A trillion dollars now moves in a loop among the handful of companies building artificial intelligence. Here is how the circle turns, why it is powering and quietly straining the market that leans on it, and what an asset with no counterparty looks like beside it.

Follow a single dollar. It begins at Nvidia, the company whose chips sit at the center of the artificial-intelligence boom. Nvidia invests that dollar in an AI lab, like OpenAI, Anthropic, or xAI. The lab spends it renting computing power from a cloud provider like Oracle or CoreWeave. The cloud provider spends it buying chips from Nvidia. The dollar has traveled in a full circle and landed back where it started, and on the way it has been booked as Nvidia's investment, the lab's funding, the cloud's revenue, and Nvidia's sales. One dollar, four headlines.

Now multiply by a trillion. Over the past year the companies at the heart of AI have bound themselves together in a web of overlapping commitments that, by most counts, runs north of a trillion dollars. It is the largest and fastest infrastructure build in the history of business, and it has minted staggering paper wealth. It is also starting to resemble something investors have seen before: a market where the same money moves in a circle, and where the line between real demand and its own reflection has begun to blur.

This week's brief is about that circle. How it turns, why it is at once both the engine and the source of the fragility in today's market, and what it means for anyone whose savings, through an index fund or a pension, are quietly tied to it. And at the end, a contrast worth sitting with: in a market built from counterparties, leverage, and accounting judgment, what does an asset with none of those look like when juxtaposed?

The biggest build in history

Start with the scale, because it is hard to overstate. After first-quarter results this spring, the four largest cloud companies, Microsoft, Alphabet, Amazon, and Meta, are on course to spend roughly $700 billion on capital projects in 2026. That is up about 77% in a single year and close to triple what they spent two years ago. Three of the four raised their forecasts in the same week. Most of the money flows into data centers and the chips inside them, and a large share of those chips will be outdated within a few years.

For most of the past decade these were the most profitable, cash-generating businesses ever built. They earned far more than they spent and returned the difference to shareholders. That has flipped. The spending is now so large that it is reshaping their cash flows, their balance sheets, and the way the whole market trades. When four companies make up a growing slice of the S&P 500 and lash their fortunes to one technology, the index is no longer as diversified as it looks.

The bull case is simple and not unreasonable: this is the build-out of a general-purpose technology, the best chips are truly scarce, and whoever owns the most computing power will own the future. Perhaps. But the way the build is being financed is where the story turns from big to interesting.

How the circle turns

At the center sits OpenAI, a company that is not yet public and was, until recently, a nonprofit, yet now sets the tone for the entire market. In 2025 it signed compute and cloud deals that together approach or exceed a trillion dollars: roughly $300 billion with Oracle over five years, about $100 billion tied to Nvidia, near $90 billion with AMD, tens of billions more with Amazon's cloud, a large multi-year commitment to Microsoft, and a $500 billion data-center venture, called Stargate, alongside Oracle and SoftBank. Set those promises against OpenAI's revenue, on the order of $13 billion last year, and the gap is the whole story. OpenAI is not paying for this out of profit. It is paying with capital, much of it supplied by the very companies it has promised to buy from.

That is the loop. Nvidia invests in OpenAI; OpenAI commits to buy Nvidia chips and to rent cloud capacity from Oracle and its peers; those cloud providers turn around and buy Nvidia chips to meet the contracts. The chipmaker funds its customer, the customer funds the cloud, the cloud funds the chipmaker. The same names keep appearing on both sides of the table.

When the same dollar can be counted as a chipmaker's revenue, a startup's funding, and a cloud provider's backlog all at once, the headline numbers stop meaning quite what they seem to.

Why a circle is dangerous

There is nothing inherently wrong with a customer and a supplier helping to finance each other; it happens across industries. The danger is what a circle does to information and to risk. Take information first. In an ordinary market, a sale tells you that an independent buyer wanted something enough to pay for it. In a circle, a sale can tell you only that money completed a lap. Demand and its own funding grow hard to tell apart, and the headline figures, revenue, backlog, bookings, can swell without a single new customer arriving. The appearance of demand is manufactured by the structure itself.

