Thesis

As labor costs rise and computing costs fall, platforms that do essential work earn in all market conditions.

We believe that in the next decade machines will do a growing share of necessary physical work, for which demand is evergreen. The skilled workers who perform this labour are retiring whilst the new workforce entering the market to replace these jobs is shrinking. Meanwhile, the difficulty and danger of complex tasks create high-margin work that few workers are available to do in the first place. Production platforms that displace this skilled but aging labor force lower their customers' total cost of the work, create organisation workflow changes, generate high margin business, and in the process raise switching costs for their customers. When those customers keep buying services irrespective of macroeconomic outlook, the underlying demand for the compute and energy powering these machines keeps growing. Aexodus invests into this vertically integrated segment of the physical AI stack, where next generation startups are expected to continue to grow whichever way the AI cycle turns in the short term.

For the company that builds the machine, a customer that stays is one of its lasting competitive advantages. Software companies are losing theirs, because AI has dramatically cut the cost of creating software. Meanwhile, the advantages in physical work are of a different kind and include defensible supply chains, integrated hardware and software systems, economies of scale, customer workflow integration with the switching costs that follow, and in some systems, powerful network or marketplace effects. A capability commoditizes because everyone eventually improves, and Aexodus views many hardware and software technology advantages as temporary. What no vendor can acquire, even with equivalent or marginally better technology, is years of field data recorded under real conditions at a customer's site. That field data starts a flywheel: each deployment improves every machine, and operating knowledge accumulates alongside the data. We expect that, as the flywheel turns, the tenth site will cost less, or perform better, than the first.

The forces that make the case now arrived in the same decade, and we read all three in public data. The shortage of skilled labor in US manufacturing is structural: the sector could need as many as 3.8 million additional workers from 2024 to 2033, and 1.9 million of those jobs could go unfilled if the skills and applicant gaps are not closed (Deloitte and The Manufacturing Institute, April 2024). Average hourly earnings in the sector rose 32 percent from January 2020 to October 2026, while the energy one chip, measured in every MLPerf Tiny round, needed to classify an image correctly fell 82 percent from June 2021 to July 2026 (US Bureau of Labor Statistics; MLCommons MLPerf Tiny). Policy now points the same way, because productive capacity is strategic again. The leaders of the next decade are being founded now, and we back such companies early, after the technology begins to work and before its commercial value is obvious.

Hourly earnings in US manufacturing and the energy one chip uses per correctly classified image, June 2021 = 100
Labor: average hourly earnings in US manufacturing, monthly from Jan 2020, up 32 percent to Oct 2026 AI: microjoules per correctly classified image on one chip, measured in every MLPerf Tiny round, Jun 2021 (reference code) to Jul 2026 (maker's code), down 82 percent
Source: average hourly earnings in US manufacturing, monthly, Jan 2020 to Oct 2026, nominal, US Bureau of Labor Statistics (CES3000000003). The latest months are preliminary. The AI line is the energy, in microjoules, that the STM32L4R5 chip (Arm Cortex-M4) used per image it classified correctly in each of the 7 MLPerf Tiny rounds. The test is image classification on small devices (ResNet-8 on CIFAR-10, closed division), a fixed task with a fixed accuracy target. Each value is measured energy per image divided by measured accuracy. It is 33,765 microjoules in June 2021 and 6,036 in July 2026. The June 2021 value is MLCommons' reference code, and later values use the chip maker's own software. From September 2025 the board and the part are their P versions (NUCLEO-L4R5ZI-P, STM32L4R5ZIT6P). Other chips reached lower values in single rounds, the lowest 31.9 microjoules in November 2022, but none was measured in every round. Both lines are 100 in June 2021. MLPerf Inference's ResNet-50 test has no closed-division power result after August 2024, so the AI line uses MLPerf Tiny, the same organisation's measured-energy test. Results: MLCommons, MLPerf Tiny v0.5 to v1.4 (first round, latest round, GitHub).

Aexodus views the increasingly leveraged AI value chain as a structural risk: model labs raise capital and commit it to cloud providers, which commit it to capital expenditure on chips and data centers. Each company's growth depends on a counterparty's spending, so a shortfall at one link can cascade through the chain. Against that outcome, we back companies that can keep going through a break in the AI financing cycle, and we invest only where a company's private valuation does not far exceed what public markets would pay for it. We want a resilient company, one that keeps its pricing power because its product is hard to substitute, can hold its revenue through difficult conditions and holds its value whichever AI model or lab wins. The physical AI companies that become the incumbents of the next decade will perform necessary work, reduce human exposure to dangerous tasks and keep essential systems running in good markets and in bad.