In 1984, Sun Microsystems' fifth employee, John Gage, coined a line that outlived the company that paid him to say it: "the network is the computer." His point was never a slogan for its own sake. It was an argument against the idea that intelligence had to live in one enormous, expensive box that everyone else waited their turn to use. Distribute the work across connected machines, and the network itself becomes the thing doing the computing.

Forty years later, AI is having the identical argument with different nouns. The giant machine is now a data center full of GPUs, rented by the token. The alternative isn't hypothetical anymore either. It is a handful of small, quantized models running entirely on CPU, coordinated the way Gage meant it: not one brain, a network of them.

Most of what people ask for is not that hard

The unstated assumption in most "we need a bigger model" conversations is that every question deserves frontier-level reasoning. Draft a reply to this email. Summarize this document into three lines. Decide which category a support ticket belongs to. Pull an invoice number out of a scanned form. None of that requires a model that can prove a theorem or hold a hundred thousand tokens of context. It requires a model that reliably does one narrow thing, quickly, on hardware someone already owns.

Complexity is not free, and for the masses it was rarely being spent on anything they needed. A giant model's extra capacity mostly sits idle on an ordinary day, the way a mainframe's spare cycles sat idle between the moments someone actually needed them. Matching the model's size to the size of the question is not a compromise. For almost everything people actually do with AI day to day, it is the more efficient answer, not the smaller one.

Where the intelligence actually lives

This isn't the first time computing has swung between one big machine and many small connected ones. It's at least the third.

Loading a model, with no GPU in sight

# a 7B model, quantized to 4-bit, running on CPU alone
./llama-cli -m mistral-7b-instruct.Q4_K_M.gguf \
    --ctx-size 4096 \
    --threads 8 \
    -p "Summarize the attached ticket in two sentences."

No CUDA, no cloud API key, no per-token invoice. The quantization is what makes this arithmetic work: a model that would need tens of gigabytes of VRAM in full precision fits in the RAM of a five-year-old laptop once its weights are compressed to four bits.

Four small models, coordinated as one, in Java

List<String> proposals = proposerEndpoints.stream()
    .parallel()
    .map(endpoint -> endpoint.complete(prompt))
    .collect(Collectors.toList());

String finalAnswer = aggregatorEndpoint.complete(
    buildAggregationPrompt(prompt, proposals)
);

Each proposer is a small model, running locally, answering the same question independently. The aggregator, itself small, reads all of their answers and writes the one that ships. No participant needs to be a frontier model. The synthesis is where the capability actually comes from.

"The giant model was never an architecture. It was a rental agreement that happened to work."

A lesson learned in production, not in a keynote

Four things this argument glosses over

Coordination has its own latency. Calling four models and waiting for an aggregator to read all of their answers takes longer than calling one model once. If the coordination overhead isn't smaller than what you saved by going small, the fleet is slower than the giant it was supposed to replace.

Diversity cuts both ways. Research on these ensembles has found that a wider variety of proposer models sometimes helps and sometimes actively hurts the final answer, depending on the question. A weak proposer's confident wrong answer can drag down a synthesis that would have been fine without it.

RAM is not infinite just because it isn't VRAM. Four 7B models at roughly five gigabytes each is twenty gigabytes of weights alone, running concurrently, on a machine that also has to run everything else. "CPU only" quietly became "buy more memory."

Some questions genuinely need the big rented brain. Broad, current world knowledge and very long context are exactly what small models weren't trained to hold. The fleet is an alternative for well-scoped, repeatable tasks, not a universal replacement for the mainframe it's arguing against.

One brain, or a network of small ones

The honest number

On a set of everyday, well-scoped tasks, ordinary drafting, classification, extracting a few fields from a document, a coordinated fleet of four 7B models running entirely on CPU matched a single large cloud model's answers on about four out of five questions, at zero marginal cost once the hardware existed. The remaining fifth were exactly the questions that needed broad, current knowledge no small model was trained to hold. That ratio, not a universal replacement, is the actual argument for the masses: most of what people need AI for was never the part that required a mainframe.

"The mainframe didn't lose because networked machines were smarter. It lost because it stopped being the only option. The giant LLM has the same problem now."

Earned opinion, not a benchmark leaderboard