Then risk. When the same few firms are one another's customers, financiers, and suppliers, their fates are welded together. A stumble at one becomes a stumble at all. If OpenAI cannot finance its commitments, Oracle's backlog is suddenly in doubt; if Oracle's revenue softens, its orders to Nvidia follow; if Nvidia's sales wobble, so does the single stock propping up a large part of the market. The circle that moves money so efficiently in good times moves trouble just as efficiently in bad ones.

There is a subtler twist still. The circle is not only money; it is also valuation. Each new deal lifts the share prices of the companies announcing it, and those richer valuations become the currency and the collateral for the next deal. Rising stock begets bigger commitments, which beget rising stock. On the way up it is a marvel of momentum. The trouble with a machine that runs on its own reflection is that the process works just as forcefully in reverse.

We have watched a version of this before. In the late 1990s the companies that made telecom equipment, names like Lucent, Nortel, and Cisco, lent their customers the money to buy that equipment. For a while it looked like roaring demand. When the customers could not pay, the loans soured and the losses cascaded back through the suppliers. The broader crash erased trillions in value and took even the strongest survivors years to climb back. The internet was real and transformative; the financing structure around it was not sound. Both were true at once.

One wrinkle makes this cycle different, and it is not reassuring. The companies at the center are not speculative startups, as many dot-com darlings were. They are the largest, most widely owned businesses on earth, and their shares sit in nearly every index fund and retirement account in the country. That makes the build steadier in one sense, these are real companies with real revenue, and more dangerous in another. If the circle breaks, it does not just take down a handful of speculative firms, it reaches the index funds and pensions that hold them.

The accounting underneath

The profits that justify all this spending rest, in part, on a choice. When a company buys a chip it does not record the full cost at once; it spreads that cost over the years it expects the chip to be useful. Stretch the useful life and this year's expense shrinks and this year's profit grows, with nothing real having changed.

Here is the tension. Nvidia ships a meaningfully faster chip every twelve to eighteen months, yet the big cloud companies depreciate their chips over five or six years. The investor Michael Burry, who made his name on the 2008 crisis and closed his fund last autumn after betting against Nvidia, argues that the mismatch lets the industry understate its true costs by roughly $176 billion between 2026 and 2028, flattering reported profits by something like a fifth. His critics answer fairly: a chip retired from cutting-edge work can still earn its keep for years on lighter tasks, so a longer life is not necessarily dishonest, and revising an estimate is permitted accounting, not fraud. Yet the companies themselves are not of one mind. Under the same facts, Amazon recently shortened the assumed life of some servers from six years to five and took a charge for it, while Meta lengthened its own. When the people closest to the hardware cannot agree how long it lasts, the profits built on that assumption deserve a careful eye.

From cash machines to borrowers

The most telling shift is in the cash. For years these giants drowned in it. Now, for the first time, their combined capital spending is outrunning the cash their businesses produce. Across the four, free cash flow is estimated to have shrunk by more than 40% over the past year and change. Amazon's has collapsed toward zero, and analysts model it turning negative this year; Barclays expects Meta's to fall by as much as 90% and to stay negative into 2027 and 2028, and called that the likely fate of every company in the race.

When internal cash runs short, you borrow. Companies that almost never needed debt issued bonds last autumn at roughly twice their decade-long pace, and the borrowing is accelerating into 2026. The balance sheets are still among the strongest in the world, and none of this is distress. It is a deliberate, all-in wager that the demand will arrive before the bills do. But it changes the character of the trade. A self-funded boom can pause and wait. A debt-funded one has payments to make whether or not the revenue shows up, and leverage is the thing that turns a disappointing cycle into a painful one.

