Free Chinese AI and the $2 Trillion Machine: Why This Feels Like 2006 Again

When I first heard that a Chinese lab had released an AI model matching the best American ones — for free, open-weight, download and run — I shrugged it off. Just another open-source release, no big deal. But then I started digging into the numbers, and the deeper I went, the more I felt a strange chill of déjà vu. It felt like 2006 all over again.

Back then, everyone was looking at record home prices and assuming they’d never stop climbing. But beneath the surface, a system built on rolling over debt — not paying it down — was quietly dying. I see the same structure now, except the houses are data centers, the mortgage brokers are venture capitalists and cloud providers, and the borrowers are startups promising trillions in future value while burning cash today. The uncomfortable part: sitting on top of this pyramid are the pensions and retirement savings of half the world, invested through index funds that don’t advertise how concentrated they’ve become.

Here’s what I’ve pieced together from earnings reports, regulatory filings, and reporting from Bloomberg, CNBC, and others. This isn’t investment advice — just the view of someone who’s spent too many hours staring at charts, and who’s tried hard to get the numbers right rather than just repeat what’s trending on social media.

Three Borrowers, One Carousel

I see three layers to this, and each one runs on the same principle: the next check has to be bigger than the last. If growth slows, the mechanism breaks — not because prices fall, but because the refinancing stops working.

1. The Homeowner of 2006

We remember 2008 as the crisis, but the real breaking point came two years earlier. The most popular mortgage for risky borrowers back then was the “2/28” — two years at a low teaser rate, then 28 years at a much higher one that almost nobody actually intended to pay. The plan: the house appreciates, you refinance at the new, higher value, pay off the old loan, and lock in another two cheap years. The debt was never really repaid — it was replaced.

That worked as long as home prices kept rising at double-digit rates. But in 2006, price growth slowed from roughly 15% to 8% a year. Prices were still near all-time highs, yet delinquencies on subprime mortgages nearly doubled, because the new loan wasn’t large enough to cover the old balance plus fees. The refinancing ladder had snapped quietly, a full year before prices actually fell and two years before the public panic. Fed Chair Ben Bernanke told Congress in September 2007 that a significant number of subprime borrowers were already in serious difficulty — a full year after the mechanism had effectively broken, and still a year before most people were paying attention.

2. OpenAI, Oracle, and the Circular-Financing Bubble

Fast-forward to today. OpenAI — a company that has never turned a profit — was valued at roughly $852 billion after completing a funding round in March 2026, with Amazon, Nvidia, and SoftBank among the backers (Amazon alone committed about $50 billion of the larger, roughly $110 billion round announced that February). By August 2026, Bloomberg reported OpenAI’s annualized revenue run rate had topped $40 billion — real and growing fast, but still dwarfed by the scale of the spending commitments the company has signed.

Some of the most interesting money in that round comes from OpenAI’s own suppliers. Amazon provides cloud capacity; Nvidia supplies the chips. So a meaningful share of OpenAI’s investors are also the companies that profit most from what OpenAI spends — a supplier funding its own customer, both sides booking revenue growth, both sides watching their valuations rise. Michael Burry — who famously shorted the 2008 housing bubble, and who wound down his own hedge fund, Scion, in late 2025 but hasn’t stopped posting — summed up a related deal (reports that Nvidia might backstop roughly $250 billion of financing for one of OpenAI’s data-center projects) with four words on X in July 2026: “Around and around we go.”

Then there’s Oracle. Larry Ellison, who once dismissed cloud computing from a conference stage, signed a $300 billion cloud contract with OpenAI in September 2025 — the largest in Oracle’s history. Oracle cut roughly 30,000 jobs in early 2026 as it redirected spending toward that build-out. Its contracted backlog reportedly swelled to somewhere around $500–600 billion, with OpenAI accounting for a large share of it — and by Oracle’s own disclosures, only a fraction of that backlog is expected to convert into cash in the next twelve months. In mid-2026, S&P downgraded Oracle’s credit rating to BBB-, one notch above junk, explicitly citing its concentration risk from OpenAI. Oracle’s stock had its worst week since the 2001 dot-com crash and, by some measures, fell more than 60% from its 2025 peak before partially recovering.

