AI Generation Is Free. Judgment Costs Everything

The Bar That Sells Beer and Gives Away Intelligence

A bar in Beijing will sell you a $1.50 glass of foam and throw in a supercomputer for free.

I keep coming back to this image because it perfectly captures the pricing inversion nobody in professional services wants to talk about yet. The AGI Bar sits on a startup street near Tsinghua, Peking, and DeepSeek's own offices. Its registered Chinese name is a bilingual pun on "distillation" — the liquor kind and the AI-model kind, because of course it does. You order the house pour, a 9.9 yuan beer off a menu that includes one literally named "AGI Bubble," connect to the WiFi, and get unlimited DeepSeek tokens to vibe-code between sips. There's a "Drinking Plan" that gets you free beer for a year.

The tokens? Always free.

Read that tab again. The beer costs money. The intelligence doesn't. They're charging you for the froth and giving away the model.

That's the whole story of AI pricing right now, poured into a pint glass.

When Generation Becomes Free, What Are You Actually Selling?

I was on a call last week with a firm's innovation committee. They were debating whether to upgrade their AI subscription tier — $30 per seat versus $20 per seat. The entire conversation centered on cost-per-token, utilization rates, and whether they could negotiate volume discounts.

Nobody asked the question that matters: If generation is racing toward free, what business are we actually in?

Because that race is already over. Generation — the first draft, the code, the memo — is racing toward the cost of the WiFi it rides on. DeepSeek's R1 model costs fractions of a cent to run. Llama is open-source. The Beijing bar isn't a gimmick; it's a preview. In twelve months, the commodity part of AI will be as unremarkable as electricity.

Which brings me to the uncomfortable historical parallel everyone wants to ignore.

Electricity Didn't Stay Expensive Either

Electricity did this first. In the 1880s, having access to power was a competitive advantage. Factories paid consultants to wire their buildings. Early adopters promoted their "electric lighting" like we promote "AI-powered" today.

Then the grid arrived. Power became infrastructure. Once electricity stopped being scarce, nobody won by having access to it. They won on what they built with it.

The firms that survived weren't the ones who negotiated the best rates with the power company. They were the ones who redesigned their factories around continuous operation, who built new products that required motors, who hired engineers instead of asking their accountants to optimize the electric bill.

We're at that hinge point now with AI. The conversation I'm hearing in most professional services firms is still about the cost of the electricity — cost-per-token, utilization metrics, subscription tiers. But what do I know — I've only watched this movie four times.

The Expensive Part Was Never the Drafting

Here's what hasn't changed in regulated work, and won't: the expensive part was never the drafting. It's the review. The controls. The provenance. The person willing to put their name under the output and say "ship it."

AI can pour a perfect first draft of an audit memo in three seconds. It can generate a tax position, draft a disclosure note, structure a compliance framework. It cannot tell you it's done. That's judgment, and judgment doesn't come with the WiFi.

I see this every time I advise clients on AI implementation. The technology works beautifully. The breakdown happens at the human layer: Who reviews the output? What's the control framework? How do you document that a junior associate didn't just accept the AI's answer because it sounded confident? How do you explain to a regulator that the model gave you guidance, but you made the call?

This is the $40 billion fat-finger moment waiting to happen. Not because the AI is wrong — because someone forgot that free intelligence still requires expensive judgment.

What You're Actually Paying For

Let me make this concrete. If you're a tax partner reviewing a return, the AI can:

• Pull relevant code sections (free)

• Draft position memos (free)

• Generate supporting calculations (free)

• Flag potential issues based on pattern matching (free)

What it can't do:

• Decide whether the position is defensible given your client's risk tolerance (judgment)

• Know which regulator is likely to challenge it based on their recent enforcement patterns (experience)

• Weigh whether technically correct advice might still damage the client relationship (context)

• Put your professional license behind the recommendation (accountability)

The tokens are on the house. Judgment still runs a tab.

Your competitors have the same free tokens you do. The differentiation — the part clients actually pay for — is everything that happens after the AI stops typing. How fast can your people review? How good are your controls? How much do you trust your team's judgment when the model gives them something that's 95% right?

The Question Your Monday Morning Meeting Should Start With

So if your AI budget fight is still about the price of generation, you're arguing over the cost of the beer. The real question is: What's your firm actually paying for — the draft, or the signature at the bottom of it?

Because here's what I'm seeing in the firms that are getting this right: They've stopped optimizing AI costs and started investing in AI judgment. They're training people to review faster, building control frameworks for AI-assisted work, documenting their oversight processes, creating new roles that didn't exist two years ago.

They've accepted that the model is infrastructure, like electricity or WiFi, and reorganized their business around a world where intelligence is free but accountability isn't.

The firms still negotiating their token budgets? They're the ones who spent 1885 trying to get a better rate on electric lighting while their competitors were redesigning the factory.

What to Do Monday Morning

Here's your specific action: Stop asking "How much are we spending on AI?" Start asking "How much are we investing in the humans who review AI output?"

Those are different questions with different answers:

• What training have we built for AI-assisted review?

• What controls have we documented when work product includes AI generation?

• How do we measure reviewer judgment, not just reviewer speed?

• Who owns the decision when the AI and the human disagree?

• What's our answer when the regulator asks about our oversight framework?

The bar in Beijing isn't coming to your city. It's already here. The model is free. The WiFi is fast. The only thing still running a tab is the judgment that turns a draft into a deliverable.

What are you charging for?