An a16z podcast exposes the unsustainable bubble of consumer AI
By Klaus Berger · Reporting from Frankfurt ·
As partners on an a16z podcast discuss how users spend money, the high cost of serving them reveals a subsidized bubble that must face economic reality.
On a recent episode of the a16z podcast, titled "The Current State of Consumer AI," host Elena Burger sat down with investing partners Olivia Moore and Josh Elman. They discussed how consumers spend money on artificial intelligence. The most striking revelation is the financial fragility of this nascent market. While roughly half of Americans report using AI, only a few pay for a subscription. More alarming still is the concentration of this spending. The top 1 percent of paying users spend an average of $93 per month on their personal credit cards. Meanwhile, the computational bills to serve them can cost startups hundreds or even thousands of dollars per user.
The unsustainable balance sheet of compute
On the podcast, Moore explained that their latest report incorporated consumer card spend data for the first time. This revealed a stark power-user economy. She noted that many active users employ these tools for coding and technical automation. This explains why server costs for some assistant products run into the thousands of dollars. Elman argued that this high cost of goods sold (COGS) represents a fundamental departure from previous internet eras. In the Web 2.0 era, startups focused on user density. The marginal cost of serving an additional user was near zero. Today, growing too fast without strict monetization meters can lead to out-of-control liabilities.
From an ordoliberal perspective, scale does not alter the laws of arithmetic. If your marginal cost of production exceeds your marginal revenue, expansion merely accelerates bankruptcy. The current venture-backed model of consumer AI is essentially a massive subsidy program. Investor capital bears the risk of expensive, energy-intensive queries. This creates a severe moral hazard. Developers are incentivized to build resource-heavy "agents." Meanwhile, consumers are shielded from the true market cost of the computing power they consume.
The unnatural inversion of the internet economy
To cope with these high infrastructure costs, consumer AI has turned to subscriptions. Moore noted that 85 percent of the web-based AI products on their list monetize through direct subscriptions. She called this an "unnatural inversion" of internet history. Historically, the largest consumer platforms have relied on advertising to keep access free. Elman countered that as inference costs fall, traditional models like advertising will become viable. He pointed to OpenAI’s advertising run rate as evidence that commercial intent can subsidize these costs.
This optimism ignores how institutions enforce their boundaries. The ablest advocates of the AI boom argue that computing costs will decline sharply. They believe subscriptions can be discarded for high-margin ad networks. But this assumes consumer expectations will remain static. As models grow more sophisticated, the computing power required to run them will scale exponentially. OpenAI’s ad run rate is impressive. However, it is a drop in the bucket compared to the capital expenditure required to maintain its infrastructure. Relying on advertising to bridge this gap is a grand gesture that ignores the underlying balance sheet.
The software layer cannot hide the liability
Elman argued that the ultimate value of consumer AI is moving back to the software layer. Here, startups build rich, bespoke experiences that leverage underlying models rather than acting as mere "wrappers." In his view, this context and community will become the defensible asset. This will allow startups to survive even as base models are commoditized.
This thesis fails the test of long-term incentives. A startup's operational viability remains entirely dependent on the pricing and stability of a few massive model providers like OpenAI or Anthropic. It holds all the liability and none of the structural power. History shows that when the underlying infrastructure of an industry is concentrated in a tight oligopoly, intermediaries are squeezed out. No bespoke interface can protect a startup from a sudden hike in API fees. Nor can it prevent direct feature replication by platform providers.
The market for consumer AI is built on subsidized illusion. We are witnessing a classic misallocation of capital. Investors bear the risk, while consumers enjoy the free lunch. The basic rules of fiscal discipline are suspended in hopes of a future miracle. A stable economy is built on boring competence, exact pricing, and rules that outlast technological hype. Until consumer AI products are priced to reflect their true marginal costs, this boom remains a speculative bubble. It is waiting for its balance sheet to catch up with reality.