The demand for AI is real – but the boom is burning through capital, electricity, and strategic security faster than its returns can be proven

The hidden bottleneck of the AI ​​boom: Why electricity and water are suddenly more important than new chips

China's unexpected response in the chip war: Why the West's strategy could backfire

The hype surrounding artificial intelligence has sent the global semiconductor industry into an unprecedented trillion-dollar frenzy. It's no longer just about simple supply chains of software and hardware – AI and microchips are rapidly merging to form the new nervous system of the global economy. While tech giants are pumping astronomical sums into gigantic data centers and specialized hardware in a relentless arms race, the challenges are growing in the background. This unprecedented boom is burning through capital, electricity, water, and strategic security faster than these massive investments can currently pay off. At the same time, the industry's bottlenecks are shifting: High-bandwidth memory, advanced packaging, and the cutthroat fight for every megawatt of power will determine future success or failure. Added to this are sharp geopolitical tensions that are increasingly fragmenting the global chip market. Is the economy heading towards a golden age of AI, or are we witnessing the riskiest gamble in technological history, one that is about to face a painful correction? This in-depth analysis reveals how the AI ​​semiconductor complex truly works, where the real bottlenecks lie – and who ultimately benefits from these strategic constraints.

AI and semiconductors in a trillion-dollar rush: The new nervous system of the global economy: When computing power becomes a factor of production

The connection between artificial intelligence and semiconductors is far more than a typical software-to-hardware supply relationship. It is evolving into an industrial complex where data centers, chip designers, contract manufacturers, memory producers, network equipment providers, power suppliers, and cloud platforms are interdependent. Artificial intelligence requires highly specialized computing power, while the semiconductor industry increasingly bases its investment decisions on the anticipated development of AI models and their applications. This creates a self-reinforcing cycle: More powerful chips enable larger and better models, these models generate new applications and increased usage, and the growing demand, in turn, justifies new factories, data centers, and power connections.

This cycle is economically significant because computing power is evolving from a technical tool into an independent factor of production. Companies are no longer simply purchasing software licenses or server capacity, but are securing access to a scarce resource that can determine the speed of innovation, product quality, and competitiveness. In some sectors, AI computing power is already comparable to electricity, capital, or skilled labor: without sufficient access, companies cannot develop certain products, automate processes, or utilize large datasets economically. The price of computing power is therefore increasingly influencing the cost structure of digital services and, indirectly, the productivity of traditional industries.

However, this boom is not simply a normal surge in demand. It stems from a combination of real-world usage, strategic stockpiling, anticipated future applications, and intense competition among major platform companies. Part of the investment is financed by current revenues, while another part represents a bet on markets that are not yet fully developed. This very mix makes the AI-semiconductor complex both exceptionally dynamic and prone to miscalculations. The crucial question, therefore, is not whether artificial intelligence will require more semiconductors in the long run. That is highly likely. Rather, the question is whether the capacities being built up today can be utilized at the right time, in the right place, and with a sufficiently high return.

A semiconductor market of a new magnitude

The global semiconductor market will have reached a size in 2026 that was considered a long-term goal only a short time before. Current industry forecasts place annual revenue well above US$1.5 trillion. Compared to 2025, this represents an exceptionally high increase, depending on the definition and availability of data. However, this jump is not solely due to a correspondingly strong increase in the number of chips produced. A significant portion results from rising memory prices, a higher-value product mix, and the sale of extremely expensive AI systems. Revenue growth and actual volume growth should therefore not be equated.

The shift in favor of the memory business is particularly striking. The memory segment is projected to generate revenue of more than $800 billion by 2026, representing growth of approximately 250 percent. This would mean that memory accounts for more than half of the total market growth. This development is closely linked to high-bandwidth memory, but extends beyond it. Traditional DRAM and NAND products are also benefiting from higher prices, limited manufacturing capacity, and the expansion of data centers. The term "memory inflation" aptly describes the situation: not only are technologically advanced products becoming more expensive, but price increases are spreading across the entire memory market.

Logic chips are also experiencing strong growth. Graphics processors, custom AI accelerators, network processors, switch chips, and central processing units together form the foundation of modern AI clusters. Their value is not solely determined by the number of transistors. Crucial factors include the interplay of computing power, memory bandwidth, data transmission, software, and cooling, as well as the ability to reliably connect thousands of accelerators into a complete system. The market is therefore shifting from evaluating individual chips to evaluating complete computing platforms. Those who only consider the price of a processor underestimate the true economic unit: the fully usable AI system.

