SemiAnalysis's 'moratoriums only delay 2.3GW' measures delay, not exit. Its own filter scores cancelled and relocated projects as zero, while Heatmap counted at least 3.5GW of proposals cancelled in the first quarter alone.

Key takeaways

On September 15, SemiAnalysis, a research firm that covers semiconductors and data centers, published the first audit that converts US data-center moratoriums into power capacity project by project. By SemiAnalysis's count, roughly 300 local bans plus New York State's executive order genuinely delay only about 2.3GW, against the 38GW of new US data-center capacity SemiAnalysis forecasts for 2027. The figure has become an answer to whether the wave of moratoriums will drag down the US AI buildout. Its answer is no. This publication's September 16 daily brief called the figure sound for 2026 and 2027 deliveries. It added one limit: the yardstick is built so that it cannot see projects from 2028 onward that have not yet picked a site.

This piece takes that call apart and re-tests it. The result has three layers. First, for 2026 and 2027, 2.3GW is the right reading, and a low number was inevitable: of the 38GW due in 2027, 22GW is already in vertical construction and most of the rest is financed and doing sitework. SemiAnalysis writes that moratoriums cannot reach capacity under construction or fully permitted, and that this "covers all of 2026 and most of 2027." Second, the original call pointed at the right blind spot but drew the boundary in the wrong place. SemiAnalysis's model does carry early announced projects that have no financing, filed under 2028 and later. What the yardstick cannot see is not the years after 2028 but one outcome: a project leaving. SemiAnalysis's own filter says a cancelled or withdrawn project does not count as delayed, and the same report says developers usually relocate or redesign, while a relocated project scores zero delay at the original site. Third, a separate yardstick built to count cancellations reads anything but small. The data team at Heatmap, an energy and climate news outlet, counted at least twenty data-center proposals cancelled after local opposition in the first three months of this year, at least 3.5GW of electricity demand in total, and Heatmap estimates that about half of locally challenged projects end up cancelled.

Why this is worth twenty minutes: the daily brief gave this one line, 2.3GW against 38GW. Here the denominator is taken apart, and the low ratio turns out to be built in. The brief's own claim, that readings for 2028 and later are structurally zero, is also tested. Half of that claim did not survive, and the corrected blind spot is now called exit. Finally, three results are fixed in advance and listed at the end under What would prove this wrong. If any of them shows up, this piece is wrong.

If this is the only part read: the next time a "moratoriums only delayed X gigawatts" figure appears, the first question is whether cancelled and relocated projects are in the count. A delay figure answers whether next year's deliveries will fall short. Cancellation rates, and where cancelled projects go, are what answer whether a project planned for 2028 or later can still be built where it was first planned.

The site has no evolution map for this topic yet, so this piece carries no chart.

One: where 2.3GW comes from. A filter with eleven conditions

SemiAnalysis's method maps the 400-plus local instruments, actions and proposals in its moratorium database, spread across 17 states, onto the parcels of the 6,000-plus data centers in its project pipeline. That pipeline is built facility by facility from property records, permits, power-usage data, freedom-of-information requests and satellite imagery. The audit first separates two numbers:

• Capacity sitting inside a restricted boundary: about 20GW. This is what the map circles.
• Capacity genuinely delayed as a result: 1,525MW, or 7.6% of the first number, almost all of it from three projects: an AWS campus in Ohio, a campus in Pennsylvania belonging to a developer that secures grid power for land before selling it on, and a site in Colorado. Add New York's executive order, which touches about 1.4GW and in SemiAnalysis's estimate delays about 0.8GW. Together, roughly 2.3GW.

