OpenAI Loses $33B a Year. Its Fix Is the Most Dangerous Thing for Your Job.

There is a line from an OpenAI investor pitch that should be far more famous than it is:

"We have no current plans to make revenue. We have no idea how we may one day generate revenue. We have made a soft promise to investors that once we've built this generally intelligent system, we will ask it to figure out a way to generate an investment return for you."

For three years, that was not a joke. It was the business model.

And yet in 2026, OpenAI's revenue exploded by 92% to a run rate of $25 billion a year — while the company is on track to lose $33 billion. In January 2025, Sam Altman admitted they were losing money even on the $200/month Pro subscribers.

Read that again. In every normal business on earth, your best customers are your most profitable customers. In OpenAI's world, the best customers are the least profitable ones.

Wall Street looked at this and called it an AI bubble. The thesis is brutally simple: if the biggest and best player cannot make money, how will anyone else? And if nobody makes money, why is everyone pouring trillions into this?

But there was always one escape hatch. One way the bubble turns into a genuine gold rush: somebody has to collapse the cost of intelligence by 10 to 100 times.

In the last 90 days, that collapse started. And this is where it stops being a finance story and becomes a career story — because the only thing standing between most professionals and automation right now is not capability. It's price.

First, understand the actual problem: inference

The whole nightmare comes down to one word.

Think of a chef. Training a model is sending that chef to culinary school. You do it once, it costs a fortune, and the chef comes out with extraordinary skills. This is why OpenAI spent billions making the smartest models on earth.

Inference is what happens every time a customer orders a dish. The stove fires up, the food gets cooked, and it costs money — every single time. Every question you ask, every agent you run, spins up GPUs and burns cash.

Training happens once. Inference happens forever.

Now multiply that by roughly 900 million people placing billions of orders a week, and you get the number that defines the industry: OpenAI is set to burn around $14 billion on inference alone in 2026.

Why the best customers lose the most money

Take a heavy user — a software engineer who pays $20 a month and genuinely loves the product. She uses it to write code, debug, draft emails, plan trips, and think through decisions.

Analysts estimate a heavy user firing ~100 queries a day can burn through more than $30 of compute in a month.

She pays $20. She costs $30. OpenAI loses $10 on her every month — and the more she loves the product, the more she uses it, and the deeper the hole gets.

So why not just charge her $40? Because OpenAI is caught in two traps.

Trap 1: Competition is racing the price of intelligence to zero

The price collapse in this industry has no precedent in modern business. GPT-3-level intelligence cost roughly $60 per million tokens when it first shipped. By late 2024, that same level of capability cost about 6 cents — a ~1,000x collapse in around three years.

Google made Gemini free at the consumer tier. Anthropic cut prices dramatically in a single announcement. In this market, if OpenAI raises the price to $40, the customer switches to Claude or Gemini in thirty seconds.

You cannot raise the menu prices.

Trap 2: OpenAI doesn't own the kitchen

Jensen Huang describes AI as a five-layer cake: energy, chips, infrastructure, models, and applications. Look at who actually owns each layer.

Every ChatGPT answer runs on Nvidia's chips, in Microsoft's data centres, powered by someone else's electricity. OpenAI has been a restaurant that doesn't own the kitchen, the ovens, or the building — and can't raise its prices.

There is exactly one way out of that trap. Stop renting the kitchen. Build your own oven.

The 90-day comeback nobody was ready for

While everyone stared at the losses, OpenAI shipped four things in quick succession. Individually they read as product news. Together they are a single, coherent escape plan.

1. A model that matched the leader at a third of the cost

For most of 2026, the coding crown belonged to Anthropic. Claude Fable 5 was good enough that developers happily paid $10 per million input tokens, and Anthropic's ARR ran to roughly $47 billion against OpenAI's $25 billion.

Then on July 9, 2026, OpenAI shipped GPT-5.6 with Sol, Terra, and Luna.

