AI's consumer bubble is masking a much stronger business case

This article originally appeared in Capital Brief.

OpenAI's audited accounts were leaked last week, and the headline figures make for tough reading. Last year, the company spent USD34 billion and booked revenue of just over USD13 billion.

That has helped provide some context for this week's correction in SpaceX shares and the temporary market shutdown of the South Korean exchange, after memory makers SK Hynix and Samsung Electronics suffered sharp pullbacks.

Investors are asking whether there is an AI bubble, how large it is and when it will pop. Policymakers are asking the same question, as well as how far any bubble should affect the level of policy support the industry receives.

I'm not a fund manager. But I do know that Anthropic, which has just a sliver of OpenAI's much-vaunted consumer reach but far better penetration in the business market, is reportedly expecting to celebrate its first profitable quarter. That points to a two-speed economy running through the AI sector - consumer and business - and it should shape Australia's approach.

On the consumer side, adoption is wide but shallow. Around 2% of US households pay for an AI subscription, and OpenAI's own data shows just 5% to 6% of its users pay for the service.

The use cases everyone points to ChatGPT answering questions, Meta's AI selfies, song generators and Musk's chatbot talking to you for free in your Tesla - are almost all given away for free to buy user growth. For example, Mark Zuckerberg's Meta has committed USD600 billion to AI infrastructure and data centres, despite not having a clear path to returns on that investment.

On an earnings call, he told investors, "we are going to be largely focused on scaling and deepening engagement for at least the next year before we'll really be ready to start building out the business here".

These adoption rates have fed an argument that AI compute costs are much higher than the price AI labs charge for tokens. The argument is that an AI price shock is coming when the actual costs are passed on. But OpenAI's finances show clearly that pricing is not the problem. The freebies are.

Let's do some basic maths. If every one of OpenAI's 900 million weekly users paid for even the cheapest plan, ChatGPT Go, at USD8 a month, that would equate to more than USD80 billion in revenue a year. That is just the base plan, and it is clearly more than enough to cover the company's USD21 billion in annual losses.

OpenAI is not bleeding because compute costs more than the product. It is bleeding because almost no one pays for it, despite its leaked financials showing it spent a staggering USD5.7 billion on sales and marketing, much of which was undoubtedly chasing consumer adoption.

The business side looks nothing like that. Organisations pay per seat, or per token, at a price that carries a margin. Paid demand for business AI appears to be considerably outstripping supply.

Take Google Cloud's revenue backlog, which hit USD462 billion at the end of March. This is contracted customer work the company does not yet have the capacity to deliver, and Sundar Pichai attributed it predominantly to AI compute demand.

All of this makes consumer AI look like a capital-funded race for attention, while business AI looks like a transaction with a payback. It's the same industry and the same foundational technology, but the financial underpinnings are increasingly divergent.

Whether OpenAI can rein in costs and pivot towards more revenue-generating offerings remains an open question. It is also complicated by the company's impending public listing. For now, metrics such as active users may need to be maintained to justify what will likely be an eye-watering valuation for a company losing money hand over fist.

Funnily enough, one of OpenAI's competitors, SpaceX, which has pinned a large part of its future value on its AI arm xAI, may have emboldened Sam Altman to hold the line on giving products away for free, at least in the short term. After all, OpenAI looks set to list on a similarly sky-high revenue multiple and with the same absence of any near-term profit.

What we should do

But the more important question is not whether OpenAI turns out to be the Netscape of the AI era. It is what this tells us about whether there is an AI bubble about to burst, and what that would mean for Australian business leaders.

Australia's stuttering approach to AI policy should be shaped around sustainable AI business models - namely ones more likely to compensate copyright holders and follow through on data centre pledges.

Most importantly, given the question marks around the business model of consumer AI, any deals must be ironclad if a signatory changes its ownership structure as a result of market consolidation.

On copyright, the government is yet to confirm whether it will make changes to allow AI training to happen here. Reports suggest it is considering a range of options, including an extended collective licensing arrangement that would collect royalties, or a compulsory statutory licence without the option to opt out.

Whatever the solution, the government needs to structure it in a way that benefits not just rights holders, but also the broader Australian community and economy.

That means implementing a mechanism that forces AI labs seeking to train models here to invest in sustainable energy generation. It should not only cover their direct data centre consumption but also provide additional capacity for the Australian grid.

There should also be requirements for AI labs to set up local research facilities that contribute to training work, rather than simply pushing compute through Australian data centres. That would help reverse the brain drain and give Australians opportunities to work at the forefront of AI development.

Beyond policies tied to copyright reform, the government should also consider incentives and direct investment strategies like those pursued by Japan and South Korea, which have helped germinate burgeoning AI industries. And I'm not talking about policies tied to data centre builds or chip manufacturing.

It is worth noting that while the share prices of South Korean chipmakers are off about 10%, they are up almost 1000%. This is largely because Seoul has a steadfast AI policy agenda built on securing South Korea a strategic position in the AI supply chain. Australia can secure something similar.

Yes, there's a correction coming. Anyone watching AI services being given away to consumers for free can see it. The key is not to conflate the two sides of the industry, especially when Anthropic's results point to sustainable business applications that Australia can get behind.