OpenAI Burned $3.7 Billion in One Quarter. That's Why There Are Ads in ChatGPT.

I was looking something up on my phone in February, logged out of my Plus account, and there it was at the bottom of the answer: a small sponsored line, cleanly labeled, offering me a product related to what I had just asked about.

My first reaction was mild annoyance. My second reaction, about ten seconds later, was that this was the least surprising thing to happen in AI all year.

Because I had read OpenAI's Q1 numbers. And once you read the numbers, ads stop looking like a product decision and start looking like an arithmetic one. There is a gap between what this company earns and what it has promised to pay, the gap is measured in hundreds of billions, and subscriptions cannot close it. Not at any realistic price. Not with any realistic number of subscribers.

So let me walk through the money, because the answer to why does OpenAI need ads is sitting right there in the filings, and it is not complicated. It is just large.

OpenAI Made $5.7 Billion Last Quarter and Burned $3.7 Billion Doing It

In June 2026, The Information reported OpenAI's first-quarter figures based on documents the company shared with shareholders. Revenue: $5.7 billion. Cash burned: $3.7 billion. Both roughly tripled year over year. Gross margin improved to 39% from 33%. R&D spending came in at $8.6 billion for the quarter. The company ended March with more than $73 billion in cash and marketable securities.

Read that burn figure again. It is not 10% of revenue. It is 65% of it. For every three dollars that came in the door, roughly two went out and did not come back.

Ed Zitron, working from the same reporting, put the non-GAAP operating margin at negative 122%, which implies an operating loss around $6.95 billion for the quarter. His own caveat is worth repeating: those margins appear to exclude large line items like stock-based compensation, and it is not clear how training costs are treated. So treat the $6.95 billion as a well-reasoned estimate rather than an audited fact. The $3.7 billion cash burn is the number I would put weight on.

The 2025 picture is better documented. Zitron published what he described as OpenAI's audited financials, and the Financial Times verified them: revenue of $13.07 billion for the year, total costs of $34 billion, an operating loss of $20.92 billion, and a headline net loss of $38.5 billion. That net loss number gets quoted a lot without the asterisk it deserves. Most of the difference between the operating loss and the net loss came from a $41.55 billion non-cash charge tied to the nonprofit-to-for-profit restructuring. That is accounting, not money leaving a bank account. The $20.9 billion operating loss is the one that describes the business.

None of this means the business is fake. Revenue tripling year over year is genuinely extraordinary, and I say that as someone who pays for these tools every month and would be worse at my job without them. But growth that fast against costs that large only works if something eventually scales faster than the compute bill.

The Number That Actually Drives Everything Is $665 Billion

Here is the line that reframes the whole conversation. According to the same reporting, OpenAI has roughly $665 billion in long-term computing procurement commitments through 2030, spread across deals with Oracle, Microsoft, Amazon, Nvidia, and Broadcom.

Not $665 billion of spending they might do. Commitments.

Spread that across roughly five years and you get an average of about $133 billion per year. Now annualize Q1 revenue: $5.7 billion times four is $22.8 billion. That is every revenue line the company has, consumer subscriptions, enterprise, API, everything.

So current total revenue covers something like 17% of the average annual compute commitment.

I want to be careful here, because that comparison is rougher than it looks. The commitments are not evenly distributed, revenue is growing fast, and a chunk of that compute buys capacity that generates future revenue. OpenAI's own forecast, per the same documents, is $25 billion of burn in 2026 and $57 billion in 2027, which is the company telling investors plainly that it does not expect this to get better before it gets worse. Reporting from January said internal projections put cumulative losses at around $44 billion through 2028 before a profit in 2029.

Those are forecasts. Forecasts from a company raising money. Weigh them accordingly.

Subscriptions Cannot Get There. Here Is the Napkin.

At the end of Q1, OpenAI had roughly 55 million paying subscribers against an average of about 905 million weekly active users. That is a conversion rate near 6%.

Six percent is not bad. For a consumer product, six percent is actually quite good. If you want the single sentence answer to why does OpenAI need ads, it is this ratio multiplied by that commitment. Spotify and Dropbox would recognize that shape. The problem is not the conversion rate. The problem is what you have to multiply it by.

Take the $133 billion average annual compute commitment. Take a $20 per month subscription, which is $240 a year. To cover compute alone, ignoring salaries, ignoring research, ignoring everything else, you would need roughly 554 million people paying $20 every month.

Against about a billion weekly users, that means more than half of everyone who touches ChatGPT would need to become a paying subscriber. Today it is six in a hundred.

You can improve that napkin. Enterprise seats and the Pro tier pull average revenue per paying user well above $20, so the real subscriber count needed is lower. Fine. Cut it in half and you still need 275 million paying subscribers, five times the current base, just to break even on the machines. And every honest read of consumer software says the 6% who pay are the 6% most willing to pay. The next hundred million are harder, not easier. They are the ones who already decided the free version is good enough.

This is the same structural squeeze I wrote about in Your Favorite Tool Is Getting Worse. That's Not Bad Product Management, That's Panic., except the numbers have three more zeros on them.

Advertising Is the Only Revenue Line With a Ceiling That High

So what business models actually produce $100 billion a year from consumers who mostly will not pay?

There is exactly one proven answer, and it is about thirty years old. In 2024, Google's advertising business brought in over $260 billion. Meta brought in over $160 billion, almost all of it ads. No consumer subscription business on earth is in that weight class.

