Speed to Power — Why the most expensive resource in AI isn’t energy, it’s the date you can switch it on
Executive summary
The AI infrastructure market has gone through a shift in its binding constraint that most investment materials have not yet registered. Between 2021 and 2024, the bottleneck was hardware: servers, routers, chips, with lead times that had stretched to 52 weeks. From 2025 onward, the bottleneck moved from the server rack to the substation.
The practical consequence: capital is abundant, demand is confirmed by contracts, land exists, chips arrive — and yet billion-dollar projects sit still. Not because anyone said no, but because the grid operator has no slot to connect them.
In Europe’s traditional hubs, the connection queue averages seven to ten years, while the actual construction of a data center takes 18–24 months. In the United States, the median time to commercial operation is approaching five years, and for data centers it can reach twelve.
Romania is, right now, one of the few European markets where the equation can be closed in under twelve months. Not because it is more advanced, but precisely because it is less loaded: with 77 MW of installed IT capacity in 2025 versus 4,260 MW in Germany, the Romanian grid still has headroom where the rest of Europe no longer does.
The TerranOS Energy Equation
Alex Hormozi taught a generation of founders that value is not a list of features — it is an equation. Dream outcome times likelihood, divided by time and effort. Change one variable, and the whole thing moves. AI infrastructure works the same way, and almost nobody prices it that way.
Inspired by the Hormozi Value Equation ○ TerranOS.com
The TerranOS Energy Equation — six factors multiply in the numerator; time divides everything.
Why this equation matters to investors and energy developers
It is an underwriting tool, not a slogan.
One line prices any energy or AI infrastructure project. Rule 1 kills bad deals in minutes: if any of the six factors is zero — no signed power purchase agreement (PPA), no connection permit, no land title, no water permit, no contracted fiber, no term sheet — the asset’s real capacity is zero, whatever the pitch deck says. Ireland’s 5.8 billion euros of stranded projects would have failed this test on day one. Rule 2 prices the deals that survive: the same six factors delivered in 12 months are worth an order of magnitude more than the identical asset stuck behind a 7–10 year queue, because in AI, capacity available now trains models and capacity available in 2035 trains nothing.
For energy developers, it explains why “cheap MW” pitches fail.
The market does not buy electrons; it buys dates. Developers who document all six factors and compress the denominator sell at premium multiples — the build-versus-market spread of 400,000 euros per MW against 500,000–800,000 and above exists precisely because someone else closed the equation.
TerranOS.com is where the equation becomes investable.
The platform lists energy projects at three stages — ready to build, ready to implement, and already in operation — each with the six factors documented and the time to energization stated, and makes them accessible through asset tokenization. Instead of underwriting the equation alone, investors enter projects where it has already been solved; instead of hunting capital for years, developers list assets the market can actually price. That is the entire function of the platform: turning speed to power into a tradable asset.
The demand side of the equation is not a forecast anymore — it is a wall:
Redrawn by TerranOS from public data. Original data and charts: IEA — Energy and AI (charts under CC BY 4.0) and industry capacity projections. ○ TerranOS.com
The Energy Wall: global new power capacity demand reaches 1.6 TW by 2030 and ~1,650 GW by 2035; data centers alone need ~945 TWh by 2030 (IEA). Demand only grows from here — the numerator race is on, and the denominator decides who wins.
Part I. The six conditions that multiply
Sellable AI capacity is not a sum of components. It is a product. Six factors must exist simultaneously, and none compensates for the absence of another.
Orange = the binding factor today in Western Europe and the US ○ TerranOS.com
Fig. 1 — The six factors and the document that proves each one.
Each factor has a document that proves it: firm power — a signed PPA or permitted own generation; grid connection — an issued technical connection permit and a connection contract; land — a title deed or a registered option; cooling and water — a water management permit; fiber — two physically separate routes, under contract; capital — a signed term sheet and a team with a track record.
The rule is brutal and simple: if a single factor is zero, the product is zero. A perfect plot with fiber and water but no grid connection produces zero sellable MW. A fully permitted project, on purchased land, that cannot connect, produces zero — and that is not a theoretical example, it is the exact description of 5.8 billion euros of stranded investment in Ireland.
This is also the most frequent error in how energy projects are marketed to the AI industry: sellers pitch factor 1 (“we have cheap MW”) when the market pays for factor 2 (“when can I use them”). A grid connection with an energization date is the only factor that cannot be bought with money. That is why it has become the price of entry, and the industry has already named it: speed to power.
Part II. The size of the gap, in numbers
Sources: CBRE IM, JLL, EnkiAI, Data Center Watch via Forbes ○ TerranOS.com
Fig. 2 — The market in four numbers.