Most coverage treated it as another version bump. The efficiency numbers say otherwise:

Sol reaches near-frontier intelligence at roughly half the cost and 60% less time. Terra beats Claude Opus 4.8 on several axes. And Luna does a comparable task for about seven cents.

The market reaction told the real story. Anthropic had Fable 5's subscription access set to expire on July 7. It got extended to July 12. Four days after Sol landed, it was extended again to July 19. Read between the lines: it looks a lot like Anthropic had planned to pull its best model out of subscription tiers to protect margins — and then a competitor shipped something close enough in quality and dramatically cheaper, and the plan changed.

2. Codex turned questions into work

For two years the deal with AI was simple: you type a question, it gives you an answer.

Codex breaks that. It writes the code, builds the tool, ships the site, runs the multi-step job, and comes back when it's finished. Active users exploded by 500%. Steve Jobs called the computer a bicycle for the mind — Codex is closer to an autopilot for it, and OpenAI confirmed it stays included on every ChatGPT plan.

Marketers, PMs, lawyers, analysts, researchers: anyone can now commission software instead of requesting it.

3. GPT Work folded the desktop into one screen

Email, docs, files, and Codex in a single surface. The significance isn't the UI — it's that the agent now sits where the work actually lives, with the context it needs to act without being asked twice.

4. Jalapeño: OpenAI finally owns a second layer

Jalapeño is OpenAI's first custom silicon — an inference-only chip designed with Broadcom. Broadcom's CEO has said early tests show roughly 50% lower cost per token than current Nvidia GPUs, with performance on par with Nvidia's best Blackwell parts.

If that claim holds, the $14 billion inference bill gets cut close to half over the next few years.

So the model, the agent, and the chip are not three stories. They're one:

Here's the part that concerns your career

Every analysis of this story stops at "will OpenAI become profitable?" That's the wrong question for anyone who isn't an investor.

The right question is: what happens to the automation decision when a task costs seven cents?

Right now, most jobs are not protected by capability. They're protected by cost and friction. A manager doesn't automate a workflow because the model can't do it — they don't automate it because the migration is a project, the tokens add up, and the ROI case is annoying to write. That friction is the actual moat, and almost nobody admits it out loud.

Cheap inference dissolves it. Consider what a 10x cost collapse does to the arithmetic:

At $3.15 a task, automating you is a project. At 7 cents, it's a default. Nobody convenes a meeting to approve a rounding error — and there is no procurement process to slow down, no sponsor to lobby, and no ROI review where a human advocate might speak up for the team.

Then there is Codex, which changes the shape of what gets automated. Chat-era AI replaced *sentences* in your workflow. Agent-era AI replaces the *workflow*. That's why one Codex task can cost what a hundred ChatGPT questions cost — and why cost per token becomes the single most important number in the labour market.

The trap in the comeback: Jevons paradox

There is one more wrinkle, and it's the reason none of this stabilises.

Everyone assumes cheaper AI means OpenAI saves money. It doesn't. When you make something 10x cheaper, people don't use less of it — they use insanely more of it. That's Jevons paradox, and it's been true of coal, electricity, bandwidth, and cloud storage.

Which means the cost curve never gets to rest. OpenAI cuts inference costs, usage explodes past the savings, and the pressure to cut costs again returns immediately. The whole industry is locked into a permanent race to push the price of thinking toward zero.

For your career, the implication is uncomfortable but clear: this is not a wave that crests. There is no version where the cost of automating knowledge work goes back up.

Who is most exposed

The defensive playbook

The honest closing

There is a genuinely optimistic reading here, and it deserves stating. Trillion-dollar companies are burning their balance sheets to demolish the barrier between an idea and a working thing. If you have ever wanted to build something and been stopped by cost, time, or not knowing how — that constraint is dissolving in real time. When thinking is nearly free and building is nearly free, an infinite canvas opens up for anyone willing to use it.

But that canvas only rewards people who shift from operating to directing. The same price collapse that lets you build a company from a prompt lets your employer replace a function with one.

Same curve. Two completely different outcomes. Which one you get depends on whether you spend the next twelve months proving you can execute — or proving you can decide.