Wave 11: AI Industry Economics

The following report summarizes responses from 191 experts, in addition to 52 superforecasters and 625 members of the public collected between Jul 13, 2026 and Aug 11, 2026. Expert respondents include 33 computer scientists, 36 industry professionals, 43 economists, and 71 research staff at policy think tanks.

Our wider website contains more information about LEAP, our Panel, and our Methodology, as well as reports from other waves.

Insights

1. Experts expect private investment in data center structures to grow by 76% in real terms from 2025 to 2028.
   We asked forecasters to predict the total annual U.S. private fixed investment in data center structures (in 2025 dollars) by 2026, 2028, and 2035. In 2025, this figure was $40.8 billion. Despite recent restrictions and public opposition to data centers, experts expect investment in data center structures to continue to increase, rising by 29% by the end of 2026 and 76% by the end of 2028, compared to the 2025 level. By the end of 2035, experts predict that investment in data center structures will be 2.2-times its 2025 level. These translate to total private investment of $52.6 billion in 2026, $71.8 billion in 2028, and $89.8 billion in 2035, all in constant 2025 dollars.

   This same trajectory is evident in superforecasters' predictions. They expect private fixed investment in data center structures to grow by 30% by the end of 2026 and 70% by the end of 2028, compared to the 2025 level. By 2035, superforecasters expect investment to be 2.5 times its 2025 level. This translates into expected investment in 2035 of $102 billion for superforecasters versus the $89.8 billion predicted by experts. The public expects considerably lower levels of investment, forecasting a 48% increase by the end of 2028 and an 84% increase by the end of 2035.

   In their written rationales, forecasters with higher investment forecasts argued that announced data center projects are likely to translate into large increases in investment by 2028. Low-investment respondents cited public opposition and state-level moratoria on data center construction, such as the recent moratorium imposed in New York State, as factors that may slow data center buildout and, in turn, investment.

2. Experts also predict significant increases in investment in IT equipment and electrical and communication structures, although they expect slower growth than in data center structures
   We also asked forecasters to predict total annual U.S. private fixed investment in electrical and communication structures, and IT equipment. These cover the increased electricity generation and transmission needed to power data centers and the servers, networking gear, and other electrical infrastructure that keep them running, although neither of these two categories exclusively covers spending related to data centers.

   Experts forecast considerable growth in investment in IT equipment compared to a 2025 baseline of $514.8 billion. They expect this to grow 20% by 2026, 48% by 2028, and 90% by 2035—implying real investment of $978.1 billion by 2035. Superforecasters predicted similar growth, though they forecast that investment will top $1 trillion by 2035.

   In contrast, both experts and superforecasters expect slower growth in total annual private investment in electrical and communication structures. From a baseline of $146.4 billion in 2025, experts forecast 5% growth in investment in 2026, 15% growth by 2028, and 32% growth by 2035. Superforecasters forecast similar levels of growth.

   In their written rationales, forecasters noted that the electrical and communication structures category is a broader part of the economy affecting several non-AI sectors, and is thus less sensitive to a data center boom. Also, investment in this sector has been relatively flat over the past few years. By contrast, forecasters expected the short life of hardware to drive investment in IT equipment, pointing out that data center structures last for decades while servers become outdated after a few years.

3. Experts forecast that the value of US-listed semiconductor equities will rise through 2028, but at a slower pace than in the last 2.5 years. Experts similarly expect growth in the value of North American software equities through 2028, but at a slightly faster rate than in the last 2.5 years.
   The VanEck Semiconductor ETF (SMH) tracks the 25 largest US-listed semiconductor companies (Nvidia, Broadcom, AMD, TSMC, and similar) and is the most direct liquid proxy for the AI compute/hardware buildout. The iShares Expanded Tech-Software Sector ETF (IGV) tracks roughly 110 North American software companies (Palo Alto Networks, Microsoft, Palantir, Oracle, CrowdStrike, Salesforce, and others). Over the past 2.5 years, SMH has more than tripled in value, rising to 3.37 times its starting level (the price of a share rising from $180.31 on January 18, 2024 to $607.73 on July 9, 2026), while IGV has increased to just 1.14 times its starting level (the price of a share rising from $82.19 on January 18, 2024 to $93.88 on July 9, 2026).

