A small AI-themed bar in Beijing has turned local inference into a customer perk—and, according to its owner, a money-losing one. Tom’s Hardware reports that AGI Bar, located in Zhongguancun’s Haidian district near Inno Way, gives visitors free and unlimited access to a house DeepSeek-based AI agent through the venue’s WiFi code. The bar was opened last year by Song De, described by Tom’s Hardware as an independent AI developer in his thirties who runs the venue in his spare time. Its signature drink is a 9.9 yuan, or roughly $1.50, glass of foam also called AGI. Tom’s Hardware says Song told Reuters that the bar is losing money, with about 10 times more drinks being given away than sold. The compute setup is part of the pitch. According to Tom’s Hardware, AGI Bar runs inference locally on two Nvidia DGX Spark mini-PCs, which are kept on display. That makes the bar less a conventional hospitality business than a physical demo of AI access bundled into a low-priced drink. The location is also part of the story. Tom’s Hardware places the bar near Tsinghua University, Peking University and the Beijing offices of DeepSeek and Microsoft. It also reports that the venue has hosted parties for Chinese AI labs including Z.ai, and that a gong tied to Z.ai’s January Hong Kong listing sits outside the entrance. The bar’s registered Chinese name reportedly translates to “knowledge distillation,” a pun that fits both liquor and large language models. Tom’s Hardware says the venue has also automated parts of its own operations. AI agents reportedly handle inventory, reservations and memberships. Song is said to be planning to add humanoid robots later this year, though the source gives no further detail on timing, vendor or deployment scope. The hardware details matter because the bar is advertising free DeepSeek access while using a relatively small local setup. Tom’s Hardware says each DGX Spark pairs a 20-core Arm CPU with a Blackwell GPU on Nvidia’s GB10 and includes 128GB of unified LPDDR5X memory. Nvidia says that is enough for models up to 200 billion parameters at FP4, and linking two units over ConnectX-7 network interface cards can pool 256GB. Tom’s Hardware compares that capacity with DeepSeek’s larger models, reporting that V3 and R1 are 671 billion-parameter models and that V4 is 1.6 trillion parameters. On that basis, the outlet argues that the bar is likely serving a distilled or heavily quantized version rather than a full-scale model. The cluster does not include an independent technical confirmation of exactly which DeepSeek model variant the bar is running. The article also places the bar against a broader, more politically sensitive backdrop: Chinese AI developers’ reliance on Nvidia hardware. Tom’s Hardware reports that DeepSeek suspended a second fundraising round in late July after remarks attributed to founder Liang Wenfeng circulated on Chinese social media. It says the round had targeted a pre-money valuation of about 480 billion yuan, or $71 billion. Those related claims are less settled. Tom’s Hardware says reports citing Chinese outlet Yicai described leaked meeting minutes in which Liang discussed DeepSeek’s reliance on Nvidia chips and estimated China’s AI labs trail leading U.S. labs by 12 to 18 months while using about one-twentieth of their compute. The same item says Bloomberg, which had covered the transcript, had not verified its authenticity. That makes the AGI Bar story firm as a local-inference anecdote, but the broader DeepSeek-policy frame should be treated as reported context rather than established fact. Who benefits: Visitors benefit most directly: according to Tom’s Hardware, anyone with the WiFi code can use the house AI agent at no charge. Nvidia also gets visible placement in the story through the DGX Spark systems, though no commercial relationship is reported. Who's exposed: Song De and the bar bear the immediate financial exposure if free AI access and giveaway drinks do not convert into paying demand. The story also highlights the sensitivity around Chinese AI projects that still depend on Nvidia hardware, but the broader DeepSeek claims remain reported and partly unverified.