Ryan — the part of this that really jumps out is the distinction between project representation and earned reliance.
A developer can enter a queue, post security and clear an administrative milestone without necessarily establishing that the underlying load, customer, power strategy, timing or operating assumptions have become durable enough for planners to rely upon.
The same issue shows up with flexibility: demonstrated capability is evidence, but it is not permanent assurance that the capability will remain available under the conditions in which the grid later depends on it.
Project maturity determines whether a project has earned advancement. Revalidation determines whether the evidence that earned advancement remains true.
Good piece, but from the operator's chair — 35 years in energy and finance, gas assets in Texas — it treats a real-estate phenomenon as a grid phenomenon.
The 474 GW isn't a load forecast; it's a listing count. Most of the "developers" in that queue have never built, bought, or run a megawatt of generation. A load-study letter from Oncor or AEP became the marketing document for "powered land," and the queue filled up the way an MLS does. Auditing it is fine. Freezing everything punishes the few applicants who can actually energize.
Behind-the-meter in West Texas isn't a bridge, either. Waha gas has traded negative. A producer with stranded gas and a buyer whose economics, by your own numbers, could bear $5,600/MWh aren't building a plant to shut it in year five. The grid tie becomes the backup and the export path; the plant stays primary. Your two facts — GPUs cost far more than power, and on-site power costs more than grid — point the same way: on-site is permanent.
The $50k/MW deposit isn't thin for the population causing the problem. A 500 MW request is $25 million posted. Speculators can't post it; hyperscalers won't notice it. That's the filter working, and the forfeit percentage is beside the point.
ERCOT also doesn't build anything. Oncor, AEP and CenterPoint do, and their crews, transformer deliveries and 765 kV schedules are the gate. No rule change at ERCOT adds a lineman.
Last, the Legislature isn't in regular session until January 2027. SB 6 already gave the state deposits, disclosure, backup-gen reporting and curtailment authority. The freeze is an executive overlay on a statute that was already doing the filtering, and the office that imposed it can narrow it tomorrow. It doesn't need to wait for October 19.
Your figures are great! The 'it depends on who pays' line is the crux. In PJM the answer crosses state lines: household power prices rose 38% in New Jersey over three years and 56% in DC, against 15% in Virginia, where the data centers are. Should large-load tariffs be set by states or by PJM?
You've done a nice job summarizing the issues behind the trends in the AI-power industry. I'd push back on the idea that developers "spam" the interconnection queue with "speculative" projects. Understanding why developers file multiple interconnection requests deserves the same insightful approach you take with your other subjects.
The fundamental issue a developer faces with a power project is overlapping uncertainties that have to be derisked in parallel over a 3-5 year timeframe -- e.g. site control, entitlement, interconnection, engineering, fuel supply, and power offtake. They all need to resolve into a bankable package of interlocking agreements with a project finance lender's credit standards.
Interconnection presents very unique uncertainties on cost and schedule. It is one of the "long poles" in the tent. It takes 2-3 years of study time to get to a firm estimate on cost and schedule.
Interconnection applications are studied in cohorts or clusters and the process is typically broken out into phases (usually 2 or 3 phases, but each ISO/RTO differs). At each phase's conclusion, the developer is presented with an estimate of cost and schedule. The developer then chooses whether to continue or withdraw based on the results. Continuing obligates them to fund the next study phase. Each successive phase increases in cost over time and, theoretically, increases the precision of the estimate.
The critical input that drives cost and schedule is the number of other applications in the cohort that rely on the same facilities to deliver power. In other words, interconnection cost is not simply a construction estimate based on a fixed BOM--the upgrades assessed for an interconnection reflect the capacity available in the study scenario to all active applications.
You'll immediately recognize that the interconnection process is therefore an economic game. The study process is providing information to each participant about their piece of the game. When choosing to continue or abandon an application, game theory dominates the developer's decision. No single player has all the information that defines the game space, not even the ISO/RTO, because they don't know what cost is viable for a developer's project.
When presented with an initial study result, the developer faces a game theory problem. If you believe the other developers will abandon, you may want to stay in, because your cost will go down if enough other applications drop out. On the other hand, if you think the others are likely to stay in, abandonment may be the best choice.
