Model sourcing and using open-source models

Watch on YouTube • Listen on Spotify • Listen on Apple

Here is something that actually happened: A company put advertisements inside a coding tool—the kind of tool used by senior software engineers at serious enterprises who, as a demographic, are famously resistant to being marketed to—and instead of the internet erupting in the predictable conflagration of outrage, the company generated somewhere between $5 million and $10 million in annualized revenue within a month.

Quinn Slack, co-founder and CEO of Sourcegraph, announced something called “AMP Free” in October, watched his X post accumulate 900,000 impressions, and—this is the part that made me sit up—had advertisers signing six-figure contracts between 6 PM and 8 AM the night before launch. Which is to say, companies were so eager to advertise inside a coding agent that they were doing paperwork at 2 AM.

The implications here extend considerably beyond one startup’s monetization experiment, though if you’re the kind of person who tracks startup monetization experiments (and if you’re reading this, you probably are), it’s a pretty interesting one.

Quinn is explicitly building towards an ad network that could serve other coding agents and developer tools, which is another way of saying he wants to become the Google of developer attention. His thesis—and it’s the kind of thesis that sounds either visionary or delusional depending on which assumptions you accept—is that ads are the only business model capable of driving the scale necessary to justify the truly staggering datacenter buildouts currently underway, and that ad-supported agents become what he calls a “sponge” for unused GPU capacity.

Advertisers, meanwhile, are apparently willing to pay $500 to $1,000 per “qualified action,” which in this context means a developer who not only clicks on an ad but actually implements an API key. If Slack is right about all this, the entire AI infrastructure stack just found its demand floor. If he’s wrong, well, at least someone tried something genuinely weird.

For operators and investors this interview functions as a kind of masterclass in contrarian positioning. While Slack’s competitors are racing to win enterprise deals by discounting their products 85-100% (which, just to be clear, means they are sometimes giving away their product entirely in exchange for the privilege of having a customer), Slack is deliberately taking only 10% of deals in order to preserve what he calls “product velocity.” The bet is that staying on the frontier matters more than short-term market share, and that ads let you answer to users voting with their feet rather than disconnected enterprise buyers who want commitments to things like “self-hosting” and “model choice.” It’s either brilliant discipline or extremely sophisticated cope—and the next twelve months will tell us which. But $5-10M ARR in the first month is, as they say, pretty baller.

Ideas & Analysis

1. Ads Are the AI Demand Floor

Thesis: Ad-supported AI agents create guaranteed demand for GPU capacity, de-risking datacenter investments and enabling more ambitious infrastructure buildouts.

Quotes:

“An ad support coding agent is essentially like a sponge for any unused tokens from good models out there. And so it lets the whole world be bolder in all these data center build outs.” — Quinn Slack

“Ads are the only way to truly drive the kind of scale that we need to get every single person using a coding agent. And that ultimately drives economies of scale that is going to support all of these data center build outs and that can make it so that everyone’s CapEx plan can be 25% more ambitious because they know that the moment all those GPUs come online, they’re going to have a use.” — Quinn Slack

Analysis:

The quiet insight here—and it took me a moment to catch it—isn’t really about advertising at all. It’s about infrastructure finance, which is to say it’s about the thing that makes all the other things possible. Consider that hyperscalers and model providers are currently making trillion-dollar commitments on datacenters that will not, cannot, be fully utilized on day one. The traditional playbook assumes demand will materialize through paid subscriptions and API usage, but that’s fundamentally a bet on conversion rates and price elasticity, which is another way of saying it’s a bet on human behavior, which is notoriously difficult to predict. What Slack is arguing is that ad-supported agents flip the entire equation. Instead of hoping users will pay, you guarantee utilization by making the marginal cost to users zero. The GPU hours get monetized through attention rather than direct payment. It’s the same economic logic that made broadcast television work, back when broadcast television was a thing people cared about: you don’t charge viewers; you charge Procter & Gamble for access to their eyeballs.

