A sandwich, bag of crisps and a drink for £5 is an actual deal. Sandwich alone in U.S. would be $10 and the “$15 Meal Deal” just doesn’t have the same ring to it.
I work from Starlink regularly and it’s just fine for meetings and everything else I do.
Meetings are usually under 10Mbps of downstream and 2Mbps upstream.
It’s simply false you need 1Gbps of “low latency” internet to work from home effectively unless your job is something like content production and working with huge files.
Producing for OnlyFans? Benefit from 1G to send raw to your editor.
Doing most other forms of work? Doesn’t require remotely close to this amount of bandwidth.
Amazon is probably deploying in the range of 250,000 racks per year into AWS. That’s millions of square feet before you get into all the infrastructure around them so they’re powered, cooled and operated the way they need.
Figure on ~10 million square feet of conditioned DC space per year, approximately 5-10GW of additional power consumption to power those 250k cabinets (depending on the exact mixture of what’s in the racks — Compute, Storage, Network, Accelerators), and that’s just for one hyperscaler.
There are at least 5-7 companies in the hyperscaler weight class although likely none individually meaningfully larger than AWS, they’re the 800lb bear and everyone else is in the 500-750lb range.
It’s a lot. Datacenters also take long enough to build that a hyperscaler is pouring concrete today for shells they expect to serve real workloads in 2029 - 2031. What you’re seeing come online today in response to customer demand really started being built in 2021 - 2023.
Building on grade is much cheaper. There is in general plenty of surface area on Planet Earth.
Datacenters aren’t built next to Nordstrom. Theres just no reason to spend on engineering and construction that increases density like underground parking.
It’s true it costs money to then hopefully make it. :)
Architecture, civil engineering, and other design and permitting fees can easily be 7–10% of the overall cost of smaller projects. Even in large projects they’re often 4-5% or more, and the number of billable hours for complex impervious surface and stormwater management adds up fast, as do engineering stamped plans for structural and other factors.
Most cost here is also incurred fairly early before you have any vertical construction done — Phase 2 in a program after land acquisition. So you feel like you’re spending a ton on paperwork and you can’t see anything yet.
You CAN spend a bit of money even before land acquisition on quick feasibility studies but in U.S. terms for something like a residential, small commercial or light industrial project every parcel you go “I like that, can it work?” you are dropping $15-50k during a 45-120 feasibility period. Should it 100% not work out you are NOT getting reimbursed that by the selling party. You’re out the money. Even within 90 days you may find some uncertainties like SEPA approval won’t close before you have to say deal or no deal on the parcel acquisition. This is quite unlike massive companies doing business where the land may not change hands until essentially every approval is locked in (but should it not work out the buyer may be out millions in engineering fees paid to try and make it work).
Borrowing to buy land and then borrowing more to build something is also treated very differently by most lenders. It carries tremendous uncertainty versus you buying a preexisting lot and structures which they know how to value. That’s fundamentally something that causes unwillingness to lend, or changes the rates and down payment or security terms. In contrast with a conforming mortgage for a SFR (single family residence) at 6%, borrowing to build (a construction loan) can be 10-14% APR, often secured via personal guarantee and other assets you possess, and then you have to convert to a personal or commercial mortgage after you complete building what you wanted.
Borrowing to buy the land is even more complicated again — you very often must be able to pay cash for land, and then just borrow to do the construction. Borrowing for both especially with limited assets to secure against will always be a polite “Sorry; we can’t help.”
Datacenters rarely get mothballed after 10 years, they’re usually 15-year depreciation schedules in terms of the building and the plant. Equipment within them like servers typically depreciates over 72 months but is rarely removed by hyperscalers promptly after 6 years. More typically it’s around Year 8.
So you won’t see the building cease being useful for about 20 years. You’re getting usually two full cycles of “the servers inside” before a renovation program (or potentially asset/building disposal to another party, or demolition, depending what changes in those 20 years really).
What’s often happening around these tax agreements is they are a mixture of incentive and an offset of prepaid improvement costs to the city/county for developing mains water, sewer, roads, schools, fire and police, and other infrastructure, sufficient to support the 100s of families who may move here to take the post-commissioning jobs, etc.
Often money out the door for the hyperscaler is about the same over 20 years, it’s just some number of $M’s is paid upfront as an “Contribution to Improvements”. That’s actually good for the municipality involved, too.
