Part 4 of a five-part series on the structural opportunity in modular AI infrastructure

There is a pattern in technology infrastructure cycles that recurs reliably enough to be useful as a planning heuristic. The infrastructure built during the first phase of any major compute paradigm — the burst phase, the experimental phase, the phase where the model architecture itself is still being figured out — is almost never the infrastructure that turns out to be optimal for the second phase, the production phase, the phase where the workload settles into a steady-state economic equilibrium.

This was true of the mainframe-to-minicomputer transition. It was true of the on-premise-to-cloud transition. It was true of the desktop-to-mobile transition. And it is becoming true, right now, of the training-to-inference transition that the AI infrastructure market crosses through over the next eighteen months.

The investor implication is straightforward but worth stating directly: the infrastructure being built today, on the assumption that frontier-model training is the dominant workload, will not be the infrastructure that monetizes the production AI economy from 2028 onward. The geography, the form factor, the latency profile, the customer mix, and the unit economics of inference-dominant AI compute are materially different from those of training-dominant AI compute.

DCXPS’s modular distributed architecture is not optimized for the training-burst world. It is optimized for the inference-continuous world that follows it. And the 6-year SPV horizon described in Article 2 is calibrated to the exact transition curve where the value reallocates.

Here is the analysis behind that positioning.

The flip

The market-sizing data points that frame this transition:

• In 2025, training workloads accounted for approximately 62% of AI compute demand, with inference at ~38%.

• By mid-2027, inference workloads cross training in aggregate compute consumption.

• By 2030, the projected mix is inference at ~60–65%, training at ~35–40%.

• On energy footprint specifically, multiple independent estimates (IEA, McKinsey, Schneider Electric) converge on inference reaching 60% of AI energy consumption by 2027, despite using less per-query compute than training.

The shift from training-dominant to inference-dominant is not a marginal rebalancing. It is the largest workload-mix transition the data center industry has experienced since the mobile-first transition reshaped consumer compute in 2010–2015.

The mechanism driving it is straightforward and not particularly controversial. Frontier model training is, by its nature, a periodic event — a single model is trained once (or a small number of times for fine-tuning), and that training run consumes enormous compute for weeks or months. Inference is the productive deployment of those trained models. Every query, every API call, every agent step, every chatbot interaction, every image generated, every line of code completed — each one consumes inference compute. As enterprise AI deployment moves from proof-of-concept to production scale, inference volume grows by orders of magnitude even as training cluster size grows linearly.

The clearest signal that we are already inside this transition. NVIDIA’s most recent earnings disclosures and capital markets commentary explicitly call out inference as the dominant near-term growth vector for Blackwell deployment. AMD’s MI300X/MI350 product positioning is explicitly inference-first. Hyperscaler custom silicon (Google’s TPU v6, AWS Trainium-3, Meta MTIA v3) is increasingly inference-optimized rather than training-optimized.

The capital market is pricing the transition. What most operators have not yet internalized is what it changes about where and how you build the infrastructure.

What changes when inference takes over

Six characteristics of AI compute infrastructure change materially as the workload mix shifts from training-dominant to inference-dominant. Each has direct implications for site selection, deployment architecture, and unit economics.

1. Single-node performance becomes more important than aggregate cluster performance

A frontier training run requires tens of thousands of GPUs operating in synchronized parallel, with extreme demands on inter-node bandwidth and latency. Training infrastructure is dominated by the question of “how do I keep 25,000 GPUs lockstep at 99.9% efficiency for six weeks.” This is why hyperscale training clusters live in single-building campuses with InfiniBand fabric and meticulous topology engineering.

Inference is overwhelmingly single-node or small-cluster. A typical LLM inference request runs on 1–8 GPUs. A computer vision inference runs on 1 GPU. A recommendation model inference runs on a fraction of one. The aggregate compute is enormous, but the unit of compute is small. This means:

• Infrastructure can be physically smaller and geographically distributed without performance penalty.

• Inter-node bandwidth requirements are dramatically lower (you do not need 800 Gb/s InfiniBand for a single-node inference query).

• Hardware failures have local rather than systemic impact (one container down ≠ entire training run lost).

2. Geographic distribution becomes a feature, not a cost

Training infrastructure benefits from physical concentration. Inference infrastructure benefits from physical distribution. The reason is latency.

Modern agentic AI applications, real-time copilots, autonomous-systems control loops, robotics control planes, and consumer-facing AI products operate on sub-50-millisecond round-trip latency budgets. At the speed of light in fiber, that’s roughly 4,000 km of round-trip distance under ideal conditions — and real-world routing typically delivers 30–40% of theoretical, meaning a sub-50ms budget translates to 1,000–1,500 km of actual reach in practice.

The implication: a hyperscale data center campus in northern Virginia cannot serve a real-time inference workload in Prague, Madrid, or Helsinki at the required latency. Not even if the data center is fast, not even if the network is well-engineered. It is a physics problem, not an engineering problem.

This is why the inference geography looks structurally different from the training geography. Training can be done at the most power-efficient site available, anywhere on the planet. Inference must be done near where the inference is consumed — which means distributed, regional, and increasingly metro-edge.

3. Utilization curves change shape

Training workloads exhibit a “burst” utilization profile. A new model is announced; training begins; cluster utilization runs at 95%+ for weeks; training completes; utilization drops. The capital model has to monetize the high-utilization period sufficiently to amortize the deep-cycle equipment cost.

Inference workloads exhibit a “continuous” utilization profile. Customer queries arrive 24/7. The diurnal pattern is real (utilization is higher during business hours) but the dynamic range is much narrower — typically 60–85% utilization on a continuous basis, rather than 0–100%.

For modular operators, this is the friendlier curve. Continuous utilization is easier to forecast, easier to bill against contractual commitments, and produces more predictable EBITDA. The 96% margin profile described in Article 2 is structurally easier to defend against an inference-continuous mix than against a training-burst mix.

4. Hardware specifications diverge

Training-optimized hardware (H100, B200, B300 in their training configurations) is dominated by high-bandwidth memory (HBM3E, soon HBM4), extreme inter-node interconnect (NVLink, NVSwitch, InfiniBand), and maximum FLOPS per dollar.

Inference-optimized hardware emphasizes memory capacity (to hold larger models in single-GPU memory), memory bandwidth (for token generation throughput), and energy efficiency (because the workload runs continuously rather than in bursts).

The H200, in particular, is interesting in this context. Its higher HBM3E capacity (141 GB vs. H100’s 80 GB) makes it materially better than H100 for inference of larger models that did not fit in H100 memory. The B300 inherits this profile and extends it. Both of these chips are inference-relevant in a way that the original H100 generation was not. This is part of why our SPV unit composition is 14 B300 + 35 H200 — it produces a mix that is well-positioned for the workload transition over the SPV’s 6-year term.

5. Customer mix broadens

Frontier model training is concentrated among a small number of customers: OpenAI, Anthropic, xAI, Meta AI, Google DeepMind, a handful of Chinese labs, and the major European frontier labs. The customer count is in the low double digits.

Inference customers are every enterprise running AI in production. The customer count for the inference economy is in the millions. The pricing dynamic is different (smaller average ticket but more buyers), the procurement cycle is different (faster, transactional, often self-serve), and the loyalty profile is different (price-sensitive, performance-sensitive, multi-cloud by default).

For the Chapek platform — DCXPS’s bare-metal GPU cloud monetization layer — this transition is operationally favorable. We are built for the inference-dominant world: self-service provisioning, transparent hourly pricing, no orchestration tax, geographic distribution that puts compute near demand.

6. Energy economics change

The training workload favors absolute lowest energy cost regardless of location — which is why hyperscale training facilities locate near hydropower, nuclear, or stranded gas, and why Crusoe’s bitcoin-flare model worked.

The inference workload trades energy cost for latency. A site that is 20% more expensive on power but 200 km closer to demand may be the better economic site for inference because the value of the latency reduction exceeds the energy cost premium.

This is the dynamic that makes Central and Eastern European positioning particularly attractive for inference workloads serving the EU enterprise market. Czech power costs are competitive with German and French rates. The latency profile from Prague to Frankfurt, Vienna, Munich, Warsaw, Budapest, Berlin, and Amsterdam all fits within real-time inference budgets. The regulatory positioning (EU member state, EU AI Act jurisdiction, GDPR-compliant) is a structural advantage for the EU-sovereign demand pool described in Article 3.

What this means for capital deployed in 2026

The investment thesis question is not “is inference becoming bigger than training” — that is settled, the data are clear, the only debate is timing. The question is: what does an infrastructure portfolio look like that is positioned for the post-transition equilibrium, deployed inside the transition window?

Five propositions follow from the analysis above:

Proposition 1 — Distributed beats concentrated, in 6-year horizon terms

Capital deployed into single-campus hyperscale builds is exposed to the risk that the workload mix being optimized for is not the workload mix that exists at the operating end of the lifecycle. Capital deployed into geographically distributed modular capacity has the option to serve either training (with intra-site clusters of multiple containers) or inference (with single-container deployments serving regional demand) — and the architecture is flexible enough to shift between them as the mix evolves.

Proposition 2 — The training cluster premium is compressing

Hourly pricing for top-tier training clusters has been declining since mid-2024. The squeeze comes from two directions: hyperscaler internal capacity ramping up, and inference-economy demand absorbing the marginal GPU at lower per-hour rates than the peak training rate. The peak per-hour pricing for training-burst configurations may not return to 2024 levels. Operators whose capital model assumes those rates are exposed.

Proposition 3 — Hardware-fungibility matters

The B300 and H200 configurations in our SPV structure are not training-specific or inference-specific. They are hardware platforms that can be rebalanced toward either workload as customer demand evolves. The Chapek platform supports this rebalancing through its bare-metal provisioning model — capacity can be allocated to training tenants or inference tenants as the market mix dictates. Capital partners are not exposed to a hardware bet on one workload type.

Proposition 4 — EU latency geography is undervalued

The market is currently pricing AI infrastructure substantially on its training capability (power cost, cluster scale, interconnect bandwidth). It is not yet pricing AI infrastructure on its inference geography (latency reach, regional sovereignty, regulatory positioning). As the workload mix transitions, the EU-distributed-and-sovereign positioning will reprice. Operators positioned ahead of that reprice capture the value transfer.

Proposition 5 — The 6-year SPV horizon spans the transition perfectly

A unit deployed Q4 2026 begins commercial operation in 2027 — the year inference crosses training in aggregate compute. The unit operates through 2032. The first half of its operating life monetizes the late-training era; the second half monetizes the established-inference era. This is not a coincidence. The 6-year term was calibrated specifically to span the transition.

The technology arc — what 2028 looks like

For investors thinking 5+ years out, the technology arc that will shape the second half of the SPV term:

HBM4 deployment (2026–2027). NVIDIA Rubin, AMD MI400 series. ~15 TB/s memory bandwidth, vs. ~9 TB/s for current HBM3E. The performance uplift is meaningful for both training and inference, but the inference impact is more economically significant — it dramatically improves per-query throughput for large models.

Inference-specific silicon at scale (2027–2028). The economics of dedicated inference chips improve as the inference market matures and reaches a size that justifies dedicated silicon investment. Google TPU v7, AWS Inferentia-3, Meta MTIA v4, plus startup entrants (Cerebras, Groq, SambaNova) scaling production volume. This is a risk for general-purpose GPU revenue per hour — and a reason to think carefully about lifecycle planning at the unit level.

Mixture-of-experts and model efficiency (continuous). Architectural innovations are reducing per-query compute requirements by 30–50% for many workloads. This is offset by expansion of the addressable user base (Jevons paradox) but should be modeled as a moderating factor on aggregate accelerator demand growth.

Edge inference (2028+). 5G/6G-native AI workloads, autonomous vehicle compute, industrial IoT, smart-city infrastructure. This is a market growing from a small base but at significant CAGR. It is structurally aligned with the distributed modular thesis — these workloads cannot be served from hyperscale campuses.

The DCXPS roadmap accounts for each of these technology arcs:

• Generation 1 (current): 1 MW MADC units. Brownfield grid. Human technicians. Revenue from month five.

• Generation 2 (2027–2028): MADC+ at 2 MW. On-site CHP. First-generation robotics. 85%+ grid independence.

• Generation 3 (2028–2030): DDCU at 4 MW. Dark, distributed, dynamic. Lights-out compute. Software-defined energy management.

• Generation 4 (2030–2032): 10 MW DDCU. Humanoid robots in operational role. Integrated agricultural co-loads. EBITDA margin profile expanding toward ~93%.

• Generation 5 (2032–2034+): 100–200 MW “Robotic DC City” campuses. 200–500 humanoid robots. 5–10 human orchestrators.

The Generation 1 SPV that a capital partner commits to in 2026 is not exposed to the Generation 5 build — but the operator’s roadmap, the unit-level technology refresh paths, and the residual-value pathways do tie into the broader trajectory. Capital partners hold the option to participate in the next generation, but their exposure is locked at the unit-level economics of the generation they entered.

The operator framework

For DCXPS as operator, the workload mix transition creates a specific set of execution priorities for the next 24 months. I am laying these out openly because capital partners ought to be able to verify, in diligence, that the operator has thought about the right questions:

Priority 1 — Customer mix construction. The Chapek platform must build a customer mix that balances training tenants (for peak utilization revenue) with inference tenants (for utilization-stability revenue). Our target mix at 18 months of operation is 35% training / 65% inference, weighted by revenue. This is a deliberate mix; pure-training mix exposes us to the burst-cycle compression risk, pure-inference mix gives up the upside of peak training rates.

Priority 2 — Geographic expansion ahead of inference reprice. The Site 01 Kladno deployment is the anchor. Subsequent sites prioritize EU-distributed latency geography — central Czechia, Hungary, Poland, additional Western European positions where co-located power exists. Each site is structured as its own SPV pool.

Priority 3 — Hardware refresh discipline. The SPV operating term is 6 years, but hardware generations turn over in 18–24 months. The operator’s job is to optimize the workload mix on each generation as it ages — moving an H200-heavy unit toward inference workloads as B300/Rubin take over the training tenant base, and so on. This is the lifecycle management that capital partners hire us to do.

Priority 4 — Customer development at the regulated-enterprise tier. The inference customer mix that monetizes the EU-sovereign positioning is BFSI, healthcare, government, automotive, frontier AI labs (per Article 3). Each of these is a long-cycle enterprise sales motion. We have started; we are not done.

