The $662 Billion Question Hanging Over the AI Capex Boom

Hyperscaler capital expenditure has passed $700 billion for 2026, and that is only the part investors can see. Microsoft, Amazon, Alphabet and Meta, plus Oracle, have guided to a combined $635 billion to $690 billion this year, up 67% to 74% from 2025. Sitting alongside those figures is a number that appears on no debt schedule and in no leverage ratio. Roughly $662 billion of signed data-center leases have been committed to but have not yet commenced, which under standard accounting rules means they are not yet liabilities.

That is not a scandal, it is a rule working exactly as written. It also means the visible balance sheet understates what has already been promised, and the gap is larger than the annual capex budget of every hyperscaler combined.

The 2026 AI capex picture

• Total hyperscaler capital expenditure for 2026 has topped $700 billion

• Combined guidance from the big four plus Oracle sits at $635 billion to $690 billion, up 67% to 74% year over year

• Roughly $662 billion in signed-but-not-commenced data-center leases sits off balance sheet

• UBS projects capex growth of 76% this year, then 25% in 2027 and 6% in 2028

• Amazon's trailing-twelve-month free cash flow has fallen 95% to $1.2 billion

A week ago this was an accounting footnote. Here is why it became the number that decides the next leg for every risk asset you hold.

Where the $700 Billion in AI Capex Is Actually Going

The spending is concentrated in five companies, and the individual figures are large enough that they distort national economic statistics.

Alphabet reports on Wednesday July 22, 2026, and the capex line in that release matters more than earnings per share. A raise tells you the buildout is accelerating into the deceleration forecast. A trim tells you the largest spenders have started reading the same math the bears are reading. Either way the number moves the whole complex, including NVIDIA and the memory suppliers sitting behind it. Oracle is the swing factor in the group, since a larger share of its committed capacity is contracted to AI tenants rather than funded from its own diversified cash flows.

Price levels as of Monday's session, before the US close, put AMZN at $247.63, MSFT at $394.30, GOOGL at $348.78, META at $643.75 and NVDA at $203.48.

What $662 Billion Off Balance Sheet Actually Means

Under GAAP lease-commencement accounting, a lease does not become a recorded liability when it is signed. It becomes one when the lease commences, meaning when the lessee actually gets control of the asset. For a data center that is being built to order, signature and commencement can be separated by years.

Think of it as signing a mortgage on a house that has not been built yet. You are contractually bound, the builder is drawing on that commitment to break ground, and the payments are coming. Until you get the keys, though, the loan does not appear in your net-worth calculation.

That is the entire mechanic, and it explains the whole gap. When an analyst pulls up net debt for Microsoft or Meta and concludes the balance sheets are fortress-grade, the conclusion is accurate on the numbers presented and incomplete on the obligations actually incurred. The $662 billion is disclosed, it is auditable, and it is sitting in the commitments notes of filings that most people never open.

Two things follow from that. The first is that leverage is being understated across the sector right now, and it corrects mechanically as those leases commence over the coming years. The second is that this is normal. Airlines, shipping companies and retailers have run the same treatment for decades. The difference is scale and speed, because no industry has ever committed this much lease capacity this fast.

The Deceleration Cliff That Arrives Before the Leases Do

UBS has modeled the forward curve, and the shape is the actual bear case. Capex rises 76% this year to roughly $673 billion, grows 25% in 2027, then grows just 6% in 2028.

A 6% growth year is not a collapse, it is a perfectly healthy mature business. The problem is that almost nothing in the AI supply chain is priced for a 6% growth year, because revenue expectations at the chip designers, the equipment makers, the memory vendors and the power developers were all built on the current rate continuing.

And the timing is the part that gets underappreciated. The leases signed in 2026 commence through 2028 and 2029, which means the fixed obligations peak exactly as the growth rate flattens. Rising committed costs meeting decelerating incremental revenue is the classic setup for a margin squeeze, and margin squeezes at the top of the market are how capex cycles have historically ended.

Amazon offers the early evidence. Trailing-twelve-month free cash flow reported in the company's quarterly results has fallen 95% to $1.2 billion, because the buildout is consuming the cash the retail and cloud businesses generate. The company can afford it. The direction is still worth watching, since a business converting almost none of its operating cash into free cash flow has less room to absorb a demand disappointment than the headline balance sheet implies.

