Name the Workload Before You Sign
Most companies probably don’t need a dedicated server, and the ones that do can usually name the job that pushed them there. That’s the most useful test to run before anyone signs a hardware contract. A general feeling that the cloud bill has grown too large rarely qualifies on its own. A specific workload that runs all day, every day, on a machine you share with strangers often does. Seven such workloads follow, ranked by how much harm the wrong hosting choice tends to do to the business behind them, and the ranking isn’t kind to a couple of the arguments people make most often.
The Order Database Earns First Place
The database that takes orders and keeps the books comes first, and in most businesses it isn’t close. Every other system in a company leans on it, and it works in a steady rhythm that suits hardware nobody else touches. What tends to hurt it on shared infrastructure is inconsistency more than raw speed. In a 2020 study, researchers from universities in Amsterdam, Utah, California and Delft measured how much cloud network performance drifts after pushing more than nine petabytes across commercial clouds, and found the swings persisted even where providers had controls meant to stop them. Sometimes those controls made things worse. Their subject was large data-processing jobs rather than order systems, but a database can only be as steady as the network and hardware underneath it.
A machine with a single tenant removes the neighbors from that equation, even if it can’t remove every source of delay. Atlantic.Net‘s dedicated servers in the USA give a fair picture of what a buyer gets in exchange, which is a private physical server built to the processor, disk and memory the customer picks, with no resources shared with other customers. For an order system that has to answer just as quickly at three in the morning as it does during a holiday rush, that separation is most of the point.
None of this makes the cloud’s managed databases bad products. For a team with nobody who wants to own backups and upgrades, they’re a sensible purchase. A company that already pays someone to do that work, though, is arguably buying the same skill a second time inside the provider’s price.
Anyone weighing the move should start by measuring the right thing. Averages flatter shared infrastructure, so pull a month of query times and look at the slowest one percent, because those slow moments are what a customer at checkout notices. It’s just as important not to put the whole business on one box. In Uptime Institute’s 2026 outage analysis, 57 percent of respondents to its latest annual survey said their most recent major outage cost more than $100,000, and a single server with no standby is an easy way to join them. A second machine in another data center, kept in sync, belongs in the budget from day one.
Software licenses deserve the same scrutiny as hardware, since they can move the comparison more than the machine does. Britain’s competition regulator found that Microsoft charges rival clouds much higher wholesale prices for some of its key business software than it charges customers on Azure, and in March 2026 it said it had seen no material progress on Microsoft’s cloud licensing practices. The same database can cost noticeably different amounts depending on whose hardware it runs on, so it pays to get a written license quote for each option before comparing monthly totals.
What the Cloud Was First Sold On
The cloud’s early pitch, worth recalling here, was never that it would be cheaper for everything. It was that a startup could get going without buying a single machine, and that promise still holds. Somewhere along the way, the argument for starting in the cloud turned into an argument for staying there indefinitely, and not many finance teams seem to have noticed the switch.
The people paying the bills are noticing now. Flexera’s 2026 State of the Cloud survey found 85 percent of respondents naming cloud spending as a top challenge, and its estimate of wasted cloud spend rose to 29 percent, the first increase in five years, with AI workloads blamed for the jump. The same survey found 73 percent of organizations already running hybrid setups. Waste on that scale isn’t an argument for dedicated hardware by itself. It does suggest that plenty of companies no longer know which of their workloads are steady enough to be priced differently.
AI That Never Clocks Off
Second comes the AI work that never switches off, like a support assistant answering tickets overnight or a fraud check that scores every card payment as it arrives. Stanford’s 2026 AI Index shows how widely that kind of work has spread, and the survey behind its adoption figures found 88 percent of organizations using AI in at least one business function in 2025. The same chapter carries a detail that cuts against this column. The computing bills of the leading AI labs mostly reflect rented cloud capacity, not data centers they own.
If the biggest AI companies rent, why should a regional insurer buy? Because the labs rent at a volume, and on terms, that a regional insurer is unlikely ever to be offered. Hourly billing charges for the freedom to walk away at short notice, and a model that answers customers around the clock has little use for that freedom.
The honest objection is price. Researchers at Epoch AI estimate that the cost of reaching a given level of AI performance has fallen about 13-fold a year since 2023, which makes any hardware purchase look like a bet against a very steep curve. Their own caveat matters, though. Real users rarely switch models often enough to capture every saving, and the price objection weakens further once the data flowing into the model includes customer records a company would rather keep on hardware it controls.
Machine learning now turns up in hospital diagnostics and bank fraud screening alike, and most of those uses run continuously once they go live. Owning the hardware for them still means guessing at capacity months ahead. Buy too little and the support assistant slows down every Monday morning when the ticket queue fills. The opposite mistake leaves a very expensive set of graphics chips idle through the weekend, a cost that renting would have absorbed.
A practical habit helps here. Log how many hours a week the model is genuinely busy, for a full quarter, before pricing anything. Chips that would sit idle most nights are a strong hint to keep renting. The case for owning gets stronger as that weekly log flattens out, and round-the-clock AI work tends to produce a flat one, which is why this category sits second rather than further down.
Large Files and the Price of Leaving
Third is any business that stores or serves a lot of large files, whether that’s a library of training videos or an archive of medical scans. The big clouds typically charge little or nothing to bring data in and bill for moving it out, so the bigger the pile gets, the more expensive it can become to leave or even to use. The same British regulator put the switching problem plainly. Its 2025 final decision on the cloud services market found that fewer than one customer in a hundred switches provider in a given year, and that fees for moving data out weigh heaviest on smaller customers and on those holding large amounts of stored data.
