The global shortage of artificial-intelligence computing capacity is moving beyond the training of large models and into a new phase dominated by inference, reasoning workloads, data-center power constraints and alternative compute markets. Cerebras Systems reported a $25 billion backlog in chip orders, saying its available capacity has already been taken by customers seeking faster hardware for advanced AI tasks.
Chief Executive Andrew Feldman said demand is no longer being driven only by companies building foundation models. The next wave, he said, is coming from businesses that need to run those models at scale, especially for reasoning tasks that can require massive volumes of token processing after a user submits a prompt.
The shift is creating pressure across the AI infrastructure chain. Chipmakers, cloud providers, data-center developers, power companies, open-source AI labs and decentralized compute networks are all competing for a place in a market where demand is now arriving before supply can be built.
Cerebras said its reasoning-optimized systems are fully booked by major U.S. technology companies. Feldman said the company’s hardware can run certain reasoning workloads up to 15 times faster than conventional chips, turning work that might take weeks into a single day of processing.
That speed advantage is becoming more important as AI systems move from simple text generation to multi-step problem-solving, planning, coding, research, search and agentic workflows. These tasks can require models to process far more information at inference time than earlier chatbot-style systems.
Reasoning workloads become the bottleneck
For much of the AI boom, the main constraint was training. Companies raced to acquire graphics processors and specialized chips to build ever-larger models. That phase is still significant, but the industry’s center of gravity is shifting.
Inference is the process of running an AI model after it has already been trained. In a basic use case, inference may involve answering a question or summarizing a document. In advanced reasoning systems, however, the model may break a problem into steps, test possible answers, revise its approach and generate long chains of output before producing a final response.
That process can consume huge numbers of tokens. Tokens are the small units of text or data that AI models read and generate. The more tokens a system must process, the more compute it requires. As AI applications become more interactive and more deeply embedded in business operations, inference demand can rise quickly and continuously.
Feldman said this is where specialized hardware can reshape economics. If a system can complete reasoning workloads far faster, companies may be able to reduce latency, serve more users and lower the cost of running complex AI applications.
The practical effect is that AI infrastructure is no longer a one-time requirement for model development. It is becoming an ongoing operating expense for companies that deploy AI at scale. That has made guaranteed access to compute capacity increasingly valuable.
Data centers reshape power demand
The surge in demand is also changing the data-center industry. New facilities are being planned and built across North America, Europe, the Middle East and Central Asia as companies try to secure enough electricity, land, cooling capacity and network connectivity to meet contracted AI workloads.
Some new data-center campuses are expected to draw power comparable to a mid-sized city. Industry executives have warned that the total electricity demand from projects now in development could exceed the combined power consumption of data centers built over the past several decades.
That scale is forcing a closer link between AI infrastructure and energy markets. Power contracts, grid access, long-term leases and local permitting have become central issues. In some regions, the ability to secure electricity is now as important as access to chips.
The article’s source material pointed to a recent long-term power and infrastructure lease valued at $19 billion as an example of how large these commitments have become. It also cited a proposed $9 billion acquisition of a major mining group, driven largely by access to large-scale data sites, as evidence that existing power-connected facilities are becoming strategic assets.
Mining sites are especially relevant because many were built with access to significant electricity supply. As AI companies search for locations that can support high-density computing, some former or current mining facilities are being considered for conversion into AI data centers.
Prepaid capacity changes the market
Feldman said a notable change in the current cycle is that major companies are paying in advance for compute capacity. In previous technology infrastructure cycles, supply often came first. Companies built capacity and waited for demand to arrive. In AI, customers are increasingly contracting for capacity before it is available.
That reversal lowers utilization risk for hardware suppliers and data-center operators. If capacity is already booked, companies can justify aggressive expansion. But it also creates a more difficult environment for smaller technology teams that cannot compete with the largest buyers.
When the biggest U.S. technology companies reserve chips, power and data-center space years in advance, smaller AI developers may face higher prices, longer wait times and fewer choices. That is pushing some teams toward alternative approaches, including open-source models, smaller specialized models, cloud spot markets and distributed compute networks.
The backlog reported by Cerebras suggests that the market for high-performance AI chips remains tight even as new suppliers enter the field. It also explains why hyperscale technology companies are investing heavily in in-house chip development. By designing their own AI accelerators, large cloud providers and platform companies can reduce dependence on external suppliers and improve control over cost, availability and performance.
Still, in-house chip development is expensive and technically demanding. Only a limited number of companies have the capital, engineering staff and manufacturing relationships needed to compete at the highest end of the market.
Open models gain ground in regulated sectors
The shortage of compute is not the only force reshaping AI deployment. Robin Rombach, head of Black Forest Labs and a leading figure in generative AI research, said regulated industries are increasingly interested in open-source and locally deployed models.
Finance, healthcare, government and other sensitive sectors often need greater control over data, compliance and model behavior. For these organizations, sending proprietary or personal information through external systems can create legal and operational risks.
Open models can be deployed inside a company’s own infrastructure, giving users more control over how data is handled. They can also be customized for local languages, internal documents and specific workflows. That flexibility is one reason sovereign AI initiatives are gaining momentum in several countries.
Sovereign AI refers to the development or deployment of AI systems under local control, often with national data, domestic infrastructure or region-specific governance. Governments and large enterprises are increasingly treating AI as a strategic technology, not just a software tool.