However, these high growth rates cannot be extrapolated linearly. A market can nearly double in size within a year due to an exceptional price cycle and a one-off investment surge, without being able to maintain this pace in the long term. As soon as new manufacturing capacities become available, memory prices ease, or hyperscalers smooth out their investment programs, growth will normalize. This would not necessarily be the end of the AI ​​boom, but could mark the transition from an extreme development phase to a more mature investment cycle.

HBM is becoming a key strategic resource

High Bandwidth Memory (HBM) has become the most visible example of how artificial intelligence is shifting value creation within the semiconductor industry. Large models move enormous amounts of data between memory and processing units during training and inference. If the processor can perform calculations faster than the data is being provided, some of its theoretical performance remains unused. The economic bottleneck then lies not in a lack of calculations, but in memory bandwidth. HBM addresses this problem through vertically stacked memory chips, very wide interfaces, and a tight connection to the accelerator.

Manufacturing is complex. Multiple memory layers must be stacked, interconnected, and tested with high precision. Even minor defects can reduce the yield, while the tight integration places high demands on heat dissipation, packaging, and quality assurance. This limits the number of suitable suppliers and increases the market power of the leading manufacturers. SK Hynix was able to expand its strong position early on, while Samsung and Micron are investing significant resources to gain additional market share in HBM3E and HBM4. While this competition offers customers some diversification, supply remains tight in the short term.

Economically, HBM is changing the traditional landscape of the storage industry. This industry has long been characterized by strong price fluctuations, relatively interchangeable products, and recurring overcapacities. With HBM, long-term purchase agreements, joint product development, and early capacity reservation are taking center stage. Storage is thus transforming from a cyclical component into a strategic element of system architecture. Suppliers can achieve higher margins, but at the same time, they must invest significantly more capital and bear the risk that technological generations will become obsolete more quickly or that a major customer will change its architecture.

The critical weakness lies in the simultaneous occurrence of multiple bottlenecks. A finished AI accelerator requires a suitable wafer process, sufficient quantities of HBM, an available advanced packaging method, substrates, and high-performance testing capabilities. Simply increasing front-end production will not solve the problem. Companies must coordinate all components in terms of timing. This interconnectedness increases the value of sound supply chain planning and favors large customers who can reserve capacity years in advance. Smaller providers and independent cloud operators, on the other hand, are more likely to fall behind.

Packaging beats the individual transistor

For decades, miniaturizing transistors was considered the primary source of increased performance and reduced unit costs. This model still works, but it is becoming increasingly expensive. Modern manufacturing processes require highly complex lithography, extremely pure materials, precise process control, and factories with investment volumes in the tens of billions. At the same time, the economic benefits of a smaller node no longer increase automatically to the same extent as before. The cost per transistor is not decreasing reliably, and the development of large monolithic chips is reaching its limits in terms of yield, power consumption, and manufacturability.

This is why advanced packaging is gaining importance. Several specialized dies (chip components) are combined in a single package to form a high-performance system. Processing units, input/output chips, and memory can originate from different processes and be manufactured where cost and performance are best aligned. Chiplets enable more flexible product development, reduce certain risks associated with large monolithic designs, and, under favorable conditions, shorten time to market. Competition is thus shifting from the pure production of the smallest structures to the mastery of heterogeneous integration.

CoWoS and related processes are therefore no longer downstream, secondary technologies, but strategic capabilities. In practice, a customer may have secured wafer starts but still be unable to deliver finished accelerators if packaging slots or suitable memory stacks are lacking. This situation explains why contract manufacturers and specialized packaging companies are aggressively expanding their capacities. It also explains why established suppliers with experience in testing, substrates, and assembly are gaining new growth opportunities, even though they are not at the forefront of transistor miniaturization.

In the long term, system performance is expected to increase significantly through three-dimensional integration, faster connections, and improved heat dissipation. Silicon photonics can help transfer data more energy-efficiently between computing units and storage. Direct copper connections and new interposer techniques reduce latency. Consequently, economic value shifts to areas that have often been grouped together as the backend. This is important for investors and industrial policymakers: those who focus solely on new wafer fabrications overlook a significant portion of the strategic bottlenecks.

