Five Companies, One Technology: The Power Concentration Problem

Frontier model training requires money, chips, data and specialised talent at a scale available to a very short list of organisations. Everything downstream of that fact — pricing, safety policy, what gets refused, which languages get served well — is decided by a small number of boards.

Where the chokepoints are

Compute. Advanced accelerator design is dominated by one company, and leading-edge fabrication by essentially one foundry in one region. That is a narrower dependency than the model layer itself and a geopolitical exposure at the same time.

Capital. Frontier training runs and the data centres behind them are financed at a scale that effectively requires hyperscaler backing. The major labs are consequently entangled with the major cloud providers through investment and compute commitments.

Distribution. Assistants reach users through operating systems, browsers, office suites and app stores owned by a handful of the same firms. Being technically competitive is not the same as being reachable.

Talent and data. Mobile, concentrated, and increasingly locked behind exclusive licensing deals that favour whoever can pay for them.

Why people worry

A few firms end up setting the effective norms for how hundreds of millions of people write, research and get answers — through pricing, refusal policy and default behaviour, with no democratic accountability. Whole industries build on APIs that can change terms, prices or capabilities with limited notice. Languages and markets that are not commercially attractive get worse service. And well-resourced incumbents can absorb compliance costs that would sink a challenger, which is why safety regulation and market structure are impossible to discuss separately.

The counter-argument

It is not weak. Capability is diffusing faster than the concentration story predicts: open-weight models from several countries now sit close to frontier performance on many tasks, inference costs have fallen dramatically, and small fine-tuned models handle a large share of real work. Switching between providers is genuinely easier than switching cloud or mobile platforms, because the interface is text. There are more credible labs today than five years ago, across more jurisdictions.

On this view, the market looks like an expensive oligopoly at the frontier with vigorous competition just behind it — closer to semiconductors than to search.

Where the disagreement actually sits

Not on the facts, mostly, but on which layer matters. If the frontier is what determines the future, concentration there is decisive. If most economic value comes from good-enough models close behind, diffusion wins the argument. If the chip and energy supply chain is the real bottleneck, then both camps are debating the wrong layer.

There is also a live split on open weights themselves: one side sees them as the main check on concentrated power, the other as the removal of the only practical control point over misuse. That trade-off is real and does not resolve neatly.

What you can do

Not much about market structure. Something about your own exposure: keep prompts and evaluation portable, avoid designing systems that only one provider's quirks can satisfy, and try alternatives regularly. Our platform pages exist partly for this — we cover Claude, Gemini, Grok, Midjourney, Higgsfield and Figma Weave without being paid by any of them.

Related: AI regulation around the world.