The Bull Case Deserves More Weight Than the Bears Give It

Here is where most AI-bubble takes fall apart. They assume the hyperscalers are building into demand that does not exist yet, and that is the opposite of what the operators are reporting.

All five major players describe themselves as supply-constrained, not demand-constrained. They cannot get enough compute to serve customers who are already asking for it and already paying. Amazon has said as much repeatedly on earnings calls, and the pattern shows up in Alphabet's own investor materials and across the group.

That distinction changes the risk profile entirely. A company signing enormous lease commitments into empty demand is reckless. A company signing them because it is turning away paying customers is behaving rationally, and the correct criticism of rational behavior is that it can still be wrong about how long the demand lasts, not that it is irrational today.

The accounting deserves the same fairness. Off-balance-sheet lease commitments are standard, disclosed and legal across every capital-intensive industry. Nobody is hiding anything. The honest framing is that the disclosure sits where few people look, and the market has been valuing these companies off a leverage picture that is temporarily flattering.

The trigger for all of this landing at once was Friday July 17, 2026, when Moonshot AI released Kimi K3 at 2.8 trillion parameters and the PHLX Semiconductor Index entered a bear market. If frontier-grade models can be free, the market has to ask what the compute is worth, and that question is covered in depth in our companion piece on the semiconductor selloff.

Why a Crypto Trader Should Care About Hyperscaler Capex

The AI buildout has quietly become one of the primary drivers of overall market risk appetite, and crypto sits at the far end of that risk curve.

When capital decides the AI trade is over-levered and decelerating, the de-risking does not stop at semiconductor stocks. It moves outward through every long-duration, high-beta asset, and crypto is the highest-beta asset class most traders hold. Bitcoin around $64,785 and flat, with Ethereum near $1,878, is currently trading quietly while the equity complex works through this, and quiet has never been the same thing as insulated.

The mechanism running underneath all of it is liquidity. Hundreds of billions of dollars flowing into physical infrastructure is capital leaving the pool that funds speculative positioning everywhere else. It has been offset so far by the wealth effect from rising AI equity valuations, which is exactly why a semiconductor bear market matters to someone who has never traded a chip stock. Break the equity leg and the offset disappears.

So watch the capex guidance the way you watch funding rates. It tells you what the liquidity backdrop looks like for the next two quarters, and it moves before crypto does rather than after.

Frequently Asked Questions

Is the AI capex boom a bubble?

The spending is real and the demand is currently real, since all five major hyperscalers report being supply-constrained rather than short of customers. The bubble risk sits in the 2028 growth curve, not in 2026 demand, because the supply chain is priced for growth that UBS models at 6% by then.

Why are $662 billion in data-center leases not on the balance sheet?

GAAP records a lease liability at commencement, meaning when the tenant takes control of the asset, and not at signature. Data centers built to order can take years between the two, so the obligation is fully committed and fully disclosed in the filing notes while remaining absent from reported debt.

Which AI stock is most exposed to a capex slowdown?

The pure infrastructure suppliers carry the most direct exposure, since their revenue is a straight function of hyperscaler spending rather than of end-user software demand. The hyperscalers themselves have diversified cash flows to absorb a slowdown, though Amazon's 95% free-cash-flow decline shows the cushion is thinner than the headline balance sheet suggests.

When does the AI capex slowdown actually start?

UBS models the sharp deceleration in 2028, with 2027 still growing 25%. The market will price it long before it arrives, which is why the Alphabet guidance on July 22 and the subsequent hyperscaler prints carry more weight than usual.

The Bottom Line

The number to watch is Alphabet's capex guide on Wednesday July 22, 2026. A raise above the $185 billion top end says the operators still see demand outrunning supply, and the semiconductor drawdown was a sentiment event rather than a fundamental one. A trim, or any language about "optimizing" the buildout pace, says the largest spenders have started pricing the same deceleration UBS modeled, and that is the signal to reduce exposure across the entire high-beta complex.

Between now and then, treat the $662 billion as what it is. Not fraud, not a smoking gun, just a large committed obligation that becomes visible on a schedule nobody is watching closely enough. It converts into reported debt over 2028 and 2029, which is the same window where growth flattens to single digits.

For a crypto position, the practical version is simple. The AI trade is now the liquidity engine underneath your risk appetite, and it has started sputtering. Size accordingly.

This article is for informational purposes only and does not constitute financial or investment advice. Cryptocurrency trading involves substantial risk. Always conduct your own research before making trading decisions.