From model training to inference economics

The first phase of the AI ​​boom was primarily characterized by the training of increasingly larger models. Training is visible, capital-intensive, and technologically prestigious. It requires large clusters connected via high-speed networks, extensive datasets, and specialized teams. The next phase, however, will be more strongly driven by inference—that is, the actual application of trained models. Every answered query, every automated translation, every image analysis, and every control decision made by a robot generates computational effort. With millions or billions of such operations daily, the cumulative inference load can significantly exceed the training requirements.

This shift changes the economic viability. In training, the focus is often on completing a model as quickly as possible. High hardware costs can be accepted if an early market launch offers strategic advantages. In inference, however, the price per usable output is crucial. Companies must optimize latency, energy consumption, memory usage, and utilization. A technically powerful model can be economically unattractive if each query consumes too much computing time or only functions reliably with expensive hardware.

This is driving growth in the market for specialized accelerators. Cloud providers are developing their own chips tailored to their data centers, software, and typical workloads. Such application-specific circuits can be more cost-effective and energy-efficient than general-purpose graphics processors when used in stable, high-volume environments. At the same time, they reduce dependence on a single dominant supplier. However, this advantage comes at a price: Developing custom chips requires significant development costs, long-term planning, suitable software tools, and sufficiently large production volumes. For smaller providers, purchasing standardized platforms usually remains more economical.

Model architectures are also increasingly driven by cost considerations. Quantization, sparsely activated models, smaller specialized models, caching, and intelligent selection between different models reduce computational requirements. The market is therefore unlikely to develop solely in the direction of ever-larger systems. A division of labor is more probable: particularly powerful models will handle complex tasks, while smaller and more affordable models will manage the majority of recurring processes. For the semiconductor industry, this broadens demand but simultaneously increases price pressure in the inference business.

Nvidia's moat is software

Nvidia's exceptional position cannot be explained solely by the performance of individual graphics processors. Over many years, the company has built a comprehensive ecosystem of programming interfaces, libraries, development tools, reference systems, networking technology, and support. CUDA has become the de facto standard for numerous AI applications. Developers, research institutions, and companies have invested considerable time in this environment. Switching to competing hardware therefore incurs not only procurement costs but also adaptation efforts, technical risks, and productivity losses.

This software-based competitive advantage strengthens pricing power. Customers aren't just buying a chip, but the prospect of running existing code, finding skilled personnel, and quickly deploying new models. Nvidia is expanding this position with complete systems, high-speed interconnects, and standardized data center architectures. The more control a vendor has over the entire platform, the more difficult a direct price comparison at the component level becomes. A more expensive accelerator can be economically superior if deployment is faster or utilization is higher.

Nvidia's dominance is not unassailable, however. AMD is improving its hardware and software, while hyperscalers are developing their own accelerators. Broadcom and other specialists benefit from custom designs. In inference, the advantage of universal platforms may be less pronounced than in training. Furthermore, large customers will try to establish at least a second source of supply to mitigate price and delivery risks. Even if Nvidia gradually loses market share, the company can continue to grow strongly as long as the overall market expands faster than its share declines.

The greatest strategic risk, therefore, does not necessarily stem from a single competitor. A more dangerous scenario would be a combination of more efficient models, slower capacity expansion by hyperscalers, falling prices for inference, and the successful adoption of alternative software stacks. This could lead to a decline in willingness to pay for top-of-the-line hardware. Conversely, every new application that is initially optimized on the established platform strengthens the network effect. Competition thus hinges on the inertia of a powerful ecosystem and the economic incentive for large customers to reduce this dependency.

The hyperscalers' trillion-dollar bet

Amazon, Microsoft, Alphabet, Meta, and other major platform companies are increasing their capital expenditures at a rate that surpasses even previous technology waves. Estimates for the major hyperscalers combined, depending on the definition used, range into the high hundreds of billions of dollars for 2026. Including additional cloud providers, energy infrastructure, and subsequent years, this investment wave quickly reaches several trillion US dollars. The capital is flowing into accelerators, servers, networks, land, buildings, cooling, power generation, and grid connections.

From a business perspective, this initiative is rational as long as three conditions are met. First, the demand for AI services must increase quickly enough. Second, the platforms must be able to monetize a significant portion of the resulting value creation. Third, the technical lifespan of the infrastructure must not be significantly shorter than assumed in the depreciation models. This third point, in particular, is often underestimated. Data centers and power connections are long-lasting, while accelerators can rapidly lose relative competitiveness with new generations. A system can continue to function technically but still be economically obsolete.

