AI Labs Have a $7 Doritos Problem

When PepsiCo's Doritos rose to $7, consumers immediately switched to substitutes, and Frito-Lay suffered slowing sales and shrinking shelf space. The author argues that today's AI labs could face the same fate if they fail to answer the same question. The core issue isn't performance, but whether the product is a necessity or a discretionary good.

For most consumers, a $20/month AI subscription competes as an optional choice against existing subscriptions like Netflix, Spotify, and iCloud. Enterprises face a similar situation: various surveys and meta-analyses show that many organizations are not seeing sufficient ROI from AI investments, and there's a growing perception that strategy is more about appearances than actual execution. So AI labs have yet to prove the 'necessity' that would justify higher prices.

The structure of absorbing GPU costs to grow the free user base is becoming increasingly unsustainable. OpenAI claims over 900 million weekly active users, but paid subscribers number only about 50 million, a conversion rate of just 5.5%. The author points to OpenAI's $5 billion loss in 2024 and a projected $14 billion loss by the end of this year, summarizing that shrinking free tiers and limiting features amount to de facto shrinkflation.

Anthropic, by contrast, places more weight on enterprise API deals and large contracts than consumer subscriptions. Anthropic's annualized revenue reached $30 billion this month, surpassing OpenAI's $25 billion, with 70-80% of revenue coming from enterprise API contracts. Over 500 companies spend more than $1 million annually, 8 of the Fortune 10 are Claude customers, and Claude Code alone has annualized revenue of $2.5 billion.

This gap isn't simply a matter of product preference—it's a difference in unit economics. OpenAI's costs surge as it absorbs massive consumer traffic, while Anthropic's monetization is relatively straightforward, centered on expanding existing enterprise customers and signing new contracts. The author believes this leaves OpenAI facing greater long-term capital expenditure burdens and profit-and-loss pressure.

Microsoft and Google are moving in different directions. Microsoft is raising Microsoft 365 prices starting July 1, 2026, bundling in Copilot features to protect ARPU, even as Copilot usage sits at just 3.3% of eligible users. Google, on the other hand, is leading with TPUs to lower the cost structure itself: TPU v6e claims up to 4x better dollar-for-dollar performance than NVIDIA H100 by internal metrics, and the new generation Ironwood (TPU v7) is designed specifically for inference.

Ultimately, the contest isn't about how expensively labs can sell more expensive AI, but about proving greater utility at lower cost. At the same time, alternatives in local models are maturing rapidly—Ollama's monthly downloads grew to 52 million in Q1 2026, and 135,000 GGUF models have accumulated on HuggingFace. As consumers and enterprises begin shifting to substitutes, AI labs, like Doritos, will need to prove their irreplaceability before raising prices.