Priority 5 — Platform investment in audit, observability, and compliance tooling. The AI Act enforcement window (August 2026 onward) will commercially differentiate operators with mature compliance tooling from operators offering raw compute. Chapek’s roadmap includes auditability, training-data-traceability, model-versioning, and human-oversight integration tooling as differentiation points for the regulated-enterprise customer tier.

Where this leads

For capital partners, the inference-transition thesis is a confirmation argument for the modular distributed structure, not a contradiction of it. The 6-year SPV horizon is structurally aligned with the workload mix transition. The hardware composition (B300 + H200) is structurally aligned with the workload profile. The geographic positioning (EU-distributed, latency-optimized) is structurally aligned with where the inference economy actually consumes compute.

The DCXPS Site 01 deployment at Kladno is the deployment vehicle. Nine SPV positions. $45M unit size, $15M minimum participation. 2.75× modeled multiple, ~29.2% simple yield, 85/15 EBITDA waterfall. First units online October 2026.

For data-room access: investors@dcxps.com.

The final article in this series steps back from the operating thesis and looks at the asset-class framing. Why hardware-backed exposure to AI infrastructure is structurally different from platform equity, from tokenized compute claims, from listed neocloud equity, and from traditional infrastructure asset classes. And why, for a specific class of capital partner with a specific risk-return mandate, the SPV-owned modular structure is the only one of these vehicles that fits.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek bare-metal GPU cloud platform. Previous in this series: “Power Is the New Silicon,” “The 195-Day Data Center,” and “The CLOUD Act Conflict.” Final: “Own the Metal — A Framework for AI Infrastructure as an Asset Class.”

Every fifteen years or so, European technology policy produces a regulation that the rest of the world initially dismisses as continental bureaucracy, then quietly internalizes as global commercial reality. GDPR was the last one. The EU AI Act is the next one. Both were underestimated in their year of enactment. Both became, within five years, the binding compliance ceiling for any company that wanted to do business in the European single market.

The EU AI Act enters its decisive enforcement phase on 2 August 2026. As of this writing, the binding question is not whether the regulation will be enforced — it will be — but whether the European Commission’s Digital Omnibus proposal, which would defer the high-risk obligations until 2 December 2027, will be adopted before that deadline. The second political trilogue on the Omnibus, held on 28 April 2026, ended without agreement. The default outcome — the one prudent organizations are planning against — is enforcement from August.

Most coverage of the EU AI Act treats it as a compliance project: a set of documentation, audit, and governance obligations that EU-operating AI companies must satisfy to avoid penalties of up to €35 million or 7% of global turnover for prohibited-AI violations, or €15 million or 3% of global turnover for high-risk system breaches.

That framing is true, but it misses the larger structural force that the EU AI Act unlocks. The Act, combined with GDPR and the still-unresolved US CLOUD Act conflict, creates a market — not a compliance burden — for genuinely EU-sovereign AI infrastructure. That market is being capitalized in real time, the precedents are forming now, and the operators positioned inside it before August 2026 will be the ones that European enterprises and governments default to for the next decade.

This is the third pillar of the DCXPS thesis. It is why our first site is Kladno, not Atlanta. And it is the reason that the time-to-deploy question — the 195-day modular thesis from Article 1 — becomes more than an efficiency advantage: it becomes a regulatory window advantage.

The CLOUD Act problem, in three sentences

The US Clarifying Lawful Overseas Use of Data Act of 2018 (the “CLOUD Act”) permits US law enforcement agencies, with appropriate legal process, to compel US-incorporated cloud providers and their subsidiaries to produce data in their possession, custody, or control — regardless of where that data physically resides.

The EU General Data Protection Regulation (GDPR) requires that personal data of EU residents be processed only in conditions that protect fundamental rights, with restrictions on transfers outside the EU/EEA and protections against unauthorized governmental access.

A US-incorporated cloud provider operating an EU data center cannot, as a matter of statutory construction, fully satisfy both regimes. Data residency — i.e., the physical location of the server — is not the same as data sovereignty. The provider’s parent jurisdiction is the binding one.

This is not a theoretical concern. It has been litigated. The Court of Justice of the European Union’s Schrems II ruling (2020) invalidated the EU-US Privacy Shield specifically on these grounds. The EU-US Data Privacy Framework (2023) was constructed as a replacement but is now under renewed legal challenge for the same underlying reason: US surveillance law is structurally inconsistent with EU fundamental rights protections.

For most enterprise workloads, this is a manageable risk. For workloads involving regulated personal data — healthcare records, financial transactions, biometric identifiers, defense and intelligence applications, government services — it has become a binding commercial constraint. The customer cannot, as a matter of law, use a US-incorporated provider for that workload. Even if the data never leaves Frankfurt.

The EU AI Act, by introducing audit, traceability, and governance obligations on top of this, makes the question more acute. Who can the auditor compel to produce training data, model weights, system logs, and governance documentation? If the answer is “a US-incorporated entity subject to US legal process,” then the AI Act compliance position is structurally compromised — irrespective of where the GPU physically sits.

This is the gap that EU-sovereign AI infrastructure operators exist to fill.

The market is capitalizing this gap, right now

The clearest evidence that capital markets have priced in the sovereign-infrastructure thesis is the Mistral AI debt financing announced on 30 March 2026. The structure is worth examining in detail because it is the precedent that the next two years of EU AI infrastructure financing will be built on.

Mistral raised $830 million ($722 million) in debt — its first-ever debt financing — to fund a 44 MW data center at Bruyères-le-Châtel, south of Paris. The capital purchases 13,800 NVIDIA GB300 GPUs, with operations starting in Q2 2026.

The composition of the lending syndicate is the part most worth attending to. Seven banks: BNP Paribas, Crédit Agricole CIB, HSBC, MUFG, Bpifrance, La Banque Postale, Natixis CIB. Six European institutions plus one Japanese partner. No US bank participation.

That is not coincidence. That is structural signal. European banks are willing to underwrite AI infrastructure debt against EU-sovereign assets, EU-resident management, and EU-resident customer revenue. They are not extending the same terms to assets with US-jurisdiction exposure. The capital structure of European AI infrastructure is being deliberately ringfenced from US legal reach by the institutions that finance it.

For context on what this represents: Mistral’s annual recurring revenue grew from approximately $20 million to $400 million in a single year, with a $1 billion target by end-2026. The customer base includes the French armed forces, BPifrance, ASML, SAP, IBM, Cisco, Stellantis, and Accenture. These are not customers that can run sensitive workloads on US-incorporated cloud providers. They are the prototype of the European AI demand profile.

The broader pattern. In the first four months of 2026 alone, European AI infrastructure raised, by independent count:

• Mistral AI: $830M debt + ongoing Series C (€1.7B at €11.7B valuation in Sept 2025)

• Nscale (UK): $2 billion equity, alongside a reported $23 billion Microsoft contract for 200,000 GB300 GPUs

• Wayve: $1.2 billion (autonomous driving, GPU-intensive training)

• AMI Labs (Yann LeCun’s France-based lab): $1 billion

• Ineffable Intelligence (London, David Silver / DeepMind alum): $1.1 billion seed — Europe’s largest-ever seed round

That is over $6 billion of fresh capital flowing into EU-sovereign AI capacity in a four-month window. Add the European Commission’s €15 billion AI Factories program as the public-sector anchor, and the total addressable capital pool for EU-jurisdiction AI infrastructure stands at well over €20 billion in 2026 alone — against an installable capacity that is currently a fraction of what the demand profile requires.

The demand side — who is buying

The customers that the sovereign infrastructure thesis serves fall into five categories. Each has different price sensitivity, different latency requirements, and different procurement timelines, but they share the structural fact that they cannot use US-incorporated providers for their AI workloads.

Category 1 — National governments and defense

The European Commission’s Joint Procurement Office for AI Factories, France’s DGA and CNRS, Germany’s Bundeswehr Cyber and Information Domain Service, the UK’s MoD AI Lab, the Czech NCSA. These customers buy compute on multi-year frameworks with strict jurisdictional requirements. They are not price-sensitive in the conventional sense; they are sovereignty-sensitive. The contracts are large, multi-year, and dominated by relationship rather than spot-market pricing.

Category 2 — Financial services and insurance

European retail banking, insurance, capital markets infrastructure, payment processors. Workloads include fraud detection, risk modeling, anti-money-laundering compliance, algorithmic trading, credit scoring (specifically named as a high-risk category in Annex III of the EU AI Act). GDPR exposure on customer data is the binding constraint; AI Act compliance is the emerging binding constraint. Both push these workloads out of US-incorporated providers.

Category 3 — Healthcare and life sciences

EU-resident pharmaceutical companies (Sanofi, Bayer, Roche, Novartis, AstraZeneca’s EU operations), national health services, clinical research organizations, genomics platforms. Patient data sovereignty is the binding constraint. The AI Act adds requirements around medical AI as a regulated product (Annex I). This is one of the highest-growth verticals in EU AI demand.

Category 4 — Industrial and automotive

The European automotive sector is rebuilding its AI infrastructure stack as the autonomy and digital-twin compute requirements converge. Mercedes, BMW, Volkswagen, Stellantis, Renault. None of these companies can route training data through a jurisdiction that may compel disclosure of competitive engineering data. Stellantis is already a Mistral customer for exactly this reason.

Category 5 — Frontier and applied AI labs

European AI labs (Mistral, Stability, Aleph Alpha, Helsing, Black Forest Labs, Wayve, Ineffable Intelligence) that have specifically chosen to build their commercial position on EU jurisdictional positioning. These are the customers most aligned with the sovereign infrastructure thesis because their own customer pitch depends on it.

If you sum the addressable spend across these five categories at a 2027–2030 horizon, you arrive at a market materially larger than European AI infrastructure can currently serve. McKinsey’s base-case projection of 125 GW of incremental global AI capacity through 2030 implicitly assigns roughly 20–25 GW to Europe, against a current pipeline that can realistically deliver less than half of that on conventional grid-connected timelines.

This is the gap. It is bigger than any single operator can fill. It is also denominated, almost entirely, in workloads that must sit in EU-jurisdiction infrastructure.

The regulatory mechanics — what August 2026 actually changes

For the operators, customers, and capital partners thinking about this market, the technical content of the AI Act’s August 2026 enforcement matters. Let me name the parts that have direct infrastructure implications.

Article 6 / Annex III — high-risk AI systems

The categories of AI applications that become high-risk on 2 August 2026 include:

• Biometric identification systems

• AI used in critical infrastructure (energy, water, transport)

• AI used in education and vocational training admissions and assessment

• AI used in employment, worker management, and access to self-employment

• AI used in access to and enjoyment of essential private and public services (including credit scoring and insurance pricing)

• AI used in law enforcement

• AI used in migration, asylum, and border control

• AI used in administration of justice and democratic processes

For each, the AI Act mandates: risk management systems, data governance, technical documentation, automatic event logging, transparency provision to deployers, human oversight, and accuracy and robustness requirements. These are not boilerplate. They imply infrastructure capability — the platform must support detailed logging, audit access, traceability of training data, model versioning, and reproducibility of inference outputs.

Article 50 — transparency obligations

Providers of generative AI systems and deployers of AI systems that generate or manipulate audio, image, video, or text must label outputs as AI-generated. The platform-level implication: provenance metadata, watermarking infrastructure, audit trails.

Articles 53–55 — General-Purpose AI Models

GPAI obligations have technically applied since 2 August 2025; the penalty regime begins in August 2026. For systemic-risk models (currently a small set, but the threshold will not stay where it is), additional obligations apply: model evaluation, systemic risk assessment, incident reporting, cybersecurity protections.

Article 99 — penalties

The penalty structure that the market is pricing against:

• Up to €35 million or 7% of worldwide annual turnover (whichever is higher) for prohibited AI practices.

• Up to €15 million or 3% of worldwide annual turnover for high-risk system violations.

• Up to €7.5 million or 1% of worldwide annual turnover for providing incorrect information to authorities.

For context: a US hyperscaler with global revenue around $250 billion is exposed to a maximum single penalty of $17.5 billion under the high-risk provisions. That is not a theoretical exposure number that gets priced at zero in board-level risk reviews. It is a structural reason to ensure that AI workloads with EU-exposure are running on infrastructure where the compliance position is genuinely defensible.

Why modular wins this race

The combination of (a) accelerating EU sovereign demand, (b) 2 August 2026 enforcement deadline, (c) ongoing capital availability for EU-jurisdiction infrastructure, and (d) the universal grid-bottleneck problem detailed in Article 1, creates a competitive landscape with a specific shape:

• Hyperscaler greenfield builds: too slow. The 36–60 month timeline does not survive the August 2026 deadline.

• US-incorporated neoclouds: structurally unable to claim sovereign positioning regardless of where the assets sit.

• Conventional EU colocation operators (Equinix EU, Digital Realty EU, Interxion): legally EU-resident but parent-jurisdiction-exposed; partial solution, not complete.

• EU-incorporated frontier AI labs (Mistral, Aleph Alpha): building their own infrastructure for their own workloads; not a marketplace.

• EU-incorporated operators with modular deployment capability at sites with existing power: the architecturally correct fit for the window between August 2026 and the next major capacity expansion.

DCXPS sits in the last category by deliberate design. Czech entity (DCXPS a.s.). EU-resident management plane. Czech and broader CEE operational footprint. Site 01 at Kladno, co-located with an existing 200 MW power plant — meaning we are not exposed to the multi-year transmission upgrade timeline that gates most greenfield builds in Western European data center markets.

Most importantly: the 195-day deployment timeline matches the regulatory window. A capital partner committing to a DCXPS SPV in May 2026 sees first units online in October 2026, full operating cadence by mid-2027. That timeline lands inside the AI Act enforcement window, ahead of the demand surge that the enforcement deadline triggers, and at a moment when the cost of capital for EU-sovereign AI infrastructure is still being established.

Two years from now, that cost of capital will be established, the premium that early operators captured will be visible in cap tables, and the question of “should we have moved on this in 2026” will have an answer.

