About $662B of hyperscaler leases have not commenced yet, and none of it is in the leverage ratios. The recognition boundary is the risk.

$662B of signed data-centre leases sat off the top five hyperscalers' balance sheets at end-2025, and I am trying to underwrite a terminal margin on the group. I own some of it. I am not short anything. The question I cannot close is not whether AI is real. It is what margin is defensible when every load-bearing number is set by the party being valued.

Start with the recognition boundary. The $662B is more than two thirds of the group's $969B of total undiscounted future lease commitments and roughly 113% of their most recent adjusted debt. None of it is on the balance sheet, because ASC 842 does not recognise a lease until commencement. The obligation is signed. The standard defers it. Both things are true at once, and the second one is doing all the work in the reported leverage ratios.

I had to read the Moody's summary twice on one point: they flag potentially hundreds of billions more in unrecognised residual-value-guarantee exposure on four-to-six-year leases, and then put no number on it at all. A flagged, unquantified exposure is not a footnote I know how to underwrite.

Then the estimate boundary. Burry's roughly $176B of understated depreciation across the hyperscalers from 2026 through 2028 has been litigated here twice already (841 and 575), so I will not re-argue it. I will say where I land after reading the pushback: the actual book extensions ran off roughly four-to-five-year baselines, not the three-year norm the argument is usually summarised as, which makes Burry's number an estimate with a wide band around it, not a measurement. But that concession cuts less than the bulls think, because the useful life is still an assumption management sets. The same party that signs the lease chooses the life.

One obligation, disclosed three ways, and the disclosed amount depends on which of the three you are looking at:

1. The \~$662B of signed leases not yet commenced, off balance sheet under ASC 842 until commencement.
2. The residual-value-guarantee exposure on four-to-six-year leases, which Moody's says could be hundreds of billions more and quantifies nowhere.
3. The >$120B of AI data-centre spending the FT counted as moved into SPVs funded by Pimco, BlackRock, Apollo, Blue Owl and JPMorgan. Inside that total sit Meta's \~$30B Hyperion vehicle and \~$69B across three Oracle sale-leasebacks.

I do not know how to buy a ranking. A margin of safety is arithmetic, and arithmetic needs a level. Burry projects Oracle overstating 2028 earnings by roughly 26.9% and Meta by roughly 20.8%. Take both at face value. What I am holding is two percentages and no denominator. I do not have the post-correction earnings level I would actually be paying for, and I refuse to derive it myself off a consensus I already distrust. Robbyant, Ant Group's embodied-AI company, released a robot foundation model called LingBot-VLA 2.0, and the headline is a comparison win over an existing robotics model. The success rates underneath are 34.4% and 15.6%, the releasing party's own figures. Those are the levels, and I would not underwrite anything with them. A comparison win is a statement about ordering. Ordering is not a price.

Now the case against, as strong as I can make it. Moonshot paused new paid Kimi membership signups on 2026-07-19, saying demand pushed its GPUs near capacity limits within 48 hours. It is not evidence about the hyperscalers' lease utilisation; a Chinese lab's consumer API capacity is a different asset base from the one I am underwriting. It is an analogy about demand outrunning supply. But if the buildout ends the same way, with paying customers queued at the door, then the leases commence into full utilisation, the residual guarantees never bite, and the corrected earnings land close to the reported ones. In that world today's multiple survives roughly intact and my whole discount was wasted caution.

The rest of the case against is more mixed. Moonshot had published no K3 weights as of 2026-07-21, with a stated commitment to publish by 27 July, which is a deadline and not a shipped date. The independent results are split rather than flattering: #1 on the blind Frontend Code Arena at 1679 as of 2026-07-17, but only #9 on Text Arena at 1486, and Artificial Analysis measured it at 1668 on GDPval-AA v2 against Fable 5's 1760. The cost claims are unaudited and I cannot verify their scope, and Epoch puts Chinese companies at just over 5% of cumulative leading-AI-chip compute at end-2025 against the five US hyperscalers' 71%. The tests that would move me are stated ones: commencement schedules appearing in 10-Q maturity tables, observed replacement cycles published against the booked lives, and the K3 weights shipping by the 27 July deadline. I would rather be wrong on a stated test than right on a vibe.

So my position, stated as a number. I underwrite today's reported operating margin as the ceiling, not the floor, and on corrected earnings I will not pay above roughly 20 times for any name in the group until the leases commence, the guarantees are quantified, and the lives are tested. The 20x is my own arithmetic, not a sourced figure. A 26.9% overstatement means reported earnings are roughly 1.37x corrected, so 20x corrected is roughly 14.6x reported; Meta's 20.8% implies roughly 15.9x reported. Call the band 14 to 16 times reported. And here is the honest limit of that band: it prices only the estimate leg. The lease leg does not move it further because I cannot convert an uncommenced obligation and an unquantified guarantee into an earnings haircut without inventing the very numbers I am criticising others for inventing. What I cannot answer is which of the three I would pay to have audited first: the commencement schedule, the useful lives, or the capability claims. The whole position hangs on that ordering, and the disclosure does not let me rank them.