Developers have to work on a portfolio basis, derisking multiple projects over time because there is no such thing as a rifle-shot in power project development. Decisions to stay in our abandon a project are driven by game-theory-driven decisions that cannot be solved in the opening moves. Developers prospect multiple sites and file multiple interconnections because it is the only way to ensure enough viable projects at the end of the process.
Are there unserious or stupid developers out there clogging the queues? Undoubtedly a few. But I'd argue that most of what is in the queues reflects serious attempts to solve a very complex game that assumes from the start that only a minority of projects will succeed. It's not "speculation" it's optimization in a complex gamescape.
The jump from 63 GW to 474 GW is hard to read as real demand when developers file the same project at several sites. A queue that counts each request as load will freeze permits for projects that were never going to energize, and also for the ones that would. I'd want the queue to collapse duplicate or customer-less requests before it sums gigawatts, otherwise the pause is reacting to the filing pattern more than to the grid.
Nan — exactly. That is the distinction I was getting at with project representation versus earned reliance.
Before aggregate requested MW becomes an input to consequential planning, I think the underlying records need to establish at least whether the demand is unique, attributable to a real project/customer, and sufficiently mature to justify being counted for the particular decision at hand.
Otherwise a queue can be accurate as a record of requests while still being misleading as evidence of future load.
The Reported Request → Sustained Demand ladder is the right framing. In practice the hard part is the revalidation trigger: you need something that notices when the facts behind a project state go stale, which is basically the same problem as keeping a warehouse table's freshness guarantees honest. Have you seen queues actually carry that state, or does it live outside the ISO's records?
From what I can see, queues often carry pieces of the state, but not necessarily the full decision-reliance state as a persistent object.
ERCOT’s current process, for example, does track things like study phase, deposits, phased milestones, eligibility, curtailment obligations and whether a project has fallen far enough behind that capacity can be reassigned. Ryan notes that Batch Zero can reassign capacity if a project is two years late, and that projects can advance through phased connections and specific flexibility commitments. Pasted text
What I have not seen in the material here is an explicit record that says, in effect:
Current project state
→ evidence supporting that state
→ effective-as-of date
→ last revalidated date
→ material-change status
→ decision(s) currently relying on it
That is the difference between maintaining a queue record and maintaining the evidentiary currency of the state the queue represents.
Your warehouse-table analogy is a good one. A record can still exist and be technically accurate as a historical entry while no longer being fresh enough for the decision being made.
So my view is that the revalidation trigger probably has to sit across both the ISO’s operational records and the external facts the ISO does not fully control — customer status, site control, financing, power strategy, construction, permitting, ownership, and similar dependencies.
The interesting design question is whether the ISO should merely ingest those updates, or whether it should also carry an explicit “current reliance state” that tells planners when the prior evidence has gone stale.
Worth noticing what that 474 GW actually counts: applications, not buildings. It's over five times the state's record peak, and holding a spot costs a fraction of what a delayed campus costs, so the line fills with maybes. Texas isn't short on electrons so much as short on a way to tell a real project from a placeholder — and until the audit sorts that out, the sorting is what's waiting.
The supply side of ERCOT tells the same story from the other end. As of Oct 7, DC Hub's power index has every tracked Texas metro at CAUTION (Dallas 48.2, Austin 49, Houston 49.2, Abilene 48.4 out of 100), with roughly 19.6 GW of generation additions queued for the next 12 months against the 474 GW of large-load requests you cite. Even Abilene, where Crusoe is building behind the meter, shows about 631 MW of behind-the-meter industrial headroom, so a 20% forfeit on a 1 GW request is cheap next to what can actually be delivered. Source: DC Hub (dchub.cloud), as of Oct 7, 2026. Live scores: https://dchub.cloud/connect
Absolutely excellent piece. Comprehensive and thoughtful. We are very focused as A16z LPs, LPs in other funds and as direct investors in the opportunity for microreactors (Radiant is fantastic) and the need for wires and interconnect. Texas as a power island is such a curious situation, such is America! Look forward to tracking your thinking and connecting in person at some point. Gavin
Ryan — the part of this that really jumps out is the distinction between project representation and earned reliance.
A developer can enter a queue, post security and clear an administrative milestone without necessarily establishing that the underlying load, customer, power strategy, timing or operating assumptions have become durable enough for planners to rely upon.
