Seven data center projects along forty miles of the Columbia Gorge have asked for about 5,450 megawatts. The dam they sit beside makes about 800 on an average day. Airbrx can't change either number. What it can change is how much of that power ends up doing work somebody asked for.

The short film. The article below is the same story, with the numbers and sources.

The river

Bonneville Dam went into service in 1938, and the Columbia started making electricity. Nearly ninety years later, that cheap hydropower, along with cold water and high-speed fiber, is why the data centers came to The Dalles. Google has run a campus there since 2006.

Now six more projects on the Washington bank have asked for power.

ProjectLoad requestedWhere it stands
Lakeside Industries, Dallesport1,260 MWRequested from BPA in March; residential parcel, needs a conditional use permit
Dallesport Server Farm1,000 MWIndustrial orchard site; PUD records cite an initial load of 300–350 MW
Former Goldendale smelter958 MW2022 request; air-cooled; built in small increments, 2028–29
Energizer II1,000 MWBPA queue L0702
Vines Rest, east of Alderdale1,000 MWBPA queue L0704; Phase 1 scoping
Energizer Data Center78 MWBPA queue L0493
Google, The Dalles (existing)~154–260 MWOperating since 2006; about 40% of the city's water in 2025

That's 5,296 megawatts requested on the Washington side. With Google's campus, about 5,450.

The ask

Put that next to the river. According to BPA's 2026 White Book, The Dalles Dam makes about 2,080 megawatts flat out. On an average day in a median-water year, about 800. The data centers want more than 5,000, and they want it around the clock.

That's more than all four lower Columbia dams make together in a typical year: about 3,780 megawatts, on average.

And the region is already in the red. In a low-water year, Bonneville's own forecast has the Northwest roughly 770 megawatts short by 2029, and about 3,500 short by 2036, before any of these projects count. When the January 2024 cold snap hit, the Northwest leaned on 4,900 megawatts of imports to keep the lights on.

Follow one query

So where does a megawatt go? The fastest way to find out is to follow a single dashboard refresh.

It's 9:00 on a Monday. A sales manager opens the weekly revenue dashboard. The BI tool sends a SQL query to the company's warehouse, Snowflake or Databricks. The warehouse wakes up a cluster in a cloud region somewhere, maybe in a building like the ones above. It scans the tables, computes the answer, and sends it back. The meter runs.

At 9:04, her director opens the same dashboard. Same query, same data, all over again. At 9:11, the regional VP. At 9:30, six AI agents build a summary for the Monday meeting, and because agents double-check their numbers, each one asks three times.

Nothing in the underlying tables changed all morning. The warehouse computed one answer, and billed for it 21 times.

Fake load

Multiply that by every dashboard, every scheduled report and every agent loop, in every company.

That repeat work is real load. It lands on the customer's cloud bill every month. Then it lands again in the provider's demand forecast, as proof that the next building is needed.

The Northwest Power Council noted back in 2008 that data centers have flat hourly load profiles. Flat doesn't mean needed. It means the machines run whether or not anyone wants a new answer.

One hostname

The Airbrx gateway sits between a company's tools and its warehouse. Adopting it means changing one hostname. The SQL, the drivers and the logins stay the same.

Go back to Monday. With the gateway in place, the 9:00 query runs on the warehouse once, and the answer is saved under a key. At 9:04 the same question arrives, the gateway recognizes the key, and hands back the saved answer. Same at 9:11. Same at 9:30, for every agent and every double-check. The warehouse never woke up.

A sleeping warehouse doesn't bill, and it doesn't draw power.

When the data does change, through a pipeline run, a schedule, a rule edit or a write, only the answers it touches are cleared. The rest of the cache stays warm. The next ask for a cleared answer runs fresh on the warehouse and is saved again. A few more design choices make that safe enough for production:

  • Deterministic keys. The same user, statement and tenant always produce the same key. A hit returns byte-identical results, not a guess at a similar query.
  • Credential pass-through. Every query runs under the caller's own token, so existing row-level security and masking still apply.
  • Your storage. Cache, logs and rollups land in the customer's own bucket.
  • It scales to zero itself. The gateway is stateless. When nobody is querying, nothing is running.

Two kinds of megawatt

Back to the Gorge. Every building out there runs two kinds of megawatt. One kind answers a question somebody actually asked. The other keeps a server awake to answer the same question again, for nobody new. The buildout is being sized for both.

Take out the repeats, and the buildings that go up do more real work. A rack that spent its morning recomputing the same dashboard is a rack not serving new demand. Without the repeats, the same hall, the same cooling and the same grid connection carry work somebody asked for. The community gave up land, water and grid capacity for that building. What runs inside it should be work, not idling.

And the next building gets pushed out. Developers size buildings from forecasts, and forecasts follow what customers consume today. When thousands of customers stop paying for repeat compute, consumption drops and the forecast drops with it. At about $11 million a megawatt for shell and core, "later" is worth a lot. Sometimes "later" becomes "not needed."

The developers in the Gorge already build in steps. The smelter project plans to go up in small increments. The Dallesport Server Farm starts at 300 to 350 megawatts against a 1,000 megawatt ask. Scaling to demand rather than to the queue is the same logic Airbrx applies one query at a time. The less waste in the demand, the smaller each next step has to be.

Tuned while it runs

Finding the waste is the hard part. Without a gateway it's guesswork: the warehouse sees a query only after it has paid for it, and the fix lives in every dashboard and every agent, one at a time.

With the gateway in front, every request lands in one log. Iris, the Airbrx agent, reads that log around the clock. She spots repeats that aren't being cached, answers that expire too soon, and keys split too fine to ever hit.

Every fix she proposes becomes a rule, tested against your own last 30 days of traffic before it goes live. She catches patterns a person would miss until the bill arrived. But she sits beside the request path, not in it, and you decide how much she controls.

The waste keeps shrinking. Only new questions reach the warehouse, which is the work you want it doing.

What it doesn't do

An argument that pretends there's no trade-off is marketing. Here's what Airbrx can't touch.

  • AI training. A big share of the new load is training clusters running flat out. There's no repeat query to cache there. Airbrx works on analytics, dashboards, reporting and the agents that query warehouses.
  • First-time questions. A new query still runs on the warehouse. The savings come from repetition, and how much repeats varies by company. That's why the scan exists: measure before believing anyone's percentage, including ours.
  • Rebound. Cheaper compute invites more uses, and some saved megawatts will be spent on new work. That's fine if it's work someone wanted. The goal is zero wasted compute, not zero compute.
  • Scale. One company's savings don't move a hyperscaler's plan. Thousands of companies do. This only changes the Gorge in aggregate.
  • Sunk cost. Once a building is up, it draws power and needs cooling. The biggest leverage is before the concrete is poured.

Why is that machine on at all?

The Gorge is a balance sheet with hard limits: dams with fixed output, a region already forecast short, and the Dog River, The Dalles' water supply, running low in August. Software is the only part of the system that can bend.

FinOps asks what we're spending. The county asks how much they'll take. Airbrx asks a simpler question: why is that machine on at all? Answer it across enough companies, and the first two questions get smaller answers.

They're going to build data centers in the Gorge. Let every megawatt be one somebody needed. Scan your own query history to see your share.

Sources