   We asked respondents to forecast the closing value of these two ETFs by the end of trading at the end of 2026 and 2028, rebasing the value so that 100 corresponds to the value of each ETF on July 9, 2026. If each repeated the historical trends from the past few years, it would suggest that by the end of 2028, SMH would have a value of roughly $2,048 (rebased value: 337, i.e. a 3.37x increase from July 9th levels) and IGV would be valued at roughly $107.23 (rebased value 114). The median expert instead forecasts SMH to reach 150 and IGV to reach 126 by the end of 2028. Superforecasters and the public make similar forecasts: SMH at 136 and 145, respectively; IGV: 120 and 126.

   Using the 10th, 50th, and 90th percentile forecasts provided by each respondent, we construct an average distribution across all experts and superforecasters. On average, experts and superforecasters forecast less than a 10% chance (8% and 7% respectively) that SMH will reach the trend-extrapolated value of 337 by the end of 2028. In fact, they predict that it’s more likely (experts: 17%, superforecasters: 27%) that SMH will decrease in value. While predicting similar probabilities that IGV will decline from its current value (experts: 25%, superforecasters: 29%), the average expert and superforecaster gives a 63% and a 58% chance the index will surpass the trend-extrapolated value of 114 by the end of 2028.

   

   

4. The average superforecaster predicts a 34% chance that the combined annualized revenue run rate of OpenAI and Anthropic exceeds $400 billion by the end of 2030 while forecasting an 18% chance that it's less than $100 billion.
   Recent reporting suggests that by 2030, OpenAI and Anthropic combined revenues could reach over $470 billion. In February 2026, Bloomberg reported that OpenAI expects revenue to exceed $280 billion by 2030 and in August 2026, Reuters reported that Anthropic is forecasting revenue in 2028 to reach $190 billion. For comparison, 10 companies on the 2026 Fortune Global 500 list have annual revenues above $400 billion and only about 100 companies have revenue of $100 billion or above.

   We asked respondents to forecast the combined annualized revenue run rate of OpenAI and Anthropic, assuming at least one of them exists as an independent company.¹ The median superforecaster gave a 93% probability that at least one of OpenAI or Anthropic would exist independently at the end of 2030. If one or both companies exist, the median superforecaster expects their combined annualized revenue run rate will be $300 billion in 2030, or roughly four times the $72 billion annualized revenue run rate reported in July 2026 at the time of forecasting. The median superforecaster predicts this will rise to $770 billion by 2040. Since the survey has closed, the reported combined annualized revenue run rate surpassed $105 billion (OpenAI: $40 Billion; Anthropic: $65 Billion).

   Using the 10th, 50th, and 90th percentile forecasts provided by each respondent along with their forecasted probabilities of at least one company existing independently, we construct an average distribution across all superforecasters. On average, superforecasters predicted a 34% chance that the combined OpenAI and Anthropic annualized revenue run rate in 2030 will exceed $400 billion while forecasting an 18% chance it's less than $100 billion.

   

   We chose to focus the insight above on the superforecaster predictions as we suspect a large proportion of our expert and public samples may have erred in forecasting this question such that their responses do not reflect their true beliefs or best judgment. Median forecasts for both groups were below the latest published value available at the time of forecasting. While one might hold the view that either or both companies’ revenue will collapse before year’s end, rationales showed no evidence that these forecasters intended to predict a reversal of recent trends. Instead, we suspect our elicitation design may have made it easy for forecasters to ignore this most-recent data point and anchor on earlier data. In the forecasting materials we provided two historical baselines for the combined annual revenue run rate: $30 billion from the end of 2025, and $72 billion from mid-July 2026. The $30 billion figure appeared on the visual user interface some forecasters use to enter their predictions, while the $72 billion figure appeared in an expandable information section that forecasters had to click to view. Almost half of expert forecasters gave median forecasts below the mid-July figure, indicating that they may not have seen or considered it. We are currently running an accuracy analysis of LEAP forecasts, which will include exploring the effects of elicitation designs and attentiveness on forecast accuracy.

Questions

• Frontier AI Company Revenue: What is the probability that at least one of OpenAI and Anthropic exists as an independent company at the end of the following calendar years? Given that at least one of these companies exists, what will be their combined annualized revenue run-rate (in billions of 2026 USD) at the end of the following calendar years? ⬇️

• AI-Exposed Equity Performance: What will be the closing value of the iShares Expanded Tech-Software Sector ETF (IGV) and the VanEck Semiconductor ETF (SMH) at the end of regular trading on the resolution dates below? ⬇️

• Data Center Investment: What will be the total annual US private fixed investment (in 2025 USD) in data center structures, electrical and communication structures, and Information Technology equipment for each of the below resolution years? ⬇️

For full question details and resolution criteria, see below.

