The frenzied cash burn in AI is starting to look pretty familiar
The Wall Street Journal reports that Microsoft is losing on average more than $20 per user of Copilot. We've certainly seen this before.
Author’s note: To provide an update from last week, I’m still recovering from a rough injury so I’m working at mixed capacity. There will be a paid issue coming out this week, and I am still targeting two paid issues next week. Apologies for any typos and any other assorted mix-ups.
When it comes to growth-at-all-costs spending in a massively hyped emerging technology, it turns out that everything old is new again.
We’re seeing the initial signs of that in AI. In this most recent case, The Wall Street Journal reported that in the first few months of the year Microsoft was losing on average more than $20 per month per user of Copilot—with the company losing as much as $80 per month on some users. Copilot costs $10 per month for individual accounts, while Business accounts are $19 per month.
The extreme gap there represents a really familiar consumer profile: the typical user and the power user. Copilot is essentially available on-demand despite being offered for a single monthly price. Power-users inevitably get much more of a benefit here, but its free availability offers a pretty amazing user experience for people that are playing around with it for the first time—and building it into their daily habits.
The colossal burn from these operations that deep-pocketed companies like Microsoft can afford is less about the actual usage of the product and more about user acquisition. AI has become such a ridiculously hot sector (even if the hype is dying down) that word of mouth trumps all forms of marketing.
Microsoft wants to get GitHub copilot deeply integrated with the daily activity of every developer. By doing that, it can eventually start to build a business around it—either through upselling new products, reducing its cost of operations (like pulling back on marketing), or simply raising the prices eventually.
GitHub even released some stats this week about surveys for developers that basically boiled down to customer satisfaction without releasing general new metrics about usage. Microsoft here is clearly angling to get Copilot into the hands of as many people as possible and will make whatever sacrifices necessary to ensure it becomes the home of copilot experiences across a variety of products, with its AI pair programmer being an entry ramp within organizations.
The foundation model providers, like Anthropic, OpenAI, and Inflection, have all managed to secure billions of dollars in capital. While Microsoft’s Copilot burn is one example, we could easily walk backwards into understanding how expensive all these products are. Dylan Patel and Afzal Ahmad at SemiAnalysis estimate the cost of running ChatGPT is around $694,444 per day to operate, and they are about as informed as it gets here.
But the land grab in AI—and its frenzied cash burn—is once again reminiscent of an earlier land-grab era where investors poured money into a space that was actively setting cash on fire. In this case, it was for the sake of wooing users (and drivers) to achieve growth above all else: the on-demand food delivery and transportation industry.
Parallels to the burn in on-demand
The whole on-demand industry has since coalesced into roughly two “winners” and a third-place finisher with Uber, DoorDash, and Lyft respectively. But these companies all had explosive cash burn throughout their earliest years as they tried to aggressively pull in drivers and users amongst colossal levels of competition. It helped, too, that we were in the ZIRP party decade.
While we aren’t in the zero interest era any more, the math for all these mega-deals is still a bit fuzzy due to the use of compute credits in these infusions. The breakdown for a funding round usually isn’t available, but it’s safe to say that many (most?) funding rounds that you see from the cloud providers have a subset of GPU resource allocation associated with it. This is not funny money as running—and particularly acquiring—the hardware is very expensive. But it also isn’t an on-the-spot dollar wired to a back account (and instead could be recognized in some different fashion).