Two gigawatts. That is the implied compute capacity from CPP Investments’ $1.75 billion commitment to EQT’s AI infrastructure strategy. The number looks clean on a pitch deck—enough to power 750,000 H100 GPUs, enough to make the next GPT-6 dream tangible. But I have seen this script before. In 2017, the DAO’s logic held until the oracle blinked. In 2020, Uniswap v2’s TWAP looked solid until a $50,000 flash loan skewed it. Now the oracle is the power grid, and the flash loan is a chip shortage. The numbers add up only if nothing breaks. And something always breaks.
The investment was announced with the usual fanfare: long-term capital, institutional conviction, AI demand as a secular trend. EQT, a European private equity firm, will develop or acquire data centers optimized for AI workloads—high density, liquid cooling, direct connections to hyperscalers. CPP, managing over $600 billion, is treating this as a modest allocation, roughly 0.3% of its portfolio. But modest in size does not mean modest in signal. Every pension fund that dips a toe into AI infrastructure validates the narrative that the current compute paradigm is permanent. That is the assumption worth dissecting.
The Core: A Systematic Teardown of the Assumption
First, the power constraint. A 2GW facility demands continuous electricity that rivals a small city. In regions like Northern Virginia (the world’s largest data center market), Dominion Energy has already warned that new connections face years of delay due to grid capacity. EQT is presumably targeting Nordic or Canadian locations for cheap hydro and political stability, but even there, permitting and transmission lines create friction. The logic held until the oracle blinked—the oracle being the local utility’s transformer. Based on my analysis of energy contracts in DeFi staking, I know that long-term power purchase agreements (PPAs) are the only hedge, but they lock in costs that can become uneconomic if GPU efficiency improves faster than expected. Solidity does not lie, it only omits; the whitepaper for this investment omits the Power Purchase Agreement’s fine print.
Second, the chip dependency. Every H100 GPU requires 700W under load, plus networking and cooling overhead. A 2GW facility can host roughly 500,000 to 750,000 GPUs, depending on the mix of training and inference nodes. That means a single investment is effectively a bet on NVIDIA’s continued dominance—and on the absence of a chip supply shock. In 2021, I audited a DeFi protocol that assumed smooth oracle price feeds; it collapsed when one exchange paused withdrawals. Today, the oracle is TSMC’s fabrication capacity. A single geopolitical event—a Taiwan blockade, a new export control—can freeze the chip supply for a year. Entropy finds its way through the gap; the gap here is the South China Sea.
Third, the technology risk. The entire AI infrastructure thesis rests on the assumption that larger models will always require more compute. But the history of computing is a history of efficiency gains. From mainframes to microprocessors, from Bitcoin ASICs to proof-of-stake, demand does not scale linearly forever. New architectures—state-space models, liquid neural networks, neuromorphic chips—could dramatically reduce the compute required per unit of intelligence. If a startup ships a training algorithm that is 10x more efficient in 2027, the 2GW facility becomes a stranded asset. Ape gold was built on glass foundations; the gold here is the rental income, and the glass is the assumption that Transformer-scaling is the only path forward.
Fourth, the centralization vector. A 2GW data center is a single point of failure—not just physically, but for the entire AI ecosystem that relies on it. I spent weeks in 2022 modeling the Terra-Luna death spiral using differential equations. The same pattern applies here: if one or two hyperscale tenants go bankrupt or consolidate, the facility loses its anchor lease. Pension fund capital, by its nature, chases stable cash flows. But stable cash flows require stable tenants, and the AI industry is anything but stable. Silence in the logs speaks louder than noise; the silence is the absence of a signed 15-year lease from a proven hyperscaler. Without that, the investment is speculation on future demand.
Contrarian: What the Bulls Got Right
To be fair, the bulls have a point. AI demand is real and growing. OpenAI, Google DeepMind, Anthropic—they all need compute, and they have shown willingness to pre-pay for capacity. Microsoft has already committed tens of billions to new data centers. The secular trend is genuine, not a narrative. My experience with DeFi told me that if an investment has a clear revenue model and a long track record, it can survive the volatility. Data centers, unlike algorithmic stablecoins, have a proven business model: wholesale colocation with long-term contracts. In a sideways market, they offer bond-like yields. CPP is not stupid; they are simply taking a calculated risk on a macro trend.
But the error is in assuming that the trend is linear and that current technology will persist. The contrarian insight is that the risk is not in the demand but in the fragility of the supply chain and the pace of innovation. The same engineers who build bigger GPUs are also working on optical interconnects, wafer-scale chips, and analog computing. The infrastructure built today might look like a bet on horse-drawn carriages just as the automobile arrives. Precision is the only shield against chaos, and this investment lacks precision in its technology forecast.
Takeaway: The Code Remembers What the Whitepaper Forgot
The $1.75 billion is not a mistake; it is a delayed consequence of the AI hype cycle. But the code—the actual physics and economics of power, chips, and cooling—remembers what the whitepaper forgot: that scaling laws have physical limits. In 2017, I traced the reentrancy bug in the DAO to a single line of Solidity. In 2025, I trace the fragility of this AI bet to the power substation and the fab. When the next disruption comes—a chip shortage, a regulatory clampdown on energy, a breakthrough in efficiency—the capital will flow out as fast as it flowed in. Entropy always finds its way through the gap. The only question is which gap will break first.