Results

In this section, we present each question, and summarize the forecasts made and the reasoning underlying those forecasts. More concretely, we present background material, historical baselines, and resolution criteria; graphs, results summaries, results tables; as well as rationale analyses and rationale examples. We analyse these rationales alongside predictions to provide significantly more context on why experts believe what they believe, and the drivers of disagreement, than the forecasts alone.

Frontier AI Company Revenue

Question. 1. What is the probability that at least one of OpenAI and Anthropic exists as an independent company at the end of the following calendar years?

2. Given that at least one of these companies exists as defined in 1.a., what will be their combined annualized revenue run-rate (in billions of 2026 USD) at the end of the following calendar years?

Question (I). What is the probability that at least one of OpenAI and Anthropic exists as an independent company at the end of the following calendar years?

Results. All three groups are near-certain that at least one of OpenAI and Anthropic will still be an independent company at the end of 2026. The median forecast is 99% for experts, 100% for superforecasters, and 95% for the public. Confidence declines over time in every group. By 2030 the medians fall to 90% for experts, 93% for superforecasters, and 75% for the public, and by 2040 to 70%, 72%, and 50% respectively. Superforecasters are slightly but significantly more confident than experts in 2026 and 2030, though the two groups converge by 2040, where their forecasts are statistically indistinguishable. The public is significantly lower than both groups at all three dates. Disagreement widens at longer horizons for both experts and superforecasters. The public is the least homogeneous group throughout — even for 2026, a quarter forecast at or below 50% while a quarter forecast 99% or more.

Rationale analysis 1.a

• Acquisition: Forecasters who think it likely that at least one of the two companies will remain independent frequently point to high current valuations as being protective. “They would be too expensive to acquire for all but one or two companies,” writes one optimist, and another that “at current valuations…no feasible acquirer exists.” Pessimists tend to argue that current valuations could fall, and that this could facilitate an acquisition: “If anything happens where investors suspect future revenues might not be as incredible as AI promises, they may decide to pull the plug on these companies.”; “Displaced technology companies almost never end in liquidation; they end in acquisition.”; “The most probable reason is that they will be bought out by another company (e.g., Google or Microsoft). This occurs with some frequency especially in the tech world.”
• Nationalization: Both poles often point to state intervention as a potential threat to independence, but optimists tend to expect any nationalization instruments imposed will be of the lighter-touch variety that would leave the companies independent (according to the resolution criteria): “Hard nationalization seems quite unlikely, as we will likely see policy moves like golden shares, Defense Production Act orders, security agreements, board observers, and more before a government takeover is even viable.”; “Governments are likely to take non-dominant stakes but let companies remain independent.” Pessimists more often expect the harder version to prevail—i.e., “[a] world where frontier labs become strategic national assets and a government takes control”—and several note this outcome would likely affect both firms at once: “A government that takes one frontier lab in such a world will very probably take the other.”
• Access to capital: Several optimists cite the companies’ cash position, and upcoming IPOs, as virtually ensuring near-term independence: “Both companies are currently extraordinarily well capitalized and have filed for IPOs, making the probability that at least one remains independent through 2026 and 2030 very high.”; “Both Open AI & Anthropic have cash reserves to maintain their burn rate.” Pessimists tend to look at the same balance sheets and instead emphasize massive “compute commitments due over five years” and the fact that OpenAI and Anthropic are not well diversified: “Both OpenAI and Anthropic have no existing business supporting them, nor a proper plan B.”
• High bar set by resolution criteria: Optimists typically focus on elements of the criteria that allow for restrictions on independence without triggering a resolution: “The definition of independence is more permissive than it first appears. An IPO, Chapter 11 reorganization, close partnership with a cloud provider, or intensive government oversight would all leave a company independent.”; “It does not ask whether both remain independent. It asks whether at least one does, and even an OpenAI and Anthropic merger still counts as survival.” Pessimists tend to either not weigh these factors as much or to focus on all the ways the question could resolve; as one observes, “Conservatorship counts as ceasing to exist here.”
• Base rate for large-company disappearances: Many forecasters consider the corporate default rates supplied in the question details and note that—given that only one of the two companies need survive—this suggests very high survival probabilities, particularly for the shorter horizons: “Look at the base rate of replacement for companies. Say it’s about 4%. Formula is (1 - 0.04)^(N years). Apply N = 0.5, 4.5, 14.5. Result: roughly 98%, 83%, 55% for one company. Then apply the OR probability.” Pessimists tend to express skepticism that the base rate is relevant, given the nature of the business: “Totally possible that the other AI companies (which haven't even been founded) eat these for breakfast and a few more newer ones for lunch.”