The second-order effect is competitive. If Quinn is right, companies without an ad-supported tier are leaving demand on the table and ceding scale advantages to those who have one. The counterargument is obvious and has probably already occurred to you: developers hate ads, or at least they say they do, and the market will punish anyone who degrades the experience. But Slack’s 95% advertiser close rate and the 900K-impression post suggest the market is at least curious, which is not the same as enthusiastic but is considerably better than hostile.

2. Coding Agents Have Google-Like Ad Potential

Thesis: Coding agents uniquely combine high-intent search behavior with always-on engagement, creating an ad surface superior to either Google or Instagram alone.

Quotes:

“The cool thing about a coding agent is you actually have the opportunity to do both because it’s up on their screen, it’s in their workflow at all times, kind of like people are on Instagram all the time... but then they’re also going to it with high intent. And you don’t really go to Instagram with high intent for specific purchases or actions, you go to Google. But yeah, the coding agent has both.” — Quinn Slack

“There’s no other kind of ad that delivers customers that are so well-qualified and far along on implementing... We’ve heard from our ad customers that they would be willing to pay $500, $1,000 for that kind of really highly qualified lead.” — Quinn Slack

Analysis:

Slack is making a structural claim about attention quality, and it’s the kind of claim that’s either profound or tautological depending on how generous you’re feeling. Here’s the argument: Google wins on intent (you’re searching for something specific, you want an answer, you are a person with a problem), Instagram wins on time-on-platform (you’re scrolling endlessly, you have perhaps forgotten why you opened the app, time has become somewhat irrelevant), but neither platform has both. A coding agent, by contrast, is open all day and captures discrete moments of high commercial intent—specifically, the moments when a developer decides they need authentication, payments, or database infrastructure. The targeting data is also considerably richer than what you’d get elsewhere: not just “what did they search for” but “what’s actually in their codebase” and “how many people are working on this repo.”

This is an idea I strongly believe in and endorse. Ads integrated into B2B workflow products is a market no one has really cracked, and if AI finally makes it possible, hundreds of billions of dollars are suddenly available.

The $500-$1,000 CPL quote is, I think, the buried lede here. That’s not display ad pricing—that’s lead-gen pricing for enterprise SaaS, which operates according to entirely different economics. If AMP can actually deliver a qualified lead who has already implemented an API key, the comparison isn’t Google Ads; it’s a sales development rep who never sleeps and never misses a buying signal and doesn’t require health insurance.

The risk, obviously, is execution: can they actually build the attribution and action layers necessary to prove that value, or will advertisers pay for a few months, see murky ROI, and churn? (Advertisers are famously patient about murky ROI, which is to say they are not patient about it at all.)

3. Trust Is the Moat for AI Ads

Thesis: Separating ads from AI recommendations—like Google separates organic from sponsored results—is essential to maintaining user trust in agentic products.

Pull Quotes:

“It’s really important to know we do not inject ads to steer what the AI is doing or saying or recommending to you. It’s entirely separate, just like in Google, how the organic results are separate from the sponsored results.” — Quinn Slack

“Trust is so important. And this is why Google, you could say, they have all the temptation in the world to just start showing ads as organic results. But you can stop going to Google if they do that. And coding agents, it’s a really highly competitive space.” — Quinn Slack

“I think there’s a lot of other coding agents that if they came out with ads, people would say, that’s just going to be junk. But because that trust we had with AMP, we were able to try this and ultimately build something I think is valuable.” — Quinn Slack

Analysis:

The Google analogy is doing a tremendous amount of work here, and it’s clever framing—perhaps too clever, in the way that things designed to be persuasive sometimes are. Google’s entire business depends on users believing organic results are unbiased, even as the company makes $200+ billion per year from ads displayed alongside them. (This is one of those arrangements that sounds impossible when you describe it but has somehow persisted for decades.) Quinn is arguing that the same church-and-state separation can work for AI agents: the model gives you unbiased recommendations, and the ads are clearly labeled and shown separately, and everyone understands which is which, and nobody feels manipulated. The mechanism that enforces this arrangement is competition: if users feel manipulated, they switch to Cursor or Claude Code or Codex. The switching cost for coding agents is low enough that trust violations get punished quickly, which is either reassuring or terrifying depending on whether you’re building one.

