Bubble or not? The money loop, the token revolt, and why the AI demand is still real

In part one I laid out the two numbers that make this whole AI buildout feel like a dare: Big Tech spending around $725 billion in a single year, and JP Morgan's math saying the industry needs $650 billion in fresh revenue every year, forever, just to earn a 10 percent return. If you have not read that one, start there, because it sets up the stakes. This is the half where I stop describing the bet and start poking at whether it holds.

The money that keeps going in circles

Here is where the bubble comparison gets sharp. A cluster of the biggest names in AI are investing in each other in ways that make demand look larger than it might actually be.

• Nvidia has committed money to OpenAI.
• OpenAI commits enormous sums to cloud providers like Oracle.
• Those providers turn around and buy Nvidia chips to build the capacity.
• Amazon invests in Anthropic, which runs heavily on Amazon's own cloud.

By 2026, analysts had counted more than $800 billion of these interlocking arrangements. Supporters call it a virtuous circle that locks in scarce supply. That is a fair reading. If you know demand is coming, tying up chips and capacity early is smart. Critics, including the investment firm GMO, see something else. It looks a lot like the vendor financing that inflated the internet bubble, where companies bought each other's services to make growth look organic right up until it did not.

The deciding question is easy to say and hard to measure: how much of the revenue comes from real customers outside the circle? When one hundred billion dollars can show up as a chipmaker's revenue, a lab's funding, and a cloud's backlog all at once, the headline numbers stop meaning what they appear to mean. And when a single stalled negotiation between Nvidia and OpenAI rattled three of the largest companies in the world back in February, it was a reminder of how tightly wound this all is.

Enterprises noticed the bill

Now the part that I think matters more than people realize. Through the first half of 2026, the companies actually buying AI at scale started saying out loud that it costs too much.

The unit of AI is the token, the small chunk of text a model reads and writes. Agentic tools that plan and write code and click around burn through tokens frighteningly fast. As vendors move toward metered pricing, where you pay for exactly what you use, those bills started to sting.

Then DeepSeek showed up. The Chinese lab's V4 Flash model runs at about $0.14 per million input tokens and $0.28 per million output tokens. In May it made a roughly 75 percent price cut to its heavier V4 Pro model permanent. At comparable context lengths, that lands somewhere between 20 and 100 times cheaper than the flagship Western models.

The moment it got real was June, when reports surfaced that Microsoft was weighing a self-hosted version of DeepSeek's V4 for parts of its Copilot lineup, specifically because token costs were pressuring the economics. When the company most deeply tied to OpenAI is quietly pricing a Chinese alternative, the cost problem is not hypothetical anymore.

But the demand is not imaginary

It would be easy to stack all this up and call it a bubble. That would also be wrong, or at least incomplete, because the counterevidence is strong.

Cloud revenue is genuinely surging. Google's cloud business grew more than 60 percent year over year in early 2026. Nvidia's data center revenue keeps setting records on real orders.

And on the software side, the money is arriving faster than almost anyone guessed. Anthropic, the company behind Claude, passed OpenAI on annualized run rate in April 2026, reaching about $30 billion, up from roughly $9 billion at the end of 2025. Around 80 percent of that comes from enterprises and developers, not free consumers. The company projected its first operating profit as early as the second quarter. More than a thousand companies reportedly spend over a million dollars each, per year, on it. That is not the profile of hype. That is real workflows getting replaced by software that pays for itself.

Meanwhile OpenAI is the counterweight. It booked about $13 billion in revenue in 2025 and hit roughly a $25 billion run rate by February 2026, which is impressive, and yet it is projected to lose around $14 billion this year, does not expect positive free cash flow until near the end of the decade, and has committed to enormous infrastructure obligations. It filed confidentially for an IPO in June. A business losing more than a dollar for every dollar it earns is asking public markets to fund years more of exactly that.

Both things are true at the same time. The technology works, the revenue is real and growing fast, and the spending may still be running ahead of what near-term returns can justify. Reality does not owe anyone a clean verdict.

So, where do I land?

The most honest answer is the one JP Morgan buried in the same report with the eye-watering numbers. "Even if everything works," they wrote, "there will be spectacular winners and probably equally spectacular losers, given the sheer amount of capital involved and the winner-takes-all nature of parts of this market." That is not a bubble call. It is something more useful. It says the real question is not whether AI is real. It is who survives the race to build it.

History rhymes here. The railroads were transformative and also bankrupted a generation of the companies that laid the track. The fiber optic boom of the late 1990s wired the world for the internet age and wiped out the firms that overbuilt it. The internet itself changed everything, exactly as promised, several years after the market that bet on it crashed. Transformative technology and a brutal financial reckoning are not opposites. They tend to show up together.

So here is what I would actually watch, as a developer and not a trader:

• Free cash flow at the hyperscalers, as capex eats into it.
• Debt loads and credit spreads at the companies leaning hardest on borrowed money.
• Whether enterprise revenue keeps compounding once the novelty wears off and the token bills come due.
• How much of the demand comes from genuine outside customers, rather than the same dollars circling a small group of firms.

The tech is real. The revenue is real. Whether they justify $725 billion in a single year is the trillion-dollar dare, and nobody actually knows the answer yet - not JP Morgan, not the CEOs, not the people writing confident takes in either direction. What we will find out, probably sooner than the timelines suggest, is not whether AI matters. It is which of the giants placing this bet are still standing when the bill comes.

If you jumped straight here, go back and read part one for the two numbers this whole thing rests on. It is the setup that makes this half land.