Municipalities typically know or are told these aren’t tax rebate for 10 years and then it’s getting bulldozed. They’re sophisticated enough and well advised to understand this is a 20+ year investment in their town.
I mean, you’re right, but when the alternative is all that investment and those construction jobs AND the post-commissioning operational jobs go to a different community and a different economy, …
What would you prefer? To me, local communities tend to benefit in multiple different ways during and after these projects, poorer communities become richer, communities with little opportunity now have more opportunity. I’m always a bit baffled by someone saying “Please don’t invest $5B and create 100s of jobs and taxable improvements in my back yard”.
This is a common argument: wanting 1000s of jobs during construction and 1000s of jobs after construction, but this isn’t a car manufacturing plant. That’s a “we want our cake and we want to eat it too” argument — not saying it’s your argument just that this comes up frequently.
It’s unusual to buy the land and take a gamble on its utility, at least whenever datacenter construction is involved. Purchasing parties are risk averse to this exact scenario and work hard to craft contracts that reduce risk.
Often the choices are —
1. Buy land at $/acre that reflects very little premium, based on a short feasibility study, but without any ultimate contingency that permitting will occur. This is your example. But problematically all permitting applications are typically public record, so when you fail, the land can’t be sold on to someone else as if that didn’t happen, any sophisticated buyer will know the exact issues the city/county had with your usage. Land often transacts onward at firesale prices under these circumstances.
2. $/acre for land is bid upon at a substantial premium reflecting the future value as a datacenter, it remains under contract for potentially years pending outcome of approvals, then it transacts. Permitting being denied usually results in either no money changing hands or a small termination fee reflecting the carrying cost of the land during that period. If permitting works out the seller of land walks away very happy as the $/acre was extremely lucrative.
Building a datacenter typically employs thousands of people in the trades, often for hundreds or a thousand plus hours, per person, from the start of a big build to the campus being complete. It’s quite literally millions of billable hours in trade labor.
Modern datacenters also require very high standards of construction and are complex, so these projects create jobs and also represent a real training, upskilling and work experience opportunity for labor. There are many examples of electricians, plumbers and groundwork teams who did Microsoft’s site getting future work from Meta, Google or Amazon in the same part of the state because the experience has value.
It’s easy to dismissively say datacenter is bad, or that it consumes too much water (despite many datacenters accused of this being a closed-loop cooling system), and ignore the billions of dollars spent during the project on labor which supports that local economy, or the improvements negotiated for the local area and paid for the hyperscaler, bundled in by the city/county planning as part of the permits and approvals.
It’s also rare the tax for a campus is fully rebated, although it’s normal for the improvements to be partially rebated for some period (this is an investment incentive). Viewed over 20-40 years these sites are often tremendously lucrative in tax for the county/city as well.
AWS publicly stated I think in 2021 that the larger availability zones in US East 1 consisted of 17-18 datacenters each. It’s likely grown a lot since, and they recently announced AZ7 will be online in Maryland soon, so they must be running out of ability to grow the ones in NoVA.
I can’t find a link now but it was one of the re:Invent talks like Peter DeSantis briefly explaining AZs before he dug into how Amazon optimizes their concrete mixtures to be more environmentally friendly or something…
All things point to that being the biggest region any hyperscaler has in the world, and several gigawatts of power consumption.
James Hamilton also gave a talk in 2021 about AWS having crossed 20 million Nitro cards deployed and 12GW power consumed —
Most hyperscalers now prefer to build larger sites as “campuses” which may consist of many buildings each consuming 40-100MW, and then yes each building needs most of what you mentioned, so it adds up.
A few sites are now also contemplating BTM or ‘behind the meter’ power generation which takes additional space.
Then some sites like Microsoft’s Fairwater design are optimized for a very large number of Accelerator cabinets — think GPU, TPU, etc. Those cabinets are each consuming 140kW today and with a path to 700-1000kW cabinets soon, so that’s one super dense building instead of a campus of less dense buildings filled with Compute.
Just as a heads up to the author, some of the commits against Luxury Yacht aren’t attributed to a GitHub account because Git wasn’t configured to use an email that’s associated —
From 62953f68b94e552596a149474c632c0ea0a05bf3 Mon Sep 17 00:00:00 2001
From: John Jeffers <[email protected]>
Date: Thu, 15 Jan 2026 21:07:51 -0700
Subject: [PATCH] add linux troubleshooting info