A Beijing bar called AGI Bar is running local artificial intelligence inference on expensive Nvidia hardware to serve customers unlimited free AI tokens with discounted drinks, despite the owner acknowledging the business model is unsustainable. The venue, located in the Zhongguancun tech hub near Tsinghua and Peking universities, pairs two Nvidia DGX Spark mini-PCs to run DeepSeek's large language models (LLMs) directly on the premises, eliminating the need to send requests to remote servers.

Song De, an independent AI developer in his thirties who opened AGI Bar last year on Inno Way startup street, sells a signature 9.9 yuan glass of foam (about $1.50) and lets anyone connected to the bar's WiFi use its house AI agent at no charge. The bar's registered Chinese name translates to "knowledge distillation," a pun that works equally well for liquor and large language models. According to reporting, roughly 10 times more drinks are given away than sold, making the operation financially unsustainable.

What Hardware Powers This On-Device AI Bar?

The AGI Bar's local inference setup relies on two Nvidia DGX Spark units, which represent a significant hardware investment. Each Spark pairs a 20-core Arm CPU with a Blackwell GPU on Nvidia's GB10 architecture and carries 128 gigabytes of unified LPDDR5X memory. When linked together over their ConnectX-7 network interfaces, the two units pool 256 gigabytes of total memory, enough according to Nvidia to run models up to 200 billion parameters at FP4 precision (a compressed numerical format). After Nvidia raised the Founders Edition price 18 percent to $4,699 in response to memory shortages, a matched pair costs approximately $9,400 before a single free token is poured.

However, the hardware's capacity creates a technical constraint. DeepSeek's V3 and R1 models weigh in at 671 billion parameters, while V4 runs to 1.6 trillion parameters. Whatever flows over the bar's WiFi is therefore a distilled or aggressively quantized cut of the model rather than the full version, meaning customers are accessing compressed or simplified versions of DeepSeek's AI capabilities.

How Does On-Device Inference Benefit Users and Businesses?

• Privacy Protection: Running inference locally on the bar's hardware means user queries and interactions never leave the premises, eliminating concerns about data being sent to remote cloud servers or logged by third parties.
• Reduced Latency: Local processing eliminates network delays associated with cloud-based AI services, allowing the AI agent to respond faster to user requests without waiting for data to travel to distant data centers.
• Operational Independence: The bar can offer AI services without relying on external API providers or paying per-token fees to companies like OpenAI or Anthropic, though the upfront hardware cost is substantial.
• Demonstration Value: The visible hardware serves as a marketing attraction and conversation piece, showcasing cutting-edge AI technology to customers and visitors in the tech-focused Zhongguancun district.

The bar's automation extends beyond just the AI inference. Much of AGI Bar's operations have been automated with AI agents handling inventory management, reservations, and membership tracking. Song De is set to introduce humanoid robots later this year to further automate service.

Why Is This Business Model Losing Money?

The fundamental economics of AGI Bar reveal the tension between hardware costs and revenue generation. The $9,400 investment in Nvidia DGX Spark hardware must be recouped through drink sales, yet the bar's signature beverage sells for only $1.50. With roughly 10 times more drinks given away than sold, the venue is operating at a severe loss. The owner's admission that "the bar is completely losing money" suggests this venture prioritizes showcasing AI technology and building community around DeepSeek's ecosystem over profitability.

The timing of AGI Bar's operations also reflects broader tensions in China's AI sector. DeepSeek suspended its second fundraising round in late July, days after remarks attributed to founder Liang Wenfeng went viral on Chinese social media. The round had targeted a pre-money valuation of roughly 480 billion yuan, or about $71 billion. Reports citing Chinese outlet Yicai say leaked meeting minutes had Liang discussing DeepSeek's continued reliance on Nvidia chips and estimating that China trails leading U.S. labs by 12 to 18 months on around one-twentieth of their compute.

The presence of two American Blackwell boxes displayed as a Beijing bar's main attraction creates an ironic contradiction. Beijing has been pushing its AI sector to reduce dependence on American silicon, yet DeepSeek reportedly faced repeated failures when attempting to train its R2 model on Huawei's Ascend hardware before returning to Nvidia GPUs. The AGI Bar's showcase of Nvidia hardware, while serving as a technical marvel, underscores the continued reliance on American chips that Chinese policymakers have been trying to minimize.

Despite the financial losses, AGI Bar has become a notable gathering place in Beijing's AI community. The venue sits a short walk from Tsinghua and Peking universities and the Beijing offices of DeepSeek and Microsoft, and has hosted parties for Chinese AI labs including Z.ai. The bar even keeps the gong that Z.ai struck for its January Hong Kong listing displayed outside the front door, cementing its role as a cultural hub for the local AI ecosystem.