Centralized Compute Is A Brittle System

The race to build the infrastructure that powers AI is not over. Anyone telling you that centralized data centers have already won is reading the wrong scoreboard. The current moment is one phase of a much longer contest, and the side winning right now is winning because they took the easier path, not the right one. That should worry anyone who cares whether the technology being built today still works in twenty years.

Concede the centralized argument its due. Standardized hardware procurement is genuinely cheaper at scale. Specialized cooling and power delivery work better engineered as one system. Training the largest models requires GPUs to communicate at speeds consumer internet has not yet matched, and there are workloads that need to run on tightly coupled clusters connected by fiber thicker than a thumb. None of this is being disputed. The decision to centralize compute makes sense for a list of real reasons real engineers can defend.

The argument is not that centralization is wrong. The argument is that centralization is a trade. Specifically, it trades longevity for ease.

What Ease Actually Costs

Ease is what concentration delivers. One vendor relationship. One billing cycle. One support contract. One organization to call when something breaks. The convenience is real and it is exactly why companies pay enormous premiums to use hyperscale services rather than running their own infrastructure. Ease is genuinely valuable.

What gets sacrificed for it is durability across time. Most of the people making the trade are not pricing this side correctly.

Centralized systems fail. Not occasionally. Regularly. AWS published its own post-event summary page of major incidents. The page lists S3 February 2017, DNS Seoul November 2018, EC2/EBS Tokyo August 2019, Kinesis November 2020, Lambda June 2023, Kinesis July 2024, and the DynamoDB DNS race condition of October 19 to 20, 2025 that took US-EAST-1 dark for roughly 15 hours. According to Cisco ThousandEyes' incident analysis, the race condition wiped DNS records for the database service and cascaded across 58 AWS services. Snapchat, Fortnite, Roblox, Ring doorbells, McDonald's mobile orders, United Airlines bookings, the British government's tax website. According to insurance analytics firm Parametrix and Moody's, insured losses ran as high as $581 million, with total losses to affected businesses estimated between $500 and $650 million.

Nine days later, Microsoft Azure Front Door collapsed for over 8 hours on October 29, 2025, caused by a configuration change pushed to the control plane. The Azure status page assigned it Tracking ID YKYN-BWZ. It was the second Azure Front Door incident in three weeks. Microsoft 365, Outlook, Teams, Xbox Live, Minecraft, Costco, Starbucks, Alaska Airlines, and the website of London's Heathrow Airport went dark together. Outlook alone serves over 400 million users.

Then on May 7, 2026, AWS US-EAST-1 went down a third time inside seven months. A thermal event in availability zone use1-az4 damaged hardware. Coinbase was offline for nearly seven hours. FanDuel and CME Group trading were also disrupted. AWS engineers required more than 12 hours to bring damaged hardware back online. Same week, Coinbase reported a $394 million Q1 2026 net loss and cut 700 jobs. Three major centralized cloud outages in seven months. Three different root causes. Same architectural reason.

Now compare any of that to Bitcoin. According to the Bitcoin Uptime Tracker, the Bitcoin network has gone down twice in its entire history. Once in 2010 for about 8.5 hours due to a value overflow bug. Once in 2013 for about 6 hours and 20 minutes due to a database migration consensus mismatch. Both incidents were resolved by community consensus without any central coordinator issuing orders. Since 2013, Bitcoin has had zero downtime. Over thirteen consecutive years of one hundred percent uptime, running across thousands of independent nodes operated by people who have never met and have no boss. Total all-time uptime is approximately 99.988 percent, which beats the published numbers from every major centralized cloud provider over that span.

The clarification matters. Decentralized infrastructure is not the backup option for when centralized fails. It is the more durable option, full stop. The narrative that centralization equals reliability and decentralization equals chaos has the truth backwards for anyone willing to look at the actual numbers. A properly designed decentralized network has no single point of failure. Bringing it down requires breaking most of the network at once rather than tripping one wire in Northern Virginia. Failure in a distributed system is graceful and partial. Failure in a centralized system cascades. The reliability claim should run the other direction.

The Waste Problem Nobody Will Admit

There is a deeper version of the argument that nobody making the centralized case wants to engage with. Centralized compute is wasteful. Not abstractly. Measurably. While hyperscalers race to construct new megawatts of capacity, billions of consumer devices sit underutilized for most of the day. The chips exist. The power flows to them. The thermal management is handled. None of that capability is plugged into the systems doing real work. New infrastructure is being funded to handle workloads an enormous installed base could handle right now if the coordination layer existed. The gap between deployed capability and deployed utilization is the actual definition of waste.

A robust distributed network would also be cheaper for the corporations using it than building their own data centers. The hardware comes for free because contributors already own it. Power costs are absorbed by contributors who pay their own utility bills. The hardware refresh cycle takes care of itself as consumers upgrade their phones and laptops on their own schedule. None of that lands on a corporate balance sheet. The corporation pays only for work performed, never for capacity sitting idle. The cost structure flips from a massive recurring capital expenditure into a variable operational expense that scales with actual demand. By any reasonable measure of capital efficiency, that is a stronger deal than what these companies have today. The fact that they have not pursued it tells you the priority is not efficiency at all.

Why The Industry Refuses

The reason they are not doing it is not technical in the foundational sense. The technology to coordinate distributed compute exists. Active research closes the remaining gaps every year. The reason hyperscalers are not pursuing distributed infrastructure is control. Centralized infrastructure means centralized control. Centralized control means the people running it get to decide what happens, who participates, and most importantly who profits. Distributed infrastructure threatens that arrangement directly because it lets the people who own the hardware participate in the economics of the system itself. Of course the people who would lose that profit are not enthusiastic about the alternative.

The honest case requires naming what the trade looks like on the long side. Centralized infrastructure is easier to procure and operate today. The savings of distributed infrastructure show up over years, not quarters. The operational expertise required to participate in a decentralized network is not yet standard in enterprise IT. Some workloads, particularly low-latency consumer applications and regulator-bound financial settlement, are not yet a clean fit. These are real costs and they should be priced honestly.

What should not be priced is the assumption that the easy path keeps being easy. Three major centralized cloud outages in seven months. PJM grid clearing prices hit the FERC ceiling three years running. Residential electricity rates crossed 18 cents per kilowatt-hour as data center construction absorbed capacity. The bill for the easy path is moving from the cloud invoice to the utility bill to the federal grid policy debate. Whoever signed those contracts believing convenience came without a tradeoff is going to look at the next decade and realize the shortcut went somewhere they did not intend to end up.

When the centralized side claims it has already won, what is actually being said is that one phase of a long race ended with the easier path being chosen first, by companies who would rather not share. The race continues. The hardware that will power the next phase is already deployed across billions of devices. The economics work. The reliability is better. The capital efficiency is better. The only thing missing is willingness to take the harder path.

Easy is not the same as right. The longer this race goes, the worse the trade looks for the side that took the shortcut.