“Growth at All Costs” Is Dead — Long Live the AI Subscription

Set Trending Topics as a preferred source on Google.

The triumphant advance of artificial intelligence (AI) faces an economic hurdle: astronomical operating costs. As the technology becomes ever more present for end users, industry giants like Meta and Google are undertaking a strategic shift. To shoulder the massive investments in computing power and infrastructure, new tiered subscription models are moving into focus, increasingly monetizing access to powerful AI.

For a long time, the strategy of the big tech corporations was defined by “growth at all costs” — often financed through massive advertising revenue and largely free services. But the era of unlimited, free AI capacity appears to be drawing to a close. The computing power required to train and operate modern language models is so expensive that the previous business models alone are no longer sufficient.

That inference is becoming the ultimate factor in the AI rollout is also evident in the price increases at Anthropic and OpenAI, which — in light of their planned IPOs — will likely soon start pressing even harder on the monetization pedal.

Meta: From Social Media to “Meta One”

The social media giant Meta clearly demonstrates what this transition can look like. The company is currently rolling out new subscription plans for its core applications worldwide. With models such as “Instagram Plus,” “Facebook Plus,” and “WhatsApp Plus” (priced between $2.99 and $3.99 per month), Meta is attempting to diversify the monetization of its already saturated user base. While these plans primarily offer additional features for power users — such as advanced insights or profile customization — the strategic direction points to a deeper goal.

Particularly relevant to the AI debate is the new pilot project “Meta One.” Here, specialized AI subscriptions are being tested that go beyond purely social functions. While a basic version remains free for casual users, the “Meta One Plus” ($7.99) and “Meta One Premium” ($19.99) plans offer a decisive advantage: more capacity for compute-intensive requests. Premium users gain access to deeper “reasoning” and expanded capabilities for generating videos and images. Meta uses this differentiation to pass the cost of the most complex computing operations directly on to the users who make the most intensive use of these features.

Google: Massive Investments and Tiered Bundles

Google is pursuing an even more aggressive investment strategy. According to CEO Sundar Pichai, the company plans to invest between $180 and $190 billion in AI infrastructure this year — a six- to eightfold increase compared to 2022. To justify these sums, monetization is being driven forward on two tracks: on one hand, through the integration of advertising into AI answers, and on the other, through highly tiered subscription models.

Google’s strategy relies on bundling services. Across various tiers — from “AI Plus” (approx. €8) up to “AI Ultra” (up to €200) — users gain access to ever more powerful Gemini models. These subscriptions are often linked to existing services such as YouTube Premium to increase the value of the package for the end user. The scale is enormous: Google now processes more than 3.2 quadrillion tokens per month — a volume that has multiplied within a very short time (more on this here).

Why AI Is So Expensive: The Principle of Tokens and Inference

To understand why these subscriptions are becoming necessary, one has to consider the technical foundation of AI computation. The costs arise primarily from two factors: token consumption and the process of inference.

Tokens are the basic units in which AI models process information. They can be thought of as small building blocks — often word fragments or whole words. When a user asks a question, that text is broken down into a long sequence of tokens. The more complex the request or the longer the response text, the more tokens have to be processed. Since each individual processing step requires computing power, costs rise linearly with the volume of tokens.

Inference describes the actual computing process: the moment when the already-trained model receives an input and generates an answer from it. In contrast to “training” (the one-time learning of the AI), inference takes place with every single user request. This process runs on specialized, extremely expensive graphics processors (GPUs) that consume enormous amounts of electricity and require a massive hardware infrastructure.

When a user poses a complex task — such as analyzing a long document or creating a video — not only does the number of tokens rise, but so does the complexity of the inference. The model has to “think more deeply,” which causes more compute cycles and therefore higher costs. The new subscription models from Meta and Google are ultimately an attempt to cover these variable costs through a fixed fee and to separate the high-intensity users from the casual ones.