Pilbara Economics for the AI Age: how the world’s leading AI models are becoming commodities

It was a sunny day in Midtown Manhattan earlier this year when your author and several colleagues attended a lunchtime presentation by OpenAI’s CFO, Sarah Friar.

Touting many impressive statistics, Friar delivered her core ambition: “We want to price for value, not just cost–plus.”

As creators of some of the world’s most advanced AI models, pricing for the value these models create would lead to revenue–nirvana – a worthy ambition for any profit–seeking company.

But the ability to price for value is not determined by a producer – it is the result of the competitive dynamics of the industry.

Consider how this plays out in familiar territory – for instance, every producer of iron ore would love to participate in the economic value-unlock from the steel, infrastructure and cities they helped create. Just as every pen or paper manufacturer would love to participate in the economic value of the new ideas written using them.

But for commodities – products with little differentiation that can easily be substituted or switched between – pricing is usually determined by the cost structure required to supply them, not the value they unlock.

We decided to apply commodities-thinking to the new world of AI models. As we are likely only months away from the mega IPOs of Anthropic and OpenAI – producers of the western world’s most advanced frontier AI models – the competitive framing is an important one. After all, to justify near-trillion dollar valuations (for businesses that are largely profit-less today), investors need to believe in some kind of differentiating offering that can be sustained over time.

In this piece, we set out the defining characteristics of a commodity, assess how today’s AI models stack up against them, and extrapolate the competitive implications for AI model producers.

We find that, notwithstanding the extraordinary feats these companies have already achieved, most AI model companies appear to be competitively challenged. The source of durable competitive advantage, at least today, stems from owned and scaled compute infrastructure – favouring the world’s largest hyperscalers.

What is a commodity?

First of all, it is important to understand that commodity–status, or its absence, has nothing to do with the usefulness of the product. Oil, for example, is the most extraordinary substance that gives us energy, plastics, clothing, cosmetics and numerous medical products. Yet it is a commodity.

Commodity products tend to display the following key characteristics:

• Substitutability – Commodities are inherently similar enough that they can be substituted. Iron ore is iron ore – no matter where or by whom it is produced.
• Low customer switching costs – Given the high degree of substitutability, customers can quite easily switch between suppliers.

Given these criteria, the broad implications for producers of commodities are the following:

1. Pricing is set by the market, not the producer: Brands and relationships matter little because the price is usually set by the marginal cost of the producer who delivers the last unit of product demanded by the market.[1]
2. Competitive advantages arise from cost leadership and supply chain control: Commodity products rarely benefit from customer captivity – instead, their competitive advantages usually stem from lower production costs relative to other suppliers, and things like privileged (or monopolist–like) access to resources to produce said commodity. Consider Australia’s three large iron ore producers in the Pilbara: their advantages stem from their scale, combined with their access to high quality natural resources which cannot be replicated in most other places in the world.
3. Industry profits are usually cyclical: Profits are usually driven by the following dynamic:

• As demand growth exceeds supply growth, the price of the commodity increases – pushing up industry profits;
• Higher prices means higher-cost producers can now sell into the market profitably;
• Consequently, supply growth exceeds demand growth, so prices fall – pushing down industry profits;
• This knocks out the higher-cost producers from the market, and so the cycle continues.

Today’s AI models sure look commodity-like

Today, most readers will have played with at least one or two AI models. From Gemini, to ChatGPT, to Claude, to Grok, DeepSeek, Mistral, Llama, and Kimi, there are a wide range of options.

These AI models appear to satisfy the criteria that define commodities quite well:

• Substitutability – Ask the same question of any of the above models, and you’ll get a relatively comparable answer. Sure, each will have a slightly different ‘vibe’, and all will hallucinate from time to time, but for the vast majority of use cases, these models deliver a very similar experience.
• Low customer switching costs – It is trivially easy to switch between models for most use cases. A consumer using the ChatGPT app on their phone could simply open the Gemini app instead. Even for complex cases of AI models being integrated into applications, many developers agree it is possible to switch between AI models with very little disruption – many applications even give their users this choice explicitly.

With this in mind, AI models sure do look a lot like commodities… and if that is the case, we can extrapolate the competitive implications for producers.

For most model producers, this looks like bad news

The marginal cost of delivering ‘intelligence’ essentially has two parts: the cost of model hosting and inference, plus the ‘mark–up’ the producer chooses to charge above those unit costs.

On the first, model hosting and inference costs are essentially the same for all producers and outside of their control (at least for those who don’t own their own data centre infrastructure). That’s because those costs are set by the handful of large cloud providers who deliver most of the industry’s compute: Amazon Web Services, Microsoft Azure and Google Cloud Platform.

That means the only lever left is to lower the mark-up itself – which directly determines the model producer’s gross profit margin. And the bad news for model producers – particularly for Anthropic and OpenAI which have been reportedly enjoying very attractive gross margins – is that several Chinese producers have pushed the mark-up to basically zero – releasing near-state-of-the-art models as free, open-weight downloads for anyone to run. That is, they’re essentially giving them away for free!

As Bloomberg reported in , Chinese AI lab Moonshot released Kimi K3, a more advanced open-weight model that it said outperforms all rivals except for Anthropic’s Fable 5 and OpenAI’s GPT–5.6 on overall capability. Other Chinese firms, including DeepSeek, have pushed to undercut US rivals on pricing.

Just days later in response, Anthropic launched their latest model, Claude Opus 5. The company says the new offering approaches the capabilities of its best models, but at half the price.[2]

Just as it is for commodities, price competition in foundation AI models looks like it’s here to stay.

For those who own large–scale infrastructure, the news looks good

Close followers of stock prices will know that Google, creator of the Gemini AI model, has done exceedingly well over the last year or so. But if model gross margins are being competed away, how is Google succeeding?

The answer lies less in Gemini itself, and more in Google’s cost advantage to host and serve the model. Google has designed its own chips, called TPUs, to optimise the cost performance of Gemini (on which its core advertising businesses rely). The company has also optimised its entire data centre network to specifically reduce the costs of serving Gemini. This represents a structural competitive advantage that other model producers cannot replicate.

Amazon and Microsoft should also do well, but again, not because of their AI models. As the world’s two largest cloud computing hyperscalers, they will likely win if demand for model hosting is high – and are agnostic to which model consumers and enterprises demand. As price deflation in the AI model space intensifies, this should boost consumer and enterprise demand further – to the benefit of the world’s hyperscalers.

Even Meta appears positioned extremely well, but also not because of the internal AI models it has developed. Again, as one of the largest investors in data centre capacity over many years, it can now host and deliver intelligence (largely to its own core business and users of Instagram and Facebook) at a significant cost advantage to others.

Montaka continues to own Amazon, Microsoft, Meta and Alphabet on this basis: not for their AI models, but the competitive advantage their compute infrastructure gives them over rivals. We have held these positions for many years, and the explosive growth of AI only strengthens our conviction in them.

* * *

We are likely only months away from the mega IPOs of Anthropic and OpenAI and the hype surrounding these companies remains intense. But investors would do well to remember that, in the end, the Pilbara lesson will likely hold.

The hyperscalers are the iron ore miners who own the advantaged ore body and can produce at structural cost advantages. Anthropic and OpenAI are the steel mills – buying their essential input at the same price as everyone else, competing for margin on a commodity product against Chinese producers who are willing to sell for little mark–up.

OpenAI’s Sarah Friar wants to price for value.

The market may only let her price for cost.