AI-linked semiconductor shares have fallen roughly 30% to 40% from their peaks even as orders for large-scale computing capacity have expanded sharply, creating a disconnect between equity-market expectations and the demand visible in hardware procurement. EpochAI figures cited in the source material put the global backlog for major compute orders at more than $2 trillion, up from about $500 billion in early 2025.
The growing queue suggests that demand is no longer concentrated in a small group of frontier model developers. Government agencies, scientific institutions, cloud providers, product companies and individual users deploying AI-assisted software are competing for capacity, often with very different tolerance for high prices. That mix makes a simple comparison between chip production and a handful of corporate AI budgets an incomplete measure of the market.
Rental prices keep dedicated hardware economically attractive
The economics of owning computing equipment rather than leasing it from a public cloud provider help explain why demand can persist even after hardware prices rise.
A self-built inference machine priced at $15,000 was compared in the source material with a similar cloud GPU offering listed at about $3,248 per month on an on-demand basis. At that rate, the cost of buying the system would be recovered in roughly 4.6 months. A three-year cloud commitment, quoted at around $1,444 per month, lengthens the implied payback period to about 10 months, but still leaves the owned machine cheaper over a multi-year operating life.
The comparison does not account for every cost of operating hardware, including electricity, networking, maintenance, financing and data-center space. Yet it illustrates why companies with sustained workloads may continue to buy or lease dedicated machines instead of relying entirely on public cloud availability. The source material says similar calculations apply across other systems and estimates that an H200-based cluster could recover its purchase cost in less than two years under comparable conditions.
For AI deployment, the relevant workload is increasingly inference: the process of running a trained model in response to user requests. Training large models remains expensive and capacity-intensive, but a successful consumer or enterprise product can generate a continuous stream of inference demand long after the training run ends. The source material estimates inference service-provider margins at roughly 50% to 70% and says demand for available inference capacity exceeds supply.
That profitability has implications beyond chipmakers. It places a premium on organizations that can deploy power, cooling, networking and server racks quickly, particularly where cloud providers face long construction timelines.
Older chips are retaining utility
A normal technology cycle would suggest that older GPUs should become steadily cheaper as newer generations arrive. The source material argues that this has not reliably happened in AI infrastructure, where older units have held value and leasing costs have risen.
The explanation lies partly in the different components of an AI system. Raw computing performance tends to depreciate as newer processors become available. Memory can become a bottleneck in its own right, particularly for large models that need substantial capacity simply to remain loaded and responsive.
Kimi K3, cited as an example in the material, requires roughly 1.5 terabytes to 2 terabytes of memory to operate. Models of that size can make older hardware commercially useful if it has enough memory or can be combined into a viable cluster. The result is a replacement cycle shaped by memory availability, power access and workload requirements rather than processor age alone.
Open-source models could intensify this effect. Freely available models lower barriers for developers and smaller companies to deploy AI products, but they do not eliminate the infrastructure needed to serve users at scale. A model that can be downloaded by anyone may still require expensive memory, GPUs and networking to run reliably in production.
Power limits are becoming part of the supply equation
The International Energy Agency expects global data-center electricity consumption to double to around 945 terawatt-hours by 2030. The source material estimates that facilities dedicated to machine learning could consume about 200 terawatt-hours in 2026.
Those projections place electricity supply and grid connections alongside chip fabrication as constraints on AI expansion. New data centers need more than servers: they require large and dependable power contracts, cooling systems, transmission access, construction capacity and local regulatory approvals. Regions able to provide those resources can gain an advantage even if they are distant from traditional technology hubs.
The article identifies North American power availability as close to its limit for near-term buildouts and points to Canada and Australia as potential expansion locations. Both countries offer areas with substantial energy resources, though project economics would vary by province, state, grid connection and the type of generation available.
This shortage of ready-to-use infrastructure is also relevant to cryptocurrency mining operators. Large mining facilities already manage power-intensive equipment, industrial cooling and data-center operations. Some operators have sought to adapt or expand those sites for high-performance computing and AI hosting, although a mining warehouse is not automatically suitable for AI workloads. High-performance clusters generally require more sophisticated networking, redundancy, cooling design and service-level commitments than many conventional mining operations.
Crypto markets face a more selective infrastructure trade
The AI buildout offers a potential new narrative for publicly traded mining companies and for decentralized computing networks, but it does not justify treating every token linked to “distributed compute” as an equivalent beneficiary. A peer-to-peer network of home computers may be useful for smaller or flexible workloads, while high-volume inference for commercial customers often requires predictable uptime, data controls and specialized hardware.
CoinShares is cited in the source material as projecting that high-performance hosting could account for 70% of public mining-company revenue by the end of 2026. Whether individual operators achieve that mix will depend on signed contracts, available power, capital spending and their ability to meet the technical requirements of AI customers.
The more durable shift is in how compute infrastructure is valued. AI demand is turning electricity access, cooling capacity and immediately deployable data-center space into commercially scarce assets. Semiconductor shares may continue to reflect fears of a future spending slowdown, but the $2 trillion-plus order backlog described by EpochAI indicates that current demand is being driven by active deployment needs well beyond the largest AI laboratories.
To navigate AI-driven market dislocations, explore tokenized equities as a bridge between chip stocks and on-chain finance.
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.