Meanwhile Alphabet (Google) reported its first-ever negative free-cash-flow quarter in Q2 2026 — roughly $6 billion in outflow — after record quarterly capital spending near $45 billion and a raised full-year capex outlook of about $205 billion. Anthropic, OpenAI’s chief rival, used the same investor enthusiasm to push its own valuation to somewhere between $900 billion and $1.1 trillion by May 2026, briefly making it the more valuable of the two — before OpenAI’s own valuation kept climbing too. The eye-popping detail: by mid-2026, the combined contracted backlog across the major cloud providers reportedly topped $2.3 trillion.

None of this makes any single company a fraud. But it does mean a huge slice of “the AI economy” currently consists of promises to spend money that hasn’t been earned yet, funded by investors who are also counterparties, all of it feeding stock valuations that ripple into pension funds and index portfolios most people don’t realize are this concentrated.

3. The U.S. Treasury — The Biggest Borrower of All

But that’s just the visible part. The third borrower is an entire country. The U.S. runs a deficit of roughly $2 trillion a year — in peacetime, not during a crisis. Total federal debt passed $38 trillion in 2026 and is climbing by billions of dollars a day. Interest payments alone now exceed $1 trillion a year — more than the entire defense budget.

And here too, old debt isn’t repaid so much as rolled over. Roughly a third of all U.S. Treasury debt needs to be refinanced within any given year. If confidence in that process wavers, buyers at Treasury auctions get pickier, and the government has to offer higher rates to attract them. That raises borrowing costs for everyone, from mortgages to corporate debt to, yes, data-center financing.

These three layers don’t sit side by side — they sit on top of each other. Confidence in American AI props up the stock market; the stock market holds the retirement savings of half the world; and those savings, through bond funds and reserves, help absorb the next Treasury auction. Sitting on the top floor, mostly without knowing it, are ordinary pension holders, insurance company portfolios, and central bank reserves.

China Enters: When Growth Slows

That’s where China comes in. This country has run the same playbook for thirty years: first toys and furniture on cheap labor, then steel, then solar panels (China now accounts for something like 80% of global solar manufacturing capacity), then electric vehicles (BYD overtook Tesla in global battery-electric vehicle sales for calendar year 2025). Each time, American industry waved it off as low-end competition — until it wasn’t.

Now China is doing it with AI. On July 16, 2026, the Chinese startup Moonshot AI announced Kimi K3, a 2.8-trillion-parameter model that topped at least one closely watched coding benchmark (Arena.ai’s Frontend Code arena), edging out even top-tier U.S. models there, while still trailing the very best American systems on broader capability tests. Roughly eleven days later, Moonshot made the model’s weights available for anyone to download and run.

A White House official then accused Moonshot of training K3 in part on banned Nvidia chips routed through Thailand, and separately alleged the company had distilled a rival U.S. lab’s model (reportedly Anthropic’s) to shortcut development. Moonshot and Chinese officials denied both claims, and independent researchers who examined K3 have publicly expressed skepticism about the distillation accusation, noting no conclusive public evidence has surfaced either way. Nvidia’s own CEO, Jensen Huang, went the other direction entirely, calling open-source Chinese models like Kimi “excellent” and arguing they should be embraced rather than banned.

What’s not in dispute is the price and the market reaction. Open Chinese models reportedly cost 60–90% less than the top American ones — pennies versus tens of dollars per million tokens for the premium U.S. alternative. Businesses are voting with their usage: on the OpenRouter platform, where companies pick AI models the way travelers compare flights, U.S. models’ share of traffic reportedly fell from about 70% a year ago to about 30% by mid-2026, while Chinese models’ overall share climbed as high as 61% in some monthly readings. Chinese models have also been gaining fast specifically in coding workloads, one of the more lucrative categories for providers.

