The Cost Dilemma of AI: How Infrastructure Economics Will Reshape the Next Phase of the Market

Compiled and edited by: Gonka.ai

AI is expanding at an astonishing pace, but its underlying economic logic is far more fragile than it appears. With three cloud giants controlling two-thirds of the world's computing power, with training costs approaching $1 billion, and with inference bills catching startups off guard—the true cost of this computing power arms race is quietly reshaping the value distribution of the entire AI industry.

This article does not discuss who will build the most advanced model. It explores a more fundamental question: Is the current economic model of AI infrastructure truly sustainable after scaling up? And how will changes in computing power allocation mechanisms reshape the value distribution of the entire market?

I. The Behind-the-Scenes Intelligence Costs

Training a cutting-edge large-scale model can easily cost tens or even hundreds of millions of dollars. Anthropic has publicly stated that training Claude 3.5 Sonnet cost "tens of millions of dollars," while its CEO, Dario Amodei, previously predicted that the training cost of the next generation of models could approach $1 billion. According to industry media reports, the training cost of GPT-4 may have exceeded $100 million .

However, training costs are just the tip of the iceberg. The real, ongoing pressure at the structural level comes from inference costs—the fees incurred each time the model is invoked. According to OpenAI's publicly available API pricing, inference is charged in millions of tokens. For high-usage applications, this means that even before scaling, daily inference costs could already reach thousands of dollars.

AI is often described as software. But its economic nature is increasingly resembling that of a capital-intensive infrastructure – requiring both high upfront investment and continuous operating expenses.

This shift in economic structure is quietly changing the competitive landscape of the entire AI industry. Those who can afford computing power are the giants who have already built large-scale infrastructure; while startups trying to survive in the cracks are being gradually eroded by inference bills.

II. Capital Intensity and Market Concentration

According to Holori's 2026 cloud market analysis , AWS currently holds approximately 33% of the global cloud market share, Microsoft Azure approximately 22%, and Google Cloud approximately 11%. These three companies collectively control about two-thirds of the global cloud infrastructure, and the vast majority of AI workloads worldwide run on the infrastructure of these three companies.

The practical significance of this concentration is that when OpenAI's API goes down, thousands of products are affected simultaneously; when a major cloud service provider experiences a failure, services across industries and regions are interrupted.

Concentration is not narrowing; on the contrary, infrastructure spending continues to expand. For example, Nvidia's annualized revenue from its data center business has exceeded $80 billion , demonstrating the continued strong demand for high-performance GPUs.

More concerning is a hidden structural inequality. According to SEC filings and market reports, leading labs like OpenAI and Anthropic have secured GPU resources at near-cost prices of $1.30–$1.90 per hour through multi-billion dollar "equity-for-computing-power" agreements. Meanwhile, smaller companies lacking strategic partnerships with Nvidia, Microsoft, and Amazon are forced to purchase at retail prices exceeding $14 per hour—a premium of up to 600%.

This pricing gap is driven by Nvidia's recent strategic investment of a total of $40 billion in leading labs. Access to AI infrastructure is increasingly being determined by capital-intensive procurement agreements rather than open market competition.

In the early adoption phase, this centralization may appear "efficient." But once scaled up, it brings with it a triple vulnerability: pricing risk, supply bottlenecks, and infrastructure dependence.

III. The Neglected Energy Dimension

There is another often overlooked dimension to the cost of AI infrastructure: energy.

According to the International Energy Agency (IEA) , data centers currently account for about 1–1.5% of global electricity consumption, and AI-driven demand growth could significantly increase this percentage in the coming years.

This means that the economics of computing power is not only a financial issue, but also an infrastructure and energy challenge. As AI workloads continue to expand, the geopolitical significance of electricity supply will become increasingly prominent—whichever country can provide the most stable computing power at the lowest energy cost will gain a structural advantage in industrial competition in the AI era.

When Jensen Huang announced at GTC26 that Nvidia's order visibility had surpassed $1 trillion, he described not just the commercial success of one company, but the grand process by which an entire civilization is transforming electricity, land, and scarce minerals into intelligent computing power.

