Infrastructure Thesis Computing
Dormant Bitcoin, AI Film Disclosure, and Why the Infrastructure Thesis Could Break
Twelve AI models applying a Bayesian probabilistic framework collectively estimated that the probability of Satoshi Nakamoto's approximately 1.1 million dormant Bitcoin ever being moved is under 5 percent. At current prices, that Bitcoin is worth somewhere in the range of 100 to 110 billion dollars — an essentially unclaimed fortune untouched since 2010. The models considered scenarios including key loss, Satoshi's death, voluntary inaction, and deliberate philosophical commitment to never spending the coins. The result reflects the collective weight of those scenarios rather than a definitive answer, but stands as a striking exercise in probabilistic reasoning applied to one of finance's enduring mysteries.
The biographical film 'Young Washington,' released on the Fourth of July, became the first major theatrical release to voluntarily disclose using AI in exactly 100 production shots. The creative team's choice to count and disclose rather than obscure is a bet that transparency is better commercial positioning than ambiguity — a bet whose success depends on whether general audiences care, with early evidence described as mixed.
The week's most consequential analytical exercise concerns the thesis that AI infrastructure stocks — the companies supplying the semiconductor and data center layer of AI — will outperform the AI application layer for the foreseeable future. UBS formally endorsed the thesis this week, and market consensus has been building behind it for 18 months. The historical analogy is the California Gold Rush: the sellers of picks and shovels made reliable money while individual prospectors faced variable outcomes. Several developments this week, however, stress-test that logic. AI model efficiency is reportedly improving — Anthropic's Claude and similar models are achieving equivalent performance with meaningfully fewer parameters than earlier generations. If that efficiency trajectory continues, the amount of hardware required per unit of AI output could decline dramatically, flattening the demand curve that infrastructure investors are pricing in.
The customer vertical integration threat compounds the risk. Google has its TPUs, Amazon has Trainium, Microsoft has co-developed custom silicon. If hyperscalers increasingly build their own chips, the picks-and-shovels providers face the prospect of their best customers becoming their competitors. A third risk — that current GPU-based transformer architectures are not the dominant computing paradigm for the next five years — is underweighted in most infrastructure investment theses. Computing paradigms do shift, and the emergence of neuromorphic or optical computing approaches could depreciate current infrastructure investments faster than expected. The concrete signal to watch: any major language model lab announcing a meaningful capability improvement achieved with significantly less compute than the prior generation. If two or three consecutive quarters show hyperscaler capital expenditure on AI infrastructure flattening rather than growing, the thesis warrants serious reexamination.