The implementation framework

For capital partners, family offices, and strategic investors thinking specifically about how to position for the EU sovereign AI thesis, here is the framework I would apply:

Step 1 — Define the exposure. Is the goal direct asset ownership (modular SPV structures), platform equity (neocloud operator equity at the EU-resident operator level), or strategic positioning (anchor customer relationships with future capacity preference)? Each has different risk-return characteristics and different liquidity profiles.

Step 2 — Validate the jurisdictional structure. EU-incorporated parent entity, EU-resident management, EU-resident data plane, EU-resident customer revenue. Each of these is verifiable in corporate documents. Anything short of all four is partial sovereignty, not full.

Step 3 — Confirm the power pathway. As detailed in Article 1, this is where the operator’s claim of “we can deploy in months” stands or falls. Off-take agreements with documented term and indexing structure. Co-located generation or secured grid connection. Not “we are in conversations with utilities.”

Step 4 — Confirm the customer pipeline. EU sovereign demand is structurally underwriting this thesis. But not every operator has the customer development capability to convert that demand into contracted revenue. Demand structured commercial pipeline evidence — not press releases.

Step 5 — Confirm the regulatory positioning. Has the operator’s legal team mapped its position against the AI Act’s specific provisions? Is the platform’s logging, audit, and traceability capability mapped to Annex III requirements? This is increasingly a buyer’s diligence question; expect it to become a financing diligence question by Q4 2026.

Step 6 — Confirm the operator’s track record at unit-scale operations. Same diligence framework as Article 2 — years of experience, sites operated, uptime delivered. Sovereignty is necessary but not sufficient; operational competence is the binding constraint on actually monetizing the regulatory position.

For DCXPS, the answers to all six questions are documented and available in the data room. We are happy to walk through them line by line.

Where this leads

The Site 01 fleet at Kladno is the first commercial expression of the DCXPS sovereign infrastructure thesis. Nine modular unit positions, $405M of contracted capacity, 6-year operating term, ring-fenced Delaware SPV structure with hardware owned by capital partners and operated by DCXPS under the 85/15 EBITDA waterfall described in Article 2.

The Czech jurisdiction matters here. EU membership, GDPR application, EU AI Act application. The Delaware SPV structure provides US-investor familiarity with the legal vehicle while the underlying operating entity (DCXPS a.s.) ensures the EU jurisdictional positioning on the data plane.

For capital partners specifically thinking about the EU sovereign thesis — family offices with European mandate, sovereign-adjacent capital pools, strategic investors with European customer exposure — this is the deployment vehicle that exists now, operating now, with first revenue in October 2026.

The window closes on 2 August 2026. Operators that are deployed and operating before that date will be positioned for the demand surge that follows. Operators still in greenfield development will be priced out of the early commercial cycle, regardless of how attractive their long-term roadmap looks.

For data-room access: investors@dcxps.com.

The next article in this series examines the second large structural shift inside the AI compute market — the training-to-inference transition that crosses over by mid-2027, and why modular distributed deployment is structurally better suited to the inference-dominant world than the training-dominant infrastructure being built right now. That transition is what makes the 6-year SPV term horizon mathematically aligned with the next phase of the cycle.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek bare-metal GPU cloud platform. Previous in this series: “Power Is the New Silicon“ and “The 195-Day Data Center.” Next: “From Training Burst to Inference Continuous — Why 2027 Reshapes the Compute Map.”

Part 2 of a five-part series on the structural opportunity in modular AI infrastructure

In the first article in this series, I argued that the binding constraint on AI compute is power, not silicon — and that this structural shift creates a specific opportunity for modular deployment at sites where grid capacity already exists. That argument lives or dies on the unit economics. If the math doesn’t work, the macro is irrelevant.

So let’s work the math. Openly, transparently, with every assumption on the table. This is the analysis I would expect from any institutional capital partner doing their own diligence — and it is the analysis you should demand of any operator asking you to commit $15M, $30M, or $45M to a hardware-backed AI infrastructure SPV.

What follows is the actual unit-level economics of a single DCXPS modular AI data center, structured as a ring-fenced Delaware LLC SPV. The numbers are drawn directly from our Confidential Investor Memorandum (April 2026) and our public pricing at chapek.io. I will show you what the math says, where the sensitivities are, what the comparables look like, and where you should push us in diligence.

The asset

Each SPV unit comprises:

• 2 × 40-foot ISO containers, reinforced for structural and thermal envelope requirements

• 450 kW IT load at design PUE of 1.3 (total facility power: ~585 kW)

• 49 GPU servers total, split:

  • 14 NVIDIA B300 servers (Blackwell Ultra generation)

  • 35 NVIDIA H200 servers (Hopper generation, high-bandwidth-memory configuration)

• Mixed cooling: forced-air for H200 capacity, direct-to-chip liquid for B300 capacity

• Spine-leaf RDMA fabric (Cisco), high-throughput NVMe storage tier, out-of-band management

• Modular UPS, switchgear, busway (Schneider Electric), BMS telemetry, fire suppression, physical security

Total unit cost: $45 million, of which:

This is list pricing. It is also what we charge today on chapek.io for production capacity. Several layers of conservatism sit beneath that number that I want to call out explicitly, because no analyst worth their bonus believes a list-price assumption at face value:

1. Utilization assumption. The modeled revenue assumes 8,760 hours per year — i.e., 100% time availability. Our contractual benchmark is >95% utilization, but utilization ≠ availability. The math above is operationally optimistic at the revenue line, then conservative everywhere else. In practice, a blended 90–95% billable utilization is a more realistic central case; the 6-year aggregate reflects a blended ramp plus discount profile.

2. No reserved-instance discount modeling. Real customers buy multi-month and multi-year reservations at 15–30% discounts to spot. The modeled $25.28M does not net that out — meaning realized revenue at scale will sit below list. Counterbalanced by the fact that reserved revenue is contracted, predictable, and refinanceable.

3. No price decay modeling. GPU hourly rates compress over time. H100 hourly rates declined ~30% across 2024–2025 as supply came online. B300 will follow a similar curve. By year 4–5 of the SPV term, we expect blended pricing materially below year-1 list.

Net of these adjustments, our internal modeled central case for fleet revenue across the 6-year term sits in a $130–155M range against the $151.7M list-optimistic case. The 2.75× multiple holds across that range. The structure absorbs the price decay; the math does not require the upside.

1. Power cost is the dominant variable cost. At $222/MWh — a representative Central European industrial rate, embedded in a long-term off-take with a co-located generator — the unit consumes ~3.94 GWh/year. That’s $875,124 of electricity against $25.28M of revenue. ~3.5% of gross. This is the magic number that makes modular AI infrastructure economics work: power is cheap relative to the value of the compute it produces.

2. Personnel cost is near-zero at the unit level. The unit is monitored and operated through Chapek’s centralized NOC. Field intervention is exception-based, not continuous. The $108,000 annual admin overhead covers fractional NOC allocation, basic site presence, and corporate overhead.

3. There is no co-location margin layer. We are not paying a colo operator $200/kW/month for rack space. The container is the rack space, and the SPV owns it.

For comparison: a typical retail colocation revenue model runs at 20–35% EBITDA margins. Hyperscaler internal compute runs at 30–50% (estimated). The neocloud category, depending on operator, runs 35–65% on a gross-margin basis, with significant variation by lease structure. Bare-metal GPU-as-a-service at a power-co-located modular deployment, with no real estate lease and no colocation margin layer, structurally runs above 90% EBITDA before financing.

Sensitivity analysis. The number that breaks this model is not GPU pricing — it is power cost. At $222/MWh, electricity is 3.5% of gross. At $500/MWh, it becomes 7.8%. At $1,000/MWh — i.e., PJM at the cap — it becomes 15.6%, and the margin compresses to ~80%. Even at the cap, the unit economics are still extraordinary by every comparable benchmark. This is why site selection is everything.

The waterfall, in plain English

The SPV pays its capital partner first, then the operator. Specifically:

1. Gross revenue is collected by Chapek (the operating platform) and distributed to the SPV monthly.

2. OPEX (electricity, fixed admin) is deducted from gross to produce EBITDA.

3. EBITDA is split 85% to SPV capital partners / 15% to DCXPS as operator.

4. Project financing (if any) is deducted before the split in the SPV’s discretion. The model above is unlevered.

This is the headline number from the investor memorandum, derived openly. 2.75× of distributed cash on a $45M ticket over 6 years equates to ~29.2% per annum simple yield. On an IRR basis — which accounts for the cash flow timing across a 6-year build/operate cycle — the figure depends on the deployment ramp, but sits broadly in the 24–28% range for a single unit, before optionality value.

Where the optionality lives

The 2.75× multiple is the base case. There are three places where optionality sits, none of which is modeled in the headline number:

Optionality 1 — Refinancing. Once the unit is operating, the asset becomes financeable. A capital partner who put in $45M of equity can extract working capital by refinancing 50–70% of asset value into senior debt secured against the hardware and the cash flow. The cash extracted is redeployable into further units. This is the same structural play CoreWeave executed in scaling its $34B off-balance-sheet lease pipeline. At our scale, it is meaningful.

Optionality 2 — Hardware refresh and residual value. GPU hardware has measurable secondary-market value. H100s deployed in 2023 still trade actively in 2026 at 30–50% of original cost, depending on configuration. B300 will have a similar residual profile. At year 6, the SPV holds 49 servers of B300/H200 capacity with either (a) a secondary-market exit path, (b) a refresh-and-extend path with new hardware redeployed into the same containerized infrastructure, or (c) a repurpose path into inference-optimized configurations (more on this in Article 4). None of these residual paths is in the 2.75×.

Optionality 3 — Compute price inflation. The model assumes flat-to-declining GPU hourly pricing. If supply rationing intensifies — which the macro analysis in Article 1 suggests is likely — realized pricing could exceed model. We have not yet seen a market where modular co-located capacity systematically undersells the spot market over a multi-year period. The reverse is the historical pattern.

What it costs to be wrong

Every honest investor pitch should include the failure modes. Here are the four that matter for this structure, and what they look like quantified:

Failure mode 1 — GPU price collapse. If average GPU hourly pricing declines 50% over the 6-year term (vs. ~25% in our central case), the multiple compresses from 2.75× to ~1.9×. Still positive. Still better than most infrastructure asset classes. Not a wipeout.

Failure mode 2 — Power cost shock. If our co-located generator off-take is disrupted and we are pushed to grid power at PJM-cap-equivalent rates, EBITDA margin compresses from 96% to ~80%, and the multiple comes in around 2.3×. Still positive.

Failure mode 3 — Utilization collapse. If sustained utilization runs at 75% rather than 95% across the entire term — a scenario that implies a fundamental break in AI demand — the multiple sits around 2.0×. Still positive, but the IRR profile materially weakens.

Failure mode 4 — Operator failure. If DCXPS, as operator, fails to deliver contracted operational performance — uptime, billing, customer acquisition through Chapek — the SPV has remedies. The hardware is owned by the SPV. The container is owned by the SPV. The operator can be replaced. This is the structural alignment that ring-fenced ownership creates. You are not exposed to operator equity; you are exposed to the asset, with the operator as a service provider.

Where this model would break catastrophically is a scenario in which:

• AI inference demand collapses and

• GPU secondary market collapses and

• Power costs spike beyond model and

• The operator fails to execute and

• The structural EBITDA margin compresses below ~50%.

That stack of conjunctions is not, in any reasonable view, a base case. It is a stress-test scenario, and in that scenario most AI infrastructure exposure — equity in neoclouds, equity in hyperscalers, equity in GPU vendors — does materially worse than a hardware-backed SPV holding refinanceable assets.

How this compares

The 2.75× / ~29.2% yield profile sits inside a recognizable infrastructure-investment universe. To calibrate:

• Core US infrastructure (toll roads, airports, regulated utilities): 7–11% IRR, 1.5–1.8× over 7–10 years.

• Renewable energy project equity (utility-scale solar/wind): 8–13% IRR, 1.7–2.1× over similar tenor.

• Traditional retail colocation: 11–15% IRR, 2.0–2.5× over 7–10 years.

• Hyperscaler-leased build-to-suit data center: 13–18% IRR for development, lower for operated.

• Modular AI compute (DCXPS modeled): 24–29% IRR, 2.75× over 6 years.

The premium is real, and it is paid for in three forms of risk: (1) higher operator dependency than core infrastructure, (2) technology obsolescence cycle compression, (3) less established secondary-market liquidity for the underlying asset compared to commercial real estate.

That premium is what the cycle is paying right now to capital that can move at the speed of the modular deployment timeline. We do not believe it persists at this level indefinitely. That is the case for moving now rather than later.

The diligence checklist

If you are evaluating this structure — ours or any operator’s — the questions to ask, in order:

1. Who owns the hardware? If the answer is anything other than “the SPV directly,” the entire alignment story falls apart. Demand evidence of bills of sale, customs documentation, and SPV title.

2. Where does the power come from? Off-take agreements, term length, indexing structure, fallback provisions. A modular thesis without secured power is not a thesis.

3. What is the operator’s economic share, and is it paid first or second? If first, alignment is theoretical. If second, alignment is structural.

4. What is the GPU allocation pathway? Letter of intent vs. binding purchase order vs. delivered inventory. The further down that ladder, the less of a thesis you have.

5. What is the realistic revenue ramp? Linear, hyperbolic, S-curve? At what utilization, against what customer mix? “95% contractual benchmark” is a number to verify against contracts, not against models.

6. What is the exit mechanism? Refinance, resell to operator, resell to third party, wind-down with hardware liquidation. Any operator who cannot articulate at least three exit paths has not thought about your capital seriously.

7. What is the operator’s track record at unit-scale operations? Years of experience, sites operated, uptime delivered. This is not a venture bet. It is an industrial operations bet.

We are happy to answer all seven questions in writing, with documentation, in a structured data-room session. The first three are answered in the memorandum.

Where this leads

The Kladno site (Site 01) hosts 9 of these SPV unit positions. Total fleet capacity, at full deployment: $405 million of contracted infrastructure investment and a 6-year gross revenue potential of approximately $1.37 billion. First units online October 2026.

The minimum participation in any single SPV is $15 million. The full SPV size is $45 million per unit.

For data-room access, write to investors@dcxps.com.