Occupancy
North America runs at 1.6% vacancy in primary markets, with 74.3% of capacity under construction already pre-leased (CBRE Group); JLL reports 1% for the second consecutive year and 92% of new construction precommitted. Europe is heading toward 6.5% by the end of 2026, an all-time low. Preleasing sits at ~75% against a historical norm of 40–50%. The sector effectively operates at 99% occupancy — the oversupply debate has no support in the data.
The flow
The gap accumulates in the interconnection queue: 2,600 GW at the start of 2026 ○ Sources: Introl, EnkiAI, Carbon Direct ○ TerranOS.com
Fig. 3 — The mechanism that keeps the queue long: demand structurally outruns delivery.
The queues
United States: 2,600 GW in the interconnection queue at the start of 2026, a median approaching five years, up to twelve for data centers. Projects that reached operation in 2025 had spent an average of eight years in the PJM queue (regional transmission organization for 13 US states). Nearly 80% of new projects withdraw, with network upgrade costs reaching 30–37% of total budget. European Union: between two and ten years depending on the country, seven to ten in the FLAP-D hubs (Frankfurt, London, Amsterdam, Paris, and Dublin), thirteen in the most congested markets. The Agency for the Cooperation of Energy Regulators (ACER) reports that the direct cost of grid congestion in the EU was 4.3 billion euros in 2024 alone.
The waiting capital
The European Data Centre Association estimates 176 billion euros needed cumulatively over 2026–2031, at 25–26 billion per year for AI-scale colocation alone. The constraint is not capital. It is the grid. And the result, in a single image:
Sources: IEA, Colliers via Prime East, DataCenter Forum Romania 2026 ○ TerranOS.com
Fig. 4 — The central contrast: 7–10 years in the hubs, 12 months on a site with the factors closed
Part III. United States — five documented blockages
Sources: PJM, ERCOT, Sightline Climate via Bloomberg, Data Center Watch, datacenterbans.com ○ TerranOS.com
Fig. 5 — United States: the blockages, case by case.
1. PJM — Virginia and 12 more states: a queue closed for years, reopened with 220 GW.
Data Center Alley, Ashburn / Loudoun County, Virginia
The largest grid operator in North America suspended new applications under a backlog that exceeded 300 GW. Its first reformed cycle, published in May 2026, drew 811 projects totaling 220 GW. In April 2026, PJM took an unprecedented step: it asked the Federal Energy Regulatory Commission for authorization to procure 15 GW of new generation on an expedited basis, outside the normal process — a formal admission that even the reformed queue cannot keep pace. In Loudoun County, Virginia — the densest concentration of data centers on the planet — data centers will generate nearly half of local tax revenue in 2026.
2. Electric Reliability Council of Texas (ERCOT): 198 GW in a single quarter.
Large-load connection requests totaling 198 GW came in during Q1 2026 alone, with 86 GW under review — roughly the peak demand of the entire state. The total large-load queue: 410 GW, of which 87% data centers. A single utility, CenterPoint, reported a 700% increase in requests within a year.
3. xAI Colossus 2, Memphis: locked in litigation.
The gas-turbine-powered project entered a Clean Air Act lawsuit. The Department of Justice intervened in June 2026 on national security grounds. The case is unresolved and it shows what happens when you bypass the grid queue with on-site generation: the air permit becomes the new bottleneck.
4. Meta Hyperion, Louisiana: 27 billion dollars under scrutiny.
Meta’s Hyperion, under construction.
Utility provider Entergy Louisiana requested the purchase of a ~1.8 billion dollar gas plant, largely attributable to the project. The question “who pays for new generation — the industrial customer or household ratepayers” has become a project risk in itself.
5. Announcement-stage projects that never start.
A 400 MW campus announced in Texas for “early 2026” is still seeking a partner and has gone quiet. Cloudburst San Marcos, announced as a 1.2 GW project with 50 MW live in Q3 2026, slipped to Q4 — with analysts noting the original timeline was unrealistic from the start.
In aggregate: roughly half of the capacity scheduled for 2026 in the US will be delayed or cancelled — of a ~16 GW pipeline, only ~5 GW is actually under construction. Cancellations rose from 2 in 2023 to 25 in 2025. Over 75 projects worth ~130 billion dollars were blocked or delayed in Q1 2026 alone. At least 69 local jurisdictions have enacted bans, including Seattle.
The detail that matters most to an investor: many of these projects would need only 12–18 months of actual construction. They sit frozen at the announcement stage, not the construction stage. Remember that number — we’ll come back to it.
The capital exists. The grid does not.
Part IV. Europe — five documented blockages
Sources: Avanza Energy, EirGrid/CRU, TenneT, IEA, Ember Energy, ACER, Enlit ○ TerranOS.com
Fig. 6 — Europe: the blockages, case by case.
