Wyoming's Data Center Boom: Dawson Kluesner on Economic Opportunity, Public Backlash, and the Infrastructure Behind AI

Full analysis

The news here is a Wyoming economic-development official, Dawson Kluesner, making the case for the state as a data center hub, with Signal & Noise hosts Brett House and Rio Longacre cheering him on. Project Tembo, Google's planned Cheyenne campus, is the anchor: roughly 1.7 gigawatts of power and about $17 billion of initial investment, first phase targeted around 2030. Strip away the local-politics color and the question for an ad-tech operator is simple: what does the physical build-out of AI compute mean for the cost and reliability of the inference you're about to depend on?

How hard is this to undo? Nothing here is a decision you make. It's a read on where AI compute costs are heading and how fragile the supply is. Easy to revise as facts change, so no need to over-deliberate. But the planning assumption underneath it, cheap and abundant inference, is expensive to get wrong if you've built a product around it.

What's actually being decided: Whether operators should treat large-scale AI inference as a cost line that falls on a reliable curve, or as a volatile input subject to power, water, and local-politics bottlenecks that don't clear until 2030.

What sets the deadline: Project Tembo's first phase is a 2030 target. That's the horizon for when this new capacity actually serves load.

The Market Analyst. Follow the money, and it isn't flowing to anyone on this watchlist. Google is spending $17 billion to pour concrete in Wyoming so that OpenAI, Anthropic, and Google itself can run models. The value accrues to the chip and power supply chain, not to the DSPs and SSPs buying tokens downstream. For an informed outsider: the company selling you AI features is renting its compute from a landlord who just told you the building won't open until 2030. Second point: a 1.7 GW campus with a 2030 first-phase date tells you frontier compute is capacity-constrained now, which is why model pricing has been sticky rather than falling the way cloud storage did. Don't model your 2027 margins on inference getting cheap.

The Skeptic. The whole episode is two sponsored hosts talking their book. Brett House calls opposition "complete BS" and claims "up to 40% of data centers are defeated locally" with no source, plus a New York "statewide moratorium" that doesn't exist. Kluesner cites a 600,000-gallon water figure he admits he doesn't have in front of him. When the people telling you compute will be abundant are paid by the compute-adjacent economy, discount the optimism. The evidence cuts the other way: roughly 75% of registered voters concerned, 61% strongly opposed, bipartisan. That opposition is the constraint, and it's being waved away by exactly the people who need it to not matter.

The Operator. Here's the part that touches your Tuesday. Permanent staff at a hyperscale site runs 25 to 150 people, against construction crews of 1,800 to 5,000 for 18 months, per Brett House's figures. These are jobs machines, not job machines, and more to the point they take years to light up. If your product roadmap assumes a step-change in available inference capacity in 2026 or 2027, this episode is telling you the physical supply isn't there yet. Black Hills Energy's Large Load Tariff makes the data center operator fund all the new transmission and generation. That cost lands in the compute price you pay. Closed-loop cooling is real progress on water, but it doesn't speed up a power build.

The Customer / End User (the advertiser buying AI-driven media). Nobody in an agency is asking where the GPUs sit. They're asking why the shiny AI creative tool or bidding optimizer costs what it costs and whether it'll hold up under load. This episode answers both: the price reflects a capacity crunch that doesn't ease until the end of the decade, and the reliability of anything inference-heavy at scale is downstream of a power grid being fought over in town halls. For an informed outsider: the AI in your ad stack runs on electricity somebody has to build a reactor to supply.

The CFO. The interesting number isn't Google's $17 billion. It's the gap between 2030 first-phase completion and the 2026 and 2027 budgets operators are writing now that assume AI costs decline on schedule. If you've penciled in falling per-query inference costs to make an AI feature's margin work, this build-out timeline says that relief is years out. Virginia's math, per Brett House, shows $1.5 million in tax benefit per data center job against $1.2 million in incentives. That's the public-sector return. Your return as a token buyer is the opposite shape: you pay the amortized cost of a $17 billion campus baked into every API call, with no guarantee of the volume discounts that cloud buyers got a decade ago.

The tensions.

First, abundance versus scarcity. Brett House and Rio Longacre frame this as inevitable AI infrastructure flooding in. The Market Analyst and CFO read the same facts as evidence of scarcity: multi-billion-dollar campuses that don't produce power until 2030 mean constrained supply and sticky prices through the back half of the decade.

Second, local politics as noise versus local politics as the binding constraint. Brett House treats 61% strong opposition as hysteria to be managed. The Skeptic treats it as the one variable that actually decides whether gigawatts come online on time, and it's bipartisan, which makes it hard to steamroll.

Third, who captures the value. Kluesner, House, and Longacre all assume AI build-out is good for the AI economy. For an ad-tech operator specifically, the build-out is a cost you inherit, not a tailwind you ride.

What this hinges on. One belief: whether AI inference gets meaningfully cheaper per unit over the next two to three years. If yes, this episode is local color. If no, operators pricing AI features on the assumption of falling costs have a margin problem coming. The council leans toward scarcity. Gigawatt-scale projects with 2030 dates, tariffs that push all infrastructure cost onto the operator, and bipartisan opposition that slows permitting all point one way: compute stays tight and priced accordingly.

What to de-risk. If you're building anything inference-heavy into your 2026-2027 roadmap, don't assume the cost curve bends in your favor. Lock pricing where you can, keep a cheaper-model fallback, and treat any vendor pitch that promises AI costs "coming down fast" as the thing to verify, not accept.

Prediction: The published per-million-token list price for the flagship frontier model from at least two of OpenAI, Anthropic, and Google will be equal to or higher in October 2027 than it is today, measured against those companies' posted API list prices.

Confidence: Medium. Capacity is years out, but a cheap-model breakthrough could undercut it.

Why: This episode shows the physical supply of AI compute is gated on projects like Google's 1.7 gigawatt, $17 billion Cheyenne campus that don't deliver first-phase power until around 2030, while demand for frontier inference keeps climbing. When supply of the scarce input, power-backed GPU capacity, can't expand for years and demand rises, the price of the top models doesn't fall the way commodity cloud storage did. The cheaper-inference case usually comes from smaller cheaper models rather than the flagship tier, so list prices on the frontier models specifically are the harder thing to drop. The build-out timeline and the local opposition slowing permits both argue the crunch persists.

Revisit by 2027-10-09: We're right if the published per-million-token list price for the flagship model from at least two of OpenAI, Anthropic, and Google is equal to or higher than today's. We're wrong if both of those prices have fallen.

One caveat worth stating plainly: this is a call about the flagship tier. If your workload can run on a smaller model, you'll likely pay less over time regardless. The frontier is where the power crunch bites.