A Drink for Unlimited AI Compute: AGI Bar May Be the First AI Brick-and-Mortar Business Ordinary People Can Reach
From a 50㎡ Zhongguancun bar to a Sequoia-backed Shanghai shop: AGI Bar turns local DeepSeek compute into an amenity—and may be the first AI offline business ordinary founders can copy.
NewName Editorial
Editorial Team
Column: New Business Forms | Keywords: AGI Bar, DeepSeek V4 Flash, OPC, offline AI spaces
01｜The story: a bar started giving away compute
At Building 2 on Zhongguancun Entrepreneurship Street in Beijing, a bar called AGI Bar recently broke out of the circle: spend in the venue, join the shop Wi-Fi, and you get free API access to DeepSeek V4 Flash—unlimited while you are on-site.
The mechanic is almost crude. Take the shop's Base URL and API Key, drop them into any OpenAI-compatible client—Cursor, Claude Code, Cherry Studio, even SillyTavern—and start. No signup. No subscription. No quota anxiety. The model runs on the shop's own hardware.
This was never a normal bar. Founded on June 1, 2025 by Song De, the operator behind the media brand Cyber Zen Mind, AGI Bar is 50㎡ with 30 seats, open 15:00–23:00—AI showcase space by day, bar by night. It positions itself as an offline social hub for the AI industry, with a blunt slogan: "the most bubble-inflated bar in history."
Exactly one year later, on June 1, 2026, a second location opened in Shanghai's Xuhui Beiyang AI Town—ground floor of the Sequoia accelerator, 300㎡ with outdoor space. On opening day, Sequoia covered all the drinks.
From a 50㎡ grassroots shop to a 300㎡ space backed by top-tier capital—that expansion path itself is the strongest commercial endorsement of the model.
02｜Why this deserves a serious look: a founder's angle
Background that is easy to miss: free online AI quotas are already everywhere—Zhipu gives new users 20M tokens, and intermediary relays keep stacking free models. "Free tokens" alone is not news.
AGI Bar's novelty is different: it turned AI compute into an offline amenity—alongside Wi-Fi and power outlets.
But the more important question for us is: is this a business ordinary people can copy and actually start?
Break down the structure: customers pay for drinks at clear prices; the shop gives away compute as a traffic magnet. No fundraising narrative. No tech-moat narrative. No "disrupt the industry" slogan—just an offline store with an AI-era giveaway layered on top.
In a moment when AI entrepreneurship almost equals "build a model, build an agent, build SaaS," this may be the first AI-related venture form that does not require writing code, stockpiling GPUs, or raising big money.
03｜The math: why the business can work
This is the most counterintuitive part. The key is not how much a guest can use—it is how cheap the model itself has become.
Official DeepSeek V4 Flash pricing: cache-hit input ¥0.2 / million tokens, cache miss ¥1, output ¥2. For context, flagship Claude Opus charges about ¥170 per million output tokens—roughly 85× more. Two yuan of output quota is enough to let a model read the entire Three-Body trilogy and write a 10,000-word analysis for less than the cost of a bottle of water.
And that is still the API retail price. AGI Bar deploys locally on its own hardware—so for the owner, the real cost of compute is electricity plus a one-time hardware outlay. The more guests use it, the thinner the unit cost.
The more important proof: it is cheap not because it is a weakened model. The open-source agent platform OpenCode burned 8 trillion V4 Flash tokens in a single day (5 trillion from free quota, 3 trillion from paid plans)—a workhorse model that global developers vote for with real money. Cheap and strong is the foundation of this model.
So the ledger is clear: drink margins are cash flow; compute is a near-zero-cost lure. Against a ¥60–70 cocktail, giving away compute costs little more than a dish of free peanuts—and it buys conversation and repeat motives peanuts never will.