The same issue shows up with flexibility: demonstrated capability is evidence, but it is not permanent assurance that the capability will remain available under the conditions in which the grid later depends on it.
The missing layer seems to be:
Representation → Evidence → Dependency → Decision Gate → Material Change → Revalidation.
Project maturity determines whether a project has earned advancement. Revalidation determines whether the evidence that earned advancement remains true.
Really thoughtful piece.
Neil P. Osnato
Founder | Persistence Analytics Group LLC
Good piece, but from the operator's chair — 35 years in energy and finance, gas assets in Texas — it treats a real-estate phenomenon as a grid phenomenon.
The 474 GW isn't a load forecast; it's a listing count. Most of the "developers" in that queue have never built, bought, or run a megawatt of generation. A load-study letter from Oncor or AEP became the marketing document for "powered land," and the queue filled up the way an MLS does. Auditing it is fine. Freezing everything punishes the few applicants who can actually energize.
Behind-the-meter in West Texas isn't a bridge, either. Waha gas has traded negative. A producer with stranded gas and a buyer whose economics, by your own numbers, could bear $5,600/MWh aren't building a plant to shut it in year five. The grid tie becomes the backup and the export path; the plant stays primary. Your two facts — GPUs cost far more than power, and on-site power costs more than grid — point the same way: on-site is permanent.
The $50k/MW deposit isn't thin for the population causing the problem. A 500 MW request is $25 million posted. Speculators can't post it; hyperscalers won't notice it. That's the filter working, and the forfeit percentage is beside the point.
ERCOT also doesn't build anything. Oncor, AEP and CenterPoint do, and their crews, transformer deliveries and 765 kV schedules are the gate. No rule change at ERCOT adds a lineman.
Last, the Legislature isn't in regular session until January 2027. SB 6 already gave the state deposits, disclosure, backup-gen reporting and curtailment authority. The freeze is an executive overlay on a statute that was already doing the filtering, and the office that imposed it can narrow it tomorrow. It doesn't need to wait for October 19.
Your figures are great! The 'it depends on who pays' line is the crux. In PJM the answer crosses state lines: household power prices rose 38% in New Jersey over three years and 56% in DC, against 15% in Virginia, where the data centers are. Should large-load tariffs be set by states or by PJM?
Ryan,
You've done a nice job summarizing the issues behind the trends in the AI-power industry. I'd push back on the idea that developers "spam" the interconnection queue with "speculative" projects. Understanding why developers file multiple interconnection requests deserves the same insightful approach you take with your other subjects.
The fundamental issue a developer faces with a power project is overlapping uncertainties that have to be derisked in parallel over a 3-5 year timeframe -- e.g. site control, entitlement, interconnection, engineering, fuel supply, and power offtake. They all need to resolve into a bankable package of interlocking agreements with a project finance lender's credit standards.
Interconnection presents very unique uncertainties on cost and schedule. It is one of the "long poles" in the tent. It takes 2-3 years of study time to get to a firm estimate on cost and schedule.
Interconnection applications are studied in cohorts or clusters and the process is typically broken out into phases (usually 2 or 3 phases, but each ISO/RTO differs). At each phase's conclusion, the developer is presented with an estimate of cost and schedule. The developer then chooses whether to continue or withdraw based on the results. Continuing obligates them to fund the next study phase. Each successive phase increases in cost over time and, theoretically, increases the precision of the estimate.
The critical input that drives cost and schedule is the number of other applications in the cohort that rely on the same facilities to deliver power. In other words, interconnection cost is not simply a construction estimate based on a fixed BOM--the upgrades assessed for an interconnection reflect the capacity available in the study scenario to all active applications.
You'll immediately recognize that the interconnection process is therefore an economic game. The study process is providing information to each participant about their piece of the game. When choosing to continue or abandon an application, game theory dominates the developer's decision. No single player has all the information that defines the game space, not even the ISO/RTO, because they don't know what cost is viable for a developer's project.
When presented with an initial study result, the developer faces a game theory problem. If you believe the other developers will abandon, you may want to stay in, because your cost will go down if enough other applications drop out. On the other hand, if you think the others are likely to stay in, abandonment may be the best choice.
Developers have to work on a portfolio basis, derisking multiple projects over time because there is no such thing as a rifle-shot in power project development. Decisions to stay in our abandon a project are driven by game-theory-driven decisions that cannot be solved in the opening moves. Developers prospect multiple sites and file multiple interconnections because it is the only way to ensure enough viable projects at the end of the process.