In March 2026, Amazon and Microsoft offered to drop those fees for UK customers who switch, for a window of at least 180 days, and to lower them for customers spreading work across rival clouds. The concessions, as described, are about exiting or mixing providers. They don’t address the ordinary cost of sending files to the people who asked for them, which a streaming service pays every hour its viewers are watching.
Two numbers settle most of this. One is how many terabytes leave each month, which the current bill already shows if anyone reads it closely. The other is the transfer allowance bundled into a flat monthly server plan, since dedicated plans usually fold a fixed amount into the price. For a business whose product is the file itself, comparing the two can change the arithmetic quickly. Keeping a second copy of the archive outside the main provider is worth doing either way, because it makes leaving a matter of choice rather than a migration project.
Regulated Records Rank Lower Than Expected
Regulated records take fourth place, which tends to surprise people, since compliance is one of the reasons most often given for dedicated hardware. It probably deserves less weight than it gets. The large clouds sign healthcare privacy agreements and hold card-industry certifications, and plenty of regulated firms run on them without incident. What single-tenant hardware changes, more often than not, is the length of the conversation with an auditor.
Recordkeeping rules in finance and government expect records to stay unaltered and retrievable for years, and an auditor asking where a record lives and who else could have touched it is easier to satisfy when the answer is one named machine in a known building. If you’ve sat across the table from an examiner while an engineer tried to explain on a whiteboard how customers share a provider’s hardware, the appeal is obvious. That’s a real advantage. It just isn’t the one usually claimed, which is legal necessity.
The rules themselves tend to be neutral about the hardware. The SEC’s recordkeeping rule for broker-dealers requires certain records to be kept for six years, the first two in an easily accessible place, either with a time-stamped audit trail of every change or in storage that can’t be rewritten. Where those records sit on servers owned or operated by an outside company, that company generally has to file an undertaking with the Commission, and the requirement reaches a cloud giant and a dedicated host alike. In both cases the useful preparation is a short file listing where each regulated record physically lives, who holds administrative access and where the audit logs are kept, updated before the examiner asks for it.
Games and Other Businesses That Lag in Public
Fifth are multiplayer games, along with the less obvious businesses that behave like them. Players forgive a slow menu. They’re far less forgiving of a match where the lag jumps around without warning, and shared machines can be prone to exactly that kind of stutter when a neighbor gets busy.
Some of the same Dutch researchers later tested this directly. In a study first posted in 2021, a team from Delft and Amsterdam benchmarked Minecraft-style game servers on two major commercial clouds and on self-hosted university hardware, and found the cloud runs noticeably less stable. Across their repeated player tests, even the steadiest cloud result was worse than the least steady self-hosted one. They also found that the server size most often recommended by hosting companies wasn’t enough to keep the game smooth.
A common pattern is a base of owned or leased machines for the regular crowd, with cloud capacity rented for launch weekends and dropped afterward. Anyone testing that kind of setup should measure the worst moments rather than the average, such as the seconds after a player connects, because that’s where the study recorded some of its biggest spikes. The same logic reaches past entertainment. A bid that lands half a second late in an online equipment auction costs a seller real money, and the bidder rarely blames the network. They blame the site.
The Overnight Jobs Nobody Watches
Sixth, and far less glamorous, is the work that runs while everyone sleeps, whether that’s a software team’s overnight test runs or the batch job that reconciles a day of sales. Nobody needs it to scale up in seconds. That removes the main thing cloud pricing charges extra for, and it leaves a job that will happily run on the same hardware every night for years.
An architecture practice rendering walkthroughs for clients, or an agency encoding a week of video ads, knows roughly how many hours of computing each month will need. Once you can predict the hours, you can price them against a flat monthly server, and the hourly option frequently comes out behind. Growth is the one caution. A batch window that fits comfortably inside six hours today can creep toward morning as data piles up, so it’s worth leaving headroom rather than sizing the machine to this year’s load.
The Storefront, Split in Two
The storefront comes last, and this column would put it lower if the list had an eighth slot. Retail traffic spikes. Adobe’s figures for last year’s Cyber Monday put U.S. online spending at $14.25 billion, with shoppers spending $16 million a minute during the evening peak, and a curve like that is precisely the pattern the cloud’s pay-as-you-go pricing was designed for.
Many small shops never think about servers at all, because they run on a hosted platform and extend it with add-ons like the review apps Shopify merchants install to collect ratings after each order. For them, this whole debate is somebody else’s problem, and that’s a perfectly reasonable place to be.
It’s on the list anyway, because behind every larger storefront sits the order database that took first place, and the common mistake is treating the two as a single decision. The product pages can usually stay where the spikes are cheap to absorb. The ledger tends to belong on the steady machine, and a load test aimed at checkout during a simulated peak minute will usually show which of the two is struggling.
Start With This Month’s Invoice
So pull up the latest cloud invoice and try to find the person who decided that the order database should be billed by the hour. Chances are that person left some time ago and nobody wrote the reasoning down. The invoice keeps arriving each month, approved by someone who would have to find the migration money in their own budget the moment the question got asked out loud.
Ask it anyway, and keep the scope small. Go through the bill line by line, mark every workload that runs around the clock at a roughly steady level, and price only those against a flat monthly server, with a standby machine and the software licenses included. Everything spiky can stay where it is. If the database alone doesn’t justify a move once the numbers are honest, nothing else on this list is likely to, and that’s worth knowing before the next renewal rather than after it.