Rombach said the United States still has a limited range of domestic open-source options compared with the scale of enterprise interest. While several open models are available globally, many companies want systems that combine strong performance, clear licensing, security, compliance support and local deployment options.
Claims about AGI remain contested
The discussion also touched on artificial general intelligence, or AGI. Some AI leaders argue that systems have already met older benchmark-based definitions of AGI, particularly those developed two decades ago. Current models can pass or exceed long-standing human cognitive tests, perform complex problem-solving and show forms of recursive improvement.
However, the definition of AGI remains disputed. Benchmarks that once appeared ambitious may no longer capture the full meaning of general intelligence. Many researchers argue that true AGI would require stronger reliability, autonomy, long-term planning, causal understanding and real-world adaptability than today’s systems consistently display.
Even so, the rapid improvement in AI capabilities is changing expectations. Systems that once seemed theoretical are now being used in coding, research, design, robotics planning, medical analysis, legal drafting and creative production. The gap between laboratory demonstration and commercial use is narrowing.
Generative AI moves deeper into production
Rombach’s team is known for work on latent diffusion, an algorithmic approach that helped drive major advances in image and video generation. Black Forest Labs is now developing multimodal models that learn from images, audio and motion data together.
This approach allows a model to generate cinematic content while also understanding movement and physical structure. The same underlying architecture could eventually be adapted for robotics, where visual prediction must be connected to motor control.
In film and media, generative-video systems are moving from experimental tools into production workflows. Rombach described collaboration with director Martin Scorsese in which the system was used to visualize conceptual scenes from verbal descriptions. The process allowed rapid iteration, helping creators explore visual ideas before committing to more expensive production steps.
Studios and technology startups are using generative tools for marketing content, previsualization, short-form video and, increasingly, feature-related material. Some production teams report that AI-assisted workflows can reduce costs sharply and shorten timelines from months to weeks, though the results still depend on human direction, editing and quality control.
Rombach said similar models could be adapted to robotic control with relatively limited hardware-specific fine-tuning. If that proves commercially reliable, the boundary between generative media models and physical AI systems could become less distinct.
Distributed compute becomes a secondary market
As top-tier chips and data-center contracts become harder to access, smaller technology teams are looking for other ways to run AI workloads. One area drawing attention is distributed compute, where unused or underused hardware can be pooled through networks and rented to users that need processing power.
These systems may use spare graphics cards, processors, storage, bandwidth or specialized hardware from individuals, data centers, mining operators and small businesses. Instead of relying only on centralized cloud providers, users can access capacity through peer-to-peer or marketplace-style platforms.
The source material described an open marketplace for pooled technology resources that reached an estimated $50 billion in scale two years ago and has continued to grow. While estimates vary widely depending on what is counted, the direction is clear: idle hardware is becoming more valuable as AI workloads expand.
For small teams, the attraction is cost and access. If mainstream cloud capacity is too expensive or unavailable, distributed hardware markets may offer a way to test models, run inference jobs or support niche applications. For hardware owners, the appeal is the ability to earn revenue from assets that might otherwise sit idle.
These systems are not a perfect replacement for hyperscale infrastructure. They can face challenges involving reliability, latency, security, verification, software compatibility and service guarantees. But they are becoming part of the broader compute landscape.
Crypto-linked compute networks draw attention
Some distributed compute projects are tied to blockchain-based tokens that coordinate payments, incentives and access to hardware. These networks are often grouped under decentralized physical infrastructure, where participants contribute real-world resources and receive digital assets in return.
For traders, the AI compute shortage has turned these tokens into a closely watched market segment. The basic thesis is that if demand for decentralized GPU rental or compute marketplaces rises, activity on the underlying networks may increase as well.
However, trading these assets remains high risk. Token prices can move far ahead of actual network usage, and many projects still need to prove that they can deliver reliable compute at commercial scale. Regulatory uncertainty, token supply schedules, competition and technical execution all matter.
Rather than treating compute-linked tokens as a simple proxy for AI demand, traders are watching operational data. Hash rates at large mining pools, shifts from mining to AI workloads, new node growth in smaller towns, daily active users, utilization rates and hardware rental activity can all offer clues about whether real demand is developing.
If mining operators reduce hash rate to redirect power and equipment toward AI compute, that may signal stronger economics in AI workloads than in traditional mining. Likewise, a visible increase in active hardware units across decentralized compute networks may indicate that more users are renting capacity.
Still, network activity should be weighed against revenue, uptime, customer quality and long-term sustainability. A short-term spike in usage does not always translate into durable value.
The next phase of the AI build-out
Feldman said Cerebras expects its coming architectures to outperform traditional Moore’s Law scaling, with performance expected to more than double within 18 months. If achieved, that would deepen competition among chip designers at a time when AI workload growth is already stretching infrastructure.
The next stage of the AI market is likely to be defined by three linked questions: who controls the chips, who controls the power and who controls the models. Proprietary hardware, open-source systems, sovereign deployments and distributed compute networks are all becoming part of the answer.
For large technology companies, the priority is guaranteed capacity. For regulated industries, it is control and compliance. For smaller AI teams, it is affordable access. For traders, the key issue is separating real infrastructure demand from market hype.
The AI boom is no longer just about building smarter models. It is about building the physical and financial systems needed to run them every day, at global scale.
Explore how blockchain and DeFi could reshape this AI‑driven infrastructure boom in this in-depth analysis today.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
