A Beijing bar gives away DeepSeek tokens. Your per-token price list is the collateral damage.

A small bar in Beijing's Zhongguancun district has stopped advertising free Wi-Fi and started handing out free AI tokens instead, served from two Nvidia workstations behind the counter that self-host DeepSeek's V4 Flash model ^[2][3]. The gag is cheap; the thing that makes it cheap is the point, and it should worry anyone whose product is billed by the token.

The AGI Bar was opened last summer by an independent developer named Song De, in the grid of universities and start-ups that likes to call itself China's Silicon Valley ^[1]. "It's quite common for bars to provide free Wi-Fi with routers," Song told Reuters, "so I'll provide free tokens" ^[4]. Customers can prompt, tinker and code over a beer without paying a subscription to anyone ^[3]. The rest of the room commits to the bit: logos of Chinese AI labs on the walls, a house cocktail called the AGI that arrives as a glass almost entirely full of beer foam ^[5], and, according to visitors, a year-long "Drinking Plan", a brew called the "AGI bubble", and a screen cycling job openings at AI companies ^[6].

Strip out the theatre and what is left is a unit-economics disclosure. Because the bar self-hosts rather than buying metered API capacity, its cost of serving tokens is hardware plus electricity, with no per-token line item from a vendor ^[1]. That is the same shape as a router: a fixed purchase that converts a metered service into a free amenity. The threshold has moved to the point where a two-machine deployment can absorb the traffic of a bar full of engineers and give the output away with a lager, which TNW attributes to DeepSeek and its domestic rivals pushing inference costs down and continuing to cut ^[7].

The competitive consequence is already documented. DeepSeek became a global name in early 2025, when its cheap, capable models prompted what many called a "Sputnik moment" for American AI ^[8], and its ability to match far more expensive Western systems at a fraction of the cost has forced OpenAI and Anthropic to defend not only their benchmarks but their prices ^[9]. Beijing now treats the lab as a matter of national pride and strategic leverage ^[9]. If you sell an application whose price scales with tokens consumed, your customers are being trained by this market to treat that input as close to free, and to ask what else they are paying for.

The counter-signal sits in the same story, and it is the more useful half. Per TNW, DeepSeek has lately begun charging more at peak hours and softening its rock-bottom pitch ^[10], while people inside the industry quietly worry that valuations, DeepSeek's included, have run ahead of revenue ^[11]. Cheap at the margin is not the same as free at scale; someone still buys the workstations ^[7].

Watch three things. Whether peak-hour pricing spreads from DeepSeek to its rivals, which would mark the end of the flat-rate era for cheap Chinese inference ^[10]. Whether Western labs answer on price again or retreat to capability ^[9]. And whether your own pricing page still describes tokens as the scarce thing, when the demonstrated cost of serving a small fast model is now within reach of a bar's bar-snack budget ^[2][3].