Chinese models don’t need to be outright better to matter here. They just need to be close enough, at a small fraction of the cost. Some customers switch entirely; others use the competition to squeeze American providers on price. Either way, it eats into the future revenue that today’s trillion-dollar valuations are counting on. Jefferies’ widely read “GREED & fear” research note put it bluntly in late July 2026: the most likely long-term outcome of the AI story, in market terms, is “massive capital destruction” in the U.S., with market share migrating to cheaper open-source Chinese models.

Washington is now seriously weighing a ban on Chinese AI models in the U.S. market. One industry-funded estimate put the potential cost to American businesses at around $12 billion a year if a ban goes through. Nvidia and roughly two dozen other companies have publicly signed a letter opposing such a ban — notably, OpenAI, Anthropic, and Google were absent from that list. If Washington doesn’t ban Chinese models, they’ll likely keep eating into the growth that underpins American AI valuations. If it does, U.S. businesses lose access to a much cheaper tool, and the rest of the world keeps using it anyway.

What the Numbers Are Actually Telling Us

I’ve come to think that in any system where debt gets rolled over rather than paid down, the breaking point isn’t when prices fall — it’s when growth slows. In 2006, home prices were still at record highs when the refinancing mechanism quietly died. Today, the valuations of OpenAI, Oracle, and their peers don’t rest on absolute revenue; they rest on the rate of growth continuing to accelerate. The next funding round has to be bigger than the last, the backlog has to keep growing faster than the quarter before, and the build-out has to keep outpacing actual cash flow in the door.

Once that tempo slips, the whole arrangement gets much harder to sustain. And a lot of those “contracted” trillions in cloud backlogs are promises, not cash in hand.

A few more data points worth sitting with:

Alphabet’s first-ever negative free-cash-flow quarter arrived in Q2 2026, driven by record AI infrastructure spending.
Oracle’s credit rating was cut to one notch above junk in mid-2026, with S&P naming OpenAI concentration risk directly as the reason.
Anthropic’s valuation briefly overtook OpenAI’s in the spring of 2026, before OpenAI’s own number climbed again — a sign of how fast, and how reflexively, these private valuations are moving.
The combined contracted backlog across the major cloud providers reportedly passed $2.3 trillion by mid-2026, with a large share of at least one provider’s backlog tied to a single customer.

Four Things I Keep in Mind

I’m not going to give financial advice, but here’s how I personally think about it.

1. Don’t try to time the exact date. There can be a year or two between when a mechanism breaks and when the crash actually shows up. The people who spotted the 2006 problem looked wrong for months, because prices were still climbing. Green headlines and confident crowds break more people than bad math ever does.

2. Watch the rate, not the record. Don’t stare at how large OpenAI’s valuation or Oracle’s backlog is — watch whether the growth rate is accelerating or decelerating quarter to quarter. That growth rate is effectively the collateral holding a lot of this up.

3. Watch for one specific signal. The day a major AI player announces it’s cutting data-center spending and its stock rises on that news — that’s the day markets stop rewarding growth-at-any-cost and start rewarding restraint. That would mark a real shift.

4. Know what you actually own. A small handful of companies make up a large share of the S&P 500 — the index that effectively holds retirement savings for a huge swath of the world through 401(k)s, pensions, and index funds. A “diversified” fund today is, to a real degree, a concentrated bet on the AI story continuing. That’s not a reason to panic. It is a reason to understand your own exposure.

Final Thought

I don’t know when — or even if — this breaks. Maybe a year, maybe two, maybe the companies involved find a way to grow into these numbers before the music stops. But the mechanism looks familiar: debt that gets replaced rather than repaid works only as long as the next check is bigger than the last one. In 2006, almost nobody noticed the exact day that stopped being true. I’m hoping we pay a little more attention this time.

These are my own reflections, based on public reporting — not a prediction and not a call to action. But too many people have no real sense of what their savings are actually resting on right now. Watch the speed, not the height. And keep asking: is the music still playing?

Sources referenced in reporting this piece include Bloomberg, CNBC, Reuters, the Congressional Joint Economic Committee, S&P Global, Jefferies research, and public statements on X. Figures for private company valuations, funding rounds, and backlogs are based on the most recent public reporting available and can change quickly.