The next article in this series moves to the third pillar of the thesis: why the August 2026 EU AI Act enforcement deadline — and the underlying CLOUD Act conflict — creates a structural opening for genuinely EU-sovereign AI compute that US-incorporated operators cannot fully address. That’s where the geographic positioning of Kladno stops being incidental and starts being decisive.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek bare-metal GPU cloud platform. Previous in this series: “Power Is the New Silicon.” Next: “The CLOUD Act Conflict — Why EU AI Sovereignty Is a Capital Allocation Problem, Not a Compliance Cost.”

There is a comfortable story about the AI infrastructure shortage that the market has been telling itself for two years. It goes like this: NVIDIA cannot ship H100s, then H200s, then Blackwell, fast enough. Hyperscalers fight for allocation. Neoclouds raise capital, get on the wait list, and as soon as the GPUs arrive, the revenue starts flowing.

That story is wrong now. It was already becoming wrong in 2024. By the time you finish reading this article it will be obvious that the binding constraint on AI compute is no longer the chip — it is the kilowatt-hour, the substation, and the five-year median wait to connect anything material to the North American or Western European grid.

This matters because almost every dollar deployed into “AI infrastructure” right now is being underwritten against the wrong scarcity. If you are a capital partner trying to deploy $15M to $200M into this cycle, the most expensive mistake you can make is to assume that the limiting reagent is silicon. It isn’t. It hasn’t been for a year.

Let me show you the numbers.

The 2,060 GW backlog

The Lawrence Berkeley National Laboratory tracks every project sitting in U.S. transmission interconnection queues. Its most recent dataset, updated through end of 2025, shows 2,060 GW of generation and storage capacity waiting to connect — alongside 408 GW that has already signed an interconnection agreement but still has not reached commercial operation.

For perspective: total installed U.S. generating capacity is roughly 1,280 GW. The waiting room is bigger than the entire grid.

Three numbers from that dataset matter for anyone thinking about AI infrastructure investment:

1. The median time from interconnection request to commercial operations has doubled — from under 2 years for projects built between 2000–2007, to over 4 years for projects built in 2018–2024, with a median of 5 years for projects coming online in 2023.

2. The completion rate is collapsing. Of all projects that entered queues between 2000 and 2018, only 14% of capacity ever reached operation. Roughly 80% of projects withdraw before they connect.

3. Late-stage withdrawals are rising. Even projects with executed interconnection agreements — projects that look operationally certain on paper — are now pulling out in greater numbers, triggering re-studies that further delay everyone behind them in the queue.

When the AP, McKinsey, or Bloomberg writes the next “AI is using too much power” headline, this is the actual constraint they are describing. Not the megawatt-hour. The procedural, political, and physical impossibility of getting new megawatt-hours connected to where the chips already are.

The PJM signal

The clearest market price for “power, now, in the place AI wants it” is the PJM capacity auction. PJM is the regional transmission organization that covers the Mid-Atlantic, including Northern Virginia — the densest data-center market on the planet.

The PJM capacity auction price moved from $28.92 per MW-day for the 2024–2025 delivery year to $329.17 per MW-day for 2026–2027 — a more than 11× increase in two years. The 2026–2027 auction cleared at the FERC-imposed price cap — meaning the true clearing price would have been higher had regulators not intervened.

PJM is telling you, in dollars per megawatt, what every operator in Loudoun County already knows: there is no slack left. The market is rationing what cannot be expanded fast enough.

Texas tells the same story in a different language. ERCOT’s large-load interconnection queue grew from approximately 56 GW to 205 GW between September 2024 and October 2025 — a 73% increase driven primarily by data-center requests. AEP, the utility serving much of that footprint, doubled its contracted large-load pipeline to 56 GW over a similar period.

These are not the demand curves of a market that has solved its supply problem. These are the demand curves of a market that has not yet noticed it is broken.

The McKinsey math

McKinsey’s April 2025 Cost of Compute report puts a number on the gap. In the firm’s base-case scenario, data centers will require $6.7 trillion in cumulative capital expenditure by 2030, of which $5.2 trillion is AI-specific. That base case assumes 125 incremental gigawatts of AI capacity added globally between 2025 and 2030 — bringing total AI data center demand to 156 GW.

In their upside scenario, the figure climbs to $7.9 trillion and 205 incremental GW. In their downside, $3.7 trillion and 78 GW. Even the downside assumes a doubling of current global data center capacity in five years.

Now hold that 125 GW number next to the U.S. interconnection queue’s actual delivery rate. Berkeley Lab’s data implies that the U.S. can realistically deliver 60–75 GW of grid-connected new capacity on current timelines. The rest of the McKinsey forecast — the part that justifies hyperscaler capex, the part that justifies neocloud valuations, the part that justifies the $500 billion Stargate program — assumes one of three things must happen:

• Grid interconnection reform unlocks 20–30 GW of stranded capacity. (Possible, but slow, and dependent on FERC and state regulators.)

• Hyperscalers self-generate behind the meter. (Happening — BCG estimates 30–50% of new capacity will be self-generated by 2030.)

• Operators deploy compute at sites that already have power, on a timeline measured in months instead of years.

That third path is where the modular thesis lives.

Modular as a grid bypass

A traditional hyperscale data center takes 36 to 60 months from greenfield to first commercial revenue. Add the average 5-year wait for grid interconnection (where one is needed beyond what the site can already deliver) and you are looking at the better part of a decade between capital commitment and first dollar earned.

The hardware that capital was committed against will have depreciated through two full GPU generations by then. The H100 that made sense to procure in 2024 is competing with B200 in 2025 and B300 in 2026; the operator who locked in H100 capacity on a 60-month build schedule is selling vintage compute at a discount to operators who deployed Blackwell on a 195-day timeline.

This is the structural insight that animates everything DCXPS does.

A modular AI data center — two reinforced 40-foot ISO containers, 450 kW of IT load, mixed air- and direct-to-chip liquid cooling, populated with 14 NVIDIA B300 servers and 35 H200 servers — can be deployed at a site with existing power capacity in 195 days from contract to first revenue. Phase 1 is 10–12 weeks of strategic planning, site selection, container reservation, and regulatory work. Phase 2 is 15–18 weeks of system integration: mobile DC installation, hardware setup, energy-system integration. Phase 3 is 6–8 weeks of commissioning, GPU server initiation, configuration, and platform connection.

Then 15+ years of operating cash flow.

The math behind this is not exotic. It is what every modular industrial deployment has done since pre-cast concrete: pre-fabricate the complex parts off-site, ship them to where the resource is, assemble in place, commission, operate. The novelty is not the modular concept — it is that, for the first time in the history of computing, the constraint that justifies modular deployment is real, structural, and measurable in dollars per MW-day.

The economics of co-location

Here is the framework that should drive site selection in the modular era, in order of priority:

Tier 1 — Sites with existing, owned grid connection at adequate capacity. Retiring fossil generation sites, decommissioning industrial facilities, large utility properties. Power is on the meter, permits are in place, transmission is sized. Deployment timeline: 90–195 days.

Tier 2 — Sites with co-located generation. Combined heat and power (CHP), biogas, hydro, nuclear, or large renewable installations with surplus capacity. The generator becomes the anchor; the data center becomes the off-taker. Deployment timeline: 6–9 months, contingent on commercial framework.

Tier 3 — Sites with secured but not-yet-built generation. Greenfield renewable projects with executed interconnection agreements and PPAs. Higher risk, longer timeline, but available on negotiated terms. Deployment timeline: 12–18 months.

Tier 4 — Pure greenfield. Don’t bother. This is the hyperscaler trap. Capital is committed against a timeline you do not control and a counterparty (the utility) that has no commercial incentive to move quickly.

Every category-leading neocloud in 2026 — CoreWeave, Crusoe, Nscale, IREN — is now executing some version of Tier 1 or Tier 2. Crusoe built its entire thesis on stranded gas at the wellhead. IREN converted bitcoin-mining sites that already had power on the meter. Nscale anchored at Norwegian hydro. The pattern is not coincidental. It is the only pattern that works in the post-interconnection-queue era.

Why we chose Kladno

DCXPS’s first site is Kladno, in the Czech Republic, 30 kilometers from Prague. The site is co-located with a 200 MW power plant. We have direct on-site interconnect — meaning the kilowatt-hour does not transit transmission infrastructure we do not control. We are within a fully GDPR-compliant jurisdiction, EU-incorporated, and positioned to serve the demand profile that the EU AI Act will define starting in August 2026 (more on that in the third article in this series).

The site supports 9 deployable modular unit positions. Each unit is structured as a ring-fenced Delaware LLC SPV. The fleet, at full deployment, represents $405 million of contracted infrastructure capacity — and roughly $1.37 billion of gross compute revenue over the 6-year operating term, based on Chapek’s current list pricing for B300 and H200 capacity.

The first units come online in October 2026. The bottleneck, for us, is not power. It is not permits. It is not GPUs (allocations are secured). The bottleneck is capital deployment velocity — and that is the negotiation we are having with partners right now.

What to watch

A few signals that will tell you which way this market actually moves over the next 18 months:

1. PJM’s next capacity auction. If the 2027–2028 auction continues at or near the cap, the rationing thesis is confirmed and behind-the-meter / modular economics get materially more attractive.

2. FERC Order 2023 implementation. The cluster-study reforms are well-intentioned but slow. If they begin to materially shorten queue times in 2026, hyperscale greenfield re-enters the calculus. If they don’t, modular continues to take share.

3. The DATA Act. The proposed Senate exemption for off-grid data centers from FERC oversight would dramatically accelerate behind-the-meter economics. Worth tracking quarterly.

4. Hyperscaler self-generation announcements. Every gigawatt of self-generation Microsoft or Meta announces is implicit confirmation that the grid path is closed.

5. Neocloud margin compression. If sustained capacity rationing continues, the operators who locked in power before the squeeze will see margins expand, not compress. Watch CoreWeave’s gross margin trajectory.

Where this leads

The five-part series this article opens will work through, in sequence:

• Article 2 unpacks the unit economics of a $45M modular SPV in forensic detail — the math, the waterfall, the sensitivities, the comparables.

• Article 3 examines why the EU AI Act and the CLOUD Act create a structural opening for EU-sovereign AI infrastructure that US-incorporated operators cannot fully address.

• Article 4 maps the training-to-inference transition that will reshape the compute geography by mid-2027 — and why modular’s 6-year horizon is built for exactly this transition.

• Article 5 frames AI infrastructure as a distinct asset class, with a hardware-backed exposure profile that does not exist in tokenized compute claims or platform equity.

If you are a family office, fund, or strategic capital partner thinking about how to participate in the next stage of AI infrastructure deployment, the questions to bring to the conversation are not “can we get GPUs” or “is the market real.” Those are answered. The questions are: whose power, on what timeline, at what cost of capital, with what exit profile.

We’re happy to walk through ours. The Kladno site has 9 unit positions. The minimum participation is $15M. The structure is a ring-fenced Delaware LLC SPV per unit, with hardware owned by the SPV and DCXPS operating under a 15%/85% EBITDA split, paid second.

For the data room and a structured introduction to the management team, write to investors@dcxps.com.

The factory is being built. The question is whether you own a piece of it.

Jiri Fiala is CEO and co-founder of DCXPS, building Tier 3 modular AI data centers and the Chapek bare-metal GPU cloud platform. Next in this series: “The 195-Day Data Center — A Forensic Walk-Through of $45M of Unit Economics.”

I have watched the shape of this moment arrive in four different industries. Rideshare companies grew until taxi commissions organized. Crypto miners expanded until utilities and legislators noticed the bills. Short-term rental platforms scaled until cities legislated back. The pattern is always the same, and it always surprises the industry experiencing it: technology scales faster than political accommodation, and then, suddenly, all at once, the accommodation ends.

For hyperscale AI data centers, that moment is now, distributed across 31 states and counting.

Interconnected Capital, tracked by investor Kevin Xu, maintains a live dashboard of data center development restrictions. The current count: 50 separate initiatives to limit, restrict, or block new data center construction. Bipartisan — Xu describes data centers as a “bipartisan punching bag.” The left objects on environmental grounds: water consumption in drought-stressed regions, carbon footprint, grid strain. The right objects on economic grounds: electricity rate increases, rural industrial encroachment, subsidies flowing to coastal technology companies. The objection meets itself from both directions and forms a wall that lobbying cannot dissolve.

When the opposition has no party affiliation, you cannot play the political arbitrage game. The objection is local, practical, felt in the utility bill. That kind of resistance doesn’t evaporate on an election cycle.

The practical complaints are not abstract or ideological. A 500 MW hyperscale campus consumes millions of gallons of water for cooling in regions where water is not abundant. It employs thousands during construction and then relatively few people permanently. It adds enormous demand to local grids, measurably raising rates for every ratepayer in the interconnect who was not consulted. It occupies vast tracts of land that counties had other plans for. The community absorbed the impact without a voice, and now the community is the voice.

The industry’s response — site selection arbitrage, incentive package optimization, permitting before the opposition organizes — worked when permissive jurisdictions were abundant. Fifty initiatives across 31 states means the arbitrage pool is actively, measurably contracting. The hyperscale model’s expansion premise is meeting a hard geographic constraint that is bipartisan, distributed, and accelerating.

The Modular Answer

A 5MW modular DDCU deployment at an existing industrial site has a political profile that is categorically different from a 500MW hyperscale campus. Smaller footprint. Negligible visual impact. Fraction of the water consumption. Grid draw that the local utility can absorb without rate increases. No new construction footprint visible from the county road. Employment at industrial scale, not the construction boom and bust cycle of a campus build. The community impact is livable. The local politics are navigable. The permitting timeline is months, not years of NIMBYism battles. The 50-initiative backlash is not a headwind against modular compute. It is a structural tailwind — the political system eliminating the centralized model’s geographic expansion option, one county resolution at a time.

The cathedral model requires vast, politically contentious, geographically fixed campuses that take years to permit, years to build, and then cannot move when the regulatory environment changes. The modular model deploys in months, at existing industrial sites, with a footprint too small to organize against. The geography is making the architectural argument that the physics, the grid, and the power bills were already making.

Fifty fronts. Thirty-one states. The tent was always faster than the cathedral. Now it’s the only architecture that can still find a place to stand.

NVIDIA and Corning announced this week: US-based optical connectivity manufacturing expanding by 10x. US fiber production capacity growing by more than 50%. Three new facilities in North Carolina and Texas. More than 3,000 new manufacturing jobs with decade-long commitment behind them. Corning — the company that made fiber optic cable for the internet, Gorilla Glass for your phone, and glass for Edison’s lightbulbs — does not make capital commitments like this speculatively.

