# Small is a strategy

By Rea Stark, 12 November 2026
Source: https://reastark.com/writing/small-is-a-strategy

> When a machine gets small, somebody else can own it. Small language models and small modular reactors move power from institutions to owners.

The Nespresso Pixie is 11 centimetres wide. I worked on it. Into that width go a pump rated at 19 bar and a heater that is ready in 25 seconds. Espresso once needed a café and somebody trained to work the machine. The Pixie needs a shelf.

What happened to coffee is now happening to two far larger machines. Small language models run on a phone, and small modular reactors can sit beside the factory or data centre they feed. The industries that build them still worship scale and treat the small versions as curiosities. They have it backwards. When a thing gets small, somebody else can own it, which makes miniaturisation a transfer of power.

## Who owns the big one

Big machines belong to institutions, because only institutions can pay for them and wait for them. The two new reactors at Plant Vogtle in Georgia were planned for 2016 and 2017 at a cost of $14bn. They entered service in 2023 and 2024 at more than $30bn. In December 2023 the state's regulators let Georgia Power, one of the four owners, raise its rates to recover $7.56bn of its remaining costs. Only an owner with captive customers could have carried a bill like that.

Artificial intelligence has copied the arrangement. In January 2025 OpenAI and its partners announced Stargate, a plan to invest $500bn over four years in computing infrastructure. The largest models live in buildings of that kind, and everybody else rents them by the token. The landlord sets the price and can change the locks.

Computing has been here before. The mainframe sat in a machine room and belonged to the company, until the personal computer put a lesser machine on a desk and gave it to whoever paid for it. Nobody now calls the PC a weak mainframe.

## Small language models, small reactors

In April 2024 Microsoft published Phi-3-mini, a model with 3.8bn parameters that its researchers ran offline on an iPhone 14, in about 1.8 gigabytes of memory. By Microsoft's own tests it rivalled GPT-3.5, the model family ChatGPT launched on in late 2022. In August 2025 Google released Gemma 3 270M, with 270m parameters. In Google's tests on the chip of a Pixel 9 Pro, 25 conversations used 0.75% of the battery.

Models of this size do less than the giants. They also run on a device the user owns and send nothing to anyone. A law firm can run one over its own files. So can a hospital, the customer least able to post its records to somebody else's server.

Reactors are taking the same road. The International Atomic Energy Agency defines small modular reactors as those producing up to 300 megawatts of electricity per module, and counts more than 80 designs and concepts. Deep Atomic, where I am a board member, develops the MK60, which is smaller again, at 60 megawatts of electricity per unit, and is built for AI data centres. At that size the buyer can be a single company instead of a state's ratepayers.

A data centre that brings its own power stops waiting for the grid, and the company that owns its power owns its timetable. It is a fair joke that the first customers for small reactors are the buildings that house the biggest models.

## The efficiency objection

The objection is that small things are weaker and cost more per unit, and it is partly right. Economies of scale are real. In November 2023 NuScale and a group of municipal utilities in Utah cancelled what was planned as America's first small modular reactor plant, after its target price rose from $58 to $89 per megawatt-hour and too few buyers signed up. No model on a phone will match one that fills a data centre, either.

Cost per unit is the institution's measure. It counts the output and leaves out the price of admission, which is a decade of building and the capital that waits through it. An owner wants to know whether the machine is affordable at all and whether anyone can take it away, and small is a strategy because it answers both. The buyers in Utah were towns shopping for a reactor. The price stopped them, but the size is what let them into the shop.

Small things also get cheaper in a way big ones do not, because they are made in series, and whatever is built by the thousand improves with every run. Vogtle's builders got two tries. The first small reactors will be expensive, as first products are, and the comparison that matters is with the tenth.

As for weakness, most work is small work. A model that sorts invoices or answers questions about one company's manuals has no use for the whole internet. Google calls its 270m-parameter model the right tool for the job, which is a polite way of saying that the big one is often the wrong tool.

The Pixie's water tank holds 0.7 litres. That is a hopeless figure for a café and the right one for a kitchen, and the kitchen is where the espresso went.

---

Rea Stark is an industrial designer and founder in Zürich. He has designed for Sony and Nespresso, co-founded Piëch Automotive and Archaí Physics, and works on how machines see, understand and use the physical world.

About the author: https://reastark.com/about
More essays: https://reastark.com/writing