When everyone has the same models, the advantage is the data they don't have

The question: If every company can rent the same AI, where does competitive advantage go?

Dataset & sample

Source

Stanford AI Index 2026; a16z (2019, 2024); Epoch AI; Shumailov et al., Nature (2024); Google Ads documentation

Sample

Public reports and papers, linked in the text

Your competitor can rent the same AI you use. It can't rent your customer history.

That's the shift underneath "AI marketing", and it's bigger than the usual frame. The interesting change isn't that marketers have new tools. It's that the chain between a decision and the data behind it has changed shape.

The stack is flipping

For fifteen years, software for marketing and sales had three layers. Now it has four:

In the SaaS era, a person opened a dashboard, looked at charts built on data, and did the reasoning in their head. The dashboard was the product. The judgment lived in the human.

In the AI era, an agent doesn't open a dashboard. It calls tools, pulls context, reasons over it and acts. The plumbing for that already exists: the Model Context Protocol, the standard that lets an AI work directly with other software, was adopted by OpenAI in March 2025 and handed to a Linux Foundation project in December 2025, with more than 10,000 public servers. I wrote last week about building Meta campaigns through one without opening the ads manager.

Satya Nadella described where this goes: business applications are essentially databases with business logic, and that logic is moving to agents.

Notice the new layer. Between "context and reasoning" and "data" sits ground truth: what actually happened, observed rather than estimated. A person in front of a dashboard could squint, notice that a number looked off and ask around. An agent can't. It reasons over whatever it's given, at speed, with confidence. So the quality of the bottom layers decides the quality of everything above them.

The model is stopping being the edge

Two numbers explain why the advantage is moving down the stack.

The models are converging. According to the Stanford AI Index 2026, the top four models on the Arena leaderboard, from Anthropic, xAI, Google and OpenAI, are now separated by fewer than 25 points. In the previous year's report, the top four spanned roughly 97.

2025 report

97

2026 report

25

Spread between the top 4 models on the Arena leaderboard (Elo points)

Group	Value
2025 report	97
2026 report	25

Source: Stanford AI Index 2025 and 2026

The 2026 figure is an upper bound: the report says "fewer than 25".

The models are getting cheap. Andreessen Horowitz measured that, for the same level of performance, the cost of running a language model has been falling about 10x a year. GPT-3-level performance cost about $60 per million tokens in November 2021 and about $0.06 in November 2024.

Put those together and you get a simple consequence. Your competitor can rent the same intelligence you rent, at the same price, this afternoon. If you both ask the same model the same question with the same context, you get the same answer. The only part of that sentence you control is the context.

The industry has already renamed the craft. In June 2025, Shopify's Tobi Lütke and Andrej Karpathy popularized "context engineering" over "prompt engineering": the work of filling the model's window with exactly the right information for the next step. The prompt is a sentence. The context is your business.