“This partnership is proof that AI is not just a technology story. It is a manufacturing story.” — Wendell Weeks, Corning CEO, 1851. He means it in the industrial sense. So do I, in the distributed sense.

The obvious read: this expansion serves the hyperscale campus buildout. Virginia, Texas, Ohio — the data center corridors that have been consuming fiber to interconnect their GPU clusters. That’s true. But it is the less interesting half of the story.

The more structurally significant read: a distributed modular compute architecture requires optical connectivity not just within a campus but between geographically dispersed nodes. A DDC network spread across industrial sites, logistics hubs, carrier facilities, and edge deployments across a region needs the same ultra-dense fiber connectivity that hyperscale clusters use internally — but distributed across miles of geography rather than hundreds of meters of raised floor. The fiber expansion that NVIDIA and Corning are planning is not just for the centralized model. It is building the substrate for the distributed one.

The railroad analogy haunts every infrastructure cycle. The 1990s fiber boom annihilated its investors. The fiber itself carried the internet for 25 years. This expansion has a structural difference: NVIDIA is pulling demand-contracted signals into the supply chain. The factories are being built against named backlog. This is not 1999’s hallucinated future. It is a supply constraint being solved against a queue of already-purchased GPU clusters that need to talk to each other.

The Modular Answer

The modular distributed compute architecture has a fiber dependency that is, per-unit, more demanding than hyperscale — not less. Each DDCU node requires high-bandwidth optical connectivity to function as part of a distributed compute fabric. The challenge isn’t connectivity within a building; it’s connectivity between buildings, between sites, across the distributed geography of edge deployment. The Corning expansion is building the substrate that makes this architecture physically possible at national scale. The centralized model uses fiber to connect GPUs across 200 meters of raised floor. The distributed model uses fiber to connect nodes across 200 kilometers of geography. Both need the same glass. Only one of them has already built all the campuses it’s going to build.

AI of the Coast gives family offices and private equity firms an early-warning edge on AI infrastructure—so they can price risk better, see value shifts sooner, and move capital before consensus does.

ON.energy built something that should not be invisible. Their bidirectional AI UPS — a power conditioning system that sits between a data center and its grid connection — passed every stress test at the U.S. National Laboratory of the Rockies. The system protects the data center from grid instability. It also, and this is the directionally important part, protects the grid from the data center’s own oscillating demand. Bidirectional. The buffer between two chaotic systems that are currently sharing a wire and not getting along.

NERC issued a Level 3 grid alert for the same problem this technology solves — in the same week it passed its national lab tests. Timing is not coincidence. It is the market screaming at builders who are listening.

The industry’s instinct is to dismiss this as table-stakes infrastructure — a UPS is a UPS, we’ve had those for decades. That instinct is wrong and getting more expensive to maintain. The legacy UPS was a one-directional emergency backup: when the grid goes dark, the battery bridges the gap. The AI UPS is a continuous active intelligence layer: monitoring both sides, absorbing shocks in both directions, converting the chaos of frontier training workloads into signals the grid can survive. Different category. Different stakes.

I have funded infrastructure-adjacent companies through three market cycles. The picks-and-shovels play is always invisible during the gold rush and structurally unavoidable during the normalization. Grid buffers and bidirectional power conditioning are this cycle’s unglamorous essential. The companies building them are not receiving $3 billion seed rounds. They will matter more than many that are.

The Modular Answer

ON.energy’s AI UPS is not just a grid stability tool. It is the embedded intelligence layer that makes autonomous modular compute deployments viable at the edge. A DDCU deployed at an industrial site, a logistics hub, or a rural carrier facility cannot rely on the grid tolerance that a Virginia data center corridor campus takes for granted. It needs active power intelligence embedded in the unit — managing its own relationship with the local grid, absorbing local instabilities, presenting a stable demand profile to the utility. The bidirectional AI UPS is the nervous system of the autonomous modular node. ON.energy may not know they’re building the infrastructure layer for a distributed compute future. But they are.

AI training clusters are the opposite of this. They are chaos injected directly into the consensus machine. A frontier model training run can swing from near-zero to hundreds of megawatts in seconds when a job launches, then collapse back just as suddenly when a checkpoint is reached. The grid’s balancing mechanisms, built for the rhythms of steel mills and suburban dishwashers, have no analog for this.

NERC issued a Level 3 — their highest — for the PJM Interconnection. 65 million Americans. The charge is not that data centers use too much power. It is that they use it in ways the grid’s architecture was never built to survive.

NERC issued a Level 2 last year. The industry issued some reports, convened some panels, and kept building larger clusters in more concentrated geographic locations in the same handful of corridors that have been drawing data center investment since the 1990s. The Level 3 is the watchdog’s way of saying: the conversation phase is over. Summer is coming to the Mid-Atlantic. The first serious heat wave will coincide with data center load spikes and residential cooling demand simultaneously. Sixty-five million people share a grid with the machines.

The centralized model’s grid problem is not an engineering problem that can be engineered away within the centralized model. It is a concentration problem. When you put 500 megawatts of demand in a single location and it oscillates, the shock radiates outward through the grid at the speed of physics. The larger the concentration, the larger the shock. The more concentrated the buildout, the more severe the instability. This is not a fixable bug in the hyperscale architecture. It is a structural feature.

The Modular Answer

Distributed modular compute changes the grid physics. A 5MW DDCU deployment oscillating at the edge of a grid node creates a local fluctuation the node can absorb. Ten thousand such deployments distributed across the national grid create statistical regularity — the kind of aggregate load behavior the grid was designed to balance. The centralized model creates single points of grid shock. The modular distributed model creates distributed load that looks, from the grid’s perspective, more like the predictable human behavior it was built to manage. The architecture that doesn’t produce NERC Level 3 alerts is the architecture that doesn’t concentrate hundreds of megawatts in a single swinging load.

The deal happened because xAI migrated training to the adjacent Colossus 2. Colossus 1 sat humming, idle, a cathedral nobody was praying in. Anthropic’s users had been hitting usage limits for weeks — not edge cases, not power users gaming the system, but ordinary subscribers who paid for a product and found it rationed. The physics of the situation wrote the contract.

The most sophisticated AI assistant on the planet became a frustration at 2pm on a Tuesday. Not because the model failed. Because the cathedral ran out of room and there was no annex.

Strip the narrative of its drama. What the Memphis deal actually diagnoses is compute feudalism — a structural condition in which frontier AI companies do not own their critical infrastructure, cannot build it fast enough to meet demand, and therefore become tenants in someone else’s kingdom at the moment of maximum vulnerability. Anthropic discovered in May 2026 that its product destiny was temporarily hostage to the availability of a competitor’s idle cluster.

The sidebar buried in the announcement — that both companies are now discussing “multiple gigawatts of orbital AI compute capacity” — reads like the logical endpoint of the centralized model taken to its absurd conclusion. When you can’t get the land, the water, the power, or the political permission on Earth anymore, you propose building the data center in space. The escape hatch, designed in real time, in the press release.

The Modular Answer

Anthropic’s Memphis problem has a structural solution that doesn’t require renting from a rival or planning orbital infrastructure. Modular, distributed compute — deployable in months, not years — eliminates the single-point-of-failure that hyperscale dependency creates. A sovereign compute architecture built from DDCUs across multiple sites doesn’t run out of capacity in a way that hands your product quality over to whoever happens to have idle GPUs that week. The cathedral creates dependence. The distributed network creates resilience. One of these architectures had a crisis in May 2026. The other didn’t.

For builders watching this unfold: compute scarcity is not a temporary supply chain glitch that resolves in the next hardware cycle. It is the defining constraint of this era, and it rewards whoever solved for infrastructure sovereignty before the crisis arrived. The companies that will matter in 2028 are the ones that are not, in 2026, negotiating leases with their competitors.

Let me start with the historical record, because it is unambiguous. The internet created extraordinary wealth. The fiber optic infrastructure that made the internet possible — WorldCom, Global Crossing, 360networks — destroyed enormous amounts of investor capital in the collapse of 2000-2002. The companies that captured the internet’s economic value — Google, Amazon, Facebook, Netflix — came after the infrastructure was already built and largely written off.

Does this historical analogy mean AI infrastructure is a bad investment? It does not. The lesson of the fiber collapse is not “infrastructure is bad” — it is “infrastructure financed with the wrong capital structure at the wrong moment in the adoption curve is bad.” The fiber that was written off in 2002 carried the internet traffic that made Amazon and Google possible. The infrastructure was right. The timing and balance sheet were wrong.

Why This Infrastructure Cycle Is Structurally Different

The fiber collapse was driven by a specific confluence: massive capital investment ahead of demand that took 5-7 years longer to materialize than projected, financed with high-yield debt that required cash flow servicing before the demand arrived. When demand didn’t materialize on schedule, the capital structure collapsed even though the infrastructure itself was eventually proven necessary.

The AI infrastructure cycle has several structural differences that change the risk profile. First, demand is already materializing — not at full scale, but at accelerating scale with clear trajectory. The IEA reports data center electricity consumption grew 17% in 2025. That is not speculative. Second, the primary capital allocators in the current buildout are not leveraged startups — they are Microsoft, Google, Amazon, Meta, and Oracle, operating from strong balance sheets with the ability to absorb timing risk. Third, the workload characteristics of AI create stickier demand than internet bandwidth did: switching costs are higher, workload criticality is higher, and the regulatory environment is creating durable barriers.

“The question is not infrastructure vs. applications. The question is: which infrastructure, at what point in the adoption curve, with what capital structure? The answer determines whether you end up like WorldCom or Eaton Power.”

The Utility Economic Model

The strongest case for AI infrastructure investment is the utility economic model — the same model that makes electric utilities, water utilities, and telecommunications carriers some of the most durable businesses in the world. Utilities have three characteristics that compound over time: recurring revenue from metered consumption, high switching costs for customers once connected, and pricing power derived from being a critical input to economic activity.

AI compute is acquiring all three. Cloud computing revenue is already almost entirely recurring, contract-based, and consumption-metered. Customers who have integrated AI into production workflows have switching costs that grow with the depth of integration. And compute is increasingly non-optional for organizations competing in an AI-native economy — which means pricing power accrues to whoever has the available capacity.

The infrastructure operators who position themselves as utilities — providing metered AI compute capacity, building durable customer relationships through SLAs and integration depth, and deploying capital at the bottleneck points in the supply chain (energy, high-density facilities, cooling) — are the ones with the most defensible long-term economic position.

The Application Layer Is Highly Competitive, Rapidly Commoditizing

The contrast with the application layer is stark. The AI model market is intensely competitive, with Anthropic, OpenAI, Google DeepMind, Meta AI, and Mistral all competing at the frontier. Inference costs per token have fallen by roughly 99% since GPT-3’s debut in 2020. API prices continue to decline. The application layer is being commoditized from below by open-source models and from above by hyperscaler bundling.

This is not an argument that application companies are bad businesses. It is an argument that their economic moat is narrower than widely perceived, and that the infrastructure layer — power, cooling, compute delivery — has a more durable competitive position by virtue of the capital intensity, long development timelines, and scarcity of appropriately-zoned, powered sites.

The operators building modular, energy-first AI infrastructure today are not guaranteed to win. Infrastructure investments can and do fail. But the structural economics — recurring revenue, high switching costs, scarcity of appropriately powered sites, and exposure to an adoption curve that is in its earliest innings — represent exactly the kind of asymmetric infrastructure bet that has historically produced the most durable returns in technology transitions.

A standard air-cooled data center is a thermodynamic system built around a simple principle: push large volumes of cold air through hot equipment, capture the heated air, cool it, and circulate it again. This system works well up to roughly 30 kilowatts per rack. The physics above that threshold are unforgiving: the amount of air required to cool a 50 kW rack generates so much turbulence, noise, and pressure differential that it becomes thermodynamically inefficient and operationally impractical.

The CRAC units — computer room air conditioners — that cool conventional data centers are sized for the facilities they serve. A facility designed for 15-20 kW average rack density has CRAC units, underfloor plenums, and hot aisle/cold aisle configurations calibrated for that density. Doubling the rack density in that facility does not simply require more cooling. It requires a different cooling system — and in many cases, different structural supports for that cooling system.

The Three Approaches to High-Density Cooling

The industry has developed three main approaches to cooling racks above the air cooling threshold, each with different engineering requirements, cost profiles, and density ceilings.

Rear-door heat exchangers attach directly to server rack rear panels and contain liquid-cooled coils that capture heat from the servers before it enters the data center hot aisle. They extend the effective cooling range to roughly 40-60 kW per rack without requiring changes to the server hardware itself. They are retrofittable to existing facilities and represent the lowest-barrier entry point to liquid cooling. Their limitation is that they rely on facility-level liquid cooling infrastructure — chilled water loops — that conventional facilities may not have.

Direct liquid cooling (DLC) routes coolant directly to heat-generating components — processor heatsinks, memory modules, power supplies. At the direct-to-chip level, DLC can handle 80-100 kW per rack effectively and is the current standard for high-density GPU clusters including Nvidia’s Blackwell architecture. DLC requires modified server hardware with liquid cooling manifolds built in, and facility-level plumbing infrastructure to supply and return the coolant. It cannot be retrofitted to standard servers.

Immersion cooling submerges complete server assemblies in tanks of non-conductive dielectric fluid — either single-phase (the fluid does not change phase) or two-phase (the fluid boils at low temperatures, capturing heat as vapor and condensing it back to liquid). Immersion is capable of handling 100-200+ kW per rack and is the leading approach for the highest-density AI deployments. It requires purpose-built tanks, specialized plumbing, fluid management systems, and floor structures capable of supporting the weight of fluid-filled tanks — far heavier than standard racks.

The Facility Design Implications

The transition from air to liquid cooling does not just change the cooling equipment. It changes the facility design from the foundation up. Standard raised-floor data centers with underfloor air plenums are the wrong starting point for immersion-cooled deployments. The floor must support immersion tank weight — potentially several tons per tank at high densities. The plumbing infrastructure must support closed-loop coolant distribution. The mechanical systems must be redesigned around liquid circuits rather than airflow volumes.

