Seven data center projects, current and proposed, are lined up along the Columbia within reach of The Dalles Dam. Before writing a word I went and found the reporting behind every one of them.

The short version: the Washington side of the Columbia is being asked to host more electrical load than the dam across the river can produce.

Google has been in The Dalles since 2006 and now pulls a big share of the city's water. Across the river in Dallesport, two separate developers want to put 2,260 MW of data centers on an orchard and a residentially zoned parcel. Further east, the old Goldendale aluminum smelter site, the Energizer substation, and an orchard near Alderdale have another 3,036 MW in the Bonneville Power Administration's queue.

The county doesn't have a data center ordinance yet. A motion for a six-month moratorium died for lack of a second, and the other two commissioners say they can have rules in place inside 90 days.

The receipts

Six of the seven projects trace back to a specific line-load request in the public record. Here's what the reporting says.

ProjectLoad requestedWhere it standsSource
Lakeside Industries (Dallesport)1,260 MWRequested from BPA in March; 434-acre parcel zoned single-family residential, needs a conditional use permitColumbia Gorge News
Dallesport Server Farm1,000 MWOrchard zoned industrial park; owner hired Cushman & Wakefield to test market interest; PUD records cite initial load of 300–350 MWGoldendale Sentinel
Former Goldendale smelter (STACK / IPI, now Blue Owl)958 MW2022 line-load request; air-cooled; construction estimated 2028–29 in small incrementsColumbia Gorge News
Energizer Data Center78 MWBPA queue L0493Goldendale Sentinel
Energizer II1,000 MWBPA queue L0702Goldendale Sentinel
Vines Rest (east of Alderdale)1,000 MWBPA queue L0704, received July 27, 2026; Phase 1 scopingGoldendale Sentinel
Google, The Dalles (Oregon, existing)~154–260 MW (estimates vary)Operating since 2006; about 550 million gallons of water in 2025, roughly 40% of the city's useDaily Tidings, OPB

The water story is just as real. The Dalles pulls Dog River water through a pipeline into South Fork Mill Creek and down to Crow Creek Reservoir, and the replacement pipeline roughly doubled daily diversion capacity. The city now wants 150 acres of national forest to expand the reservoir to about 3,000 acre-feet. City officials say the expansion predates Google and isn't for Google. Critics point out Google's water use went from 104 million gallons in 2012 to 434 million in 2024.

Put the numbers side by side:

  • The Dalles Dam is about 1,820 MW nameplate, per EIA data.
  • The Washington side has 5,296 MW proposed. That's 2,260 MW in Dallesport plus 3,036 MW further east (78 + 958 + 1,000 + 1,000). Nearly three times the dam.
  • Google adds existing load on the Oregon side. Estimates run from about 154 MW to nearer 260 MW for the whole Dalles campus. Count it in and the seven projects come to roughly 5,450 MW or more.
  • The smelter project may grow. Another Columbia Gorge News piece describes a 1,118 MW Blue Owl data center planned nearby, against the 958 MW line-load request.

Every megawatt in that queue is somebody's idle server

Let's do the math the way the developers do it. JLL puts shell-and-core construction at about $11.3 million per MW this year, before a single server goes in. At that rate, the 5,296 MW proposed on the Washington side is roughly $60 billion of buildings. That's before the substations and transmission, which under a 2012 Klickitat PUD policy the developers pay for themselves.

Nobody spends $60 billion on a hunch. They spend it on a forecast. And the forecast is built from what customers are consuming today, extrapolated.

Here's the problem. A huge share of what customers consume today is compute nobody asked for. It's the warehouse that wakes up to answer the same dashboard query it answered an hour ago. It's the dev environment nobody turned off on Friday. It's the Redis an AI agent added because production systems have a Redis. The Northwest Power Council noticed back in 2008 that data centers have flat hourly load profiles. Flat load doesn't mean steady demand. It means the machines run whether or not anyone needs them.

That waste shows up twice. First on the customer's cloud bill, every month. Then again in the hyperscaler's capacity plan, as a reason to break ground on the next building.

The flat-stack mandate goes after the first one, and the second one follows.

