31% of Ethereum nodes are in the United States. 60% run on three cloud providers.
State root mismatch. Trust updated.
A new study from the Cambridge Centre for Alternative Finance has quantified what many L1 researchers have whispered for years: Ethereum’s physical layer is dangerously centralized. The data is clean. The implications are messy.
Context: The Numbers That Don’t Lie
The study deployed a network crawler and aggregated node metadata over a three-month period in early 2025. Key findings:
- Geographical concentration: 31% of reachable Ethereum nodes are located in the United States. Germany and Singapore follow at 15% and 8% respectively. The top three countries control over 50% of all nodes.
- Cloud provider dominance: Amazon Web Services alone hosts 28% of all nodes. Microsoft Azure and Google Cloud together account for another 32%. That’s 60% of the network running on three companies.
- Client diversity: Geth still dominates execution clients (~85%), while Prysm leads consensus clients (~40%). Single-client risk remains high, but that’s a separate threat.
The study is not a surprise to those who watch infrastructure trade journals. But having the numbers stamped by a third-party academic institution changes the conversation from “anecdotal concern” to “verified data point.”
Core: Breaking Down the Risks – Layer by Layer
Let me walk through this from the bottom up. I’ve spent the last three years auditing L2 architectures and bridge contracts. I’ve seen what happens when a centralized dependency fails. This is different. This is the foundation.
1. Network Robustness – The Single Point of Gravity
Ethereum’s consensus mechanism requires >66% of validators to be online and honest for finality. If AWS suffers a regional outage – and we’ve seen those happen – a significant fraction of nodes could go dark. In a worst-case scenario, the network could stall. No new blocks. No finality.
Bitcoin’s mining is also geographically concentrated (China, US, Kazakhstan), but miners can switch pools, and the network does not depend on a handful of cloud providers. Ethereum’s node operators, especially staking pools, choose AWS for its reliability. That reliability becomes fragility when it’s everyone’s choice.
2. Censorship – The Regulatory Lever
The United States has clear jurisdiction over nodes operating within its borders. If OFAC decides to sanction a mixer or a protocol, it can pressure AWS to stop hosting those nodes. Or it can directly order US-based node operators – including Coinbase, Kraken, and Lido’s largest node operators – to censor transactions. The Ethereum protocol cannot prevent this. It can only fork.
And forking a live network is not a clean operation. The last time Ethereum faced a censorship fork (the Merge aftermath with OFAC-sanctioned blocks), the community saw a split into “censorable” and “uncensorable” proposer sets. The centralization of nodes in the US means the “censorable” proposer set is larger than most realize.
3. Staking – The Concentration Loop
The 32 ETH staking minimum already pushes small validators toward liquid staking providers like Lido, Rocket Pool, and Coinbase. These providers operate thousands of validators on high-performance cloud instances. Lido alone controls nearly 30% of all staked ETH. Its node operators are geographically diverse but cloud-diverse? Not really. Most run on AWS, GCP, or Azure.
This creates a feedback loop: large staking pools choose centralized cloud providers for cost efficiency. The cloud providers become systemically important. Regulators can then pressure the clouds. The pools comply. The network censors.
I’ve seen this script before in the TradFi world – the “too big to fail” narrative that ends with a bailout or a shutdown. Here, there’s no bailout. Only a hard fork.
4. L2 Dependency – The Cascade Effect
Every L2 – Arbitrum, Optimism, zkSync, StarkNet – depends on Ethereum L1 for finality and data availability. If L1 stalls due to node centralization, L2s cannot confirm state roots. Their sequencers, many of which also run on AWS, become orphaned.
During my 2024 deep dive into the Arbitrum NFT bridge exploit, I traced the race condition to a latency spike in the L1 node infrastructure. That spike was caused by an AWS network partition in us-east-1. The exploit wasn’t malicious – it was a failure of centralization. We patched the wrapper, but we couldn’t patch the cloud.
5. The Bitcoin Comparison – Not a Panacea
Bitcoin’s node network is more distributed geographically but less resource-intensive. Running a full Bitcoin node costs less than $100/month on a home server. Ethereum’s current node requirements – especially for validators – demand high uptime, low latency, and significant storage. That drives operators toward professional hosting. Home validators exist but are outliers.
The Cambridge study doesn’t compare to Bitcoin, but the implication is clear: Ethereum’s node centralization is a structural byproduct of its design choices, not a temporary imbalance.
Contrarian: The Known Unknowns and the Overpriced Fear
Let me push back on my own analysis. Because the market often overreacts to academic reports.
First counterpoint: The data might be stale or incomplete. The Cambridge crawler only detects reachable nodes via TCP handshake. Tor nodes, VPN-hidden nodes, and many home validators are invisible. The real US concentration could be lower. But the cloud dependency is measurable via IP ranges, and that data is solid.
Second counterpoint: The market has already priced this in. Ethereum’s price has not moved on this study. Institutional investors already assume US regulatory risk. The study just provides the receipts. If the market was efficient, the risk is already discounted. Short-term volatility is unlikely.
Third counterpoint: L2 centralization might actually make L1 node centralization less dangerous. If most users migrate to L2s that have their own security models (e.g., zk-rollups with validity proofs), then L1 censorship only affects finality, not transaction execution. Users can still transact on L2s until the L1 recovers. This is a weak argument – because L2 security still depends on L1 for data availability and dispute resolution – but it’s worth noting.
Fourth counterpoint: The real risk isn’t node centralization – it’s regulatory capture of the staking layer. The Cambridge study focuses on nodes, but the real power lies in staking pools and their node operators. If the regulator goes after Lido and Coinbase directly, the node locations become secondary. The pools will comply regardless of where their nodes run.
So the study is a proxy for a deeper problem: the centralization of staking power. And that problem is harder to solve because it’s economic, not technical.
Takeaway: The Fork Is Coming – It's a Question of When
The Cambridge study is not a warning. It’s a confirmation. The Ethereum community has known about node centralization for years. The foundation has promoted client diversity, but geographic and cloud diversity remains ignored.
Forward-looking judgment: Within three years, we will see a regulatory event (OFAC order, state-level subpoena, or federal clawback) that targets Ethereum node operators in the US. The network will either comply and censor, or the community will fork to a version that excludes US nodes. That fork will be messy, contentious, and likely reduce Ethereum’s value as a global settlement layer.
The only technical hedge today is Distributed Validator Technology (DVT). Projects like Obol and SSV Network allow a single validator key to be fragmented across multiple nodes in different geographies and cloud providers. If the regulator targets one node, the validator can still propose blocks through the others.
But DVT adoption is low – maybe 2% of validators use it. The rest are sitting ducks.