Competition exacerbates the problem. No major platform provider can afford to significantly slow its expansion as long as competitors continue to invest. Those who cut capacity today could lose customers, developers, and data tomorrow. This creates a strategic arms race in which individually understandable decisions can collectively lead to overinvestment. The return on investment for individual companies depends not only on overall AI demand but also on how much parallel capacity competitors build and how much cloud computing prices fall.

In the short term, high order backlogs and component shortages are supporting expansion. In the medium term, however, the capital market will focus more on free cash flow, capacity utilization, and concrete AI revenues. Companies cannot justify investments solely on strategic necessity in the long run. As soon as revenue growth lags behind depreciation, energy costs, and financing expenses, pressure on margins increases. The AI ​​boom will then be measured not by technical capabilities, but by whether computing power translates into profitable customer value.

Why a correction doesn't have to mean a collapse

The discussion about an AI bubble is often too binary. The technology is either seen as the start of a decades-long productivity boom or as speculative overreaction. However, historical infrastructure cycles show that both are possible simultaneously. Railroads, electricity, telecommunications, and the internet fundamentally transformed the economy, even though investors suffered significant losses and numerous companies failed at certain stages. Overinvestment can accelerate the spread of a technology because excess capacity later becomes available at lower prices.

For semiconductor companies, the risk lies in the high operational leverage. New factories incur enormous fixed costs. If demand or prices fall, facilities cannot be easily adapted. The memory industry is particularly familiar with such cycles: scarcity leads to high prices and investments, additional capacity later meets weaker demand, and prices fall disproportionately. In AI components, the cycle is mitigated, but not eliminated, by long-term contracts and technological differentiation.

A correction could take various forms. Hyperscalers could spread out orders, use older accelerators for longer, or rely more heavily on their own chips. Model providers could make their systems more efficient, thereby requiring less computing power per request. Companies could end pilot projects if measurable productivity gains fail to materialize. At the same time, new applications could absorb the freed-up capacity. Therefore, the crucial factor is not solely whether individual investment plans are cut, but whether the overall demand for usable AI results continues to rise.

The most likely scenario is neither an uninterrupted exponential boom nor a complete collapse. A more plausible outcome is an uneven cycle with bottlenecks, price spikes, short-term overcapacity, and subsequent new waves of investment. Winners will be companies that maintain flexible costs, serve diverse customer groups, and are not dependent on a single product cycle. Particularly vulnerable are suppliers whose valuation relies on consistently exceptional growth rates, despite their business facing increasing competition and technological substitution.

A new dimension of digital transformation with 'Managed AI' (Artificial Intelligence) – Platform & B2B solution | Xpert Consulting - Image: Xpert.Digital

Here you will learn how your company can implement customized AI solutions quickly, securely and without high entry barriers.

A managed AI platform is your all-inclusive, worry-free solution for artificial intelligence. Instead of dealing with complex technology, expensive infrastructure, and lengthy development processes, you receive a ready-made solution tailored to your needs from a specialized partner – often within just a few days.

The key advantages at a glance:

⚡ Rapid implementation: From idea to ready-to-use application in days, not months. We deliver practical solutions that create immediate added value.

🔒 Maximum data security: Your sensitive data stays with you. We guarantee secure and compliant processing without sharing data with third parties.

💸 No financial risk: You only pay for results. High upfront investments in hardware, software, or personnel are completely eliminated.

🎯 Focus on your core business: Concentrate on what you do best. We take care of the entire technical implementation, operation, and maintenance of your AI solution.

📈 Future-proof & scalable: Your AI grows with you. We ensure continuous optimization and scalability, and flexibly adapt the models to new requirements.

More information here:

• The Managed AI Solution - Industrial AI Services: The Key to Competitiveness in the Services, Industry and Mechanical Engineering Sectors

Taiwan remains the vulnerable center

State-of-the-art semiconductor manufacturing is highly geographically concentrated. TSMC produces a very large proportion of the world's most powerful logic chips and is virtually indispensable for leading AI accelerators. Taiwan combines a highly skilled workforce, specialized suppliers, decades of process expertise, and an exceptionally dense industrial infrastructure. This ecosystem cannot be replicated by building a single plant. Even with substantial subsidies, new sites require time to achieve comparable yields, costs, and supplier networks.

