The number landed with the weight of a tectonic plate: $1 trillion. Headlines across financial media proclaimed the dawn of the AI build-out, a capital influx so vast it seemed to erase all previous constraints. Yet as I sat in my Copenhagen office, tracing the liquidity flows from sovereign wealth funds to cloud hyperscalers, a dissonant signal emerged. The capital is not the bottleneck. The bottleneck is concrete, copper, and kilowatt-hours. The AI build-out faces a structural challenge that no amount of money can solve: the physical world moves at a speed that defies digital acceleration. This is the $1 trillion paradox—a capital glut meets a physical scarcity, and the market has yet to price the friction.
Over the past twelve years, I have watched cycles of exuberance and pruning in crypto. The 2017 ICO mania, the 2021 DeFi bubble, the 2022 Terra collapse. Each time, capital preceded infrastructure, and the mismatch forced a reset. AI is now walking the same path, but with a degree of magnitude that makes previous cycles look like sandbox experiments. My eye is on the horizon, not the hourly candle. The horizon reveals a landscape where the race to build AI infrastructure is colliding with the immutable laws of physics, and the consequences will reshape not just technology, but the global allocation of capital for a decade.
Context: The Landscape of an Inevitable Collision
To understand the tension, one must first map the terrain. The $1 trillion figure is not a precise accounting; it is a narrative device, a rhetorical spear thrown by investment banks, tech PR departments, and infrastructure funds. In my work modeling digital asset flows, I have learned to treat such round numbers with suspicion. The actual composition—how much is cloud capital expenditure, how much is venture equity, how much is sovereign debt—varies wildly. But the directional signal is undeniable: the world’s largest technology companies, alongside sovereign funds from the Middle East to Singapore, have committed to a multi-year spending spree on AI compute.
This spending targets three layers: training clusters (the enormous GPU farms needed to train frontier models), inference infrastructure (the distributed compute required to serve applications), and the energy backbone (power plants, grid upgrades, and cooling systems). The training layer consumes the most headlines, but it is the inference layer and energy that will determine the outcome. As I wrote in my 2024 brief "The Illusion of Decentralized Yield," capital flows that ignore downstream bottlenecks tend to create bubbles in the upstream assets. We are now seeing that pattern on a global scale.
Core: The Three Hard Constraints
1. Power: The Unforgiving Physics of Megawatts
Let us begin with the most obvious, yet most underestimated, constraint: electricity. A single frontier AI training cluster, such as those operated by OpenAI or Google DeepMind, can draw 100 megawatts or more. To put that in perspective, 100 MW is the baseload consumption of a mid-sized city of 50,000 homes. When you scale to 30 such clusters worldwide, you are adding the equivalent of a new metropolitan area’s energy demand every year. The problem is not the cost of electricity; it is the availability of grid capacity.
In Virginia’s Loudoun County, the epicenter of global internet traffic, data center power requests have overwhelmed Dominion Energy’s grid, leading to interconnection queues that stretch four to seven years. Similar stories emerge from Silicon Valley, Singapore, and Frankfurt. The utilities are not designed for this pace of demand growth. Even if the funds are allocated today, the physical construction of new power plants, transmission lines, and substations takes five to ten years. This creates a temporal mismatch between capital deployment and operational reality.
During the 2022 bear market, I retreated to a cabin in Jutland and studied the energy constraints of proof-of-work mining. The same principle applies here: the price of compute is ultimately set by the price of power. AI inference costs are falling rapidly, but the underlying energy cost floor is rising. The bust was not an end, but a necessary pruning. For AI, the pruning will come when the first wave of hyperscale data centers faces power curtailments or skyrocketing electricity prices.
2. Chip Supply: The CoWoS Chokepoint
The second constraint is semiconductor manufacturing, specifically the advanced packaging known as CoWoS (Chip-on-Wafer-on-Substrate). NVIDIA’s H100 and B200 GPUs require this technology to stack high-bandwidth memory close to the processor. CoWoS capacity is limited by physical realities: the equipment is expensive, the process is delicate, and the expansion takes years. In 2023, TSMC’s CoWoS capacity was fully booked through 2025, and while the company is building new factories, the output cannot keep pace with the exponential demand from AI.
