The Black Box

At every town hall meeting this newsletter has covered, Festus, Rowan County, Monterey Park, Pine Island, Tucson, somebody eventually stands up and asks the same question.

"What's actually going to be in there?"

And at almost every one of those meetings, nobody answers it. The developer's representative talks about jobs and tax revenue. The economic development official is bound by an NDA. The opposition talks about water and noise. The building itself, the actual thing being fought over, stays a windowless abstraction.

That vacuum is where the theories live. Data centers are secret surveillance hubs. They're crypto mines in disguise. They're military installations. They're empty speculation plays. This newsletter's readers have heard all of them, because 116 municipalities' worth of public comment periods have surfaced all of them.

Issue #17 explained where the vacuum comes from: NDAs in 80% of Virginia deals, shell companies named Project Bigfoot, a Meta representative signing emails as "Ken Confidential." Secrecy manufactures speculation. That issue answered why nobody knows.

It never answered the question itself.

So this issue does. What follows is the verified, documented answer to what actually runs inside America's roughly 5,100 data centers, including the parts that sound like conspiracy theories and are simply true, the parts that sound obvious and are wrong, and a field guide to reading any facility from the outside.

No assumptions. Receipts throughout. Let's open the box.

Everything running in an American data center falls into five layers. Their relative sizes will surprise almost everyone on both sides of this fight.

Here is the single most important fact in this issue, from JLL's 2026 global outlook: AI represented only about a quarter of all data center workloads in 2025.

Read that again. After three years of AI dominating every headline, every earnings call, and every town hall fight, roughly three-quarters of what runs in data centers is everything else. Which means the honest answer to "what's in there?" is, mostly:

The overwhelming majority of data center capacity is the mundane plumbing of digital life, the stuff everyone already uses and nobody protests. A data center is mostly a warehouse for the boring parts of your own day.

But, and this is the honest caveat, that's the installed base. The new construction everyone is fighting about tilts heavily toward the next two layers. Which is why understanding the split between them explains almost everything about the map.

Training is the process of building an AI model, connecting tens of thousands of GPUs into one tightly coupled cluster and running them for months. Its physical profile, per the industry's own 2026 engineering literature:

That last point is the key that unlocks the whole map. Training campuses are the 500 MW+ giants landing in rural Texas, Louisiana, and Georgia, Meta's Hyperion campus in Louisiana, now growing toward a 5 GW AI supercluster, is a training facility. Crusoe's 900 MW Abilene build. The Stargate site. When a several-gigawatt campus lands in a county of 12,000 people, it's because training doesn't need to be near anyone. It needs megawatts, not neighbors.

In 2025, AI training consumed roughly 5 GW of data center capacity nationally. Projections put it at 23 GW by 2030.

Inference is the other half of AI, actually running the trained models every time someone uses them. Every chatbot reply, every AI search result, every copilot suggestion is an inference workload. And its physics are the opposite of training's:

Translate that: the gigawatt monsters get the headlines, but the growth story of the next five years is mid-size facilities pushing toward population centers, into exactly the metro and suburban zoning fights this newsletter covers every week. When a 40 MW facility wants to build near a neighborhood instead of in the desert, it's not because the developer is careless. It's because inference has a latency budget, and the speed of light doesn't negotiate.

One more verified number that explains developer behavior: moving a single petabyte of data out of a hyperscale cloud can cost upwards of $80,000 in egress fees. Data is expensive to move, so compute keeps moving closer to where data and users already are.

Now the part of the town hall question people are usually actually asking. Do some data centers serve intelligence agencies and the military?

Yes. Documented, contracted, and, here's the part that should permanently retire the conspiracy framing, publicly marketed.

The record:

So the accurate picture is this: classified government computing is real, large, and growing, and it lives primarily in dedicated, access-controlled regions operated by the same hyperscalers building everything else, plus purpose-built agency facilities like Bluffdale. These facilities are among the most disclosed things in the entire industry, because winning government business requires certifications, and certifications are public.