The honest other side

None of this means the AI build is a fraud, or that it must end in tears. The strongest version of the bull case deserves a fair hearing. Demand for computing power is real and, in places, desperate: cloud order books have swelled, one provider's above $460 billion and roughly doubling in a quarter, and AI revenue at the leaders is climbing fast, if from a small base. In a market where the best chips are scarce, locking in supply with long-term commitments and a little financing is a rational way to compete, not a swindle. As one cloud chief put it, the giants are buying because they have the demand, and there is nothing circular about that. A respected asset manager has described the web of deals as a virtuous circle that lines up suppliers, builders, and customers to meet a real and exploding need.

All of that may be right. The purpose here is not to call the top; it is to be clear-eyed about structure. The risk in the AI trade is not that the technology is fake. It is that the financing is interlinked, the profits lean on a flattering assumption, the cash has turned to debt, and the whole edifice is priced for everything to go right. Fragile and fake are not the same word, and it is the first that should occupy an allocator's mind.

An asset with no counterparty

Which brings us to an odd and useful contrast. Step back and ask what the circle is made of. It is made of counterparties: promises between companies, each depending on the next to keep its word. It is made of leverage: borrowed money standing in for cash that has not yet arrived. And it is made of judgment: profits that hinge on how long someone assumes a chip will last. Those are the three materials of modern finance, and they are precisely the ones that tend to fail together.

Now consider an asset made of none of them. Bitcoin has no issuer and no counterparty; it is not a promise someone else must honor. It has no vendor-financing loop behind it, no backlog that might be the same dollar counted twice. There is no debt on its balance sheet, because there is no balance sheet. Its supply cannot be expanded to meet a surge in demand, by a committee, a chief executive, or anyone else. And it settles without leaning on the solvency of a bank, a cloud provider, or a chipmaker. Its price moves, sometimes violently. But its structure is the mirror image of the circle's.

That is not a recommendation; it is an observation about what a portfolio is actually exposed to. The most risk-aware institutions on earth have begun to act on it. The US government holds about 328,000 bitcoin, a bipartisan bill would write a national reserve into law, and the state of Texas this week funded its own reserve, bought, and said it means to buy more. They are not chasing a number on a screen. They are adding something whose worth does not depend on anyone else staying solvent.

The question for allocators

So the question is not the one the headlines keep asking. It is not really whether AI is a bubble. The more useful question is quieter: how much of my portfolio's safety rests on the same few balance sheets, the same interlocking promises, and the same assumption about how long a chip lasts? For most portfolios the honest answer, once index funds and target-date defaults are counted, is more than the owner imagines.

Diversification once meant owning many different names. When the names have been welded into a single circle, and when that circle is the market's largest position, real diversification means owning something that sits outside it entirely, something that does not need the circle to keep turning. The circle will turn for as long as the music plays, and it may play for a long while yet. The quiet value of an asset with no counterparty is that it does not need the music at all.

The Radar

What matters this week across digital assets, AI, and global markets.

Digital Assets & Regulation

Strategy breaks its rule, then blinks

Strategy did something last week it had not done since 2022: it sold bitcoin. The amount was tiny, 32 coins for about $2.5 million, and the reason was telling. The sale helped cover dividends owed on its preferred stock. It was the firm's first sale in four years and only its second ever. Confidence cracked anyway. Bitcoin fell as much as 18% toward $59,000 over the following days. Then, on Sunday, Michael Saylor posted that it was 'a good time to add more dots,' signaling a return to buying, and the price steadied above $63,000. A small sale, a large reminder: the obligations stacked on a bitcoin balance sheet do not vanish when the price falls.

The worst week since February

The drop was not one company's doing alone. Bitcoin fell about 13% on the week, its worst stretch since February, while spot bitcoin ETFs bled more than $3.2 billion and a long-dormant Mt. Gox wallet stirred. The asset now trades roughly 50% below its October record. What makes the week unusual is the backdrop: global stock indexes were printing fresh highs at the same moment. Liquidity did not so much leave the market as rotate, toward the chip rally and the coming SpaceX listing. For a few sessions, bitcoin traded less like a hedge and more like the high-beta end of a risk book.

