A value manager says the market broke its own math. His build log is the better evidence

Emeth Value Capital's H1 2026 letter opens by arguing that parts of the market have stopped obeying their own valuation logic: memory chip companies earning more profit in a single quarter than their entire market capitalisation eighteen months earlier, formerly ex-growth utilities repriced as high-octane growth stocks, and secular compounders whose advantages looked insurmountable months ago now being, in the manager's phrase, sentenced pre-trial, with OpenAI and Anthropic cast as judge, jury and executioner ^[4][5][6]. The letter's second half matters more than its first, because it documents, in operating detail, how that sentence actually gets carried out.

Take the memory claim at face value and the arithmetic is blunt. A quarter of profit exceeding the old market capitalisation implies an annualised run rate of roughly four times what the equity was worth, a trailing multiple under about 0.25 on that earlier price ^[17]. Since results are struck to 30 June 2026, the reference point is roughly the end of 2024 ^[16]. That is not a rerating. That is the market having been wrong about the level of earnings, not the multiple on them.

The compounder derating is the part with a mechanism attached. The manager says he has largely ignored Mr Market's daily re-underwriting of entire industries ^[8], while conceding that AI capability is advancing at a pace that is difficult to comprehend ^[7], and claiming his own process improved more in three months than in the prior decade ^[9]. Then he shows the receipt. Insider filing data, he writes, is available in some form across Bloomberg, FactSet, CapIQ and AlphaSense, but is often incomplete, siloed inside the terminal and sometimes limited to US markets, and alerting on insider purchases is effectively nonexistent on most platforms or drowned in hundreds of daily notifications that are almost entirely noise ^[10][11].

So he built it himself. The workflow points directly at regulatory sources of truth, SEC Edgar Form 4 filings in the US, the FCA NSM in the UK, BaFin in Germany, ingesting every new filing nightly ^[13]. Volume averages a few thousand filings a day across inconsistent field structures, DOC, PDF, HTML, XML and CSV formats, and multiple languages ^[14]. He says a year ago this needed an internal developer and a six-figure budget, and that he instead built it on a Saturday night with ten cups of coffee and a $200 per month Anthropic subscription ^[12]. That is about $2,400 a year, or roughly 2.4 percent of a $100,000 budget ^[18]. He describes the reading accuracy as near perfect at a fraction of a cent per filing ^[15], though the excerpt reports no measured error rate ^[20].

The consequence is the honest answer to the model-provider risk question. Anthropic is not selling cross-jurisdictional insider-purchase alerts. It sold a subscription that let one analyst stop needing them. That is what a moat looks like when it fails: not displacement by a competing product, but a customer who no longer buys the feature.

Two cautions. This is one manager's self-reported account, and the published excerpt promises updated returns since a 31 December 2015 inception without printing the figures ^[19][3], while insisting short-term results carry little significance anyway ^[2].

Worth tracking: whether the incumbent data platforms respond on alerting and coverage rather than price, and whether the next letter attaches numbers to the process claim.