What the record does not support is the town-hall version: that the anonymous facility proposed outside your town is secretly one of them. Government and classified workloads demand specific security postures, compliance regimes, and contracting trails. They are the least likely tenants to show up unannounced behind a shell company, because their procurement is federal, documented, and auditable. The shell company hiding the tenant's name (Issue #17) is almost always hiding a brand-name commercial hyperscaler avoiding land-price spikes and PR attention, not a spy agency. The spy agencies, ironically, disclose more.

The rest of the inventory, briefly, each documented, each explaining a corner of the market:

Financial exchange colocation. The New York Stock Exchange's matching engines run in a data center in Mahwah, New Jersey; Nasdaq's in Carteret; CME's futures markets outside Chicago. Trading firms pay premium rents to place their servers in the same building, because physical proximity equals speed. The detail that tells you everything about this world: exchange colocation facilities famously provision equal-length cable runs to customer cages, literally coiling extra fiber, so that no firm gains a nanosecond advantage from being racked closer to the matching engine. Entire microwave tower networks exist between New Jersey and Chicago because light through air beats light through glass. This is one of the highest-revenue-per-square-foot land uses in America, and almost nobody at a zoning meeting has heard of it.

Content delivery. Netflix, Google, Akamai and others operate distributed caching layers, thousands of server deployments, many placed inside internet providers' own facilities, so the same popular video doesn't cross the country a million times a night. A meaningful share of "data center" capacity is just the internet's short-term memory.

Crypto and its conversion. Yes, some facilities were built as cryptocurrency mines. The 2026 story is that many are being converted: former mining operators have pivoted sites to AI hosting because the same cheap-power locations serve both, and AI pays better. The crypto data center is increasingly a transitional species.

Put the five layers together and the composite answer to "what's actually in there?" is:

Mostly the ordinary internet. Increasingly, AI, split between remote gigawatt training campuses and metro-bound inference. A real but publicly documented government layer concentrated in dedicated facilities. And specialty floors, finance, content, converted crypto occupying profitable corners.

Two verified numbers to anchor the picture:

The section for everyone who's ever driven past one and wondered. No two facilities are identical, but the physical signatures narrow the possibilities dramatically. Six things to look at:

1. Size and location. A 500 MW+ campus in a rural county with new transmission lines: almost certainly AI training or hyperscale cloud. A 20-100 MW building near a metro area, close to fiber routes: inference, cloud region, or colocation. A modest facility hugging a business district: enterprise colocation or network interconnection.

2. The cooling plant. Massive evaporative cooling towers and heavy water infrastructure: traditional hyperscale. Dense rows of dry coolers/air-cooled chillers and visible liquid-cooling infrastructure with minimal water draw: a strong signal of GPU-era AI design (remember: 80-100 kW racks made liquid cooling the baseline). The cooling tells you the rack density; the rack density tells you the workload generation.

3. The substation relationship. A dedicated new substation, or the facility built directly against a transmission corridor: hyperscale/training. Sharing existing distribution infrastructure: smaller, older, or lighter-duty use. (Issue #16 readers already know: the substation is the tell for everything.)

4. The security posture. Standard commercial fencing, badge gates, and a visitor lobby: commercial cloud or colo. Anti-ram perimeter barriers, guard forces, clear-zone setbacks, and no visible corporate branding: government or defense-adjacent work, which, again, will correspond to a documented government cloud region or agency facility, because that's how federal security requirements physically manifest.

5. The fiber story. Multiple diverse fiber entrances and proximity to major long-haul routes or internet exchange points: network-dependent workloads (cloud regions, inference, interconnection). A training campus can live comparatively fiber-light, its heaviest traffic is internal, GPU-to-GPU.

