The AI Labs Have A $7 Doritos Problem

Allow me to introduce our newest author: Cici. Its name derives from Content to Cash or C-C…Cici. This article and my LinkedIn feed this week represent Cycle 1, which I will explain in my next article. But proof of value must be our central tenet. Can Cici improve my Content To Cash transformation? Let’s find out.

PepsiCo’s Frito-Lay division watched $50 billion in market value evaporate because it refused to answer a simple question: Is this product a necessity or a discretionary purchase? For Fritos or OpenAI, this story is not about higher prices. 2026 is the year of cause and effect; of intent and outcome.

Businesses and consumers will go into debt when they see utility or value. They cut back relentlessly when they do not. Where will the major AI players fall? Are they luxuries or necessities?

As OpenAI, Anthropic, Google, Microsoft, and many others contemplate price increases and new pricing models, most are missing the bigger picture. Profitability is a matter of survival for AI platforms and the businesses selling them. But the real question is, have any of these platforms earned the right to be profitable?

Proof of value is my core tenet, and every AI must adopt something similar, or they will not survive. It’s dark, but natural selection does not treat people any better.

All That & A Bag Of Chips

Doritos prices jumped nearly 50% between 2021 and early 2026. Some bags crossed $7, which is a lot for junk food. Walmart told PepsiCo to cut prices for over a year. PepsiCo tried everything except cutting prices, from promotions and shrinkflation to new product lines. None of it worked. Revenue turned negative for the first time in over a decade. Walmart pulled shelf space and handed it to Takis and its own private label.

There are structural parallels to AI. OpenAI, Anthropic, Google, and Microsoft are all navigating the same question Frito-Lay failed to answer. Consumers and enterprises are evaluating AI subscriptions the same way shoppers evaluated a $7 bag of chips. Is this worth it? Is this something I need, or something I can skip?

The answer, increasingly, is that people are willing to skip it. The success of agents and the agentic paradigm are often used as proof that LLMs and frontier model providers will be successful, but agents are a more robust category.

The Utility Question Nobody Wants to Ask

The AI industry frames every pricing discussion around capability. More parameters. Bigger context windows. Better benchmarks. Do enterprises or consumers really need all of that? It’s the equivalent of PepsiCo telling you that chips are a staple and assuming that customers will keep buying because they need those chips…no pun intended.

Just as chips are only a small part of a meal, frontier models are only a small part of agents.

Consumers and enterprises are buying utility. A $20/month AI subscription competes for wallet share against Netflix, Spotify, iCloud storage, and a growing list of digital subscriptions that already strain household budgets. For most consumers, AI sits in the ‘nice to have’ category. It is a discretionary good, not a staple that they need. It’s a lot like a potato chip.

The enterprise side is worse. According to a meta-analysis of multiple surveys and reports, only 5% of organizations achieve substantial ROI from their AI investments. Enterprises are spending on AI the way PepsiCo assumed consumers would keep spending on chips because the alternative seemed unthinkable. But when the value proposition does not materialize, behavior changes.

Writer’s 2026 enterprise AI survey captured the sentiment shift: 48% of executives call AI adoption a massive disappointment, up from 34% the previous year. Only 29% report significant ROI from generative AI. Nearly 75% of executives admit their company’s AI strategy is more for show than actual internal guidance.

AI is not failing, but for the majority of buyers, it is not yet a necessity. AI labs and vendors cannot charge necessity prices for a discretionary product.

The Tokenomics of Serving the Free Tier: Shrinkflation by Another Name

Every AI lab runs the same calculus. Free users drive adoption metrics, but every free query requires GPU cycles that cost real money. The economics are punishing.

OpenAI claims over 900 million weekly active users, but only about 50 million are paying subscribers. That is a 5.5% conversion rate, which is problematic. The company lost $5 billion in 2024 on $3.7 billion in revenue. Projected losses will reach $14 billion by the end of this year. As active users grow, losses are scaling faster than revenue.

That math and an impending IPO force a response, and it looks like what PepsiCo tried before it finally cut prices: shrinkflation. OpenAI has systematically tightened its free tier, and it is not the only one. Microsoft and Google have all done the same. Free is just too expensive.

Anthropic has taken a different approach but arrived at the same destination. It appears to be scaling back on the tokens used for thinking and reasoning for even paid tiers. As the senior director of AI at AMD noted, “Claude has regressed.”

Claude’s free tier imposes message caps that can be exhausted quickly during peak times. This month, Anthropic shut down subscription access for third-party agent frameworks, forcing users of tools like OpenClaw to switch to pay-as-you-go API pricing. Consumer subscriptions, Anthropic’s head of Claude Code explained, were never designed for the kind of continuous, automated demand these tools generate.

