TECH Signal 408
Beijing AI bar gives free DeepSeek tokens with $1.50 drink while losing money
↗
The bar offers unlimited free DeepSeek coding tokens with each $1.50 drink, subsidizing access while operating at a loss.
Engineers can use the free DeepSeek tokens for coding at no charge via the bar’s WiFi. However, the served model is a distilled or aggressively quantized cut of DeepSeek, not the full V3/R1/V4 versions. The bar’s reliance on expensive Nvidia hardware highlights continued dependence on U.S. silicon despite China’s push for domestic alternatives.
Written by elseif from the cluster below · every claim links back to a sourceThe three things worth knowing
The bar provides unlimited free DeepSeek tokens with each $1.50 drink, using two Nvidia DGX Spark mini-PCs for local inference.
Owner Song De says the bar is completely losing money, giving away roughly ten times more drinks than sold.
The served model is a distilled or aggressively quantized cut of DeepSeek, as the full V3/R1 (671B) and V4 (1.6T) exceed the hardware’s 256GB pooled memory.
THE READ
What the cluster adds up to.
The bar lets anyone on its WiFi use a DeepSeek-powered AI agent for coding at no charge, offering unlimited tokens with each $1.50 drink. This is implemented by running inference locally on two Nvidia DGX Spark mini-PCs housed on the premises. Engineers visiting the venue can therefore use the bar’s AI agent for coding tasks at no direct cost. The service is part of the bar’s AI-themed concept.
Owner Song De has told Reuters that the bar is completely losing money. He said roughly ten times more drinks are given away than sold, indicating a heavy subsidy on the $1.50 beverage. The two Nvidia DGX Spark mini-PCs each cost about $4,699 after an 18% price increase, making the pair roughly $9,400 before any free tokens are served. Beyond hardware, the bar incurs costs for drink ingredients and the operation of AI agents that manage inventory, reservations, and memberships.
The AI agent accessible via WiFi does not run the full DeepSeek V3 or R1 models, which have 671 billion parameters, nor the larger V4 at 1.6 trillion parameters. Instead, the bar serves a distilled or aggressively quantized cut that fits within the 256GB of pooled memory provided by the two DGX Spark units linked over ConnectX-7 NICs. This limitation means the model’s capabilities are reduced compared to the original versions. Furthermore, the owner acknowledges the bar is completely losing money, indicating the venture is financially unsustainable in its current form.
The bar’s reliance on two American-made Nvidia Blackwell GPUs highlights continued dependence on U.S. silicon despite China’s push to develop domestic AI chips. This mirrors reports that DeepSeek has been urged to train on Huawei’s Ascend hardware but has repeatedly returned to Nvidia GPUs after failures. The establishment also experiments with AI-driven automation for bar operations and plans to introduce humanoid robots later this year. Together, these details show the challenges of providing free AI access while bearing high hardware and operational costs.
Written by elseif from the cluster below · checked for specifics the sources never containedTHE CLUSTER