This is why new AI data centers are increasingly being designed from the ground up as liquid-cooling-first facilities — not retrofitted conventional data centers with liquid cooling added. The engineering economics favor starting fresh over retrofitting, particularly at the highest density tiers.

“The best data center for AI is one that was designed knowing from the start that it would never use air cooling. Every facility designed for air cooling first and liquid cooling second has the wrong architecture for the highest-value AI workloads.”

Modular Cooling as a Competitive Differentiator

The modular data center model has a specific advantage in the cooling transition: cooling systems can be specified and integrated at the module level, rather than as a facility-wide infrastructure. A modular AI compute unit can be designed with immersion cooling integrated into the module — the tanks, the coolant management, the heat rejection — as a self-contained assembly that deploys alongside the compute hardware.

Belden and OptiCool presented exactly this approach at Data Center World 2026: integrated rack-level infrastructure pairing connectivity, power, and modular cooling in a single deployable system. The key insight is that the cooling architecture follows the compute hardware specification, rather than the facility architecture determining what compute hardware can be deployed.

This is the opposite of the conventional model, where the facility’s cooling capacity is a fixed constraint that determines maximum rack density. In a modular liquid-cooled architecture, the cooling specification is set first based on the GPU hardware requirements, and the module is designed around it. The result is infrastructure that can match the density of whatever hardware generation is being deployed, without being constrained by a facility-level cooling decision made years earlier.

The cooling transition is not optional and it is not gradual. Above 30 kW, air cooling degrades performance. Above 50 kW, it is inadequate. Above 120 kW, it simply does not work. The hardware is already at 50+ kW for GPU clusters and heading toward 120-200 kW. The industry is not choosing liquid cooling as an efficiency upgrade. It is adopting it as a physical requirement for running the hardware that defines the frontier of AI capability.

In February 2026, Anthropic reported that Claude had reached $25 billion in annualized revenue — up from roughly $100 million in early 2024. In the same month, OpenAI reported revenues exceeding $25 billion annualized. The AI application layer is generating real money at real scale. The growth rates are extraordinary.

And yet — and this is the fact that most AI infrastructure analysis ignores — less than one percent of enterprises globally have meaningfully integrated AI agents into their core business operations. Not AI tools. Not AI assistants. Not employees using ChatGPT occasionally. Agents: autonomous AI systems that perform multi-step tasks, access live data systems, take actions in external systems, and operate without human intervention on each individual decision.

Why the 1% Number Matters More Than the Revenue Numbers

The distinction matters because of the compute differential. A copilot — an AI assistant that helps a human complete a task, like drafting emails or summarizing documents — is relatively compute-light. A single inference call generates a response and the compute is done. The typical enterprise copilot deployment adds perhaps 50-200 API calls per user per day to an organization’s compute footprint.

An agent is categorically different. A genuine AI agent — one that can autonomously research a topic, write code, test that code, identify errors, look up documentation, revise its approach, and deliver a working solution — may make hundreds to thousands of API calls to complete a single multi-step task. It runs planning loops. It uses tools. It checks its own output. The compute requirement per task is not 2x the copilot’s requirement. It is 10-100x, depending on task complexity.

The Math of Full Adoption

There are approximately 330 million knowledge workers globally. If 10% of them — 33 million people — are using AI copilots today, they are generating something on the order of 3-7 billion API calls per day collectively. That is already a meaningful compute load, and it explains a significant portion of the AI infrastructure buildout.

Now model what happens when agent adoption reaches 10% of knowledge workers — 33 million people using agents that make 1,000 API calls per task and complete 5-10 tasks per day. That is 165 billion to 330 billion API calls per day. Roughly 50-100x the current compute load. And 10% adoption of agents is still in the early adopter category. Full mainstream adoption at 50-60% would imply compute loads 250-500x what exists today.

“We are not in an AI bubble. We are in the pre-adoption phase of a technology whose full compute requirements, when realized, will dwarf everything currently built or planned.”

This is why I say the infrastructure buildout is not speculative — it is structurally necessary. The debate is not whether the compute will be needed. The debate is on what timeline, in what configuration, and with what architectural requirements. The people who say the hyperscale buildout is “overbuild” are confusing the current adoption curve with the eventual adoption ceiling.

The Compute Requirements of a World With Agents

A world where AI agents are genuinely pervasive — where every knowledge worker has personal agents, where every enterprise process has autonomous AI assistance, where every application has an agentic layer — is a world that requires infrastructure orders of magnitude larger than what is being built right now.

The IEA’s projection of 1,000 TWh in data center electricity consumption by end of 2026 already sounds enormous. It will look trivial by 2030 if agent adoption reaches even 20% of knowledge workers. This is not speculative modeling. It is arithmetic applied to the compute differential between current copilot deployments and genuine agent deployments.

The infrastructure imperative is not about serving the AI market as it exists today. It is about building for the market as it will exist when adoption reaches the S-curve inflection point. That inflection point is not here yet — the sub-1% enterprise adoption figure makes that clear. But the infrastructure to serve it when it arrives takes years to build. The window to build infrastructure that will capture the adoption wave is now, not after the wave has already started.

This is precisely the dynamic that played out in telecommunications in the 1990s. The buildout that seemed excessive for 1999’s internet traffic was exactly right for 2005’s internet traffic. The problem was not the infrastructure investment — it was the timing and the balance sheet structure of the companies that made it. The infrastructure was needed. The capital structure wasn’t right.

Subscribe now

Northern Virginia is the largest data center market in the world. The concentration of hyperscale infrastructure in a region sometimes called “Data Center Alley” represents decades of accumulated investment, fiber network density, technical talent, and operational expertise. It also represents, as of 2025, a grid that is so saturated that Dominion Energy publicly acknowledged it could not meet power demand until at least 2026 despite adding over 3 GW of data center capacity.

The problem with Northern Virginia is not a failure of planning. It is a success of conventional planning applied too long past the point where the conventional model broke. When every data center in the region is competing for the same grid capacity, the same transformer supply chain, the same interconnection queue, the grid constraints become a ceiling on total deployment that no individual operator can break by deploying capital faster.

The Energy-First Inversion

The most interesting development in AI infrastructure strategy in 2025-2026 is the inversion of the site selection logic. Instead of asking “where is cheap land near fiber?” the leading operators are asking “where is available power?” And critically: not available power on the grid, but available power that can be accessed without joining a multi-year grid interconnection queue.

Stranded energy — power generation capacity that exists but cannot reach load centers economically — represents one of the most undervalued resources in the AI infrastructure buildout. West Texas has enormous stranded wind generation that cannot be economically transmitted to Houston or Dallas. The Permian Basin has associated gas that is currently being flared rather than used. Parts of the American Southwest have stranded solar generation that exceeds local transmission capacity. Hydro-rich regions of the Pacific Northwest, Scandinavia, and Iceland have generation capacity that exceeds grid export capacity.

Each of these stranded energy sources represents a potential AI compute deployment site where power is available without grid interconnection constraints. Crusoe Energy’s model — build AI data centers directly at stranded energy sources, bypassing the grid entirely — is the clearest expression of this thesis in commercial operation.

“The question is not ‘where do we want to put the data center?’ The question is ‘where is the power?’ Build the compute there. The data doesn’t care about geography. The electrons do.”

The Nuclear Pivot

For AI workloads that require 24/7 firm power — which is most serious AI training and inference deployments — intermittent renewables create operational complexity. You cannot train a large language model on a schedule that depends on whether the wind is blowing. This is driving a parallel development: the hyperscaler pivot to nuclear power as the cleanest, most reliable, highest-density power source available.

Microsoft has signed agreements with Constellation Energy to restart nuclear generation capacity. Google has signed agreements for power from next-generation small modular reactor deployments. Amazon has acquired nuclear-powered data center sites. The nuclear pivot is not primarily an environmental statement — it is an engineering statement. For 24/7 firm power at densities that AI workloads require, nuclear is the best match between power source characteristics and compute requirements.

The implication for infrastructure deployment strategy: the future AI compute clusters are not going to be in Northern Virginia. They are going to be in West Texas, the Permian Basin, the Columbia River corridor, Iceland, Norway, the areas surrounding planned nuclear deployments, and any other location where reliable, abundant, accessible power exists without a three-year interconnection queue.

Modular Infrastructure Enables the Energy-First Model

The energy-first deployment model requires infrastructure that can be deployed quickly at locations that were not previously data center sites — no existing fiber, no existing electrical infrastructure, no existing operational staff. That requirement is incompatible with hyperscale construction approaches. You cannot build a 500 MW campus in a West Texas stranded wind corridor in 18 months using conventional construction methods.

Modular data center infrastructure is the enabling technology for the energy-first model. Pre-fabricated modules can be manufactured at scale in controlled factory environments, shipped to site, connected to on-site power generation, and operational in 6-18 months from commitment. The Crusoe/Energy Vault powered shell model in West Texas is exactly this: modular compute co-located with dedicated power generation, operational in months rather than years.

Eaton and Siemens Energy’s collaboration on parallel-constructed power generation and IT infrastructure takes the same principle further: the gas turbine and the data center infrastructure are designed together, manufactured in parallel, and installed simultaneously. The result is a deployable, self-contained AI compute facility that can be sited anywhere accessible power generation is available — independent of grid constraints, fiber concentration, or conventional real estate economics.

This is the architecture of the next generation of AI infrastructure. Not cathedrals in data center alleys, but tents in energy corridors. Deployed where the power is, at the speed the demand requires, with the density the workloads demand.

Share

There is a number that almost nobody outside the data center engineering community is paying attention to, and it might be the most important number in AI infrastructure right now. Not the $400 billion in capex. Not the 100 GW of projected US capacity. The number is kilowatts per rack — and it is moving faster than any other metric in the infrastructure landscape.

In 2021, the average data center rack drew about 8 kilowatts of power. By 2024 that had risen to 17 kW. By early 2026, AI-specific racks were already exceeding 50 kW. Nvidia’s current Blackwell architecture is pushing toward 120 kW per rack. The next generation, anticipated for 2027-2028, is targeting 200 kW and above. That is a 25x increase in power density over seven years.

Why This Is an Architectural Discontinuity, Not an Upgrade

Consider what has to change when you go from 17 kW to 120 kW per rack. The power delivery infrastructure — cables, busbars, PDUs — must be redesigned from the floor up. At 120 kW, a standard rack draws roughly 500 amps at 240 volts. The cabling alone for a 1,000-rack AI cluster at that density represents a fundamentally different electrical engineering problem than a conventional data center.

The cooling systems change entirely. Standard computer room air handlers, which move large volumes of cool air through hot aisles, become physically inadequate above 30 kW per rack. At 50 kW, you need rear-door heat exchangers at minimum. At 120 kW, you need direct liquid cooling — coolant pipes routed directly to the heat-generating components. At 200 kW and above, the leading approach is immersion cooling, where the entire server assembly is submerged in a non-conductive dielectric fluid.

Immersion cooling requires a flooring system capable of supporting the weight of immersion tanks full of dielectric fluid — significantly heavier than standard server racks. It requires plumbing for the fluid circulation. It requires different sealing systems for the building envelope. A conventional raised-floor data center cannot simply be “converted” to immersion cooling. You would need to essentially rebuild the facility interior.

8 kW

Average rack 2021 — air cooling, standard facility

30 kW

Air cooling limit — above this, problems begin

120 kW

Nvidia Blackwell — requires liquid cooling

200 kW+

Next-gen target — immersion or direct-to-chip only

The Stranded Capital Problem

Here is where this becomes an investment thesis rather than just an engineering observation. The hyperscale data center buildout currently underway represents the largest capital investment in the history of computing. Microsoft is spending $80 billion in 2025 alone. Amazon, Google, Meta, and others are each committing tens of billions annually. A significant portion of that capital is being deployed into facilities that are being designed and built today, at 2024-2025 density assumptions, for facilities that will not be operational until 2027-2030.

By the time those facilities open, the workloads they will need to serve will require 120-200 kW per rack, not 20-30 kW. The facilities will be operational but structurally under-equipped for the frontier AI workloads. This is not a small mismatch. A facility designed for 25 kW average density will have roughly one-fifth the effective AI compute capacity per square foot of a facility designed for 120 kW density. That is the difference between serving 1,000 frontier AI racks and 200 frontier AI racks in the same building footprint.

“The density gap doesn’t just affect efficiency. It determines whether a facility can physically serve the workload at all. Air cooling a 120 kW rack doesn’t make it run slower. It makes it fail.”

The Modular Advantage at High Density

A modular data center has a fundamental advantage in a high-density world: it can be designed specifically for the density requirements of the hardware being deployed, rather than averaged across a multi-decade facility lifecycle. When a hyperscale operator builds a campus that will serve 50,000 racks over 20 years, they have to make assumptions about average density that will inevitably be wrong.

A modular deployment for a specific AI workload starts with the GPU specification and works backward to the facility design. What does a cluster of 10,000 Blackwell GPUs actually need? Specific power delivery. Specific cooling architecture. Specific network topology. The modular facility is designed to those specifications, deployed, and operated. When the next GPU generation arrives with different specifications, new modular units can be deployed alongside or in replacement, without stranding the existing infrastructure.

This is the engineering logic that the market is independently converging on. At Data Center World 2026, multiple vendors were presenting integrated rack-level solutions — power, cooling, and compute in a single deployable module — designed explicitly for the density requirements of current and next-generation AI hardware. The modular model is not a workaround for organizations that cannot afford hyperscale. It is the architecturally correct answer to a density problem that hyperscale cannot solve without effectively rebuilding itself.

The 200 kW Horizon

The 200 kW per rack figure is not science fiction. It is what Nvidia and its OEM partners are currently engineering toward. At 200 kW, a single rack of AI accelerators draws enough power to supply roughly 150-200 average US homes. A modestly sized AI data center with 500 such racks would require 100 MW of dedicated power — equivalent to a medium-sized gas-fired power plant running at full output.

This is why the most forward-looking AI infrastructure operators are no longer looking at grid interconnection as their primary power strategy. They are building dedicated power generation — natural gas turbines, small modular reactors, fuel cells — co-located with the compute. Because at 200 kW per rack and thousands of racks, you cannot simply ask the local utility for another 100 MW. The grid was not designed for that kind of point load. The transformer lead times alone would delay your deployment by three years.

The density cliff is the physical forcing function that makes modular, energy-first infrastructure not just attractive but necessary. Hyperscale was designed for a world of 8-17 kW racks. That world is ending. The next world requires a completely different architectural approach.

