Technology

Consumer AI Pays Like a Niche Product, Spends Like Infrastructure

Martin HollowayPublished 4h ago3 min readBased on 6 sources
Reading level
Consumer AI Pays Like a Niche Product, Spends Like Infrastructure
Image by Lalmch from Pixabay

Only 2.2% of consumers were paying for AI services as of May 2026, spending an average of $31 per month. That pairing of low conversion and modest ARPU defines the consumer problem in one line, according to data cited by TechCrunch.

The scale contrast is stark. Netflix counts 325 million subscribers, in the same reporting. Consumer AI has usage, cultural presence and daily habit formation, but it does not yet have a comparable base of people willing to pay directly for access.

The math is tight. Enterprise demand has grown quickly, with Anthropic's revenue surging 12-fold in 2025 to nearly $4.6 billion, as reported by Reuters. Earlier in that ramp, the company had hit $3 billion in annualized revenue, a figure reported in May 2025 by Reuters. OpenAI had projected at that time that it would end 2025 with more than $12 billion in total revenue.

Costs have grown faster. Anthropic spent $7.33 billion on compute and infrastructure last year. Its operating loss widened to $8.06 billion, according to Reuters. Those are infrastructure-scale numbers attached to software margins that have not yet materialized in consumer.

Distribution adds a second toll. Anthropic pays over $350 million in distribution fees to cloud platforms, as reported by Reuters. For consumer AI delivered through app stores, cloud marketplaces and bundled assistants, paid acquisition and revenue share compress an already thin $31 per month.

Pricing power is eroding at the same time. Cheap open models have caught up on capability with frontier models, according to Ars Technica. Paying for frontier models buys a 4-month head start at 5x the cost, in the same reporting. For engineers deciding between API tiers, that tradeoff is familiar. A short-lived quality edge rarely sustains a 5x premium once good-enough alternatives are available for inference at scale.

The macro return so far is small. AI-related investment has lifted U.S. GDP by less than 0.4% since 2022, according to Reuters. Capital expenditure arrived first. Broad productivity effects have not.

The broader context here is a mismatch between cost structure and revenue model. Training and serving large models require continuous capital outlay for accelerators, power, networking and staff. Consumer subscription revenue arrives in small monthly increments and only from a narrow slice of users. Enterprise contracts can bridge the gap, and the Anthropic revenue trajectory suggests they are doing much of the work, but consumer remains a high-volume, low-yield channel with infrastructure costs priced in dollars and revenue collected in tens of dollars.

In my view, the PC era and the early commercial internet offer a useful parallel, not because the technology rhymes but because the monetization sequence does. General-purpose platforms often lose money on broad access while the paid use cases concentrate in workplaces where time saved has a direct price. Consumers benefit enormously, just not always through a direct subscription relationship. Free, ad-supported, bundled and employer-paid access carried earlier waves long before household spending caught up.

Looking at what this means for builders, the pressure points are inference efficiency, differentiation that survives a four-month window, and distribution without a platform tax. If open weights deliver near-frontier quality for routine summarization, coding assistance, search and drafting, then paid consumer tiers need durable advantages in memory, agency, integration or trust, not raw benchmark leads. Otherwise $31 per month becomes a ceiling that only power users will pay, while most users stay on free tiers subsidized by someone else.

Worth flagging is how this could still resolve optimistically. Low willingness to pay now does not fix willingness to pay later. Once models are embedded in workflows where they reliably complete multi-step tasks, conversion can shift from novelty subscription to utility billing. The infrastructure being built at a loss today lowers latency and cost per token for that future. The near-term economics look ugly because they are. The long arc still points toward cheaper inference making capable assistance widely available, even if the path to getting paid for it runs through enterprises first.