Between 20GW and 1.5GW sits a filter. SemiAnalysis lists eleven conditions, and all of them have to hold at the same time before a moratorium counts as having delayed a project. In the report's words: "If you miss just one of these, the moratorium is irrelevant." The conditions that remove the most projects:

• Whether the moratorium reaches the parcel. County moratoriums generally apply only to unincorporated land. A campus inside an incorporated city within that county is outside the ordinance.
• Whether the project already holds the permit being frozen. Moratoriums almost always freeze new applications rather than revoke approvals already granted.
• How far the project is from breaking ground. For a project delivering in 2028 or later, a moratorium lasting three to six months expires before the approval is needed, and the developer waits it out without moving the delivery date.
• Whether the project is still active. This condition comes back below.

A financial news site reported that SemiAnalysis's research team first circled 77 campuses, about 20GW, and that three survived the screening. That account is secondhand, and the original wording has not been reviewed.

The method itself is solid. SemiAnalysis does not count moratoriums; the audit asks, for every moratorium, which permit on which parcel it blocks. That is the main difference from every earlier count. Earlier readings mostly counted instruments, from 127 in July to more than 300 in September. A count of instruments says how many places are objecting, and nothing about how much power is stuck.

Two: take the denominator apart first. Why 2.3GW against 38GW had to be small

The original reading put the two numbers side by side: about 2.3GW delayed, 38GW of new capacity in 2027, so moratoriums touch about six percent of next year's additions. That division has to be taken apart, because the numerator and the denominator measure different things.

The 38GW in the denominator was almost beyond a moratorium's reach to begin with. SemiAnalysis's own words: "Of that total, 22GW is under vertical construction, while the remaining 16GW “planned” is, in majority, already financed, doing siteworks and preparing for vertical construction." The same report says that capacity already under construction, or fully permitted and about to start, "is out of reach altogether", and that this "covers all of 2026 and most of 2027." The 2027 cohort had already cleared the only gate a moratorium can close. Most of it cleared that gate before the wave of moratoriums arrived at scale. The ratio is low because the denominator is the one year the moratoriums cannot bite.

The 2.3GW in the numerator carries no delivery year. It is a stock of capacity currently delayed. Some of that may be 2027 projects, some later. The Pennsylvania campus that is stuck was due to start its first building in the fourth quarter of 2026, and the local approval has now slipped into 2027. A multi-year stock divided by a single year's flow gives a percentage with no clear meaning.

None of this says 2.3GW is wrong. For the question of whether 2026 and 2027 deliveries should be discounted for moratorium news, 2.3GW is the right reading. Goldman Sachs forecast in May that 2027 additions would reach 36.3GW, close to SemiAnalysis's number. But Goldman's article does not say whether its figure counts server power or grid-connected capacity. Whether the two numbers measure the same thing has not been established, so the Goldman figure cannot serve as corroboration. More useful is Goldman's estimate in the same article: "Only about 50-60% of data center capacity scheduled for the next one to two years is expected to come online on time amid delays and cancellations." The real risk to 2027, then, lies in equipment lead times, grid connection and financing. Moratoriums are a small piece of that. For next year's deliveries, SemiAnalysis's conclusion stands. What the report cannot answer is anything beyond next year, and the report says so.

Three: what the filter removes. Delay, relocation and cancellation are three different things

When a moratorium lands on a project, delay is not the only outcome. There are at least three:

"Zero, by rule" is not this publication's inference. The phrase restates a condition SemiAnalysis wrote down: "A cancelled, withdrawn, downsized or otherwise abandoned project should not be counted as moratorium-delayed simply because it sits inside the restricted area." The same report says that even where a moratorium is the binding constraint, "the developer can usually relocate, redesign, or challenge the project rather than abandon it."

Those two sentences together mark the edge of the yardstick. The yardstick measures how long projects that are still alive and still in place have been held up. Projects that moved and projects that died are off its books. For the question SemiAnalysis set out to answer, that design is right: the forecast is national delivery, and a project rebuilt somewhere else subtracts nothing from the national total, so it should not be booked as a loss. But it also means that "moratoriums barely matter" rests, in part, on the premise that projects can move.