Between late 2024 and March 2025, Microsoft canceled or deferred data center capacity agreements totaling more than 2 gigawatts of planned electricity consumption across the US and Europe. TD Cowen’s analysts discovered this through channel checks and reported it in a series of notes that sent Microsoft’s share price down and sparked widespread debate about AI overcapacity.

Most of the coverage got the story wrong. They framed it as evidence that AI demand was softer than expected — that Microsoft had overbuilt and was pulling back. That framing is superficially plausible but misses the actual signal entirely.

What Actually Happened

Microsoft’s data center cancellations were not primarily about demand being lower than expected. They were about architectural mismatch. The majority of the canceled capacity was associated with training workloads for OpenAI — specifically, facilities that had been planned or commenced before Microsoft’s relationship with OpenAI evolved and before Stargate, OpenAI’s joint venture with SoftBank, was announced with a $100-500 billion commitment to build its own infrastructure.

When OpenAI reduced its dependency on Microsoft Azure for training workloads, Microsoft was left holding capacity commitments for infrastructure that had been designed for a specific purpose — large-scale model training — that no longer needed that capacity. The cancellations were not a demand signal. They were an alignment signal: the demand existed, but it was going to a different piece of infrastructure.

The key distinction: Microsoft didn’t cancel because AI demand fell. It canceled because the specific workloads the canceled capacity was designed for — OpenAI training runs — moved to a different infrastructure provider. Global AI compute demand continued to grow throughout this period. The hyperscale architecture’s rigidity is what created the mismatch.

The Overcapacity Narrative Is Wrong, But the Structural Concern Is Right

Here is where the analysis gets genuinely complicated. Microsoft CEO Satya Nadella acknowledged in early 2025 that “there will be an overbuild” of AI infrastructure. CFO Amy Hood described Microsoft as working from a “capacity-constrained place” with shortages of both power and space. These two statements seem contradictory — overbuild and constraint simultaneously — but they make sense if you understand that different types of capacity are constrained in different ways.

The capacity that is being overbuilt is general-purpose cloud infrastructure optimized for the 2022-2024 workload mix. The capacity that is constrained is high-density AI inference infrastructure optimized for the 2025-2027 workload mix. You can have a massive oversupply of the former and a critical undersupply of the latter simultaneously — and that is approximately the situation the market is in.

“The hyperscale model’s rigidity is not just an engineering problem. It is a capital allocation problem. When you build a cathedral for one liturgy, the sunk cost prevents you from redesigning it for the next one.”

The $80 billion Microsoft committed for 2025 data center spending is real. The demand for AI compute is real. But a significant portion of that capital is going into infrastructure that will be structurally misaligned with the highest-growth segments of AI workload demand by the time it is operational. That is the actual risk — not that AI demand is softer than expected, but that the capital is flowing into the wrong architecture at enormous scale.

What the Pullback Reveals About Infrastructure Flexibility

The 2 GW cancellation reveals something important about the hyperscale model’s fragility: it cannot gracefully absorb demand changes.

When workloads shift — as they inevitably do in a technology transition this rapid — hyperscale operators are left with capacity commitments that cannot be repurposed without significant cost.

The cancel provisions, the lease break clauses, the wasted site preparation — these are the hidden costs of architectural inflexibility.

A modular deployment has a fundamentally different risk profile. If the workload for which a modular facility was designed changes, the modular units can be repurposed or redeployed with significantly less sunk cost.

This is not a marginal advantage. In a technology environment where workload requirements are shifting every 12-18 months, the ability to adapt without stranding capital is a structural competitive advantage.

The Microsoft pullback is the clearest evidence yet that the hyperscale model’s rigidity is not just an engineering constraint but a financial one.

Two gigawatts of canceled capacity represents billions of dollars in projects that were committed, designed, and in some cases begun before the demand signal changed.

The operators who avoid that scenario — by maintaining flexibility at the infrastructure layer — will compound capital more effectively over the AI infrastructure build cycle.

The market read this as a demand signal. It was actually a flexibility signal. And the implication is exactly the opposite of what the bearish coverage suggested: the demand is there, and growing.

The architecture that can actually capture it — modular, flexible, energy-first — is the one that needs to be built.

Microsoft, Google, Amazon, and Meta are spending hundreds of billions building hyperscale data centers. Most of that investment will be structurally misaligned with the AI workloads it was designed to serve before it comes fully online. This is not a minor inefficiency. It is a foundational architectural mismatch.

I want to be precise about what I mean when I say hyperscale data centers are architecturally obsolete. I do not mean they will stop operating or become worthless. I mean that the design assumptions embedded in the facilities being built right now — the power density per rack, the cooling architecture, the modularity, the time-to-deployment — are calibrated for a world of cloud computing workloads that is being rapidly superseded by a world of AI inference workloads. Those are fundamentally different physical requirements.

What a Hyperscale Facility Is Actually Designed For

A hyperscale data center is an extraordinary engineering achievement. We are talking about facilities that can house millions of servers, deliver gigawatts of power, maintain sub-millisecond internal latency, and operate at 99.999% uptime. The largest facilities — Northern Virginia, Singapore, Dublin, Tokyo — represent decades of accumulated infrastructure development, billions in sunk cost, and engineering excellence at genuine scale.

They were designed for traditional cloud workloads. Email. Storage. CRM systems. Web applications. Virtual machines. The kind of computing where a rack might draw 8-12 kilowatts of power, where air cooling is sufficient, and where you can pack thousands of general-purpose servers into a standard facility footprint.

Between 2021 and 2024, average data center rack power densities rose from 8 kW to 17 kW. That was already pushing the limits of air cooling in standard facilities. By early 2026, AI-driven racks frequently exceed 50 kW — and Nvidia’s Blackwell architecture is pushing toward racks that draw 120 kW, with the next generation aiming for 200 kW and beyond.

8 kW

Average rack density 2021

17 kW

Average rack density 2024

50+ kW

AI racks in early 2026

200 kW

Next-gen AI rack targets

A facility designed for 8-17 kW racks cannot simply be retrofitted to serve 120-200 kW racks. The power delivery infrastructure, the cooling systems, the structural load requirements — everything changes at those densities. You are not upgrading a car. You are trying to turn a highway into an airport runway.

The Construction Timeline Problem

Here is the second dimension of the cathedral problem, and in some ways it is more damaging than the density mismatch. Hyperscale data centers take a long time to build. From site selection to full operational status, you are looking at 4-7 years under normal conditions. Under current conditions — with transformer lead times exceeding 160 weeks and switchgear timelines similarly extended — you could be looking at 7-10 years for a large facility with complex power requirements.

The AI model generations cycle is currently running at roughly 12-18 months. By the time a hyperscale facility designed for today’s AI workloads comes online, the architectural requirements will have shifted at least two full generations. The Nvidia H100 was the dominant AI training chip when many of the facilities currently under construction were designed. By the time those facilities open, the industry will be on H200, Blackwell, or Rubin — each generation requiring fundamentally different power and cooling configurations.

This is not a hypothetical risk. Microsoft, one of the most sophisticated capital allocators in technology history, already canceled or deferred more than 2 GW of data center capacity in late 2024 and early 2025, according to TD Cowen’s analysis. The explicit reason: the facilities were designed around training workloads for OpenAI models. When that relationship evolved, the facilities became misaligned with actual demand. Two gigawatts of stranded capacity. That is not rounding error.

The Cooling Problem Is Not Solvable With Air

Standard air cooling systems work adequately up to roughly 30 kW per rack. Above that threshold, air cooling requires so much airflow that it consumes significant additional power itself, creates hotspots regardless of airflow volume, and begins to compromise the physical reliability of the hardware it is trying to cool. At 50 kW per rack — which is already common in AI deployments — air cooling is marginal. At 120 kW, it is non-functional.

AI data centers at the frontier are now deploying liquid cooling: direct-to-chip cooling where coolant pipes route directly to GPU heatsinks, rear-door heat exchangers, and immersion cooling where servers are submerged in dielectric fluid. Each of these approaches requires fundamentally different facility design. You cannot retrofit a standard air-cooled hyperscale facility for immersion cooling without effectively rebuilding it.

This means that a significant portion of the hyperscale capacity being financed and constructed right now will be structurally unable to serve the AI workloads that will dominate computing demand by 2028-2030. The facilities will not be worthless — traditional cloud workloads are not disappearing. But for the highest-value AI use cases, they will be the wrong tool for the job.

Why Modularity Is the Structural Answer

If the cathedral is architecturally locked to its original design, the answer is not a bigger cathedral. The answer is a tent — or more precisely, a deployable, reconfigurable modular infrastructure that can be designed, built, and commissioned in parallel with the hardware evolution, co-located with power sources rather than constrained by grid interconnection queues, and upgraded or reconfigured as workload requirements evolve.

A modular data center can be deployed in 6-18 months from commitment to operational status. It can be designed around the specific power and cooling requirements of the GPU generation being deployed, rather than averaged across a multi-decade facility lifetime. It can be positioned where the power is, rather than where the real estate is cheapest. And when the next GPU generation arrives with different requirements, the modular units can be retired or reconfigured without stranding billions in fixed infrastructure.

Crusoe Energy and Energy Vault announced exactly this approach in February 2026 — a partnership to deploy modular “powered shell” data centers designed for high-density AI compute, co-located with stranded energy resources. Eaton and Siemens Energy announced a similar collaboration, combining modular power generation with modular IT infrastructure designed for parallel construction and deployment.

The market is arriving at the same conclusion independently. The cathedral model is breaking. The tent model is emerging. The only question is who builds enough tents fast enough to capture the demand that hyperscale is structurally incapable of serving.

The key insight: Hyperscale facilities are not being abandoned — they are being outpaced by a workload revolution they were not designed to serve. The fastest-growing category of AI compute demand (high-density inference, edge AI, distributed training) requires exactly the characteristics that hyperscale cannot provide: speed of deployment, flexibility of density, and proximity to power rather than population centers.

Capacity market clearing prices for the 2026-2027 delivery year hit $329.17 per megawatt-day. The price for the 2024-2025 delivery year had been $28.92 per megawatt-day. That is a 1,037% increase in one year. In a capacity market that had been largely stable for over a decade. Rapid data center growth was explicitly identified as a major contributing factor.

$28.92

PJM capacity price 2024-25 delivery year ($/MW-day)

$329.17

PJM capacity price 2026-27 delivery year ($/MW-day)

1,037%

Price increase in a single year

160 wks

Current substation transformer lead times (up from 140 in 2023)

Why Grid Interconnection Is the Real Bottleneck

Building a large data center requires connecting it to the electrical grid at transmission voltage. This interconnection process — securing a queue position, completing the interconnection study, negotiating agreements with the utility, waiting for utility infrastructure upgrades — routinely takes 3-7 years under normal circumstances. In regions with high data center demand, the queue is measured in gigawatts of pending requests.

Dominion Energy, the dominant utility in Northern Virginia, publicly stated that it faced challenges meeting data center power demand until at least 2026 despite adding over 3 GW of data center capacity. That is a utility that has been preparing for this demand wave for years, with full visibility into the pipeline, acknowledging that it cannot keep up.

The physical hardware constraints compound the timeline problem. Substation transformer lead times have stretched from roughly 140 weeks in 2023 to more than 160 weeks in 2026. Switchgear timelines remain elevated. The electrical equipment market for data centers is projected to grow from $20 billion in 2026 to $65 billion by 2030 — but that equipment takes years to manufacture and install. You cannot accelerate the grid by writing a larger check.

The On-Site Generation Pivot

The most sophisticated AI infrastructure operators are responding to grid constraints by effectively exiting the grid for their largest deployments. Not partially — entirely. They are building dedicated on-site power generation, signing long-term offtake agreements with power generators, and in some cases financing the construction of new generation capacity themselves.

Cleanview’s February 2026 report projects that 30% of anticipated data center energy capacity will come from on-site generation sources — up from effectively zero just two years ago. The company’s founder forecasts that figure rising to 50% as the grid constraint becomes increasingly binding. This is not a fringe development. Microsoft has piloted using data center batteries for grid services in Europe. Constellation Energy has signed major deals to supply nuclear power directly to data centers. Exxon Mobil is supplying natural gas for on-site generation at scale.

“The smartest infrastructure operators have stopped waiting for the grid. They’re building power plants. The AI data center of 2027 will generate its own electricity the same way a 19th-century factory ran its own steam engine.”

What This Means for Infrastructure Deployment Strategy

If grid interconnection is the binding constraint — measured in years, not months — then the competitive advantage in AI infrastructure deployment shifts dramatically toward operators who have either secured power already or have the technical capability to deploy on-site generation alongside compute.

A modular data center co-located with a natural gas turbine, a small modular reactor, or a stranded renewable energy source does not need to join the interconnection queue. It generates its own power, connects to its own generation, and operates independently of the grid constraints that are crippling conventional data center development timelines. Crusoe Energy’s West Texas deployment with Energy Vault’s modular powered shells is exactly this model — AI compute co-located with energy sources that have no grid queue problem because they bypass the grid entirely.

This is a genuine first-mover advantage. The operators who figure out the energy-first deployment model — site selection based on power availability rather than real estate convention, modular facilities designed for on-site generation integration, the operational expertise to run both computing and power generation infrastructure simultaneously — are building a competitive position that cannot be replicated quickly by hyperscale operators locked into their existing grid-dependent development model.

The grid cannot keep up. That is not a temporary bottleneck that will resolve as utilities invest. It is a structural condition that will persist for at least the rest of this decade, given transformer lead times, interconnection queue depths, and the pace of AI compute demand growth. The operators who treat it as a permanent constraint to route around — rather than a temporary problem to wait out — will have a significant structural advantage.

Every economy runs on an invisible infrastructure nobody thinks about. We flick switches without thinking about power plants. We stream video without thinking about fiber. AI compute is becoming the third entry on that list — and the people who build the infrastructure, not the applications, will capture the structural value.

Nobody in 1890 was thinking about the power grid. They were thinking about the lightbulb. Edison’s invention was dazzling — but the money wasn’t in the bulb. It was in the infrastructure required to power the bulb at scale. The generating station. The transmission lines. The substations. The meters. The whole invisible system that made the light possible.

