Hook Canada Pension Plan Investment Board just committed $1.75 billion to EQT’s AI infrastructure strategy. That capital will build data centers optimized for GPU clusters. Not a single Bitcoin ASIC in sight. The sum exceeds the combined market cap of every decentralized compute token listed on Coinbase. Survival is the ultimate metric of a robust system — and centralized AI data centers are winning the capital allocation game, hands down.
Context CPP Investments manages over CAD 600 billion. This allocation represents ~0.3% of their portfolio. EQT, a Swedish private equity firm, will deploy the funds across hyperscale data centers designed for high-density AI workloads. Think liquid-cooled racks, InfiniBand interconnects, and multi-year contracts with cloud giants like Microsoft Azure or CoreWeave.
The global data center capex is on track to exceed $200 billion in 2024 alone, with AI-related spending growing fastest. The narrative is simple: large language models need more compute. More compute requires more concrete, copper, and kilowatts. Pension funds like CPP extract a bond-like yield from these assets while riding the AI demand curve.
This is not a crypto story. Or is it? Every dollar flowing into centralized AI infrastructure is a dollar not flowing into decentralized alternatives like Akash Network, Render Network, or io.net. It also competes directly with Bitcoin mining for energy and real estate. The macro watcher in me sees a capital structure shift: institutional money prefers predictable, rent-seeking toll roads over permissionless compute markets.
Core Let’s stress-test the assumptions behind this investment. CPP and EQT are betting that AI compute demand will remain exponentially growing for the next 10-15 years. That rests on three premises: 1) the Transformer architecture remains dominant, 2) scaling laws hold, 3) power supply constraints don’t cap growth.
From my 2026 AI-agent economy protocol design work on Solana, I know that machine-to-machine transactions require latency measured in milliseconds, not seconds. Centralized data centers inherently have lower latency for colocated GPU clusters. A decentralized network of consumer GPUs across thousands of households introduces too much variance. For the training of frontier models, centralized wins. For inference of low-stakes queries, decentralized can compete.
The investment will convert into roughly 2 GW of IT load capacity. That’s enough to power ~700,000 H100 GPUs. But construction takes 24-36 months. The new supply hits the market in 2026-2027. By then, NVIDIA will have shipped B200 and likely B300. The depreciation cycle for data center providers is brutal: they must write off servers after 3-4 years. CPP’s return depends on securing long-term leases that lock in rent escalation clauses, otherwise, the margin compresses as newer, denser facilities come online.
Compare this to the crypto mining industry. A typical Bitcoin mining data center costs $500-700K per MW to build. An AI data center costs $800K-1M per MW due to liquid cooling and high-speed networking. But the revenue per MW is higher for AI if contracted properly. The catch: AI GPU utilization is 30-50% during training (batch processing), while ASICs run 24/7 near 100% utilization. Mining is a thermodynamic floor — electricity in, heat out, hash power sold. AI is a speculative compute market — demand fluctuates with hype cycles and model releases.
I audited 40 ICO whitepapers in 2017 and saw how projects raised hundreds of millions for “decentralized compute” without a single working node. Today, decentralized compute tokens like Akash have a market cap of roughly $600M. That’s one-third of CPP’s single bet on EQT. The market is voting with real dollars: centralized infrastructure is the default.
The contrarian angle emerges. Decentralized compute networks have structural advantages: they can amortize idle GPU cycles from gaming PCs and retail data centers. They also avoid the energy procurement headaches. A pension fund cannot write a check to a DAO. But they can buy shares in EQT. The capital markets are biased toward corporate structures, not protocols. Until crypto can issue dividend-like yields and pass through tax-efficient returns, institutional capital will prefer traditional infrastructure plays.
Contrarian The bullish view on centralized AI data centers is already priced into EQT’s fund terms. Capitalization rates have compressed from 8-10% to 6-8% over the past two years. That’s a sign of froth. Meanwhile, decentralized compute protocols offer higher risk-adjusted yields if measured by the spread between token inflation and compute demand growth. For example, io.net’s utilization rarely exceeds 20% despite billions in token market cap. The underlying assets exist, but orchestration and trust are weak.
What if the next generation of AI models uses MoE (Mixture of Experts) or SSM (State Space Models) that are more efficient? The demand for dense GPU clusters falls. Centralized data center owners are stuck with 100MW facilities that cannot be broken up. Decentralized networks can reallocate compute to other tasks like rendering, scientific simulation, or even Bitcoin mining. Flexibility is an insurance policy against technological disruption.
The real signal from CPP’s investment is not about compute. It’s about energy. Data centers are becoming the new industrial load. In Virginia, data center demand is driving utilities to build new gas plants and explore nuclear restarts. Bitcoin miners have already figured this out — they strategically locate near stranded gas or hydro assets. The smart money will not only own the compute but also the power purchase agreements (PPAs). EQT likely has PPAs locked for their data centers. CPP gets exposure to power markets through the assets.
Crypto-native projects could replicate this at the protocol level. Imagine a DAO that buys power capacity from a wind farm, deploys GPUs for AI inference, and sells excess capacity back to the grid. The technical architecture exists. What’s missing is institutional-grade legal wrappers and insurance. That gap will close as more regulated entities enter the space.
Takeaway CPP’s $1.75 billion is a vote of confidence in centralized AI infrastructure. Crypto must stop pretending it can win on latency or trust for high-end training. Instead, focus on the residual niches: inference at the edge, uncensored compute, and energy arbitrage. The pension funds will own the toll roads. Crypto should own the side streets and the bridges to the off-grid power plants. Survival of a robust system depends on knowing where you can compete — and where you cannot.