The Era of Cheap AI Is Over
For years, AI companies promised their product would become a democratized and abundant utility. But as the sector pivots toward business clients and financial and environmental costs rise, the question is whether its unequal gains justify the price.
Attendees watch a demonstration of Anthropic PBC's Claude Code software at a conference in London, UK, on May 19, 2026. Anthropic is in early talks with investors to raise at least $30 billion in fresh financing, setting the stage for what could be its largest funding round yet. (Chris Ratcliffe / Bloomberg via Getty Images)
When OpenAI released ChatGPT in late 2022, it quickly broke records as the fastest growing technology product in history. AI model providers such as OpenAI, Anthropic, and Google initially used artificially low flat-rate pricing to drive adoption and capture market share, trusting that they could burn investor capital to manufacture dependency and then monetize a captive user base. But compared to the other consumer-facing platforms that had run this playbook since the 2000s, such as Facebook, Uber, or Instagram, generative AI differs in two critical respects.
First, each additional user generates a huge ongoing cost per query at a scale no social network has ever approached. Specifically, the chatbots consume an immense amount of computational power, which relies on electricity, water for server cooling, land for data centers, and billions of dollars in hardware investment. Second, as the models get more advanced, they also become more expensive to run. In that sense, they are closer to cloud-computing technologies such as Amazon Web Services.
By 2023, the research firm SemiAnalysis estimated that ChatGPT was already costing roughly $700,000 a day to run. The models have only gotten more complex and resource-intensive since. As of early 2026, with eight hundred to nine hundred million weekly active users users and only thirty-five million paying subscribers, the cost of sustaining global access to ChatGPT at this scale is around $17 billion a year, or close to $47 million a day.
In the words of Harvard business professor Andy Wu, most people don’t realize how “ridiculously expensive” AI is. Most are aware of the high fixed costs, but not the variable inference costs incurred every time the model generates an image. OpenAI expects to spend more than $150 billion on inference costs alone through 2030. While the vast majority of users continue to access the platform for free, the question is how the gap between resources and revenue will eventually close, and who will bear the costs.
AI’s Realization Problem
From 2022 to 2025, the sector struggled with a realization problem. This Marxist term refers to the moment in the circuit of capital (from money to investment in productive capacity and back to money) where goods are converted to profitable sales. A realization problem emerges when firms can produce enormous quantities of goods or services, yet cannot find enough buyers to recover costs.
In other words, a realization problem occurs when productive capacity expands faster than demand. In the first years of AI, tech companies invested enormous sums into generative AI — server farms, model training, engineering labor, and so on — but demand was insufficient. In late 2025, Wu said the pool of people willing to pay $20 a month for generative AI is smaller than that willing to pay $20 a month for Netflix. To generate demand for a product their employers were financially entangled with, tech companies pushed employees to use the chatbots regardless of whether that demand produced commensurate output.
For example, Meta and Shopify created internal leaderboards to track and reward token use. Nvidia CEO Jensen Huang said he’d be “deeply alarmed” if an engineer wasn’t using at least $250,000 worth of tokens in one year, and that “this is no different than one of our chip designers who says, ‘guess what? I’m going to use paper and pencil.’” This is, of course, after Nvidia invested $30 billion in OpenAI, financing the demand for its own product.
In January, Accenture told senior staff that they must regularly use AI tools to be considered for promotions. As an anonymous Accenture employee told Jacobin, “I wish there was more transparency on the enforcement of AI usage across senior staff. The policy is vague. I am worried it’ll encourage people to just use AI for the sake of hitting some target.”
The realization problem was thus deferred through a combination of fictitious demand creation and subsidized pricing. It was further obscured by the sector’s circular financing logic, where a small set of firms insulated themselves from market discipline by funding, supplying, and selling to one another in a closed loop.
AI Class Stratification
In late 2025, Anthropic released its Claude Opus 4.5 model, an agentic AI model aimed at “knowledge workers.” Opus 4.5 represents a genuine technological achievement by any conventional measure. The model is also dramatically more expensive to run and has led to both a shift in pricing strategy and a break from the earlier rhetoric of AI as a universal utility.
Tokens are the basic data unit that AI models process. Chatting with a chatbot uses several hundred tokens per paragraph. Agentic AI tasks, in which models autonomously browse the web, write and execute code, or manage complex workflows, can quickly consume millions. According to Silicon Valley entrepreneur and investor Vasudev Bhandarkar, the main reason agentic AI is exponentially more expensive is because of its multiplicity (the ability to handle many calls at once), the amount of context it can carry, its verification capacity, its use of external tools, and the high cost when it fails. The shift from conversational AI to agentic represents a huge leap in resource intensity.
Since early 2026, Anthropic has progressively introduced token-based surcharges, premium inference tiers, separate billing for autonomous agents, and credit-based metering for tool use and integrations. In early May, the company announced that Claude subscribers would face a separate monthly credit meter for agent tools and third-party harnesses (tools that wrap the Claude model) billed at full application programming interface (API) rates starting mid-June. Reached for comment, an anonymous researcher specialized in large language models (LLMs) told Jacobin:
Since Claude Opus 4.5, model providers are increasingly prioritizing agentic capacities and business-to-business (B2B) deals, while deprioritizing end consumers. Companies such as Google and OpenAI, though still interested in mass adoption, are pivoting. Anthropic prioritized B2B from the start.
The direction is clear: the most compute-intensive and powerful forms of AI are the priority and will be increasingly rationed to white-collar workers at big companies whose employers can pay the high price for them. In early March, Claude briefly overtook ChatGPT in daily active users, reporting a 1,487 percent surge in usage. The demand is there. The pricing model to sustain it at democratic, reachable access levels is not. In the words of Bhandarkar, “the question is, Will AI become like electricity or private jets?”
The Transition No One Voted For
For now, the industry’s answer to this contradiction is that the AI will pay for itself through gains in “productivity.” Uber’s recent experience suggests this may not be so simple. After spending $3.4 billion on AI in 2025, Uber put strong internal pressure on its five thousand engineers to adopt Claude, with leaderboards encouraging maximum usage. The company exhausted its entire 2026 AI budget by April. Uber’s chief technology officer admitted he was “going back to the drawing board because the budget I thought I would need is blown away already.”
The use value of AI and its costs to capital have diverged. The productivity gains are likely real but insufficient. For example, per-developer consumption at Uber has increased five- to twentyfold, but no public benchmarks show a matching increase in “output value.” According to the LLM researcher, “whether you’re a model provider or tech company in the AI economy, the competition is fierce and you feel pressure to keep the spending up, even if you’re investing more than you should or can.” For now, there is no evidence the math is adding up, and more importantly, no evidence it needs to. The same circular financing allows the sector to keep spending without reconciling cost and return.
Even if AI eventually pays for itself at a firm like Uber, what does that do for the majority of workers who are excluded from these gains? Put differently, even if AI is productive in the aggregate, for whom is that value of productivity realized? The efficiency gains are primarily for knowledge work, but the costs — higher prices, rationed access, labor precarity, and most urgently, environmental devastation — are shared unequally.
While the AI companies that pushed AI consumption onto us have no clear plan for how to pay for rising costs, they have demonstrated a remarkable capacity to secure innovative financing and defer the reckoning through capital markets. The primary concern is not the balance sheets of the AI economy, but whether this is a bargain the rest of us want to passively accept. Is it possible to renegotiate the terms of this transition?