How many are cancelled, then? Someone does count that, and the reading is not small:

• Heatmap, May report (an energy and climate outlet): "At least 20 proposed data center projects were canceled after local pushback during the first three months of 2026", and "These canceled projects accounted for more than $41.7 billion in investment and represented at least 3.5 gigawatts of electricity demand." Heatmap's data team phones counties across the US at regular intervals to check. The July update: "More than 50 data centers have been canceled so far this year after facing local pushback of some kind", and "Today, a data center project that’s challenged by local activists has about a 50% chance of being canceled." The rate varies widely by state: 71% in Michigan, 17% in Texas.
• Data Center Watch (a research group that tracks local opposition to data centers): in the first quarter of 2026, "At least 75 data center projects worth approximately $130 billion were blocked or delayed in a single quarter". When this publication recorded a Bloomberg reporter's "75 in the first three months" in August, the note said the method behind the figure was entirely unknown. The number and the time window match this report, which is very likely the source, and the report's definition counts both local denials and proposals that developers withdrew on their own.

The two yardsticks cannot be added, and neither refutes the other. The units differ: Heatmap counts electricity demand, SemiAnalysis counts IT capacity, the power drawn by the servers themselves, and for the same campus the first is usually larger. The populations differ: Heatmap and Data Center Watch count proposals, and many proposals are speculative. Goldman's May article says plainly that developers often file in several jurisdictions at once and advance only the one with the best terms. SemiAnalysis wrote in June that the queue of large power users waiting to connect to the US grid had passed a terawatt, and that "The vast majority are highly speculative". So cancelled gigawatts are an upper bound on the loss, not the loss itself: some were never going to be built, and some will be built elsewhere.

One comparison, though, connects the two yardsticks. SemiAnalysis says Michigan has more enacted local moratoriums than any other state, 45, and that Michigan "carries effectively zero pipeline exposure", exposure being the capacity inside a restricted boundary from section one. Heatmap says 71% of contested proposals in Michigan are eventually cancelled. Our read: these two numbers describe two sides of the same thing. In Michigan, opposition mostly kills projects at the proposal stage, so no project survives into the pipeline to be delayed by a moratorium later. A delay reading of zero does not mean the moratoriums had no effect. The zero may mean the effect came too early for this yardstick to see. This paragraph is this piece's inference from placing the two reports side by side; neither firm says it.

Four: where the original call has to change

The September 16 call had two halves. The first said that 2.3GW allows "the moratorium wave drags down 2026–27 deliveries" to be marked down; section two above finds that half holds. The second said the yardstick sees only projects already in the pipeline, so that projects from 2028 onward that have not picked a site, and therefore sit in no database, read structurally zero. This piece tested that half. The direction was right; the boundary was wrong.

The years after 2028 are not the blind spot. A June report from SemiAnalysis says of projects announced on nothing more than a preliminary power estimate, with no financing and no interconnection study: "Instead of falsely declaring them as 2026 projects, they rather show up in 2028+ in our model." So the pipeline does hold early announced projects, and the 20GW of exposure includes projects for 2028 and later. The condition in section one, that a project delivering in 2028 or later can wait out a moratorium of three to six months, exists to handle exactly those projects; SemiAnalysis's wording is that "shorter moratoriums may expire before 2028+ projects need the affected approval". The original call's first condition for proving itself wrong was that SemiAnalysis disclose a layer of its model for projects proposed but not yet in the pipeline, and that the reading for that layer also be low. The first half of that condition has now been met. The second half, the reading for that layer, has not been published.

The real blind spot is exit. That half held, and it is thicker than first thought. The same June report says: "We do not put low probability datacenters in our datacenter timelines." Add the rule from section three, that cancelled projects do not count, and the yardstick cannot see four kinds of project: those judged low-probability and never admitted, those proposed and then cancelled, those moved elsewhere, and those never proposed because a moratorium was expected. The first three are partly counted elsewhere, by Heatmap and Data Center Watch. The fourth has no reading anywhere. The core sentence of the original call therefore moves from a period to an outcome: "only 2.3GW has been delayed" and "fewer projects are willing to build where they first planned" can both be true at once, and the first is no evidence against the second.

