The risk factor they hope you skip is the one about IP dependence. Boilerplate 10-K language warns generically that a company may need third-party intellectual property — and in AI, that boilerplate has sharp teeth, because the foundational building block of every modern model is the attention mechanism, and the patents around it are concentrated among a small set of incumbents.
Consider Microsoft's granted patent US12260338B2, "Transformer-based neural network including a mask attention network" (issued 2025-03-25, assigned to Microsoft Technology Licensing, LLC). It is one specific, granted claim in the attention family. The point is not this patent alone but the pattern it represents: the companies that built the transformer era — Microsoft, Google, NVIDIA, and a few others — also hold much of the granted IP around its core operations.
Disclosed, but quietly: companies acknowledge IP risk in the abstract while almost never naming the attention-mechanism concentration specifically. That is the buried exposure. A smaller AI company building on transformer architectures is, in principle, operating in a space densely populated by incumbent patents — and the filings discuss this only as generic "we may require licenses" language, not as the concrete structural fact it is.
Compare the framing to where the value sits. The competitive moat in AI has partly shifted to data, distribution, and integration — but the foundational architecture IP did not evaporate; it sits granted on incumbents' books. A granted patent like this one is an asset that could, in some future, be asserted or licensed. Whether it will is unknowable from the filing; that it exists and is held by an incumbent is documented fact.
Distinguish a held patent from an asserted one, rigorously. There is no suggestion here that Microsoft is asserting this patent against anyone — the column's house rule forbids implying infringement not alleged in a filing. The risk being flagged is structural concentration of foundational IP, not any specific litigation. Held is not asserted, and this is squarely the former.
The early-warning takeaway: read the generic IP risk factor in any AI company's 10-K as quietly load-bearing. Foundational attention IP is concentrated among the incumbents who built the field, and grants like this mask-attention patent are the receipts. It is a disclosed risk dressed in boilerplate — exactly the kind this desk exists to un-bury.