serving software free and portable across Nvidia and Huawei (Global Times). The cost edge is now in the serving stack, not in the weights. Spend your product time there.

Both: Western labs publish no per-GPU numbers. DeepSeek's disclosure is the first public yardstick, and every inference business will now be measured against it.

DeepSeek open-sourced DeepInfer, its inference engine for Nvidia and Huawei Ascend chips, and published production economics that nobody else publishes. One cluster serves 155 trillion tokens a day (36Kr) and earns $38 of revenue per GPU hour (about RMB 258) at off-peak prices, with a 96% input cache hit rate (财经/21世纪经济报道). (DeepSeek technical report and GitHub, 20 Oct — company disclosure)

The gap to everyone else is the story. On generic serving stacks the same model earns a fraction. On 财经's estimate from SemiAnalysis throughput data, at a 96% cache hit rate a B300 card earns $14.88 an hour off-peak, about 40% of DeepSeek's $38, and an H200 earns $0.79, less than the card's own rent (财经). (财经, 21 Oct — reported estimate)

Some cloud vendors don't bother self-hosting DeepSeek's models; they forward API requests to DeepSeek's own servers. Tencent's WorkBuddy agent sends its DeepSeek calls straight to DeepSeek because deploying the model itself costs more than DeepSeek's API price (财经). (财经, 21 Oct — reported)

DeepSeek is reportedly planning a large compute cluster in Ulanqab, Inner Mongolia, and has been hiring senior staff from Chinese internet companies and IDC operators. For now it still rents clusters from two large Chinese IDC providers (财经). (财经, 21 Oct — reported)

DeepSeek's latest round raised about RMB 100 billion (about $15 billion) at about RMB 500 billion (about $75 billion) pre-money (Bloomberg). (Bloomberg, 22 Oct — reported)

Meta Q3: revenue was $64.601 billion, up 26% year on year. GAAP EPS was $3.83, which includes a $9.89 billion legal charge; excluding it, EPS would have been $7.40 (Meta). (Meta press release, 21 Oct — company disclosure)

Meta spent $34.91 billion on capex in the quarter and guided Q4 revenue of $73–76 billion (Meta). It narrowed 2026 capex to $135–145 billion from $130–145 billion (AP). Zuckerberg: "Our investments are paying off, the core business is strong." (Meta press release, 21 Oct — company disclosure)

Meta created a new non-voting Class C common stock. Each Class A and Class B share gets one Class C share as a one-for-one stock dividend, effective 19 January 2027 and effectively a two-for-one split, so voting control doesn't change (Meta; CNBC). (Meta press release, 21 Oct — company disclosure)

Netflix Q3: revenue growth slowed for a third straight quarter, to 11.8% (Netflix). It guided Q4 at $13.44 billion versus $13.51 billion expected (Yahoo Finance), and the shares fell about 6% after hours (Investing.com). (Company letter and coverage, 20–21 Oct)

DeepSeek raised about $15 billion (RMB 100 billion) at about $75 billion (RMB 500 billion) pre-money (Bloomberg). (Reported)

No other new funding rounds were disclosed in today's news flow.