One more sentence of the original call has to narrow, because the SemiAnalysis estimate and Arvind Narayanan's estimate are not fully independent. The call said two estimates corroborated each other. One was SemiAnalysis's project-by-project audit. The other came from Narayanan, a computer science professor at Princeton, who converted blocked capacity into the efficiency gains that would make it up. The two used entirely different methods and pointed the same way. The methods are different. But Narayanan's arithmetic states its assumption that ninety percent of blocked capacity gets built elsewhere: "Assuming 90% "leakage", even a big state like NY banning all new data center construction only slows AI by less than a day." SemiAnalysis scores a relocated project as zero delay and says developers usually relocate. On the assumption that matters most, both rest on the same one: projects can move. What the two corroborate is only each other's arithmetic. The assumption itself has not been corroborated by either. If one day projects cannot move, for instance because neighboring counties in the same region enact moratoriums in sequence, a chain SemiAnalysis itself records in metro Atlanta, both readings go wrong together.

Five: from 2028 on, which readings to watch

By the breakdown above, a moratorium's effect splits into four questions, each with its own right yardstick:

The last question is the gap this piece most wants to point at, and the data already exist. SemiAnalysis says its moratorium database tracks every instrument "from proposal through enactment, expiration, lifting, replacement, or rejection", and its project pipeline tracks cancellations and withdrawals facility by facility. Placing the cancellation rate of projects inside restricted boundaries next to the rate for projects outside them would show whether moratoriums kill projects or only move them. Similar rates mean moving. A clearly higher rate inside, with the national forecast marked down to match, means killing. The comparison needs no new data, only someone to cut it.

Until then, what can be seen is the shape changing, and every signal points to "move" rather than "stop":

• Stepping outside the jurisdiction. Brownsville: the parcel was moved outside the city limits, and the city's moratorium no longer reaches it.
• Freezing only new applications. Foley & Lardner, a law firm, published a summary of New York's July executive order. By that summary, the order pauses only applications for state discretionary permits for data centers of 50MW or more. Applications the state environmental agency had already deemed complete are untouched. Permits issued by local governments are outside its scope, and the order sets no fixed end date. Loudoun County, Virginia, moved the same way on September 15. Local media report that the board directed staff to draft a pause on new applications of up to twelve months. Approved projects keep their approvals, and a final vote is scheduled for October.
• Skipping the grid. SemiAnalysis says Texas's review of its grid-connection queue "means nothing at all" for projects that never needed the grid and is a net positive for developers generating power on site; in early September SemiAnalysis counted 75GW of firm orders for on-site generation equipment, more of it bound for Texas than for any other state.
• Changing location. Peter Freed, formerly head of energy strategy at Meta, told Heatmap: "The impact is increasingly that developers are stopping efforts in those communities [that pass restrictions] and shifting their attention elsewhere." Microsoft's annual report filed at the end of July lists "state and local moratoriums" and "increasingly coordinated opposition to infrastructure development across jurisdictions" among its risk factors, with the consequence written as possible project delays, narrower development opportunities and higher costs.

Our call

"Moratoriums have only delayed about 2.3GW" is the right reading for 2026 and 2027 deliveries, and a low number was inevitable: the 38GW due in 2027 had cleared the one gate a moratorium can close before the wave arrived, and SemiAnalysis writes that moratoriums cannot reach "all of 2026 and most of 2027". What this yardstick cannot see is not the years after 2028 but exit: its rules score a cancelled project as zero and treat a relocated project as unaffected, while Heatmap, which counts cancellations, found at least 3.5GW of electricity demand in proposals cancelled in the first quarter of this year alone, and about half of challenged projects cancelled. Part of what gets cancelled was speculative to begin with, so that is not a reading of the loss but an upper bound on it. The real effect of moratoriums turns on two things: how far the cancellation rate inside restricted boundaries differs from the rate outside, and whether cancelled projects come back somewhere else. Every signal visible today points to "move" rather than "stop", but no one measures how much has moved. Once neighboring jurisdictions in the same region enact moratoriums in sequence and projects can no longer move, the delay figure goes wrong. So does the estimate by Arvind Narayanan, a computer science professor at Princeton, that even a ban on new data centers across New York State would slow AI by less than a day. That estimate assumes ninety percent of blocked capacity gets built elsewhere.

