LLMs as Grocery Stores

What the fastest growing grocery store in America can teach us about structuring LLMs.

Costco, ALDI

Aldi, a German grocery store, is currently experiencing rapid expansion in the United States, with over 125 stores added in 2025 alone¹. For reference, Target opened 23 stores in 2024², and Walmart only opened 9 stores in 2025³. Part of the reason for this rapid expansion is ALDI's M&A blitz, acquiring 220 Winn Dixie stores for future conversions⁴. Another factor might be their increase in marketing expenditure. Another is their private ownership structure and the subsequent ability to plan long term. However, what most economists and analysts feel is fueling their growth is their focus on less. Their stores are tiny at roughly 10,000 square feet and their product offerings are tight, carrying almost all food and few household essentials like soap and deodorant. The size and focus in their product offerings enables them to construct hundreds more stores than competitors and offer dramatically cheaper prices.

Costco is ALDI's antithesis: a warehouse sprawl where you can buy a rotisserie chicken, an engagement ring, and a cruise, then get your tires rotated, all within a half-mile radius. It would be foolish to say that Costco lacks a good business model or is inefficient because of its size (it was described as the greatest business of all time by the investment legend Charlie Munger). In comparison to ALDI's 125 stores in 2025, however, Costco opened 27⁵. It is impossible to compare a 10,000 square foot store with a warehouse that is 15 times that size, and it's unnecessary. ALDI is a semi-daily grocery store that one visits to grab a quick add-on for dinner while Costco is a once-monthly event that takes hours but provides hundreds of thousands of high-quality options.

As LLMs

The difference between these two grocery giants has parallels to the current state of LLM development. Providers like OpenAI and Anthropic are obsessively cramming more and more parameters into frontier models to see what is possible. In an April 2026 leak, Claude Mythos was reported to have been trained on ten trillion parameters⁶. These models are capable of performing incredible feats of programming talent, rivaling what a "10x developer" could barely fathom, much less perform. Simultaneously, Mythos can provide recipes for grilled cheese, plan a vacation to Denver, CO, and see if your friend's text was passive aggressive. It's easy to see how a Mythos-like model is similar to Costco: it delivers everything all in one place. That same model is slow, expensive, and time-consuming to use. It has to plan, think, delegate, and review, not dissimilar to a Costco trip. I am by no means arguing that Costco models don't have their place; they're incredibly effective at long processes that involve many inputs and differing requirements.

Future State

What I am arguing for is more ALDI-style models. Different providers have taken a stab at this style with Haiku and GPT Nano, but I want to see even more compressed models. For example, a model that is fantastic at producing Python code and nothing else. In the same way that most of the time I don't need my grocery store to provide eyeglasses, diamonds, and a pallet of peanut butter pretzel bites, I don't need my LLM to produce TypeScript and images. Trillion parameter models have their place, like Costco has their place in the grocery store food chain. What is missing from the model food chain is ALDI-style models that are good at one or two topics and that's it.