1. Ireland: a four-year moratorium and 5.8 billion euros stranded.
In January 2022, EirGrid stopped new data center connections in Dublin. No timeline, no exceptions. The reason: data center consumption had grown 531% in a decade and reached 22% of national electricity demand — more than every home in Ireland combined. One AWS project and two Microsoft projects could not proceed; Amazon moved its build to London, Microsoft began evaluating London, Frankfurt and Madrid. The moratorium was lifted in December 2025, but replaced with a condition: any new connection above 10 MVA must bring its own dispatchable generation or storage. The moratorium became a mandate.
2. Netherlands: a national ban and a grid at its limit.
Electricity pylons in the Netherlands
A national decree from January 2024: hyperscale banned across the entire country, except two northern municipalities. Separately, queues for large consumers reach ten years, with thousands of companies waiting. In February 2026, TenneT warned that the grid in Flevoland, Gelderland and Utrecht had reached its absolute limit — in Utrecht province, a full connection freeze would leave 52,500 planned homes without a connection.
3. United Kingdom: a queue tripled in seven months.
From 41 GW to 125 GW of contracted demand, before an emergency reform.
4. The FLAP-D cluster: seven-to-ten years against eighteen-to-twenty-four months.
The ratio of waiting to execution: five to one.
5. Germany: energy that exists but cannot move.
Congestion management costs tripled between 2019 and 2022; in 2023, ~19 TWh of renewables were curtailed — the equivalent of 5.6 million households for a year.
The hubs have shut themselves down.
5 of 6 factors produce nothing ○ Source: Avanza Energy, June 2026 ○ TerranOS.com
Fig. 7 — The cost of a single missing factor, reduced to one number.
The strategic consequence: capital does not wait.
Amazon Web Services (AWS) committed 15.7 billion euros to Aragon, Spain — a secondary market chosen precisely because it had available grid capacity. AWS’s head of energy markets for Europe, the Middle East, and Africa said it publicly in February 2026: there is a mismatch between the ambition to grow within two years and real connection timelines. Microsoft added 7.16 billion dollars in Aragon and 2.1 in Madrid. The Nordics attracted 15 billion dollars in 2025 alone. Ember Energy projects that secondary markets — Nordics, Iberia, Central and Eastern Europe — grow 110% by 2030, versus 55% for the hubs. By 2035, half of European capacity will sit outside FLAP-D.
Sources: IEA, EnkiAI, Carbon Direct, Colliers via Prime East ○ TerranOS.com
Fig. 8 — Same product, different calendar: years to energization, by market.
This is not geographic diversification. It is capital migrating to where the equation can be closed.
Part V. Why Romania, and why now
The argument is not that Romania is more advanced. It’s that Romania is less loaded.
Sources: DataCenter Forum Romania 2026 / Tema Energy, Transelectrica, ANRE (June 2026) ○ TerranOS.com
Fig. 9 — Romania: the assets, with numbers that survive due diligence.
The same number that shows a smaller sector — 77 MW of installed IT capacity in 2025, versus 4,260 MW in Germany — also shows a grid not yet saturated by AI-class consumers. In the west and center of the country there are energy hubs where up to 500 MW can be supplied almost immediately. The mix is also what capacity buyers ask for: over 60% clean sources, with nuclear at ~20% baseload.
The energy pipeline, read correctly. Official Romanian Energy Regulatory Authority (ANRE) data as of June 1, 2026 shows a funnel, not a single number: 1,530 projects with valid technical connection permits totaling 91 GW of approved export capacity. Of these, 998 projects with signed connection contracts (55 GW), 671 with both a connection contract and a building permit (31 GW), and 213 fully permitted, including ANRE’s establishment authorization (10 GW). The government’s April 2026 review says it plainly: permitted capacity is nearly ten times actual need, and a significant share of reserved capacity is not progressing toward implementation. This funnel is the six-factor thesis itself, demonstrated on Romanian data: a permit on paper is not capacity; only the layer with closed factors counts — the 31 GW with a contract and building permit, of which 10 GW are fully authorized.
The grid is being reinforced: Transelectrica’s 2024–2033 plan provides ~1.9 billion euros, and the 2026 allocation rules — auctions with guarantees of 20,000 EUR/MW — are designed precisely to eliminate paper reservations and accelerate real projects. And large things are already moving: the government has requested expressions of interest for a 1.5 GW AI gigafactory at Cernavodă, in a project of up to 5 billion euros, an 800 MW campus has been announced at Mischii, and Sany International has announced a data center in Timișoara.