Advertising is the only model that monetizes the 94% who will never subscribe, and it monetizes them harder the more the product knows about their intent. A conversation where someone describes exactly what they need, in their own words, before they have chosen a product, is arguably the highest-intent advertising surface ever built. Google spent two decades inferring intent from three-word queries. ChatGPT gets it in a paragraph.

OpenAI announced the advertising test on January 16, 2026, and started showing ads on February 9 for US users on the Free and Go tiers. The stated format is a labeled sponsored suggestion at the bottom of an answer, not a paid placement inside the answer itself. Plus, Pro, Business, Enterprise, and Education stay ad-free. OpenAI says conversations stay private from advertisers, that it does not sell data to advertisers, and that ads do not influence what the model says.

Then it moved fast. Thirty-one European countries on August 18. India on August 27. And on September 1, OpenAI announced ChatGPT Ads had crossed $1 billion in annualized revenue run rate in under 200 days, across more than 40 countries and tens of thousands of advertisers.

That is a genuinely fast start. It is also, against a $133 billion average annual compute commitment, about three quarters of one percent. Ads are not the solution today. They are the only line on the page whose ceiling is high enough to eventually be the solution, which is a different and more interesting claim.

What This Actually Changes for You

Two things, and I want to keep them separate.

The first is honest and mostly fine. The free tier now has a business model. Before ads, every free conversation was a pure cost with a vague hope of conversion attached. That is not a stable arrangement, and unstable arrangements end with the free tier getting worse. Ads are the price of the free tier continuing to exist and continuing to be good. If you have ever wondered how a genuinely useful free product pays for itself, I went through the same question in a different context in what Cloudflare is actually selling on its free tier.

The second thing is the one to actually watch. Right now the ad and the answer are separate: sponsored content sits below, labeled, and OpenAI states it does not affect the response. Hold that line and this is just the internet working the way it has always worked. The pressure comes later, when growth flattens and the obvious way to increase revenue per conversation is to let commercial relationships nudge what gets recommended inside the answer. Nobody announces that. It arrives as a ranking change.

I am not predicting it. I am saying the incentive is now permanently in the room, and it was not there in 2024.

For anyone whose business depends on being mentioned by these systems, that shift changes the game in ways worth their own analysis, and I have written separately about how to get cited by ChatGPT. This post is about the balance sheet.

The Part Where I Argue the Other Side

Burning cash is not the same as failing. I need to say that clearly, because the numbers above are easy to weaponize and I do not want to.

Amazon lost $1.4 billion in 2000 on $2.76 billion of revenue and did not post a full year of profit until 2003. Every one of those years, somebody wrote the obituary. The people who did the arithmetic were correct about the arithmetic and completely wrong about the outcome, because Amazon was building assets that would still be there after the losses stopped.

The scale is different, though, and the difference matters. Amazon's worst year was a $1.4 billion loss. OpenAI's cash burn in a single quarter was two and a half times that.

So here is what would have to be true for this to work out. Inference costs keep falling faster than usage grows, so gross margin keeps climbing the way it did from 33% to 39%. Enterprise revenue compounds into the largest line on the page, because seats sold to companies are stickier and higher-priced than consumer subscriptions. Ads scale from a $1 billion run rate into the tens of billions within a few years, which is aggressive but not impossible given the intent quality. And that $665 billion of compute turns out to be an asset with real utilization rather than capacity bought for demand that never showed up.

Here is what would have to be true for it not to. Weekly active users flatten, which the Q1 average of 905 million against a 920 million February peak hints at slightly. Open-weight models close enough of the quality gap that the API business gets priced toward zero. Advertisers try ChatGPT ads, cannot attribute the conversions cleanly, and quietly move budget back to channels they can measure. And the compute commitments come due on schedule regardless of any of it, because that is what a commitment is.

The contrast that keeps my attention is Anthropic. Reporting suggests their burn runs closer to a third of revenue, with positive cash flow projected by 2028. I have no way to verify those figures independently and neither does anyone outside the company, so treat them as reported rather than known. But if two companies are selling similar products and one burns 65 cents per dollar of revenue while the other burns 33, that gap is a strategic choice about how much future to buy today, and only one of those choices requires an advertising business to survive it.

What I Actually Think

OpenAI reportedly filed confidentially for an IPO in mid-2026, then according to CNBC in August was leaning toward pushing it to 2027. If you want a single explanation for the timing of everything above, that is probably it. A company preparing to sell shares to the public needs a second revenue line that is not "more subscribers," and it needs a chart of that line going up before the roadshow. A $1 billion run rate in 200 days is exactly the slide you want.

I am not cynical about it. I pay for four AI subscriptions and they have changed how I work. I have gone as far as running the numbers on buying my own hardware to run models locally, which tells you how seriously I take the cost side of this.

But I have stopped reading ads in ChatGPT as a product team's idea. They are what happens when you commit $665 billion to a compute bill and your best consumer product converts at six percent. Everything downstream, the pricing tiers, the free tier that keeps getting more capable, the ad slot at the bottom of the answer, follows from that one line on the balance sheet. So, why does OpenAI need ads? In the plainest terms I can manage: because the compute bill arrived before the subscribers did.

Watch for whether the ad stays below the answer or starts creeping into it. That is the number that will tell you how the rest of this is going.

If you are trying to work out what any of this means for your own site's visibility, that is most of what I do at Wunderlandmedia. Happy to talk it through.