6. The tenant structure. Single-tenant, single-owner, no marketed space: hyperscaler self-build. Marketed suites and "powered shell" leasing: colocation/wholesale. An LLC with a codename and no website: a brand-name commercial tenant preserving negotiating leverage, the pattern Issue #17 documented across five states.

Clip this section. It works at a zoning hearing, on a site visit, and in a diligence memo, and it's the same fact set for every side, which is the entire point of this newsletter.

Texas ratings watch, validated. Six days after our inaugural State Risk Ratings put Texas at "Build (on downgrade watch)," the state ordered a statewide audit of AI data center projects. An audit is not a downgrade trigger by itself, it's a documented step toward the scrutiny our rating flagged. Texas remains Build; the watch is now officially earning its label.

Nashville used eminent domain to block a data center near its zoo, a genuinely novel tactic in the opposition toolkit, and one to watch for spread. Municipalities that can't zone a project away may start condemning the land instead.

The spending record. U.S. data center construction starts through June: $81.5 billion, already surpassing the full-year 2025 total of $72.5 billion, with half the year remaining.

Amazon reported Q2 amid plans for "Project Eagle," a four-building Texas campus, and signaled AI infrastructure spending will rise further this year, while still expecting capacity to trail customer demand. The $80B-can't-be-served dynamic from Issue #23 is now an industry-wide admission: earnings calls this quarter shifted language from raw capex to "time-to-energy" and speed of converting infrastructure into revenue.

Georgia keeps growing both storylines at once. Columbia County officials confirmed Google will occupy a proposed 8 million square foot, 23-building campus, the same county holding a November referendum on a funding structure that could eventually eliminate residents' homestead property taxes. OpenAI's "Project Camellia" also surfaced in Georgia. Meanwhile the state Senate's moratorium-plus-NDA-ban bill remains in play. Georgia stays rated Contested; both directions keep accelerating.

Cushman & Wakefield's global comparison confirmed the frame this newsletter has argued all year: 18 global markets now exceed 1 GW of operational capacity (Virginia leads the world at 11.3 GW), and grid availability, not demand, not capital, is dictating where facilities get built next.

Today's issue is the demand-side proof for the Map's thesis. The workload shift, 7 kW racks becoming 100 kW racks, liquid cooling as baseline, inference pushing into constrained metros, training campuses building their own generation, is a re-equipping of the entire American data center fleet. Every layer in the inventory above runs on the nine categories the Map covers.

Early access list is open: reply with "MAP" in the subject line. And the "RATING" replies from last week's launch are being reviewed for the Q4 edition, keep them coming, with receipts.

The most corrosive thing in this industry isn't a moratorium, a tax, or a lawsuit. It's the unanswered question at the microphone.

For four years, the industry's default answer to "what's actually going to be in there?" has been an NDA and a codename, and Issue #17 documented what that bought: 116 moratoriums, ballot campaigns, statewide freezes, and a public that fills silence with theories.

The verified answer was never dangerous. Mostly the ordinary internet. Increasingly AI, whose two halves, remote training giants and metro-bound inference, explain the shape and location of nearly every controversial project on the map. A government layer that is real, bounded, and more publicly documented than the commercial deals hiding behind shell companies. And specialty floors most people have never heard of, coiling equal-length cables so no trader gets a nanosecond edge.

An industry that explained itself this way at every town hall would still face fights over water, power, and land, those are real. But it would stop losing fights to its own silence.

Until it learns to, this newsletter will keep doing it instead. That's the job: where the risk sits, what the machines do, who builds it all.

The box was never locked. Nobody bothered to open it.

The DC Pipeline tracks data center construction, policy, and market intelligence across North America. Home of the State Risk Ratings.

Have a tip? Hit reply.

Think a State Risk Rating is wrong? Reply with the state and "RATING" bring receipts.

To join the Power Equipment Map early access list, reply with "MAP."

Know someone still guessing what's inside the building? Forward them this. They can subscribe at thedcpipeline.com.