We are in the exact same moment with AI. Everybody is talking about the application — the chatbot, the agent, the model. Almost nobody is talking about the infrastructure required to run those applications at civilizational scale. And that is the asymmetric opportunity I have been building toward for the last four years.

The Numbers That Change the Framing

Let me give you the data that shifted my thinking. The International Energy Agency published its landmark Energy and AI report in April 2025.

The numbers are staggering. Global data center electricity consumption is on track to exceed 1,000 terawatt-hours by end of 2026 — equivalent to Japan’s entire annual electricity consumption.

In the US alone, Bloom Energy’s January 2026 infrastructure report estimates total data center energy demand will nearly double from 80 GW in 2025 to 150 GW by 2028. Driven almost entirely by AI.

Five large technology companies spent more than $400 billion on AI infrastructure in 2025. The IEA projects that number will increase by a further 75% in 2026.

This is not R&D spending.

This is physical infrastructure — steel, concrete, transformers, cooling systems, fiber, GPUs. The biggest capital deployment in the history of the technology industry.

And yet, the framing in most investment conversations is still about which AI application will win.

Which model will dominate.

Which company’s chatbot will become the interface layer.

That is like arguing about which lightbulb design is prettier while Thomas Edison is selling shares in the power station.

The Utility Analogy Is More Precise Than It Sounds

AI compute is acquiring all four of those characteristics right now!

A GPU-hour of inference compute is increasingly fungible across providers. It is already metered by every major cloud provider down to the millisecond.

The infrastructure to deliver it — data centers, fiber, power — is being built by a handful of large capital allocators.

And within five years, it will be as non-optional for serious enterprise operations as an internet connection is today.

“Nobody questions where the electricity in their office came from. Within five years, nobody will question where the compute running their AI systems came from. It will just flow from the grid.”

The difference between AI compute and electricity is the timeline.

It took electricity roughly fifty years to become a utility. AI compute will do it in less than a decade.

The reason: the demand is coming from enterprises that are already wired, already connected to cloud infrastructure, and already paying for compute.

The distribution problem is mostly solved. What isn’t solved — and this is where the real opportunity lives — is the supply side.

The Supply Side Is Structurally Broken

Here is the problem nobody in the mainstream investment conversation is grappling with honestly.

The supply side of AI compute infrastructure has a fundamental structural deficiency.

The current architecture — built on hyperscale data centers designed for traditional cloud workloads — is not capable of delivering what AI inference at civilizational scale actually requires.

Wood Mackenzie published a report in May 2026 projecting that US data center capacity needs to scale from roughly 24 GW today to 100 GW by 2030.

That is a 4x expansion in four years. Simultaneously, substation transformer lead times have stretched from 140 weeks in 2023 to more than 160 weeks in 2026.

Switchgear timelines remain elevated. Specialized cooling equipment for high-density AI racks is in short supply globally.

You cannot build a hyperscale cathedral in time. The lead times alone make it impossible.

A traditional hyperscale data center takes 4-7 years from site selection to full operational status.

The AI workload that data center was designed to serve will have evolved beyond recognition by the time it comes online.

This is not a hypothesis. It is already happening.

The Infrastructure Layer Is Where the Money Goes

I spent eleven years building enterprise solutions for Fortune 100 companies. Volkswagen, KBC, Mondelez. I have seen technology transitions from inside the organizations that have to absorb them.

And the pattern is always the same: the application layer is visible and gets all the attention, while the infrastructure layer is invisible and captures the structural economics.

Who made money from the internet?

Not the websites. The fiber. The routers. The CDNs.

Who made money from mobile? Not the apps. The spectrum. The towers. The backhaul.

Who is going to make money from AI at scale? Not the models. The compute. The power. The cooling. The interconnect.

The utility thesis is not about being contrarian for its own sake.

It is about following the physics. AI systems require electricity to train and electricity to run inference.

That electricity must come from somewhere. That compute must live somewhere. That somewhere — the physical infrastructure of AI — is the investment thesis. Everything else is derivative.

We are in the earliest innings of a transition that will make the buildout of the electrical grid look modest in ambition.

The question is not whether AI compute will become a utility. It already is one, just not yet at full scale.

The question is who builds the infrastructure, at what cost, and with what architecture.

That is what the rest of this series is about.

The DCXPS Thesis in One Paragraph: AI compute becomes a metered utility within five years. The infrastructure to deliver that utility must be modular, deployable at the speed of demand, co-located with energy sources rather than constrained by grid interconnection queues, and designed for rack densities that existing hyperscale facilities cannot physically support. The window to build that infrastructure is now. The window to invest in it is also now.

I want to be direct about something. Everything in this series — the density cliff, the grid constraints, the modular imperative, the energy-first deployment model — is not academic analysis. It is the thesis that underlies what we are building at DCXPS. I have been writing AI of the Coast for long enough that you know I call things as I see them, including when the data disagrees with my own positions. In this case, the data points in one direction and the thesis is what it is. So let me be concrete about what the destination looks like.

The DDCU Architecture

The Distributed Compute and Energy Unit — DDCU — is the architectural response to everything this series has documented. Start with the power constraints: grid interconnection takes years, transformer lead times are 160 weeks, the grid in established markets is saturated. The DDCU bypasses all of this by co-locating compute with dedicated power generation. Natural gas turbines, small modular reactors when available, stranded renewable integration — the power source is built alongside or brought to the compute facility, eliminating the grid bottleneck entirely.

Add the density requirements: current frontier AI hardware runs at 50+ kW per rack, heading toward 200 kW. The DDCU is designed from specification to power and cool that density tier. Liquid cooling is not an add-on. It is designed in from the first engineering drawing. The structural floor loading, the coolant distribution system, the heat rejection infrastructure — all specified for the density range the hardware will actually operate at.

Add the deployment speed requirement: AI demand is growing faster than any conventional construction timeline can serve. The DDCU is modular — pre-fabricated in factory conditions, shipped to site, connected to power generation, and operational. The target is 6-18 months from commitment to operational status, compared to 4-7 years for conventional hyperscale construction. That is not a marginal improvement. It is the difference between building infrastructure that serves current-generation hardware versus infrastructure that comes online into the next-next hardware generation.

The Robotic City Vision

The “Robotic DC City” framing is not marketing language. It describes an operational architecture that is a prerequisite for operating modular data centers at scale in remote or semi-remote locations where the energy is. You cannot staff a 100 MW compute deployment in West Texas at the same labor intensity as a Northern Virginia campus served by a deep technical talent pool. The economics do not work.

The answer is automation at the facility operations level. Robotic hardware replacement. AI-driven predictive maintenance. Autonomous cooling management. Remote monitoring with exception-based human intervention. The operational model of a DDCU deployment is not a data center with fewer people. It is a data center designed from the start for near-zero routine human presence — the same way a modern oil pipeline operates without someone physically checking every valve.

This is not science fiction. The technology for automated hardware management, predictive failure detection, and remote infrastructure management exists today and is deployed in hyperscale facilities at smaller scale. Applying it as the primary operational model for modular AI compute facilities co-located with power generation is the next architectural evolution.

“The Robotic DC City is not a data center with robots instead of technicians. It is an autonomous compute utility — the same way a modern power plant is not a furnace with engineers watching it constantly but an automated system with exception-based human oversight.”

The Distributed Network

The ultimate destination is a distributed network of DDCU deployments — not one large centralized facility, but many deployments each sized to the available power at their specific location, connected by a software orchestration layer that routes workloads to available compute capacity across the network. A customer accessing the network does not need to know or care which physical facility their inference workload is running on. The network routes it to available capacity, executes it, and returns the result.

This is the compute utility model made operational. The electrical grid does not tell you which power plant generated the electrons in your outlet. The compute utility does not need to tell you which DDCU facility executed your inference request. The abstraction layer handles the routing. The customer gets metered access to AI compute capacity, billed by consumption, without managing infrastructure.

The network effects compound over time. Each new DDCU deployment adds capacity to the network and makes the network more valuable to customers who need both compute availability and geographic distribution. The switching costs for customers deeply integrated into the network increase with integration depth. The capital required to replicate the network grows with each new deployment. These are the structural characteristics of a utility — and they are achievable in AI compute infrastructure in a way they were not in the fiber buildout, because the demand is arriving faster and the operational model is more capital-efficient.

The Open Question

I do not think we are guaranteed to win. Infrastructure investment is hard, capital-intensive, and takes longer than anyone projects. The hyperscalers have advantages in capital, customer relationships, and operational scale that no startup can match in head-to-head competition for the same workloads.

But the hyperscalers are not positioned to serve the workloads that the DDCU model is designed for: high-density AI compute deployed at the speed the market requires, at the locations where power is available, for customers who need compute capacity that is not available in the existing hyperscale footprint. Those workloads are real, growing, and structurally underserved by every architecture currently deployed at scale.

The utility transition in AI compute is happening. The only question is what architectures survive to deliver it. My bet — the bet I have been building toward for four years and will continue building — is that the energy-first, modular, high-density, distributed model is the architecture that serves the full demand. Not because it is elegant in theory, but because the physics and the economics leave no other option.

Your CTO signed the co-location agreement in March. The vendor promised power availability by Q3. It is now May and the utility just sent a letter.

The interconnection study will take another 14 months. Construction on the substation upgrade cannot begin until the transformer arrives. Current transformer lead times: 160 weeks.

Your AI deployment is not delayed by a competitor. It is not delayed by a model limitation or a talent gap. It is delayed by an electrical component that takes three years to manufacture and nobody ordered in time.

AI of the Coast covers the one constraint underneath all of AI that almost nobody in the mainstream conversation is tracking: where the power comes from, and who controls it.

This is the conversation happening in infrastructure decision rooms across the US right now. Not in theory. This week.

The Number That Should Have Stopped Everyone

In the spring of 2024, the largest electricity grid in North America held its annual capacity auction.

Capacity market clearing prices for the 2026–2027 delivery year hit $329.17 per megawatt-day. The year before: $28.92.

A 1,037% increase. In a single year. In a market stable for over a decade.

The grid operator cited data center growth as a major contributing factor.

Nobody treated it as the warning it was. Out of approximately 12 GW of US data center capacity announced for 2026, only around 5 GW is currently under active construction. The rest — billions of dollars of planned infrastructure — sits stalled. Not by a lack of capital. By a lack of electrons.

What Is Actually Happening Right Now

Companies are hitting an AI operations wall as projects scale from pilots to production. Technology leaders are facing an AI operational bottleneck — struggling to scale from isolated pilots to enterprise-wide implementations. In 2026, some AI workloads will not reach production at all.

Not because the models are not ready.

Because the power is not there.

Data center companies are looking for spare power capacity at every electrical utility they can find — including very non-traditional utilities and cooperatives whose doors have never been knocked on before. That is not a sign of a market finding creative solutions. That is a sign of a market running out of obvious ones.

Interconnection wait times have more than doubled over the past fifteen years. Projects now spend an average of five years in queue before reaching commercial operation. And the queue has a compounding problem: when one project withdraws or significantly changes its specifications, utilities must restudy remaining projects in the queue. This cascading effect means that even well-prepared developers can face years of additional delay through no fault of their own.

You can do everything right and still get caught in someone else’s problem.

The Three Conversations Your Peers Are Having

The co-location customer who just found out their facility cannot deliver.

They signed a contract. They budgeted for Q3 power availability. The facility operator just disclosed that the utility interconnection study is running 18 months behind schedule. The substation upgrade is contingent on transformer delivery. Transformers currently have lead times exceeding 160 weeks — up from 140 weeks in 2023. The operator is apologetic. The SLA does not cover force majeure grid delays. The legal team is now involved.

The AI workload that was supposed to go live in September is looking at mid-2028 at the earliest.

The enterprise that built its AI roadmap around cloud capacity that is not available.

High-end GPU access is becoming uneven and unpredictable — and expensive at exactly the moment when demand is exploding. While some organisations can lock in capacity years in advance, others are left refreshing dashboards and watching quotas tighten, forced to adjust their plans based on whatever compute happens to be available during the month or quarter.

The 2026 AI roadmap assumed on-demand compute availability. That assumption is now structurally incorrect in the markets that matter most.

The infrastructure team that discovered “powered land” after the fact.

For real estate developers and data center operators, securing powered land — sites with direct access to robust natural gas infrastructure — and deploying flexible gas-fired generation are no longer just options. They are strategic necessities. The teams figuring this out now are six to eighteen months behind the teams who figured it out in 2024. That gap compounds.

What the Smart Operators Did Differently

They asked the power question first. Before the site. Before the vendor. Before the contract.

Many of the facilities navigating this environment successfully rely on a mix of local grid power and independent sources — nuclear, renewable, and on-site generation. The mix is not ideological. It is operational. Whatever delivers firm power fastest, without a multi-year interconnection queue, wins the deployment race.

Cleanview’s February 2026 report projects 30% of new data center energy capacity will come from on-site generation — up from effectively zero two years ago. Forecast to reach 50% as the constraint tightens.

The model: co-locate compute with dedicated power generation. Build or bring the power source to the facility. Bypass the interconnection queue entirely. The AI data center of 2027 will generate its own electricity the same way a 19th-century factory ran its own steam engine.

Not because it wants to.

Because the grid left it no other option.

Share

The Bottom Line

Nearly half of all US data centers planned for 2026 have been canceled or delayed.

Not because AI demand softened. Not because capital dried up. Because the grid cannot deliver the power on any timeline that matches the speed of AI deployment.

The operators building durable positions this decade are the ones treating grid access as a permanent constraint to route around — not a temporary problem to wait out.

The queue is not clearing.

Build accordingly.

The DCXPS Thesis in One Paragraph: AI compute becomes a metered utility within five years. The infrastructure to deliver that utility must be modular, deployable at the speed of demand, co-located with energy sources rather than constrained by grid interconnection queues, and designed for rack densities that existing hyperscale facilities cannot physically support. The window to build that infrastructure is now. The window to invest in it is also now.

Jiří "Skzites" Fiala is a serial entrepreneur with 110+ startups built, invested in, or exited across B2B technology, enterprise software, and financial services. He is the founder of DCXPS, which builds deployable AI compute infrastructure for a grid-constrained world. AI of the Coast is the newsletter he writes because he cannot find the analysis he actually needs anywhere else — so he builds it himself. If that is the analysis you are looking for, you are in the right place.