From a business perspective, this concentration has long been efficient. Large quantities at a specialized location reduce costs, accelerate learning curves, and facilitate exchange between developers, equipment manufacturers, and material suppliers. From a geopolitical perspective, however, it is problematic. A military crisis, blockade, natural disaster, or severe disruption to energy and water supplies could affect key sectors of the global economy. The repercussions would extend far beyond technology companies because vehicles, industrial plants, telecommunications, and defense systems also rely on semiconductors.

The US, Japan, and Europe are therefore promoting new production capacities. TSMC, Samsung, Intel, and other companies are building or planning factories outside their traditional locations. This diversification increases resilience but also incurs additional costs. Construction costs, wages, permitting processes, the availability of skilled workers, and the distance to suppliers vary considerably. A politically desired factory may be strategically advantageous without necessarily being the most cost-effective solution in a purely commercial comparison.

Resilience thus has the character of insurance. It costs money during normal operations, but limits potentially catastrophic losses. The political task is to determine the appropriate scope of this insurance. Complete self-sufficiency would be extremely expensive and hardly achievable with a globally interconnected supply chain. More sensible are multiple qualified production sites, strategic inventories, transparent dependencies, and international agreements for crisis situations. Diversification should not be confused with national self-sufficiency.

The chip war is changing market logic

Semiconductors have become a commodity, a security resource, and an instrument of geopolitical power. The United States is attempting to control China's access to high-performance AI chips and certain manufacturing facilities. In 2026, the regulatory framework was adjusted again: Selected accelerators are subject to case-by-case reviews under strict conditions, while other transactions remain highly restricted. Technical thresholds, end-use controls, quantity limits, and testing make export decisions a complex component of corporate strategy.

This policy aims to limit militarily and strategically relevant capabilities, but it also produces side effects. Suppliers lose potential revenue, customers accelerate the search for alternatives, and supply chains become more complex. Controls can slow a competitor's technological progress, but at the same time create strong incentives for domestic development. The longer restrictions remain in place, the more attractive domestic designs, alternative software, and more efficient model architectures become. Technological isolation can therefore be effective in the short term and ambivalent in the long term.

For international companies, a world of diverse product lines and compliance requirements is emerging. Chips are technically adapted for specific markets, data centers are evaluated according to ownership structure and user base, and even cloud access can become relevant under export law. This fragmentation increases transaction costs. It hinders economies of scale and forces companies to consider geopolitical scenarios in procurement, research, and site selection.

A complete decoupling between China and the West remains unlikely. China is a large sales market, an important production location, and a significant supplier of mature semiconductors, electronic components, and materials. Western companies, for their part, hold key positions in design software, manufacturing equipment, specialty chemicals, and high-performance architectures. Selective de-risking is more probable: particularly sensitive areas will be separated, while a large portion of commercial trade will continue. This intermediate form is less economically destructive than total decoupling but remains politically unstable.

China's answer: Broad, not just cutting-edge technology

China is pursuing a comprehensive strategy of technological self-sufficiency. The focus is not solely on a direct leap to the world's most advanced process generation. Investments are also being made in mature manufacturing hubs, memory, chip design, packaging, power electronics, materials, plant engineering, and industrial applications. This broad approach makes economic sense. A country can significantly reduce its vulnerability and gain large market shares without being a technological leader in every single product category.

Mature nodes offer significant economies of scale. Vehicles, household appliances, industrial electronics, sensors, and energy systems do not exclusively require two- or three-nanometer chips. Large capacities in these segments can drive down prices and put pressure on foreign competitors. At the same time, they generate revenue, manufacturing expertise, and an industrial foundation for more sophisticated technologies. Western policymakers must therefore distinguish between two risks: dependence on Chinese standard chips and the competition for technological leadership.

In the field of AI, limitations in hardware access can lead to greater efficiency. Developers optimize models, memory requirements, and communication overhead when computing power is scarcer. The West should therefore not assume that more available high-end hardware automatically guarantees a lasting advantage. The crucial factor is the ability to productively combine hardware, algorithms, data, energy, and applications. A less powerful chip can be economically competitive within a better-optimized overall system.

At the same time, China's challenges remain considerable. Modern lithography, high-precision manufacturing equipment, advanced materials, and reliable software tools require decades of learning curves. Government support can provide capital, but it cannot accelerate technological competence indefinitely. Overinvestment, duplication of effort, and weak capital discipline are real risks. China's strategy should therefore neither be underestimated nor considered a guaranteed success. Its greatest strength lies in its large domestic market and the integration of manufacturing, applications, and long-term government planning.