Question (II). Given that at least one of these companies exists as defined in 1.a., what will be their combined annualized revenue run-rate (in billions of 2026 USD) at the end of the following calendar years?

Note: We suspect a large proportion of our expert and public samples may have erred in forecasting this question such that their responses do not reflect their true beliefs or best judgment. Median forecasts for both groups were below the latest published value of $72 billion available at the time of forecasting. While one might hold the view that either or both companies’ revenue will collapse before year’s end, rationales showed no evidence that these forecasters intended to predict a reversal of recent trends. Instead, we suspect our elicitation design may have made it easy for forecasters to ignore this most-recent data point and anchor on earlier data.

We are currently running an accuracy analysis of LEAP forecasts, and as part of this analysis we are exploring how the attention paid by forecasters affects the accuracy of their forecasts.

Results. Conditional on at least one of the two companies remaining independent, the three groups diverge on how fast the combined annualized revenue run-rate of OpenAI and Anthropic will grow, and on whether it will grow at all in the near term. The panel was shown baselines of roughly $30 billion at the end of 2025 and $72 billion by mid-2026. Against that, the median expert forecast for the end of 2026 is $70 billion and the median public forecast $41 billion; only superforecasters, at $90 billion, expect the run-rate to keep climbing. Over longer horizons all three groups expect substantial increases, but at very different scales. By 2030 the median forecasts are $150 billion for experts, $68 billion for the public, and $300 billion for superforecasters, rising by 2040 to $300 billion, $95 billion, and $770 billion, respectively. All three groups differ significantly from one another at every date and percentile. Superforecasters are the most uncertain group: for 2040, the median superforecaster puts the 10th percentile at $222 billion and the 90th at $1.3 trillion.

Rationale analysis 1.b

• Trendline extrapolation: Most forecasters cite the two historical baseline figures that were provided in the question details. (Roughly $30B at end-2025 and $72B by mid-2026.) High-revenue respondents typically extend it: “Both companies going from essentially nothing to a combined $72B run-rate by mid-2026 (already more than double the end-2025 figure) shows growth is still compounding fast, not decelerating.”; “Right now this is exploding exponentially.” Whereas low-revenue respondents tend to cite the recent explosive growth as evidence that a slowdown is inevitable: “You cannot sustain triple-digit annual growth against the ceiling of total enterprise software and consumer spend.”
• Commoditization: Most low-revenue respondents point to the potential for AI models to become commoditized as a result of competition: “Chinese labs and the open-weight people in the West keep putting near-frontier models out at close to the cost of running them. So customers are paying for the last little bit of capability plus whatever it costs them to switch. Neither of those looks safe to me.”; “Their models risk becoming a commodity, while most of the value ends up elsewhere: in hardware now, and in the applications built on top later.” High-revenue respondents tend to argue that an early advantage will likely prove durable, despite robust competition. “Most people are still using Microsoft and Apple despite OSS options,” notes one. Another adds that “the regulated professions are precisely where American providers are best placed to secure their position through the regulatory apparatus, and it is those clients rather than casual users who generate the revenue at issue here.”
• Level of diffusion: High-revenue respondents often argue that for the later resolution dates, revenue levels stop being just about software budgets: “The high end needs these companies to price against what a worker costs rather than what a software seat costs.”; “Once the product begins performing or materially enabling work rather than merely providing access to software, the relevant market starts to resemble a fraction of global knowledge-work spending rather than the traditional enterprise software market.” Low-revenue respondents more frequently align their revenue projections with existing budgets, implying a cap on the level of diffusion they foresee: “The largest companies today have a revenue of 300-400 billion, so I’ve capped their combined ARR to this value.”

AI-Exposed Equity Performance

Question. What will be the closing value of the VanEck Semiconductor ETF (SMH) and the iShares Expanded Tech-Software Sector ETF (IGV) at the end of regular trading on the resolution dates below?

