Hook
$1.75 billion. That’s the number CPP Investments just committed to EQT’s AI infrastructure fund. On its surface, it’s a pension fund diversifying into alternative assets. Dig deeper: this is a direct bet that the Transformer-based AI paradigm will require exponential physical compute capacity for the next decade. The signal is louder than the dollar amount. But where is the on-chain proof that this compute will be utilized efficiently? I’ve audited enough infrastructure plays to know: capital allocation without verifiable utilization is speculation dressed as investment.
Context
CPP Investments manages over CAD 600 billion. At 0.3% of AUM, $1.75B is a calculated exposure, not a reckless gamble. The target: EQT’s “AI infrastructure strategy,” which translates to building and operating high-density data centers optimized for GPU clusters. Similar moves by Blackstone, KKR, and DigitalBridge have turned this into a crowded trade. The narrative: AI training and inference will demand ever-more power and space, making data centers modern-day oil fields. The underlying assumption—that current AI architectures (mostly Transformer) will persist—is rarely questioned. My experience from the Ethereum Classic supply shock audit taught me to question assumptions built into capital flows. Back then, the assumption was that the ETC chain would recover; the code told a different story. Here, the assumption is that compute demand will outrun supply. Data tells a more nuanced story.
Core
Let’s strip the announcement to raw metrics. Industry averages for AI-optimized data center construction run $8M–$10M per MW of IT load. That puts $1.75B at roughly 175–220 MW of new capacity. At typical power usage effectiveness (PUE) of 1.2, total facility draw is ~210–264 MW. In terms of GPU density: assuming H100 GPUs at 700W each, and factoring in cooling and networking overhead (about 30% of facility power), each MW can support roughly 1,000–1,200 H100s. That yields 175,000–264,000 H100 equivalents. That’s not trivial—it could power one large cluster at Frontier-scale compute. But construction timelines? 18–24 months minimum if greenfield. That means this capacity won’t hit the grid until late 2026 at earliest. The market is pricing in a flood of compute that won’t arrive for two years.
Data doesn’t lie: the real bottleneck isn’t capital—it’s power procurement and chip availability. NVIDIA’s lead times for B200/B300 are already stretching into 2026. Even if EQT secures firm orders, they compete with hyperscalers for the same wafers. On-chain metrics > Twitter polls, but here the key metric is off-chain: the time-to-chip. I’ve tracked GPU delivery schedules since the DeFi summer when I correlated gas fee spikes to protocol exploits. The same pattern applies: infrastructure promises without execution dates are noise.
Contrarian
The consensus celebrates this as a vote of confidence in AI compute demand. The contrarian angle: $1.75B committed to centralized data centers reinforces a single point of failure—both for AI workloads and for the decentralized compute thesis. Every megawatt allocated to a traditional colocation facility is a megawatt not allocated to tokenized compute networks like Akash, Render, or io.net. These networks offer verifiable compute through on-chain contracts, enabling trustless utilization monitoring. CPP’s investment implicitly ignores the possibility that future AI workloads may demand verifiable computation (via zero-knowledge proofs or trust execution environments). I remember during the Terra collapse, systemic risk wasn’t in the code—it was in the concentration. Here, concentrating compute in a handful of facilities controlled by traditional fund managers replicates the same risk.
Moreover, the technology risk is underdiscussed. Current data center architecture assumes high power densities and liquid cooling for back-to-back GPU racks. If the industry shifts to photonic computing or neuromorphic chips that operate at lower power—or if model efficiency gains outpace scaling laws (e.g., Mixture of Experts, distillation)—these purpose-built centers could become stranded assets. My audit of the BAYC wash-trading pattern showed how market narratives can distort asset values. The narrative here is “AI needs massive compute forever.” But efficiency is improving faster than training compute is scaling. The capital may be chasing yesterday’s bottleneck.
Takeaway
Watch for three signals: (1) EQT’s specific site announcements—if they locate near renewable energy sources, they hedge against power volatility; (2) any disclosure of anchor tenant contracts—without them, this is speculative construction; (3) the emergence of tokenized compute backed by verifiable utilization metrics. If decentralized networks capture just 5% of this $1.75B flow, the crypto-AI thesis gains institutional validation. If not, we’re witnessing a walled-garden compute model that forgets the lessons of 2022: centralized infrastructure, without transparency, breeds fragility.