Europe's strength does not lie in copying Asia

Europe holds key positions in semiconductor equipment, power electronics, sensors, automotive chips, research, and specialized materials. At the same time, the continent is heavily dependent on cutting-edge logic manufacturing, cloud platforms, and large-scale AI accelerators. The European Chips Act aims to mobilize investment, strengthen research, and increase Europe's share of global production. While the strategic direction is understandable, the frequently cited goal of a significantly higher global market share must not become an end in itself.

Simply copying Asian foundry models would be expensive and could fail due to a lack of major customers. State-of-the-art factories require consistently high utilization. Europe currently has fewer domestic customers for high-performance processors than the US or Asia. Therefore, government support policies must consider demand, design expertise, packaging, skilled labor, and energy supply together. A subsidized factory without a suitable customer ecosystem cannot create sustainable sovereignty.

Europe's best opportunities lie in application-oriented sectors. Industrial automation, robotics, vehicles, medical technology, energy supply, and secure edge systems combine existing strengths with growing AI demand. Reliability, long product lifecycles, functional safety, and integration into physical processes are crucial here. European companies don't need to manufacture every training accelerator themselves to secure a significant share of the value chain. They can be leaders in sensors, control systems, power electronics, industrial data rooms, and domain-specific AI systems.

To achieve this, Europe needs faster approvals, competitive energy prices, more venture capital, and better-coordinated procurement. Regulation should foster trust and market access without unnecessarily increasing development costs. Linking research and scaling is particularly important. Europe produces many scientific results but often loses added value when companies relocate to other regions for growth. Semiconductor policy must therefore be part of a broader strategy encompassing data centers, cloud infrastructure, energy, and industrial digitalization.

Electricity is becoming the most critical location factor

The expansion of AI data centers is changing the competition for energy. Global data center electricity consumption was around 415 terawatt-hours in 2024 and could rise to approximately 945 terawatt-hours by 2030. AI-optimized data centers are driving a large portion of this growth. Globally, this share still represents only a limited portion of total electricity consumption. However, locally, new facilities can have a significant impact because they are concentrated in a few regions and require consistently high power output.

The bottleneck, therefore, lies less in the globally available amount of energy than in grids, substations, transformers, permits, and guaranteed capacity at a specific location. A data center can be planned more quickly than the necessary grid infrastructure. Waiting times for connections thus become a key competitive factor. Locations with affordable space but weak grids become less attractive. Regions with reliable energy, predictable prices, and short permitting processes, on the other hand, can attract significant investment.

Hyperscalers are responding with long-term power purchase agreements, investments in renewable energy, and a growing interest in nuclear power. Renewables offer low variable costs but require storage, grids, or supplementary secure generation. Nuclear power can provide continuous output but has long planning times and high regulatory requirements. Natural gas remains a flexible transitional solution in some markets but increases emissions and price risks. Therefore, there is no single energy source that solves the problem everywhere.

Energy efficiency is becoming a key economic differentiator. What matters is not just the performance of a chip, but the number of usable results per kilowatt-hour and per euro invested. Liquid cooling, higher operating temperatures, better utilization, and more efficient models can reduce overall consumption. However, the so-called rebound effect remains relevant: if each request becomes cheaper, usage often increases so significantly that absolute electricity consumption still rises. Efficiency is therefore necessary, but it does not replace forward-looking infrastructure planning.

Water, heat, and acceptance become relevant in terms of the balance sheet

Besides electricity, many data centers require water for cooling or indirectly for power generation. Actual consumption depends heavily on climate, cooling technology, capacity utilization, and the energy mix. Therefore, making general statements about the water consumption of a single AI request is unreliable. Nevertheless, the question is important at the site level. In arid regions, an additional large data center can compete with agriculture, households, and other industries for scarce resources.

Waste heat is also becoming more economically relevant. In densely populated regions, it can be fed into district heating networks, provided that temperature, distance, and demand are compatible. Such projects improve overall energy efficiency but require additional investment and long-term cooperation with municipalities and energy suppliers. Waste heat utilization is therefore not a universal solution, but it can increase public acceptance in suitable locations.

For operators, environmental impacts are increasingly becoming financial variables. Delayed permits, local opposition, grid fees, and regulations all affect returns. Companies that provide transparent information about electricity, water, and land requirements early on can reduce project risks. Conversely, communication that emphasizes only global innovation promises and ignores local costs leads to a loss of trust. The societal license to operate thus becomes an asset, even if it doesn't appear on traditional balance sheets.