Are there unserious or stupid developers out there clogging the queues? Undoubtedly a few. But I'd argue that most of what is in the queues reflects serious attempts to solve a very complex game that assumes from the start that only a minority of projects will succeed. It's not "speculation" it's optimization in a complex gamescape.
Best,
Arno Harris
The jump from 63 GW to 474 GW is hard to read as real demand when developers file the same project at several sites. A queue that counts each request as load will freeze permits for projects that were never going to energize, and also for the ones that would. I'd want the queue to collapse duplicate or customer-less requests before it sums gigawatts, otherwise the pause is reacting to the filing pattern more than to the grid.
Nan — exactly. That is the distinction I was getting at with project representation versus earned reliance.
Before aggregate requested MW becomes an input to consequential planning, I think the underlying records need to establish at least whether the demand is unique, attributable to a real project/customer, and sufficiently mature to justify being counted for the particular decision at hand.
Otherwise a queue can be accurate as a record of requests while still being misleading as evidence of future load.
I would separate:
Reported Request → De-duplicated / Verified Project State → Executable Load → Energized Load → Sustained Demand
And then require revalidation when the facts supporting that state materially change.
That seems especially important when the aggregate number itself begins driving infrastructure, permitting or reliability decisions.
— Neil
The Reported Request → Sustained Demand ladder is the right framing. In practice the hard part is the revalidation trigger: you need something that notices when the facts behind a project state go stale, which is basically the same problem as keeping a warehouse table's freshness guarantees honest. Have you seen queues actually carry that state, or does it live outside the ISO's records?
Nan — that is exactly the issue.
From what I can see, queues often carry pieces of the state, but not necessarily the full decision-reliance state as a persistent object.
ERCOT’s current process, for example, does track things like study phase, deposits, phased milestones, eligibility, curtailment obligations and whether a project has fallen far enough behind that capacity can be reassigned. Ryan notes that Batch Zero can reassign capacity if a project is two years late, and that projects can advance through phased connections and specific flexibility commitments. Pasted text
What I have not seen in the material here is an explicit record that says, in effect:
Current project state
→ evidence supporting that state
→ effective-as-of date
→ last revalidated date
→ material-change status
→ decision(s) currently relying on it
That is the difference between maintaining a queue record and maintaining the evidentiary currency of the state the queue represents.
Your warehouse-table analogy is a good one. A record can still exist and be technically accurate as a historical entry while no longer being fresh enough for the decision being made.
So my view is that the revalidation trigger probably has to sit across both the ISO’s operational records and the external facts the ISO does not fully control — customer status, site control, financing, power strategy, construction, permitting, ownership, and similar dependencies.
The interesting design question is whether the ISO should merely ingest those updates, or whether it should also carry an explicit “current reliance state” that tells planners when the prior evidence has gone stale.
That is the layer I think is still missing.
Worth noticing what that 474 GW actually counts: applications, not buildings. It's over five times the state's record peak, and holding a spot costs a fraction of what a delayed campus costs, so the line fills with maybes. Texas isn't short on electrons so much as short on a way to tell a real project from a placeholder — and until the audit sorts that out, the sorting is what's waiting.
why do I have a feeling that I saw something similar from a16z like a year ago or so...
The supply side of ERCOT tells the same story from the other end. As of Oct 7, DC Hub's power index has every tracked Texas metro at CAUTION (Dallas 48.2, Austin 49, Houston 49.2, Abilene 48.4 out of 100), with roughly 19.6 GW of generation additions queued for the next 12 months against the 474 GW of large-load requests you cite. Even Abilene, where Crusoe is building behind the meter, shows about 631 MW of behind-the-meter industrial headroom, so a 20% forfeit on a 1 GW request is cheap next to what can actually be delivered. Source: DC Hub (dchub.cloud), as of Oct 7, 2026. Live scores: https://dchub.cloud/connect
Absolutely excellent piece. Comprehensive and thoughtful. We are very focused as A16z LPs, LPs in other funds and as direct investors in the opportunity for microreactors (Radiant is fantastic) and the need for wires and interconnect. Texas as a power island is such a curious situation, such is America! Look forward to tracking your thinking and connecting in person at some point. Gavin