IV. Rethinking Infrastructure Mechanisms

While centralized data centers continue to expand, another type of exploration is quietly emerging—an attempt to fundamentally redefine how computing resources are coordinated.

Decentralized Reasoning: A Structural Alternative

The Gonka protocol is a representative practice in this direction. It is a decentralized network designed specifically for AI inference, with its core design goal being to minimize network synchronization and consensus overhead and direct as many computing resources as possible to real AI workloads.

At the governance level, Gonka adopts the principle of "one vote per computing power unit"—governance weight is determined by verifiable computing power contributions, rather than capital shareholding ratios. At the technical level, the protocol uses short-cycle performance measurement intervals (called Sprints), requiring participants to demonstrate real GPU computing power in real time through a Transformer-based Proof-of-Work (PoW) mechanism.

The significance of this design is that nearly 100% of the network computing power is directed to the AI inference workload itself, rather than being consumed by infrastructure overhead such as maintaining consensus and coordinating communication.

The Economic Logic of Distributed Computing Power

From an economic perspective, the value proposition of decentralized computing networks has three levels.

First is the cost layer. The pricing structure of centralized cloud service providers essentially includes huge amounts of fixed asset depreciation, data center operating costs, and expected shareholder profits. Decentralized networks can significantly reduce these costs by monetizing idle GPU resources. For example, Gonka's inference service, currently offered through its USD-billed gateway GonkaGate, is priced at approximately $0.0009 per million tokens—while centralized service providers like Together AI price similar models (such as DeepSeek-R1) at approximately $1.50, a difference of over a thousand times.

The second layer is the supply elasticity layer. The computing power supply of centralized service providers is rigid, and the expansion cycle is calculated in months or even quarters. Participants in decentralized networks can join or leave elastically according to demand fluctuations, theoretically enabling a faster response to peak demand—just as Amazon Web Services was born due to peak traffic demand during holidays, the peak and trough fluctuations of AI inference also require elastic infrastructure to handle them.

Third is the sovereign layer. This dimension is particularly prominent from the perspective of sovereign states. When a government's public services are heavily reliant on an external cloud service provider, this dependence on computing power becomes a strategic vulnerability. Decentralized networks offer a possibility: local data centers can act as nodes to access a globally distributed network, securing data sovereignty while generating sustainable commercial returns by providing computing power to the global market.

V. The Moment of Value Restructuring

Returning to the core question at the beginning of the article: Is the current economic model for AI infrastructure sustainable after scaling up?

The answer is: sustainable for the top players; increasingly unsustainable for everyone else.

AWS, Azure, and Google Cloud have built up a moat through decades of capital accumulation, and their scale advantage is almost unshakeable in the short term. However, this structural advantage also means that pricing power, data access rights, and infrastructure dependence are highly concentrated in the hands of a few private entities.

Historically, every monopoly on major technological infrastructure has ultimately given rise to alternative distributed architectures—the internet itself was a rebellion against telecommunications monopolies, BitTorrent was a subversion of centralized content distribution, and Bitcoin was a challenge to centralized currency issuance.

The decentralization of AI infrastructure may not be an ideological choice, but an economic inevitability—when the cost of centralization becomes high enough to drive a large-scale user migration, the demand for alternatives will truly explode. Jensen Huang used the analogy of "every financial crisis pushing more people to Bitcoin" to illustrate this logic, which also applies to the computing power market.

The emergence of DeepSeek has proven one thing: in a world where the capabilities of open-source models are approaching the forefront of closed-source technologies, inference cost will become a core variable determining the speed at which AI applications can scale. Whoever can provide the lowest-cost, most available inference computing power will hold the ticket to this competition.

Conclusion: The infrastructure war has only just begun.

The next stage of competition in AI will not be determined by rankings of model capabilities, but by the economic competition over infrastructure.

Centralized computing power giants possess advantages in capital and scale, but they also bear the burden of fixed cost structures and pricing pressures. Decentralized networks are entering the market with extremely low marginal costs, but they need to prove that they can meet real commercial barriers in terms of stability, ease of use, and ecosystem scale.

The two paths will coexist and exert pressure on each other for a long time. The tension between centralization and decentralization will be one of the most noteworthy structural themes in the AI industry over the next five years.

This infrastructure war has only just begun.