Physical AI is broadening the semiconductor boom

The next phase of growth is likely to see artificial intelligence more closely integrated with the physical world. Robots, autonomous vehicles, logistics systems, machines, drones, and intelligent sensors require local perception, rapid decision-making, and robust control. This physical AI places different demands on the physical world than a centralized language model in the data center. Latency, energy consumption, functional safety, cost, and resilience to heat, dust, or vibrations are becoming increasingly important.

This broadens the demand for the semiconductor industry. Not only are high-performance accelerators needed, but also camera sensors, microcontrollers, power semiconductors, communication chips, memory, and energy-efficient edge processors. Production volumes can be very high, while the value per unit is lower than that of a data center accelerator. This benefits different suppliers and regions than those involved in training large models. Japan, Europe, South Korea, Taiwan, China, and the USA each possess different strengths along this value chain.

The economic breakthrough depends less on spectacular demonstrations than on total cost of ownership. A robot must justify its purchase, maintenance, energy consumption, and risk of failure through measurable productivity. This is easier in structured environments like warehouses and factories than in open, unpredictable situations. Therefore, physical AI is likely to scale first where tasks are repeatable, safety zones are controllable, and personnel costs are high. Logistics and intralogistics are thus among the most plausible early markets.

Physical AI can simultaneously increase the demand for centralized computing power. Devices generate data, receive model updates, and utilize cloud resources for complex tasks. Edge and cloud are therefore not opposites, but complementary layers. Some processing takes place locally, while training, fleet optimization, and comprehensive analyses are performed centrally. This hybrid architecture creates a broader and potentially more stable semiconductor market than a boom dependent solely on a few training clusters.

Regulation becomes a competitive parameter

Regulation impacts the AI-semiconductor complex on multiple levels. Export controls determine market access. State aid rules guide factory investments. Security standards influence which models and systems companies are permitted to use. Energy and environmental regulations determine data center locations. Data protection, liability, and copyright affect the demand for AI services and thus indirectly impact hardware requirements.

Smart regulation can facilitate investment because it stabilizes expectations and builds trust. Uniform testing procedures, clear liability rules, and internationally compatible standards reduce legal uncertainty. Particularly in industrial and safety-critical applications, reliable regulation can be a competitive advantage. Customers are more likely to invest when they know that systems are compliant, testable, and insurable in the long term.

Poorly designed regulations, on the other hand, can generate high fixed costs and favor established large companies. If only a few providers can meet extensive documentation and testing requirements, competition decreases. Technology-neutral requirements are therefore usually more sensible than detailed specifications for individual architectures. Regulation should address measurable risks and differentiate between general office software, industrial control systems, and safety-critical systems.

The debate about slowing the pace of development cannot be resolved through regulation alone. A unilateral halt would be difficult to enforce internationally and could simply shift development to less transparent environments. More sensible approaches include tiered safety assessments, independent evaluations of particularly high-performing models, and clear lines of responsibility. Economically, the issue is the internalization of risks: companies should not reap the private benefits of rapid development while any potential societal damage falls entirely on third parties.

The winners are at the bottlenecks

In an investment boom, the companies that benefit most are those whose products are scarce and difficult to replace. These include leading chip designers, contract manufacturers, HBM providers, packaging specialists, network equipment suppliers, and manufacturers of certain production facilities. Energy and cooling technology, transformers, and data center infrastructure are also gaining in importance. The highest margins arise where technological barriers to entry, long qualification times, and limited capacities converge.

These positions, however, are not guaranteed indefinitely. High margins attract capital and competition. Customers develop alternatives, standardize interfaces, or vertically integrate value creation. Hyperscalers, in particular, have sufficient volume to make their own chips economically viable. Suppliers must therefore invest their scarcity profits in research, customer retention, and the next generation of products. Those who merely monetize the current bottleneck without anticipating future ones risk an abrupt loss of market share.

For user companies, the advantage is shifting from mere access to meaningful use. In the early stages, securing accelerator access could already be considered a strategic success. In the future, what matters is their utilization rate and the business value generated by the applications running. Companies need data quality, process knowledge, skilled personnel, and organizational adaptability. Without these complementary resources, expensive computing power becomes an unproductive asset.

States, too, should not try to be leaders in everything at once. Successful location policy identifies a country's own strengths and critical dependencies. A country can be strategically important in manufacturing facilities, materials, power electronics, packaging, industrial applications, or energy supply without achieving complete self-sufficiency. The quality of international partnerships will be just as crucial as the size of national subsidies.

