The “cheap AI” we enjoy today is, in fact, held up by massive losses across the industry. Major AI firms are serving their products below true cost to capture market share. OpenAI is projected to lose about $14 billion in 2026, and the core driver is the “inference” cost of handling each user request. Once the cross-subsidies from venture capital and cloud giants fade, AI prices are likely to normalize — in other words, to rise.

TL;DR

  • Today’s low AI prices are “subsidized” prices set below cost. By its own forecast, OpenAI expects a 2026 loss of about $14 billion (roughly $13B revenue vs. $22B spending), prompting the assessment that it “spends $1.35 for every dollar it earns.”
  • The core of the losses isn’t training but “inference” — the compute spent every time a user opens a chat, said to account for 80–90% of the load.
  • As subsidies fade, prices are set to normalize (rise). Some providers already add token surcharges, premium inference tiers and separate billing for AI agents. Surging data-center electricity costs add to the pressure.

What happened — the secret behind ‘cheap AI’

Generative AI has been cheap not because the technology suddenly got cheaper, but because frontier model makers — OpenAI, Google, Anthropic, Meta — have priced inference below cost to grab market share. Backed by venture capital and cloud-giant funding, that pricing created an “artificially low floor” in the market. As one analysis put it, “the subsidy era isn’t over, but it is ending.”

The problem is that the structure isn’t sustainable. Once capital discipline returns — that is, once investors start seriously asking “when will you turn a profit?” — prices have to adjust upward toward cost.

Why cost is the problem — losses born of inference

The crux is “inference.” Inference is the compute that generates a response to each user request; unlike one-time training, its cost keeps growing as users multiply. By one analysis, inference accounts for roughly 80–90% of AI compute load and is singled out as the biggest economic bottleneck keeping firms from profitability.

The numbers make the burden clear. OpenAI is estimated to have lost about $5 billion on roughly $3.7 billion of revenue in 2025, and by its own forecast its 2026 loss is reported to swell to about $14 billion (roughly $13B revenue, $22B spending) — the basis for the “spends $1.35 per dollar earned” assessment. OpenAI’s pursuit of huge new funding isn’t unrelated to this burn rate.

Will AI prices rise — signs of normalization

Signs of price normalization are already here. Some providers have introduced token-based surcharges, premium inference tiers, separate billing for autonomous agents, and credit-based metering. There’s also a repeated pattern of luring users with steep early discounts, then raising prices once promotions end.

“AI agents” are a wildcard that inflates costs. An agent reportedly consumes 10–50× more tokens per request than a chatbot, and Goldman Sachs projects that, as agent adoption ramps, token consumption could reach 120 trillion per month by 2030 — about 24× today’s level. As usage explodes, the actual “AI bill” companies pay grows even if unit prices hold.

It even shakes electricity bills — the data center’s shadow

The cost pressure is spilling into electricity bills. Per Goldman Sachs, U.S. electricity prices rose 6.9% in 2025 — more than double headline inflation (2.9%) — and data centers account for about 40% of the growth in power demand. Their share of total U.S. peak power demand is set to climb quickly from 4.1% in 2025. Goldman estimates higher electricity prices will weigh on household disposable income and spending and could lift core inflation by about 0.1 percentage point through 2027.

Component pressure piles on too. AI-server demand drove DRAM contract prices up 90–95% in a single quarter, and forecasts point to higher prices for PCs, smartphones and tablets. “AI cost,” in other words, is stoking cloud fees, electricity bills and device prices all at once.

What does it mean for Korea?

For Korea, there are two sides. On one, the AI adoption and operating costs borne by domestic firms and developers could rise. As overseas API prices normalize and agents push up token use, cost management — so-called “AI FinOps” — becomes a task, especially for companies adopting AI in earnest. Growing data-center power demand could also affect Korea’s power infrastructure and rates with a lag.

On the other side lies opportunity. Since much of the AI cost stems from “hardware” like memory and power, demand can work favorably for Korea’s memory-chip makers. That, however, is an industry-level trend — a separate matter from judgments about individual stocks.

The bottom line

The crux here is that the perception of “AI as a near-free tool” is colliding with the reality of cost. Today’s low fees are the price of a market-share war, not of the technology, and the gap is being filled by losses and investment. When the three threads — inference costs, agent-driven token surges, and data-center electricity bills — interlock, the cost of using AI leans structurally upward. For companies, it’s time to weigh AI’s benefits alongside its costs; for individuals, to hold realistic expectations about how long free and low-cost services will last.

※ This article is for informational purposes only and is not investment advice. Figures are estimates/forecasts and may vary by source and time.

Sources