Graphics processing units (GPUs) that power artificial intelligence (AI) computation are expected to evolve into commodities traded on benchmark and futures markets, much like crude oil or electricity. AI service pricing structures are also projected to undergo a fundamental shift from flat monthly subscriptions to dynamic pricing that moves in real time based on supply and demand.
According to a BCG report titled “Is AI Compute Power Becoming a Commodity?” published on the 23rd, the global AI compute market is projected to expand rapidly from $360 billion last year to approximately $2.3 trillion by 2030. AI compute refers to computing resources deployed for AI model training or processing user requests, encompassing GPU chips, their usage fees, and tokens billed to end users.
As demand for AI services explodes across both enterprises and individuals, demand for these resources is surging in tandem. Alongside market expansion, AI service pricing structures are expected to undergo fundamental transformation.
Until now, AI developers such as Anthropic and OpenAI have charged enterprise customers flat monthly fees for their services. When demand surged and servers hit capacity limits, they responded by throttling speeds or blocking usage rather than raising prices, forcing consumers to accept degraded service quality at the same cost.
However, this flat-rate era is effectively coming to an end. Anthropic, citing intensifying AI compute shortages, switched to usage-based billing in April this year, charging enterprise customers a $20 monthly base fee per seat plus additional charges based on API usage. OpenAI similarly shifted Codex pricing from flat message-based billing to token-based usage billing.
Experts point to “dynamic pricing” as the next evolutionary stage. Just as wholesale electricity prices spike during midsummer afternoons and fall during late-night hours, AI service costs are expected to shift to a structure that moves in real time based on supply and demand dynamics.
Opaque Pricing Systems Need Standardization Like the Crude Oil Market
The shift in pricing structures signals that the broader AI compute market is undergoing a fundamental structural transition. Currently, AI compute market pricing is opaque. Published prices and actual transaction prices diverge, and most contracts are executed through discounts, advance reservations, and enterprise-to-enterprise agreements that remain undisclosed externally. Pricing benchmarks also vary widely—the same GPU can show price differences of up to 41% on the same day depending on which index is referenced.
Price volatility is also severe. Nvidia H100 GPU hourly rental rates peaked at approximately $8 in early 2024, fell to $1.96 by the end of last year, then rebounded to $2.64 in April this year.
BCG diagnosed this as a structural market problem, emphasizing the need to move away from opaque pricing systems toward a liquid and transparent market. The crude oil market is cited as a reference case. Crude oil is also a heterogeneous commodity with varying origins, sulfur content, and viscosity, but price reporting agencies (PRAs) collected and standardized actual transaction data, enabling Brent and WTI to become global benchmark prices.
However, BCG projects that AI compute is more likely to evolve in the direction of electricity markets—which form efficient regional markets—rather than crude oil, which has a single global price. This is because computation must necessarily occur where data resides, making cross-regional arbitrage inherently difficult.
As the AI compute market becomes more transparent, potential value previously obscured by opaque transaction structures can surface. BCG defines this as “dark value”—value that exists latently in the market due to supply-demand inefficiencies in capital, goods, and services but has not yet been realized.
BCG estimates that full commoditization of the AI compute market could unlock dark value of up to $140 billion annually. The analysis suggests value creation of up to $17 billion in the AI chip market, up to $34 billion in the compute rental market, and up to $46 billion in the end-user token market.
As the market becomes more efficient, the cost gap between companies that seize these opportunities and those that do not could widen. Ultimately, cost management strategies that respond to price volatility could emerge as a new source of corporate competitiveness.
Lee Joong-hoon, Managing Director and Partner at BCG Korea, said, “South Korea is benefiting from AI infrastructure demand such as high-bandwidth memory (HBM) and power equipment, but at the same time, it is also a consumer that purchases global AI models and computing resources at high prices.” He added, “Corporate AI competitiveness will now be determined not simply by adopting good models, but by the capability to maximize cost-effectiveness.”
Lee noted, “For general-purpose tasks like document summarization or email drafting, companies should improve cost efficiency, while making bold investments in areas where AI can redesign work processes themselves and create tangible outcomes.”
He added, “Establishing a system to continuously monitor models, pricing, and quality to avoid vendor lock-in, and to flexibly switch when necessary, will also be important preparation for responding to future price volatility.”
South Korean Stock Market Hits Yearly Low in Trading Despite Semiconductor Rebound
Meanwhile, amid the structural changes in the AI compute market, South Korea’s stock market is showing signs that investor sentiment is struggling to recover despite rebounds in large-cap semiconductor stocks.
According to the Korea Exchange, the KOSPI’s average daily listed share turnover rate from the beginning of this month through the 21st was 0.56%. This means that out of every 1,000 listed shares, approximately 5.6 shares were traded on an average day—the lowest level since August last year’s 0.50%.
The KOSPI’s average daily turnover rate rose from 0.86% in January to 1.65% in February and 1.74% in March. It then declined to 1.49% in April, 1.16% in May, 0.82% in June, and 0.72% in July, before falling to the 0.5% range this month. Compared to March, which marked the yearly high, the current figure is less than one-third.
Despite share price rebounds in Samsung Electronics (005930.KS) and SK Hynix (000660.KS), trading volumes actually decreased. According to FnGuide, from the beginning of this month through the 20th, Samsung Electronics and SK Hynix posted average daily trading volumes of 24.91 million shares and 4.69 million shares respectively, down approximately 22% and 30% from the previous month.
Market analysts note a time lag between index recovery and investor sentiment recovery. Retail investors have accumulated losses during the earlier market correction, and some investment funds have shifted to U.S. equities, weakening incentives to actively trade South Korean stocks.
Kim Hyun-sung, an analyst at Sangsangin Securities, said, “The broader trend of household funds moving into equities remains intact, but within that, allocation is once again shifting from domestic to overseas markets.” He added, “Retail investors who experienced the sharp swings and crashes in the South Korean stock market appear to be starting to feel fatigue.” He noted, “This trend suggests that retail investor sentiment has weakened significantly in the short term and will be difficult to strengthen for the time being.”