DeepSeek Secretly Develops Inference Chip to Reduce Nvidia Reliance
量子位 · wechat · 2026-07-08
According to Reuters, DeepSeek is secretly developing an in-house AI chip designed for inference rather than training. The project started about a year ago and is currently in the early stages; they have contacted chip design firms, foundries, and memory suppliers, and are quietly hiring chip design engineers. The R1 base was trained on Nvidia H800s, later shifting to Huawei Ascend, and V4 has been adapted for Ascend. The UE8M0FP8 data format introduced in DeepSeek-V3.1 is believed to be specifically designed for the next generation of domestic chips. However, developing a competitive chip requires years and massive capital, and success is not guaranteed. Supporting this plan is an initial external funding round of about 51 billion RMB (approx. 7.4 billion USD) completed in June 2026, with a post-investment valuation of 52–59 billion USD. The funds will be used to expand computing centers dominated by domestic chips, develop proprietary chips, expand the team, and hire IDC planning engineers for data center construction in places like Ulanqab, Inner Mongolia.
More from Venture
- Anthropic's IPO Filing Goes Public, Revealing Strong Revenue and Massive Spending — KateClarkTweets · 2026-10-27
- Gary Marcus flags Nov 11-15: Anthropic's Q3 S-1 update and earliest IPO window — GaryMarcus · 2026-10-27
- Kimi K3 Is 'Open' Until You Succeed: License Caps Free Use at $20M Revenue — alex_verem · 2026-10-27
- Anthropic hits $16.3B H1 revenue and first-ever $974M quarterly operating profit — iruletheworldmo · 2026-10-27
- Indie dev's ASO experiment: localization via ASC CLI lifts impressions 13x to 2,853/day — rudrank · 2026-10-27
- Anthropic's S-1 reveals over half of $100K+ customers now use both Claude Platform and first-party apps — TechCrunch AI · 2026-10-27