GitHub’s AI coding assistant, Copilot, is a moneymaker

GitHub’s AI coding assistant, called Copilot, has gone from an experiment to a moneymaker, the company’s CEO Thomas Dohmke said in an exclusive interview with Semafor.

“We are happy about the growing positive margins the product has,” he said.

The revelation helps answer an important question about the business of generative AI: Because the powerful models require so much costly compute power to run, can companies make a profit on products that rely heavily on the nascent technology?

Dohmke’s answer was an unequivocal yes. “The cost of goods sold, which is all the costs to run Copilot per user, is lower than the price that we’re charging,” he said.

Copilot was the first consumer product to use AI models from OpenAI, the company behind ChatGPT. It helps software developers write code, the same way Gmail suggests the next word in a sentence while typing an email.

GitHub initially launched Copilot as a free, experimental product in 2021. Dohmke didn’t say when Copilot began posting positive margins, but he implied the company has been building infrastructure in anticipation of demand.

That is a capital-intensive investment, especially considering that AI models like OpenAI’s require graphics processors, which are not found in traditional data centers.

On Wednesday, GitHub launched several new products that significantly increase Copilot’s role in the company’s offerings. In December, it will make Copilot Chat generally available, allowing developers to treat it more like a colleague, moving it even further beyond its autocomplete roots.

Another new feature will enable companies to connect their code repositories to Copilot. For instance, if a software developer starts a new job at a company, they’ll get suggestions based on that company’s existing code.

Microsoft-owned GitHub also unveiled a new Copilot partner program, allowing third-party company integration. When developers write code, they usually connect with several third-party services that handle various aspects of software products. Now, those integrations can happen within Copilot.

All of GitHub’s announcements will likely lead to a big jump in compute costs on Microsoft’s servers. And yet, the company isn’t increasing prices. An individual account for Copilot will still cost $10 per month, or $100 per year. Enterprise accounts will be $39 per month, per user.

That suggests the company is finding a way to offer the service in a way that’s economical. If that weren’t the case, it would probably wait longer to start ramping up offerings like Copilot Chat and Enterprise.

“In a fast growing product, you always have to be a little bit ahead of the demand curve,” Dohmke said. “Otherwise, you’re running into a point where the users can no longer sign up, or that sign up to the system is down.”

The details of how GitHub has been able to ramp up capacity so quickly is probably more a question for Microsoft chief technology officer Kevin Scott, the architect of the company’s ambitious plan to give its data centers the ability to run AI models like GPT-4 at scale for hundreds of millions of users.

But Dohmke said, broadly speaking, it’s a combination of things: “The GPUs get faster and more efficient and the models get optimized to solve problems. And the applications that build on top of the model and the GPU, get smarter.”

Based on conversations I’ve had with people at Microsoft and other companies, I think of GPT-4 like the highest-paid partner at a law firm. You only want to give that partner the most complex work. For the simplest tasks, paralegals and associates are sufficient. And by the time those paralegals and associates finish their work, the partner can more efficiently do the job because all the messy busywork is out of the way.

When you prompt Copilot and other products that rely on GPT-4, I think you are actually tapping smaller models — the paralegals and associates — and they are translating that prompt in the most efficient way, then handing it off to GPT-4, the partner.

Another big takeaway from Wednesday’s announcements is that the company is building in network effects. If GitHub convinces a critical mass of third-party providers to join the partner program, the convenience will be enticing for developers. It will also be difficult for other platforms to catch up to GitHub.

There’s also the big question of where AI-assisted coding is headed and what it will mean for all of us. Eventually, non-coders could use AI assistants to help them complete complex coding tasks.

“I’m convinced that a decade from now, we’ll all have our personal Jarvis from Ironman or whatever you want to call it,” Dohmke said. “An assistant that has been with you and knows all the things about your personal and your professional life that you’ve shared with the assistant.”

GitHub is establishing a dominant position in AI-assisted coding, another reflection of how Big Tech is consolidating power around AI. New calls for regulation will only help big companies hold onto that power by creating obstacles for smaller players to challenge them.

And there are some who are not convinced that GitHub is making money. Last month, The Wall Street Journal reported that it “has been a money loser because it is so expensive to run,” citing an anonymous source. According to the article, GitHub earlier this year lost up to $80 per user, per month. The company denies those figures.

• Some software engineers, the first class of professionals to be directly affected by the generative AI revolution, are questioning whether they have a future, according to this Business Insider article that cites posts on the app “Blind,” where tech workers can post to each other anonymously.