  • Do not wake the Beast. A query answered from a predictable key costs nothing to serve again. A warehouse that sleeps doesn't bill, and it doesn't draw.
  • Build it stateless. Build it to scale to zero. No idle compute burning money and carbon while it waits for traffic that never came.
  • The balance sheet is the product owner. If a component can't justify its cost, it doesn't get a server. Here the balance sheet also includes a river.

For the company paying the cloud bill, that's the whole pitch: the bill goes down. But zoom out. Every one of those customers is also a line in somebody's demand forecast. Shrink enough of them and the forecast shrinks with them. That's how you get the industry to build only what it needs. You don't fight the developers. You stop handing them the demand.

They're going to build. Let's make it count.

I said in The Cheapest Data Center Is the One We Don't Have to Build that scale-to-zero won't magically stop construction. Some of these buildings are going up. So the second question matters as much as the first: once a megawatt is built, how much useful work does it do?

A data center full of idle servers is the worst of both worlds. The community gave up the land, the water rights, and a chunk of the grid, and half the racks are keeping warm for workloads that could have been asleep. Flat-stack customers fix that from the inside. When the waste comes out, the same building serves more real demand, and the next building gets pushed out a few years. As I said last month, "later" is worth a lot at $11 million a megawatt.

The interesting thing is that the developers are already halfway there. The smelter project plans to start in small building increments rather than jumping straight to a gigawatt. PUD records for the Dallesport Server Farm put initial load at 300–350 MW, against a 1,000 MW request. That's the flat-stack instinct applied to concrete: scale to demand, not to the queue. A 1,000 MW request is an option on the future, not a need.

Klickitat County is writing its ordinance right now. If I were sitting in that workshop, here's what I'd ask for:

  1. Phase the interconnection to real load. Build the next increment when the current one is actually busy, not when the queue position comes due.
  2. Report utilization, not just capacity. Megawatts reserved tell you nothing. Megawatts doing work tell you whether the next building is needed.
  3. Shift the heavy work off-peak. The IEA describes demand flexibility as a way to get more out of the grid you already have. A dam has a fixed output. Big batch jobs can run at 2 a.m. when the Gorge isn't running its air conditioners.
  4. Don't cool what you don't run. A KPUD commissioner told the county the best cooling technology can run a data center without water, and that could go in the ordinance. Less idle compute means less heat to get rid of in the first place.
  5. Disclose water use from day one. The Dalles fought for years to keep Google's numbers secret. Other Oregon towns already publish what Intel and Apple use. Washington can start where Oregon finally ended up.

What it gives up

There's always something given up. An argument that pretends otherwise is marketing.

  • Training runs don't scale to zero. A big chunk of the new load is AI training, and a training cluster at full tilt has no idle time to cut. The flat stack moves the needle on inference, analytics, dashboards, and the enterprise sprawl around them. That's a lot of the load, not all of it.
  • Efficiency can feed demand. Make compute cheaper and people find more uses for it. Some of the megawatts we save will get spent on new work. That's fine if it's work somebody wanted. The goal was never zero compute. It was zero wasted compute.
  • Cold starts are real. A system that sleeps takes a moment to wake. For most workloads that's a second nobody notices. For some, it matters, and those keep their warm servers.
  • One customer's discipline doesn't move a hyperscaler's plan. This only works in aggregate. One company flattening its stack saves money. Ten thousand of them change a demand forecast.
  • Sunk cost is sunk. Once a building is up, idle servers still draw power and need cooling. The leverage is before the concrete, not after.

The Gorge is a balance sheet too

The Columbia Gorge isn't an abstract cloud region. It's a dam with a fixed output, a reservoir with a fixed size, a Dog River that runs low in August, and a rural county deciding what goes on their orchards.

Those are hard limits. Software is the only part of this system that can bend. So let's bend it.

FinOps asks:

What are we spending?

The county is asking:

How much will they take?

Scale-to-zero asks:

Why is that machine on at all?

Answer that third question across enough companies, and the first two get much smaller answers. The buildings that go up stay full of real work. The ones that never needed to exist stay orchards.

They're going to build data centers in the Gorge. Let's make sure every megawatt they build is a megawatt somebody actually needed.

Sources