What it lacks — and it must be said
A material that hides weaknesses does not survive due diligence. Romania’s System Average Interruption Duration Index (SAIDI) is ~240 minutes per year versus 40–150 in Germany, with hyperscalers requiring 99.999% — the difference is covered with site-level redundancy, budgeted explicitly. The 91 GW of permits exceed real need by nearly ten times, which is exactly why ANRE reformed the system in 2026. Estimates point to at least 4 GW of storage needed by 2030 for absorption. Scaling from 77 MW to hundreds of MW requires institutional speed, not just capital. Also, the window the industry estimates for attracting major investment is roughly twelve months and the advantage of an unloaded grid disappears exactly when it is used.
Part VI. The twelve-month timeline, stated correctly
The claim “from capital allocation to operational in a maximum of twelve months” is real, but only under precise conditions. Stated in general terms, it does not hold. Stated correctly, it is the strongest commercial argument in this material.
Fig. 10 — The conditions that make the 12-month timeline credible.
Not a promise. A list of conditions.
This distinction is the product itself. In Ireland, a fully permitted project on purchased land sat for years because it lacked a single factor out of six. On a site where all six are already covered, the only remaining timeline is execution — and executing a module of tens of MW is a logistics problem, not a regulatory one. Remember the number from Part III? Many of the blocked American projects would need only 12–18 months of construction. Construction time was never the problem. The problem was always permission to begin.
Part VII. What this means, by role
For the financial investor
The thesis is not “cheap energy in Eastern Europe”. The thesis is time arbitrage. The appreciating asset is not the land and not the battery — it is the grid connection right with a firm date: administratively limited supply, structurally growing demand, no substitute. In mature markets this right has already become the dominant value component of a site; in Romania, it is not yet priced as such. To verify before committing: the status of each of the six factors with the corresponding document (not statements — documents), the published available capacity for the grid area, delivery lead times for transformers and switchgear, the redundancy budget for the SAIDI gap, and an identified capacity buyer, not an assumed one.
The main risk is not demand. It is execution and the window.
For the capacity buyer
The relevant question is not the price per kW-month. It is the date you can load the first rack, and how credible that date is. Romania means tens to hundreds of MW, in tranches, on sites with available grid in the west and center — not a gigawatt campus next year, but real capacity much earlier than any European alternative.
To demand: a contractual energization date with penalties, proof of the connection permit, the written redundancy plan, fiber routes with named operators.
To gain: between 18 and 96 months of head start versus the same capacity contracted in FLAP-D. For an AI model training run, that head start doesn’t represent cost savings, it is the difference between running and not running.
Conclusion
The AI infrastructure market has stopped being a competition for capital and become a competition for access. Money exists in excess. Demand is contracted before it is built. The chips arrive.
What cannot be bought, right now, in most of the developed world, is the permission to connect a large load to a grid that has no room left.
Romania has, for a short window, exactly that. Not forever and not at unlimited scale — but enough to turn a ten-year delay elsewhere into a twelve-month advantage there.
Whoever closes all six factors before the window closes holds the only asset that matters in this cycle.
Sources in this article: CBRE (Global Data Center Trends 2026; U.S. Outlook 2026; European Data Centres Outlook 2026; NA Trends H2 2025) · JLL (NA Data Center Report YE 2025) · CBRE Investment Management (2026) · IEA (Energy constraints / Europe data centres; Energy and AI charts under CC BY 4.0) · Ember Energy (Grids for data centres, June 2025) · ACER (EU congestion costs 2024) · EUDCA (2026 Data Center Report) · Colliers · Carbon Direct (AI Meets the Grid, May 2026) · Ascend Analytics · PJM (Cycle 1, May 2026; FERC filing, Apr. 2026) · ERCOT · Sightline Climate via Bloomberg · Data Center Watch via Forbes · Prime East · Baird · EirGrid / CRU (Dublin moratorium 2021–2025; Dec. 2025 framework) · TenneT (Feb. 2026 warning) · Avanza Energy (June 2026) · Transelectrica (2024–2033 plan; 2026 capacity allocation) · ANRE (Orders 53/2024, 79/2025, 15/2026; permit status, June 2026) · Government of Romania (permit review, Apr. 2026) · DataCenter Forum Romania 2026 / Tema Energy · DatacenterDynamics · Balkan Green Energy News · Introl · EnkiAI · Carbon Direct · Alex Hormozi. Data reflects publicly available information as of August 2026. Capacity figures and connection timelines must be re-verified at the time of any decision. This material is informational and does not constitute investment advice.
About TerranOS
TerranOS is the operating system for real-world energy assets: grid-scale battery storage, documented mineral reserves, AI computing power and carbon, brought on-chain for a global community of verified members. Built on real infrastructure in Europe. Learn more at www.terranos.com.
Follow the buildout: X (Twitter) | LinkedIn | Facebook | Medium
Article written by Razvan Laichici and edited by Rouă Denis for TerranOS