04｜Control group: why this beats "Texas Hold'em taverns"
AGI Bar reminds people of the poker tavern boom a few years ago. Both use "game/service as traffic + drinks as monetization," but the cores diverge:
| Dimension | Poker tavern | AGI Bar | | --- | --- | --- | | User spend | Chips + drinks, win/loss volatility | Drinks only, clear pricing | | Giveaway nature | Chips have quasi-cash value | Tokens have no cash value—just compute | | Compliance risk | Long gray zone | Clean: essentially "buy drinks, get Wi-Fi 2.0" | | User relation | Zero-sum opponents | Positive-sum collaborators | | Repeat logic | Gambling urge | Productivity + social gravity |
Poker taverns give away chips—real financial risk exposure. AGI Bar gives away tokens—electricity priced into the floor. For founders, that means low barrier, light assets, numbers that pencil out, and nights you can sleep through.
Poker taverns ride human nature; AGI Bar rides a trend. One ceiling is set by regulators; the other by AI adoption speed—and the latter is still accelerating.
05｜Demand side: the "loneliness dividend" of one-person companies
This business may be stepping on the largest population dividend of the past two years—OPC (One Person Company) operators and indie developers.
How big is that cohort? Wuhan has built 30 OPC communities with 400+ one-person companies. Hubei issued dedicated policy support: up to ¥3,000 free compute per OPC, up to ¥100,000 in compute vouchers, with ¥20M in first-round provincial funding. One-person companies are now an economic form written into policy documents.
But the form has a structural flaw: building in isolation. Alone at home coding, you can sit on a bug for three days with no one to ask, technical decisions with no one to shoulder-tap, and a circle that keeps shrinking. Loneliness is not only an emotion problem—it is a productivity problem.
A shop like this is the antidote: a room full of people with laptops. The bug you are stuck on, the person next to you hit last week. The stack choice you are overthinking, the peer across the table already lived through it. Tokens are the lure, drinks are cash flow, and people—are the real compute in the space.
06｜Cold water: three things to clear before you jump in
At newname.ai we do not write unconditional praise. If you are seriously considering opening a shop like this, three pits are unavoidable:
1. Compute has no moat. Anyone can give away tokens. Wuhan OPC communities already do similar things—buy a ¥13 iced Americano, get a 200K-token voucher via CSDN's TaoToken platform with 100+ models. When big platforms bring subsidies, single shops lose a pure compute war.
2. The real barrier is community density—and community is hardest to copy. What makes AGI Bar valuable is not a few servers; it is a year of AI-circle relationships—the real reason it went from a 50㎡ shop to Sequoia backing. Also see the flip side: the founder needs to already be a node in the circle. Imitators without community DNA will most likely open an ordinary, slightly weird internet café.
3. Engineering problems are real. Models on your own servers mean peak-hour throttling/queues, prompt privacy boundaries, and hardware reliability—all must be solved before opening. That is not shameful, but it needs budget and people who know the stack.
07｜Judgment: an AI-era internet café ordinary people can stretch to reach
Twenty years ago, internet cafés taught a generation to get online. A decade ago, cafés redefined the "third place" with Wi-Fi. Today, as large models become utility-like, offline venues compete on compute, not just connectivity.
AGI Bar may not be the end state—it may be too small, too founder-IP dependent. But it validates a useful business hypothesis:
When a digital resource's marginal cost approaches zero, it can be demoted from "product" to "environment." The real business is the socializing, collaboration, and deals that happen around that environment.
For the past two years, ordinary people facing the AI wave had only two roles: user, or spectator. AGI Bar hints at a third—be the landlord: put compute under the floor, and build a small shop of your own on top.
What does an AI-era internet café look like? Zhongguancun Beijing offered the first answer. The exciting part: that answer does not require you to be a genius programmer, or to raise tens of millions.
You only need a storefront, a stock of drinks, a few servers—and a neighborhood of lonely people still typing code.
Facts in this piece are from public reporting and official information; cost estimates are approximate; actual operations vary.
