A single Nvidia B200 GPU draws 700W. Scale that to a 100,000 GPU cluster, and you need 70MW of power and a cooling system that can remove 70MW of heat. Traditional data center cooling fails at that density. On April 3, 2025, Mitsubishi Heavy Industries announced it would tackle this problem by joining Nvidia's partner network for power and cooling. This isn't just a press release — it's a signal that the physical layer of AI compute is becoming the critical bottleneck for every network that claims to be decentralized.
For five years, the blockchain industry has chased decentralization at the consensus layer. We have proof-of-stake, sharding, DAGs, and zero-knowledge rollups — all designed to ensure that no single entity controls the ledger. But the physical machines that execute those proofs and validate transactions? They sit in data centers owned by Equinix, Digital Realty, or a handful of colocation providers. And those data centers depend on cooling and power systems supplied by exactly three or four industrial giants. Mitsubishi Heavy Industries is now one of them, officially endorsed by Nvidia.
Context: The DePIN Lie
Decentralized Physical Infrastructure Networks (DePINs) were supposed to change this. Projects like Akash, io.net, Render, and Filecoin market themselves as marketplaces where anyone can contribute compute or storage. In theory, a GPU miner in Iceland and one in Singapore compete on price. In practice, both of them lease hardware from a colo facility that uses Vertiv cooling units or Mitsubishi gas turbines for backup power. The physical layer is profoundly centralized.
From my 2024 audit of a ZK-rollup's proving hardware, I saw how cooling failures directly impacted proof generation latency. A single pump failure in the liquid cooling loop delayed batch submissions by three hours. The rollup’s sequencer — itself a centralized node — had no fallback. The entire layer-2 was offline because the HVAC system had a mechanical fault. Smart contracts execute. They don't audit the HVAC system.
Now MHI enters the picture. Their industrial-grade solutions — liquid cooling CDUs, gas turbine generators, heat recovery systems — are designed for ultra-high-density GPU clusters. Nvidia’s partnership means that MHI’s hardware will be pre-certified for Blackwell and future Rubin architectures. For DePIN providers, this lowers the operational cost of running high-end GPUs. But it also locks them into a supply chain that is anything but permissionless.
Core: Technical Analysis of MHI’s Stack
MHI brings two core capabilities: liquid cooling and power backup.
Liquid Cooling: The B200 GPU has a thermal design power of 700W. The next-generation Rubin architecture is expected to exceed 1000W per chip. Air cooling is physically impossible at these densities. MHI’s cold-plate liquid cooling systems use a coolant distribution unit (CDU) to circulate dielectric fluid or water through cold plates attached directly to the GPU. Their industrial experience — cooling nuclear reactors and gas turbines — gives them thermal management at scales that liquid cooling startups like CoolIT cannot match. A single MHI CDU can handle 500kW of heat load, enough for a rack of 700 B200s.
Power Backup: AI data centers require Tier III or Tier IV reliability. Grid power is insufficient; backup generators are mandatory. MHI manufactures gas turbines and diesel generators that can provide 20MW to 100MW of backup power. Their gas turbines can run on natural gas or hydrogen, aligning with net-zero goals. For a DePIN miner leasing space in a colo facility, this means near-100% uptime — but it also means they are dependent on MHI’s maintenance schedule.
The hidden value is in MHI’s heat recovery systems. Their heat pumps can convert the 70MW of waste heat from a GPU cluster into hot water for district heating or even electricity generation via an organic Rankine cycle. This lowers the overall PUE from 1.4 to 1.05. for a 100MW facility, that saves 35MW of electricity — roughly $20 million per year at current industrial rates. Math doesn’t care about narrative; it only cares about efficiency.
But here is the structural problem: MHI’s solutions are engineered as integrated systems — the CDU, the piping, the heat exchangers, the generators — all designed to work together. They are not modular in the sense that a DePIN participant can swap out a component. The vendor lock-in is deep. Once a facility chooses MHI’s stack, switching costs are astronomical. This is the opposite of the permissionless ideal that DePIN preaches.
Contrarian: The Centralization Blind Spot
The common narrative is that Nvidia partnering with MHI is bullish for AI compute in general, and therefore bullish for DePIN. More efficient cooling means lower costs for GPU hosts, which means cheaper compute tokens for consumers. This is true in the short term.
But consider the failure modes. MHI’s gas turbines are complex machinery with a supply chain that spans Japanese steel mills, German control systems, and Southeast Asian electronics. A single vulnerability — a compromised firmware update in the turbine controller, a backdoor in the CDU’s PLC — could cascade across thousands of GPU nodes globally. Smart contracts execute. They don’t have a say in whether the cooling pump fails. Community governance can’t fix a broken turbine.
We saw this dynamic in the 2021 Texas freeze, when natural gas supply lines failed and thousands of crypto miners went offline. The bottleneck wasn’t the blockchain — it was the physical infrastructure. Now imagine a coordinated attack on MHI’s supply chain. A single zero-day in the gas turbine firmware could cause synchronized downtime across every DePIN that relies on MHI-backed colo facilities. The blockchain would continue to produce blocks, but the compute supply would vanish.
This is the blind spot that DePIN projects refuse to address. They talk about staking hardware, slashing conditions, and on-chain reputation. But they never audit the cooling system of the data center that hosts their miners. They treat the physical layer as a black box. MHI’s new partnership with Nvidia reinforces that black box — it makes the most efficient cooling even more centralized.
Liquidity is an illusion until it isn’t. The same applies to compute liquidity. A DePIN that seems to have 10,000 GPUs online is only as liquid as the cooling system that keeps those GPUs from melting.
Takeaway: The Need for On-Chain Attestation of Physical Infrastructure
MHI joining Nvidia’s network is a rational business move. It solves a real problem — high-density GPU cooling — and will accelerate the rollout of AI clusters globally. But for the blockchain ecosystem, it exposes an existential gap. We cannot rely on centralized industrial suppliers for the physical layer of decentralized compute networks.
The solution is not to rebuild MHI’s factories; it is to build verifiable attestation of physical conditions. On-chain oracles need to report not just asset prices but PUE metrics, coolant temperatures, pump statuses, and turbine fuel levels. Smart contracts should automatically reallocate workload if a facility’s cooling system shows degradation. This is the next frontier for blockchain infrastructure: verifying that the physical hardware is running with adequate cooling and power.
Until then, every DePIN network is only as reliable as its centralized cooling provider. And that provider just got a lot more powerful.