The baseline scenario: Boom followed by painful normalization

For the coming years, a baseline scenario is plausible in which demand for AI hardware continues to grow, but the exceptional growth rates of 2026 decline significantly. New capacities in storage, packaging, and state-of-the-art manufacturing will alleviate some bottlenecks. At the same time, the use of inference technology increases, physical AI creates additional markets, and companies integrate AI more deeply into workflows. The market continues to grow, but price increases contribute less to revenue.

In this scenario, companies with weak differentiation or aggressive capacity plans will see corrections. Storage prices will become cyclical again, individual data center projects will be postponed, and not all model providers will be able to cover their costs. The major platforms will remain key investors, but will need to justify their spending more strongly with revenues and free cash flow. For the overall economy, this would not be a crisis, but a necessary selection process.

An optimistic scenario assumes that AI will rapidly generate measurable productivity gains. If companies achieve significantly higher output with the same number of employees, accelerate research, and create new products, demand can absorb the increased capacity. Physical AI would additionally generate large production volumes. Energy efficiency and new power generation would need to grow quickly enough to prevent infrastructure from becoming a limiting factor.

The negative scenario combines several challenges: lower willingness to pay for AI services, more efficient models requiring significantly less hardware, falling storage prices, project delays in power grids, and geopolitical escalation. This could lead to increased depreciation, a collapse in orders, and pressure on highly indebted infrastructure projects. Even in this case, the technology wouldn't disappear. However, the adaptation process would be painful for shareholders, suppliers, and locations with one-sided dependencies.

The boom is real, but the return on investment has not yet been proven

The AI ​​semiconductor complex is neither mere speculation nor a surefire path to ever-increasing profits. Real demand is evident in fully utilized production lines, limited HBM capacity, high data center investments, and growing electricity demand. At the same time, current valuations and expansion plans are based on assumptions about future usage, prices, and productivity. There is a crucial difference between technological relevance and financial return.

The most significant economic shift lies in the fact that the individual chip is no longer the center of value creation. The entire system—comprising logic, memory, packaging, network, software, energy, and application—is now crucial. Bottlenecks migrate along this chain. Today, HBM might be in short supply, tomorrow it could be the network connection, and the day after that, a company's ability to meaningfully integrate AI into its processes. Successful players, therefore, monitor not only their immediate competitors but also the weakest link in the overall system.

Artificially slowing down technological progress would be economically and geopolitically impractical. What's needed is a smarter pace: investments should be more closely tied to demonstrable benefits, security risks systematically assessed, and energy and supply chain issues addressed early on. Efficiency improvements deserve the same priority as larger models. A system that generates the same benefits with fewer chips, less electricity, and less capital is more economically valuable than mere scaling.

The rationale, therefore, is this: The long-term structural trend remains intact, but the current investment phase carries significant risks of over-expansion. There will be no simple transition from boom to sustained prosperity. A series of bottlenecks, corrections, and new applications is more likely. The winners will not automatically be those companies currently spending the most capital. Success will go to those players who most efficiently translate computing power into reliable products, more productive processes, and strategic capability. The semiconductor boom provides the nervous system for the AI ​​economy. However, whether this translates into sustainable value creation will be decided outside the chip: in business models, power grids, factories, institutions, and the ability to transform technological possibilities into tangible benefits.

The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business - Image: Xpert.Digital

Xpert.Digital is a data-driven B2B industry hub led by Konrad Wolfenstein . The company acts as an external, quasi-in-house solution for industrial partners, closing operational gaps in marketing, content, and sales – without requiring additional resources on the client side.

More information here:

• The quasi-in-house solution: How Xpert.Digital closes operational gaps in B2B marketing and sales – Smart Content-Driven Business

☑️ Our business language is English or German

☑️ NEW: Correspondence in your native language!

Konrad Wolfenstein

I and my team are happy to be available to you as your personal advisor.

You can contact me by filling out the contact form here [email protected]:or simply call me at +49 7348 4088 965. My email address is

I'm looking forward to our joint project.

☑️ SME support in strategy, consulting, planning and implementation

☑️ Creation or realignment of the digital strategy and digitization

☑️ Expansion and optimization of international sales processes

☑️ Global & Digital B2B trading platforms

☑️ Pioneer Business Development / Marketing / PR / Trade Fairs