But there’s a tension here that Slack acknowledges without fully resolving, and I found myself underlining the relevant passage. When asked why AMP doesn’t inject tool recommendations directly into the AI’s suggestions—which would, one imagines, be considerably more valuable to advertisers—he says they could do that, if it’s “clearly marked” and comes with “additional benefits to the user.” And look, maybe that’s fine. But that’s also a slippery slope, or at least it’s the beginning of a slippery slope, or at minimum it’s standing near a slope and commenting on how slippery it looks.

The line between “here’s a sponsored recommendation that’s relevant” and “the AI is now a salesperson” is exactly the line Google has been blurring for years with Shopping results and AI Overviews and all the other things that are technically ads but don’t look quite like ads.

5. Enterprise May Accept Ad-Supported AI

Thesis: Large enterprises are considering ad-supported coding agents, using Google Search as the precedent for employees seeing ads in work tools.

Pull Quotes:

“We have gotten some of our big enterprise customers that are interested in using AMP free, ad supported for their enterprise users... The analogy they have is, yeah, this sounds weird. Why would we let, you know, Big Bank let our devs use an ad-supported product, but they use Google and they see ads there.” — Quinn Slack

“If, say, JP Morgan employees go to google.com, there’s no enterprise web search contract that those two companies have signed. JP Morgan is not paying $200 a month for their employees to use Google. It’s a free product, frictionless free product, and they see ads there.” — Quinn Slack

“When you think about how much money these enterprises are spending on coding agents that has been unbudgeted and really hard to prepare for, it’s compelling in the short term.” — Quinn Slack

Analysis:

This is, I think, the most counterintuitive claim in the entire interview, and Slack seems genuinely surprised by it. The traditional assumption is that enterprises will pay almost anything to avoid ads: ads mean data leakage concerns, compliance headaches, and unprofessional optics, which are three categories of problem that enterprise buyers are exquisitely sensitive to. But the Google precedent is actually kind of persuasive, once you sit with it. Fortune 500 employees absolutely use ad-supported Google Search for work queries, every single day, and nobody—not the CTO, not the CISO, not the procurement team—has negotiated an enterprise contract to make those ads go away. If developers already see ads when they Google “how to parse JSON in Python,” is it really that different to see ads in their coding agent? I’m genuinely not sure, which is why I’m asking.

The budget pressure angle is probably underrated in all of this. Coding agents are genuinely expensive—Slack says power users spend $200-300/month, which adds up remarkably fast across an engineering organization—and that spend materialized faster than IT budgets could adapt. If a CFO is staring at an unexpected seven-figure line item for coding tools and wondering how exactly this happened, the pitch “what if some of your devs use the free tier for low-stakes tasks” starts sounding not just reasonable but prudent.

What happens next?

Here’s the thing about ads that makes people uncomfortable: they work. Not always, and not for everyone, and certainly not in the way that advertising executives would have you believe when they’re trying to sell you advertising—but they work often enough and well enough that they’ve funded most of the internet you actually use. And what Quinn is betting on, I think, is that the same economic logic that made Google and Facebook and yes, newsletters, into viable businesses can also make AI into a viable business, which is a thing it currently is not, at least not at the scale people are investing in it.

The early returns—$5-10M ARR in a month, 95% advertiser close rates, enterprise customers who should hate this but somehow don’t—suggest he might be onto something. Whether other coding agents follow suit, whether the “ad network for developer tools” vision materializes, whether users eventually revolt against the whole enterprise—these are open questions, and they’ll remain open for a while. But if you had asked me six months ago where ads would first show up in AI products, I would not have guessed “in the terminal, served to the people who invented ad blockers.” And yet here we are. Sometimes the future arrives in weird places first.