This bottleneck echoes the NFT craze of 2021, when minting capacity on Ethereum was constrained by block space, leading to gas wars and congestion. The difference is that NFTs were a speculative layer on top of existing infrastructure; AI chips are the infrastructure itself. When capacity is constrained, the price of compute rises, and only the highest-value users (frontier labs, big tech) can afford it. This squeezes out smaller players and startups, concentrating AI power in fewer hands. Based on my audit experience of decentralized GPU networks, I have seen that the secondary market for compute is inefficient and fragmented. The capital flowing into AI is not solving the supply chain; it is bidding up the price of a fixed resource.
3. Data Center Construction: The Long Lead Time
The third constraint is the most mundane yet the most binding: the physical construction of data centers. Building a hyperscale facility involves land acquisition, environmental impact assessments, water rights, local permits, and construction labor. Each of these steps is subject to local regulations, community opposition, and supply chain delays for materials like concrete and steel. A typical timeline from planning to operational is 18 to 30 months. In the current environment, where demand is doubling every six months, this lead time creates a structural shortage.
Liquid cooling, once a niche requirement, is now essential for next-generation GPUs with thermal design power exceeding 1000 watts. Retrofitting existing data centers is expensive and slow. New builds must be designed from the ground up for liquid cooling, which adds complexity and cost. The industry is effectively rebuilding the entire global data center fleet in a decade, a task that has no precedent in scale or speed.
Contrarian: The Decoupling Thesis That Nobody Wants to Hear
Here is where my analysis diverges from the consensus. The prevailing narrative is that $1 trillion will solve these problems through sheer financial force. But I argue the opposite: the capital influx is making the problem worse. How? By creating a “crowding out” effect. When the largest players commit to massive spending, they lock in long-term contracts for power, chips, and data center capacity, driving up prices for everyone else. This is not a market clearing; it is a resource hoarding. The efficient allocation of capital typically resolves scarcity, but only when supply is elastic. In AI infrastructure, supply is inelastic in the short to medium term.
Furthermore, the narrative of $1 trillion itself distorts decision-making. It encourages a “spend now, figure out later” mentality that ignores the unit economics. The infrastructure built today is based on the Scalability Laws of Transformer architectures. If a more efficient architecture emerges, such as state-space models or linear attention, much of today’s infrastructure could become stranded assets. The capital is being poured into a specific technological bet, and the bet may not pay off.
I recall the 2021 DeFi boom, where protocols competed for total value locked by offering unsustainable yields. The result was a liquidity fragmentation that served no real user. Similarly, the AI build-out is creating a fragmentation of compute resources, each optimized for a specific workload, but the underlying demand for inference is still nascent. The decoupling thesis that I hold is this: the market is overestimating the speed of AI adoption and underestimating the inertia of physical infrastructure. The bust, when it comes, will not be an end, but a necessary pruning of overbuilt capacity.
Takeaway: Positioning for the Pruning
So where does that leave us, the macro watchers, the capital allocators, the digital asset managers? The signals are clear: the next three to five years will be a period of trial by fire for AI infrastructure. The winners will not be those who spent the most, but those who managed the physics of the physical world—securing long-term power contracts, investing in energy-efficient designs, and building modular, adaptable infrastructure.
For the crypto ecosystem, this creates a unique opportunity. The energy constraints and chip shortages are driving interest in decentralized physical infrastructure networks (DePIN), which can aggregate idle compute and energy resources. While still nascent, these networks offer a hedge against centralized bottlenecks. Moreover, the regulatory clarity emerging in the EU, particularly around MiCA, provides a framework for tokenized infrastructure assets. I have been working with a small collective to explore how blockchain immutability can verify the provenance of AI-generated content, but the deeper play is in the energy and compute markets.
My advice is contrarian: reduce exposure to pure-play AI compute providers and increase allocations to energy infrastructure, grid modernization, and efficiency technologies. The market is pricing the boom, but not the bust. When the first major write-down of AI infrastructure occurs, the shock will ripple through equities, credit, and crypto. Those who understood the paradox of $1 trillion will be ready.
Silence screams louder than pumps. The market is digesting the reality of physical constraints. Watch the power grid, not the pitch deck. The horizon is clear: the pruning is coming, and it will separate the signal from the noise.