This is shrinkflation. The package and price look the same, but what you actually get keeps shrinking. The bag of chips weighs less. The free tier does less. The $20 subscription hits its ceiling faster.

With Mythos, Anthropic is marketing itself as a necessity or a staple for software development. In this week’s AI Product Management class, Vin discussed the path to a new pricing model. Anthropic committed $100M in tokens to helping partners in Project Glass Wing find and address security vulnerabilities. After that, partners will begin paying per token. It appears that there is no Mythos subscription, just a consumption-based pricing model based on tokens used.

OpenAI has yet to chart a course away from subscriptions and seats, a legacy SaaS pricing model. That puts OpenAI at a disadvantage, even with a frontier model. While Anthropic is growing the pie, OpenAI is shrinking the package.

PepsiCo tried this strategy. It reduced package sizes, offered multi-packs with fewer bags, and launched new product variants. In the end, the company’s own internal review concluded that none of it worked. Consumers noticed and switched to cheaper alternatives. The same dynamic is playing out in AI.

Anthropic vs. OpenAI: The Unit Economics Are Inverted

The biggest profitability divergence in the AI industry right now is between Anthropic and OpenAI’s unit economics, and it maps directly to the Frito-Lay lesson about who you build your business for.

OpenAI built for consumer scale. Over 900 million weekly active users. Massive brand recognition. ChatGPT became a verb, but consumer scale in AI is not the same as consumer scale in traditional software. Every additional user does not cost fractions of a penny to serve. They burn expensive GPU cycles.

The more users OpenAI acquires, the more expensive its infrastructure becomes. OpenAI’s paying users account for approximately 66% of inference spend, meaning the remaining 34%, the free tier, is a pure cost center.

OpenAI monetizes at roughly $25 per weekly active user. Anthropic monetizes at roughly $211 per monthly user, an 8x difference in monetization efficiency.

Anthropic’s revenue hit $30 billion in annualized run rate this month, surpassing OpenAI’s $25 billion, despite having roughly 5% of ChatGPT’s consumer user base. Claude has about 20 million users (best estimate, as Anthropic does not publish granular subscriber numbers), but Anthropic derives 70-80% of its revenue from enterprise API contracts.

Over 500 companies spend more than $1 million annually. 8 of the Fortune 10 are Claude customers. Claude Code alone reached $2.5 billion in annualized revenue in about a year.

The leaked WSJ financials ahead of both companies’ IPO preparations tell the story in capital expenditure terms. OpenAI projects spending $121 billion on compute in 2028, with losses of $85 billion that year. It does not expect to break even until after 2030. Anthropic projects it will reach profitability by 2028 or 2029.

OpenAI’s challenge is existential in a way Anthropic’s is not. OpenAI must convert a meaningful percentage of 900 million free users into paying subscribers while simultaneously funding the infrastructure to serve all of them. Anthropic needs its existing enterprise customers to expand their contracts and new enterprises to sign on. One of these problems scales infrastructure spending in line with revenue growth. The other gets messy very quickly.

Microsoft and Google: Different Economics, Different Traps

Microsoft and Google occupy a fundamentally different position from the pure-play AI labs. They are not building AI businesses from scratch. They are bolting AI onto existing cash machines and hoping the combined product justifies a higher price.

Microsoft’s approach is the most aggressive bundling play in enterprise software history. Starting July 1, 2026, Microsoft 365 prices increase across nearly every commercial SKU. Business Basic jumps 17%. Enterprise E3 rises 8%. The company frames this as delivering more value through expanded Copilot AI features, stronger security, and endpoint management tools.

The timing reveals the real reasoning. Microsoft is under visible pressure to demonstrate that its AI infrastructure spending translates to sustainable monetization.

Microsoft’s Copilot adoption numbers tell a cautious story. Only 3.3% of Copilot-eligible users pay for the service. Microsoft has responded by embedding Copilot into existing subscriptions. As of this year, new Microsoft 365 customers in some markets cannot buy a subscription without Copilot included.

This is bundling 101, which is what cable companies do to get customers to pay for unpopular stations. Microsoft is flattening value across tiers, eliminating cheap entry points, and using AI as cover for what amounts to ARPU defense inside an extraordinarily captive enterprise environment.

And then at the same time, Microsoft is walking back Copilot. It has taken Copilot out of multiple Windows apps. What is going on?

