Are We in an AI Bubble?

Every time I open X, I see someone asking the same question.

"How can OpenAI charge $20/month when I'm burning hundreds of dollars worth of API credits?"

It's a fair question.

If you've ever compared ChatGPT Plus, Claude Pro, Cursor, or Codex against their API pricing, the math doesn't seem to work. Heavy users can easily consume what looks like hundreds of dollars of inference while paying just $20 a month. Even the $200 "Pro" tiers appear to bundle thousands of dollars of API-equivalent usage.

So who's paying the bill?

The short answer is: not the consumer.

---

The consumer plans are bait

Calling them "subscriptions" is technically correct, but they're really customer acquisition tools.

OpenAI has roughly 50 million paying ChatGPT subscribers, yet consumer subscriptions now account for less than 60% of its business, with enterprise and API products contributing over 40% of revenue.

Think about it this way.

If you get millions of developers, students, and businesses to build their daily workflow around your model, you've already won half the battle. Once companies start integrating your APIs into production systems, switching providers becomes expensive.

The $20 plan isn't the product.

It's the onboarding experience.

---

Enterprise is where the real money is

This surprised me the most.

OpenAI is reportedly generating around $2 billion in revenue every month, while Anthropic is already operating at roughly a $14 billion annualized revenue run rate. Cursor, despite being only a coding assistant, claims over $1 billion in annual recurring revenue, and Cohere sits around $240 million ARR.

Those numbers don't come from hobbyists asking ChatGPT to explain Kubernetes.

They come from enterprises buying API access, business subscriptions, government contracts, and dedicated deployments.

One large enterprise contract can be worth more than thousands of individual $20 subscriptions.

---

The economics are still brutal

This doesn't mean AI companies have magically solved profitability.

Quite the opposite.

Training GPT-4 reportedly cost around $100 million in compute alone, and every new frontier model pushes hardware requirements even further. Meanwhile, serving those models isn't cheap either. Depending on utilization, inference on an NVIDIA H100 GPU has been estimated to cost anywhere between $0.21 and $15.25 per million output tokens.

That's before paying researchers, engineers, networking, storage, power, and data center costs.

It's no surprise these companies continue raising enormous amounts of capital. OpenAI alone recently raised funding at an approximately $850 billion valuation, giving it room to prioritize growth over short-term profitability.

---

Wait... aren't models getting more expensive?

This is where I changed my mind while reading the research.

For years we've been told AI gets cheaper over time.

That's true for older models.

It's not necessarily true for the newest ones.

Take Anthropic as an example.

Older Haiku models were incredibly cheap. Then came Sonnet. Then Opus. Each generation became significantly more capable—but also significantly more expensive.

The same pattern exists with OpenAI.

GPT-5's API pricing is dramatically higher than GPT-4-era models, and the highest-end GPT-5.2 Pro reportedly exceeds $168 per million output tokens.

We're no longer paying for text generation.

We're paying for reasoning.

---

Doesn't that make local AI inevitable?

Not yet.

Open-source models have become shockingly good, and Apple Silicon has made local inference practical for many workloads.

According to the research, an M4-class machine can comfortably run open-weight models in the 30B–70B parameter range, making local AI increasingly attractive for privacy and high-volume workloads.

But if you want the very best reasoning models, multimodal capabilities, or large-scale agents, you're still going to the cloud.

For now, local models complement frontier models—they don't replace them.

---

So... are we actually in a bubble?

I think the answer is yes—but probably not in the way people mean it.

The industry has plenty of bubble characteristics.

Companies are spending hundreds of billions on infrastructure.

Valuations are enormous.

Consumer products are heavily subsidized.

Many startups are essentially wrappers around the same handful of foundation models.

Eventually, a lot of those companies won't survive.

But that's happened before.

During the dot-com boom, thousands of internet companies disappeared.

The internet didn't.

AI feels similar.

The speculation might be excessive.

The valuations might be optimistic.

Some companies will almost certainly fail.

Yet it's difficult to argue AI has no real demand when OpenAI is generating roughly $24 billion annualized revenue, Anthropic is growing at extraordinary speed, and enterprises continue signing large AI contracts.

That's not hype.

That's a real market.

---

Final thoughts

The biggest question isn't whether AI survives.

It will.

The real question is whether today's valuations assume growth that simply can't continue forever.

If enterprise adoption keeps accelerating, inference costs continue falling, and companies eventually turn today's users into profitable customers, the current spending spree may look perfectly rational in hindsight.

If growth slows before the economics improve, we'll probably look back at this period as the AI equivalent of the dot-com boom—where the technology won, but many of the companies didn't.

Personally, I don't think we're watching an AI bubble.

I think we're watching a company bubble built on top of one of the most transformative technologies we've seen in decades.