What this changes, and for whom

• Cloud providers and hyperscale operators planning capacity: 2027 delivery assumptions need no haircut for moratorium news; the haircut belongs to equipment, grid connection and financing. For sites from 2028 on, three conditions become the first-round screen: whether a parcel sits inside a city boundary, whether its zoning already permits data centers, and whether it has room for on-site generation. These are the shapes the readings show projects escaping through. A backup parcel in the same region matters too, because how fast neighboring jurisdictions follow decides whether a project can move.
• Lenders and infrastructure funds financing data centers: the same moratorium headline is two different risks for approved and unapproved projects. A fully permitted project is exposed to the delay figure, which is small today. A greenfield project still in permitting is exposed to the cancellation rate, and with about half of challenged projects cancelled, that probability belongs in the valuation. Loan terms can make a local permit in hand, rather than construction progress alone, a precondition for disbursement. The permit is what separates the two risks.
• Suppliers of power and grid equipment: moratoriums push demand toward on-site generation that bypasses the grid, Texas most of all. The demand still exists, in a different place and form. The gain sits on the on-site generation side, the 75GW of firm orders SemiAnalysis counted in early September, most of it bound for Texas. The reading to watch is how many projects first planned for grid connection switch to on-site generation, which decides whether grid-side orders are delayed or bypassed.
• Policy and public-affairs teams: "only 2.3GW delayed", offered to local officials as proof that moratoriums are harmless, collapses at the question of what happened to the cancelled projects. The argument that holds separates the two cases: approved projects get built as planned, and new projects are a negotiation over terms.
• AI product teams that depend on cloud compute: next year's compute deliveries need no discount for moratorium news. For long-term contracts and region choices from 2028 on, whether a provider's new capacity in a region sits inside a moratorium, and whether the provider holds a backup site, is what decides whether a delivery commitment can be honored, because that is what decides whether a project can move.

What would prove this wrong

1. SemiAnalysis or another tracker publishes a comparison of cancellation rates inside and outside restricted boundaries, and the two are close. That would mean moratoriums move projects rather than kill them. This piece's "fewer projects willing to build where first planned" would still hold, but "exit is the blind spot" would stop mattering for the national total, and the last two questions in section five's table would collapse into one. Observation window to June 30, 2027, set by this publication.

2. SemiAnalysis marks down its forecast of 2028 US capacity additions by more than ten percent within the same window, and attributes the markdown to moratoriums or local opposition. That would mean exit has grown large enough to show in the national total, and this piece's "every signal today points to move" would have to be withdrawn. The threshold is this publication's own.

3. Actual US deliveries in 2027 come in clearly below SemiAnalysis's 38GW forecast, with the gap attributable to moratoriums. Then section two's "for next year's deliveries, 2.3GW is the right reading" is wrong. The verdict date follows the annual result, provisionally March 31, 2028, set by this publication.

Written from the same research and judgments as the Traditional Chinese edition. Sources are linked; we distinguish original documents from reporting and mark what we could not verify.

Sources cited in other language editions

These sources are linked in the body of this piece's other language editions but not in this one.

• x.com/kevinsxu/status/2097405506885157376
• dailybusinessgroup.co.uk/2026/09/holyrood-votes-to-pause-expansion-of-data-centr
• foe.scot/press-release/politicians-dilute-motion-calling-for-moratorium-on-ai-da
• x.com/TheStalwart/status/2077030191705080256