$116.7 billion. That is the bill two trading sessions have handed to Advanced Micro Devices, after a research firm named Hunterbrook published two reports questioning how the chipmaker books the cost of its artificial intelligence buildout and how it describes some overseas business relationships it depends on for supply.
The shares fell 6.17% in premarket trading on Nov. 20 to $459, extending a slide that has now reached 13.42% over two sessions. Selling began after the first report circulated and deepened when a second followed a day later, according to traders watching the order flow.
The first report is about counterparties. It asks whether AMD has business links to companies that U.S. export controls are meant to keep away from advanced chips, and argues that disclosure around those relationships is too thin for shareholders to judge the exposure.
The second is about promises. It counts roughly $28 billion of commitments the market has not priced: about $13.3 billion of data center lease payments, $8.5 billion of guarantee ceilings and $6.2 billion of cloud computing purchase commitments. None of those figures appears as debt on the balance sheet, which is part of the complaint.
AMD rejected the framing. The company said the allegations misstate ordinary commercial relationships, that it has consistently complied with export controls, and that obligations of this kind are routine for a business buying capacity at its present size rather than owning it.
An affiliated fund disclosed a short position, according to the reports. A short seller earns when the price falls, and these reports landed in a market already nervous about what AI infrastructure costs to build and who ends up carrying the bill when demand cools.
The accounting argument is narrower than the headlines. Lease payments, guarantees and cloud purchase agreements are not debt. They are commitments, and they sit in footnotes rather than on the balance sheet, which is why two analysts can read one filing and draw opposite conclusions about how much risk the company carries. That gap is what makes the reports effective.
For a designer that reserves wafer supply and computing time years ahead, such promises are the entry fee for growth. The bear case is that they convert into cash costs quickly if demand for accelerators slows, and that the conversion arrives before the revenue that justified them does.
Leases are the clearest example. A company that cannot wait for its own buildings rents capacity from data center operators under contracts that commonly run a decade or more, and the obligation is fixed while the revenue behind it is not. Guarantees work differently, usually backing borrowings taken on by a partner or a supplier, and they cost nothing unless the borrower fails. Investors treat them as free until the moment they are not.
Cloud purchase commitments are newer and less standard. Hyperscalers and model developers have signed multi-year compute deals at a pace the disclosure rules were not written for. The question has been put to Microsoft, Meta and Oracle this year, and none of those filings settled it.
Export controls add a second layer. Washington has tightened restrictions on advanced chip sales to China repeatedly since 2022, and U.S. suppliers must screen customers and end users before shipping. The chain runs through distributors and equipment makers, so the screening is only as good as the weakest link in it. Any suggestion of loose screening reads to investors as a governance failure rather than a compliance slip.
AMD is the only credible second supplier of the accelerators that train and run large models, and that has made its shares a proxy for the whole trade. Investors who want AI exposure without buying the market leader buy this name, and they pay a premium for the privilege.
The premium is what the reports attack. If the commitments are ordinary, the multiple stands. If they are the kind of obligation that turns into a cash call in a downturn, then the premium was borrowed from the future and the lenders have started asking questions.
Analysts said the size of the move says as much about positioning as about the reports. AMD had become a crowded long, and crowded longs fall faster when someone points at the footnotes. Two sessions do not settle an accounting question, but they show how few holders were ready to defend the stock on the way down.
What management controls is narrower than what it can explain. It can publish a schedule of the commitments, define what sits behind each guarantee and show how the leases are structured. That would not end the argument, but it would move it onto numbers, which are harder to dispute than motives.
The next scheduled test is the quarterly update, when management will have to set figures against the commitments its critics counted. Until then the burden rests on disclosure, and disclosure is the only item on that list AMD can fix quickly.