Bankman-Fried wants a pardon

Sam Bankman-Fried, serving 25 years for the fraud that vaporized FTX, is seeking a pardon from President Trump. Whatever its odds, the request marks how far the politics of digital assets have traveled in three years, from courtroom villain to clemency petition. The industry that spent 2023 distancing itself from FTX now works in a Washington far friendlier to it. The ghosts of the last cycle have not been laid to rest. They are filing paperwork.

AI & Financial Infrastructure

Nvidia comes for the PC

At Computex in Taipei, Jensen Huang unveiled the RTX Spark Superchip and said Nvidia, alongside Microsoft, would 'reinvent the PC.' The market read it as a threat: shares of AMD, Intel and Qualcomm fell on the news. For years Nvidia's story was the data center. Now it is reaching for the edge, the phones and laptops that run AI models without touching the cloud, and for a slice of every layer of the stack. A company that already captures roughly 90% of AI accelerator spending is not content to own the center of the boom. It wants the perimeter too.

Computex's quiet confession

Beneath the keynotes, the hardware vendors at Computex sounded less like chip sellers and more like electrical contractors. The bottleneck in the AI build is no longer silicon; it is power, cooling and the physical room to house it. Semiconductor lead times have stretched toward 40 weeks. The roughly $725 billion the largest cloud companies plan to spend this year cannot be poured into the ground overnight, because the racks, the substations and the water do not yet exist. The boom is colliding with the limits of the real world, and the real world is winning, for now.

SpaceX's inflated-equity playbook

Buried in SpaceX's amended IPO filing is one new sentence: the company 'may issue a significant amount of equity in connection with future transactions.' Wall Street read it as a tell. SpaceX has already folded in xAI and taken a $60 billion all-stock option on the coding startup Cursor. The bigger prize, observers suspect, is Tesla, a deal that would be the largest merger in history. By one estimate it would cost SpaceX holders 45% of their company at a price-to-earnings ratio above 400, to absorb a carmaker that earned $3.9 billion last year. When equity is the currency, the richer the stock, the bigger the deal it can buy.

Geopolitics & Markets

Hezbollah rejects the ceasefire

Israel and Lebanon agreed last week to a US-brokered ceasefire, negotiated in Washington without the one party that mattered. Hezbollah rejected it and kept firing; a UN peacekeeper was killed by mortar fire in the south. The fighting is the long tail of the February strike that killed Iran's supreme leader, and it shows no sign of ending. Brent crude pushed toward $96 a barrel. President Trump, asked about the broken truce, offered his own definition: a ceasefire, he said, is 'when you're shooting in a more moderate manner.'

Congress reaches for the war

As the Iran conflict grinds into its fourth month, the US House moved to curb the president's war powers, an effort by the legislature to claw back authority over a fight it never voted on. The measure is unlikely to bind a determined White House. But it is the clearest sign yet that the political cost of an open-ended war is starting to register in Washington, even among those who cheered its opening.

Bessent bets on growth

Treasury Secretary Scott Bessent staked the solvency of Social Security on the strength of the Trump economy this week: no benefit cuts, with 10,000 boomers retiring every day and the national debt past $39 trillion. It is a wager on growth outrunning the arithmetic. For anyone who thinks in terms of sound money, the framing is the story. When the only acceptable answer to a debt problem is faster nominal growth, the path of least resistance runs through the currency.

A Final Word

Strip away the week's noise and one episode lingers. Strategy sold 32 bitcoin to make a preferred-dividend payment, and a $1.3 trillion asset wobbled on the news. The coins themselves settled exactly as designed; nothing in bitcoin's ledger broke. What shook was the structure built on top of it, the layer of preferred stock and dividend obligations that has to be served whether the price is rising or falling. It is worth holding the two apart. The asset is one thing. The leverage wrapped around it is another. Markets that forget the difference are reminded, sooner or later, and usually at the worst possible moment.

The Week in Numbers

What to Watch This Week