The Wall Street Journal reported that Microsoft spent over $84 billion on capital expenditures in 2025. Bloomberg subsequently reported the company was canceling some AI data center leases, suggesting a pullback from what even Microsoft internally recognized as overcommitment. Azure AI sales have underperformed to the point where Microsoft slashed sales quotas in half.

Now, some of Microsoft’s leadership team are voicing regret about dialing back on data centers. There are signs of internal confusion and a lack of conviction.

Google’s position is structurally different and potentially stronger because of one asset: TPUs. It’s all about the chips.

Google’s Tensor Processing Units represent a decade-long vertical integration bet that is now paying off. TPU v6e delivers up to 4x better performance per dollar compared to NVIDIA H100 GPUs for large language model training and inference workloads (per Google’s internal numbers, which should be taken with a grain of salt). The newest generation, Ironwood (TPU v7), was designed specifically for inference.

The cost implications are significant. Midjourney reduced inference costs by 65% after migrating from GPUs to TPUs. A computer vision startup cut monthly inference bills from $340,000 to $89,000 after the same switch. Anthropic signed the largest TPU deal in Google’s history, up to one million chips, worth tens of billions, bringing over a gigawatt of capacity online this year. SemiAnalysis estimates that Anthropic achieves significantly lower (as much as 52% by some reports) total cost of ownership per effective PFLOP on TPUs compared to NVIDIA’s GB300 systems.

Google controls the chip design, cloud infrastructure, and software frameworks. This eliminates the third-party margins that inflate GPU costs, which analysts call the “NVIDIA Tax.” Google can price AI inference more aggressively than any competitor because its cost structure is fundamentally lower.

Microsoft, by contrast, is fully dependent on NVIDIA GPUs (although it is building its own line of chips). The economics only work if Microsoft can convert Copilot from a bundled feature into a revenue driver. So far, that conversion is not happening at scale.

Microsoft is raising prices on a captive installed base, hoping they will absorb the cost increase. Google is investing in structural cost advantages that could enable lower prices over time. One strategy assumes consumers will learn to live with price increases indefinitely. The other assumes the cheapest producer wins. PepsiCo learned which assumption holds.

The Substitution Problem: Local Models and Knowledge Graphs

Here is where the Frito-Lay analogy breaks in AI’s favor, but not in the direction the labs want. When Doritos hit $7, Walmart gave the shelf space to Takis and its own private label. Consumers had alternatives, and they switched. In AI, the alternatives are maturing faster than the labs acknowledge.

Ollama hit 52 million monthly downloads in Q1 2026, a 520x increase from 100,000 in Q1 2023. HuggingFace hosts 135,000 GGUF-formatted models optimized for local inference. Open-weight models like Qwen 2.5 32B achieve 83.2% on MMLU benchmarks, within striking distance of GPT-4’s reported 86.4%. Benchmarks from earlier this year show that local models deliver 70-85% of frontier model quality at zero marginal cost per request.

The hardware to run these models is no longer exotic. Consumer GPUs with 24GB of VRAM can run 7B-14B parameter models with responsive inference speeds. Workstation-class hardware, like Vin’s Dell Pro Precision with an NVIDIA RTX Pro 6000 and 96GB VRAM, can run 70B+ parameter models that provide all the capabilities required for most enterprise workflows. Local AI is more accessible than ever.

For enterprises, the math is even more compelling when paired with knowledge graphs. A local model backed by a well-structured knowledge graph can answer domain-specific questions with higher precision than a frontier model operating on general-purpose training data. The knowledge graph provides the structured context that eliminates hallucination on business-critical queries. The local model provides the language reasoning layer. Together, they replace what enterprises are currently paying per-token API fees to achieve.

The AI labs have not answered this substitution threat because they cannot. Their entire business model depends on centralized inference. Every query that runs locally is revenue that never materializes. Every knowledge graph deployment that replaces RAG-over-vector-database-over-API is a customer that does not need to scale their API usage.

This is the private label on Walmart’s shelf. It is not as flashy, and the packaging is not as polished, but it costs less. It sits right next to the premium product, and for a growing segment of buyers, it does the job.

PepsiCo’s Nicholas Fereday, a food industry analyst, said, “The company assumed consumers would absorb price increases and only now appreciates how important affordability is to the typical consumer.”

The AI labs are making the same assumption. They are assuming that capability improvements will outrun cost sensitivity and that enterprises will keep paying for API access even as open-weight models close the gap. They are assuming that the $20/month consumer subscription sits on the necessity side of the ledger.

Frito-Lay made every one of those assumptions, and they cost PepsiCo $50 billion.