Formal Verification Moves Capital

The precedent: four centuries for certainty

In 1611 Johannes Kepler proposed a conjecture about the maximum density of sphere packing. The same ancient puzzle opens this analysis. The intuitive answer was clear from the start. The rigorous proof required nearly four centuries of work.

Thomas Hales announced a proof in 1998, with an architecture that combined mathematical arguments and extensive computation. Journal reviewers spent years checking it and declared partial confidence, since the computational volume exceeded direct human verification. A margin of doubt remained. The Flyspeck project closed that margin. In 2015 a group led by Hales published a complete formal proof of the Kepler conjecture, verified by the proof assistants HOL Light and Isabelle, as documented in the paper deposited on arXiv on January 9, 2015^[1].

This detail matters. A proof checked by two independent systems eliminates dependence on trust in a single human reviewer. Certainty becomes reproducible. The conjecture concerns the densest arrangement of identical spheres in space. The problem has deep roots in the history of mathematics and physics, as explained in the dedicated encyclopedia entry^[2]. The precedent has a precise structure: an ancient puzzle, computation exceeding human capacity, a machine certifying. The same structure is active elsewhere today.

The current pattern: the machine that verifies

A mathematical result from 2015 seems distant from capital flows. Appearances deceive. The same technology, machine-executed formal verification, stands today at the center of the artificial intelligence race. Anthropic has launched work on formalizing Fermat's Last Theorem, an attempt to make controllable by a machine one of the most celebrated proofs in modern mathematics, described in its public research^[3].

Three cases suffice to call it a pattern. Kepler in 2015, progress in proof assistants over the next decade, Fermat today. The direction remains coherent. Capital pursues visible capabilities. It overlooks the foundations. This gap between what fascinates and what sustains systems defines where future returns originate. The mechanism that follows explains why.

The mechanism: trust as infrastructure

The central point concerns trust. Formal proof transforms a claim into a logical chain that a machine checks step by step. This mechanism weighs on AI more than raw model power. A model generates plausible outputs; formal verification certifies those outputs are correct, line by line, against a set of axioms. The difference separates persuasion from proof.

Whoever controls the verification layer controls trust. And trust, in financial and industrial systems, remains the scarcest commodity. The logic is causal, not correlated. The more models enter domains where error carries high cost, the more mathematical guarantee becomes a condition of access. The layer that provides that guarantee captures the value the layers below produce.

My position: verification becomes critical infrastructure

Here is the thesis of this desk. Formal verification will transition from academic curiosity to an infrastructural layer of high-risk AI systems within three years. The reasoning rests on a structural fact. As models enter finance, defense, pharmaceuticals, and critical code, demand for mathematical guarantees grows faster than confidence in probabilistic outputs. Regulators will demand proofs, never promises.

Current consensus celebrates model scale and dataset size. This view measures brute force and ignores trust. The real question concerns who guarantees output correctness. What would change this reading? One clear piece of evidence: formal verification remains too costly outside pure mathematics, and formalization costs grow faster than computing capacity. If so, the thesis falls. The limit is real and must be kept in view.

From mathematical certainty to economic certainty

History offers a precedent on the lag between discovery and economic impact. Public-key cryptography emerged in the 1970s as mathematical curiosity. It became the infrastructure of digital commerce two decades later. The pattern repeats. A verification technique matures in academic settings, then migrates to systems where trust has a price. The lag compresses with each cycle.

Formal verification follows the same trajectory. The Flyspeck project proved that a complex proof can become entirely machine-controllable. The next step is economic, before it is technical. The market has already priced model capability. It has ignored the layer that certifies their outputs. This asymmetry creates the opportunity. Those allocating capital on long horizons should study these lags. They are predictable. They signal where value will shift before price reflects it.

The bottleneck: silicon and verification

Formal verification is compute-intensive. Each checked step consumes machine cycles, and demand for certainty translates to demand for silicon. This returns the question to the bottleneck this desk considers decisive. US-China competition on AI remains a problem of semiconductor access before it is a problem of model capability. Whoever controls the fabs controls the outcome.

US sanctions on advanced GPUs have redrawn the map of computing access. Formal verification, hungry for cycles, makes that map even more relevant for those allocating capital on long horizons. Models replicate. Fabrication plants resist replication.

Three implications for capital

Value migrates toward those who own the verification layer and the computing that powers it. Here are three operational consequences, each with its own horizon.

Twelve-month horizon. Suppliers of proof assistants and formal verification tooling will attract growing venture capital. Spending follows regulatory demand, which accelerates in critical sectors.

Twenty-four-month horizon. Family offices and sovereign wealth funds should map exposure to high-performance semiconductors. Formal verification amplifies that demand and redirects attention to the fabs.

Thirty-six-month horizon. A Chief Risk Officer should insert into models a scenario absent so far. Critical AI systems lacking formal certification become a compliance risk, and the cost of compliance arrives sooner than plans predict.

The prediction

Here is the verifiable claim. By December 31, 2027, at least one leading AI laboratory will bring into production a system combining a language model with a formal proof assistant to certify mathematical or code results.

Confidence: 68 out of 100. Horizon: December 31, 2027. Verification: an official announcement from a laboratory among the top five by funding or market capitalization. The signal that would refute the thesis remains simple. At the horizon's close, no leading laboratory has such a formal verification system in production.

What to watch

Three indicators will confirm or refute this reading in the coming months.

• Funding toward formal verification startups and proof assistance
• Joint publications between AI laboratories and formal mathematics groups
• References to formal certification in regulatory documents on high-risk AI

The divergence between model capability and verifiable guarantees always resolves. The question is which layer captures the value.