Question (I). iShares Expanded Tech-Software Sector ETF (IGV)

Results. Forecasts are indexed so that 100 equals IGV's closing value on 9 July 2026. All three groups expect gains by the end of 2026 and continued growth through 2028. Superforecasters give the lowest median forecasts at both horizons, at 104 and 120, and are significantly below both other groups for 2026. By 2028 the groups are statistically indistinguishable at the 50th percentile, with experts and the public both reaching 126. Forecasters are nonetheless uncertain about the size and direction of the change: each group's median 10th percentile falls below the base period at both horizons, ranging from 73 to 78 for 2026 and from 73 to 85 for 2028. The upper tail widens over the same span, with median 90th percentile forecasts of 130 to 140 for 2026 rising to 165 to 179 for 2028.

Question (II). VanEck Semiconductor ETF (SMH)

Results. Forecasts are indexed so that 100 equals SMH's closing value on 9 July 2026. All three groups expect stronger gains for SMH than for IGV. The median forecast for the end of 2026 is 105 for superforecasters, 112 for the public, and 112 for experts, rising by 2028 to 136, 145, and 150 respectively. Experts are the most bullish group throughout and sit significantly above superforecasters at both horizons, while expert–public differences are significant only for 2028. Forecasters nonetheless remain uncertain about the size and direction of the change. As with IGV, every group's median 10th percentile falls below the base period for 2026, ranging from 70 to 75. By 2028 the groups diverge sharply in the lower tail: experts' median 10th percentile rises to 99 and the public's to 90, while superforecasters' falls to 61.

Rationale analysis

General drivers

• Respondents frequently cite historical stock market trends as a central consideration: “Applying the average growth rate of the SP500 for both for the midpoint.”; “The price of stock represents, in and of itself, a consensus aggregate estimation of the potential future value of a company. Assuming the market is efficient, this estimate is a hard benchmark to beat. My prediction is essentially a baseline estimate of how these ETFs would move based on historical average growth and variance.” Many forecasters also consider the potential for macroeconomic events to influence the question outcome as much or more than sectoral considerations. As one notes, “A recession, renewed inflation, higher real rates, geopolitical conflict, supply-chain disruption, or a broad retreat from growth equities could overwhelm excellent company-level fundamentals.” Finally many forecasters also note that both funds had drifted below the 9 July base (at the time they completed the survey) so that “simply forecasting no change would already amount to calling for a rebound.”

SMH (semiconductors)

• The compute buildout: High-valuation respondents often view the near-term buildout as showing no signs of slowing and largely already locked in: “The current chip shortage has the various companies saying the shortage won't be back to a balance between supply and demand until 2029 or 2030 at the earliest. ASML…said…that they are, essentially, sold out through 2027 and are now accepting 2028 orders. Micron has already said they have bookings into 2028.”; “Hardware demand seems insatiable in the short to mid-term.” Low-valuation respondents tend to question the durability of the current demand: “Large-scale global data center expansion wraps up between 2027 and 2028, sending the industry into a temporary Digestion Phase.”
• Current valuation: Many low-valuation respondents acknowledge semiconductor fundamentals are good, but argue future returns may suffer given the recent runup in stock prices: “The relevant question is no longer whether chip revenue rises. It is whether it rises faster than the heroic growth assumptions already embedded in the price.”; “Growth has certainly been explosive in recent years but it is starting to seem like a lot of the value in these companies is already priced into their stock prices.” High-valuation respondents tend to push back on that narrative: “I believe the market still underprices AI progress, most acutely on the hardware side, because most of the world has not internalized how large AI's economic footprint becomes across the majority of plausible trajectories.”
• Cyclicality: A common low-valuation argument is that an AI-bubble need not pop for values to fall, just for the ordinary chip cycle to assert itself: “Over capacity is the classic boom bust that drives hardware stock prices”; “Memory is a commodity which has been in a scarcity. It has already corrected some and will face oversupply at some point, even as demand continues [to be] very strong.” High-valuation respondents tend not to focus on prior cycles so much as the likely level of future demand. “Demand will be higher as AI adoption widens over time,” writes one, with another arguing that “semiconductors are critical to the modern economy; even if today's leading AI firms fail, the world will still have an inexhaustible need for semiconductors in the next 2-3 years.”
• Taiwan: Geopolitics is the largest risk respondents from both poles point to. One writes, “I place roughly a 15% probability on a serious Taiwan crisis by the end of 2028, and since that scenario would devastate this fund in particular, my bottom decile simply is that world.”; “If China attempts an attack on Taiwan, the semiconductor industry would get hit dramatically.”

IGV (software)

• Whether AI sells software or replaces it: This is the most commonly considered IGV driver, and the reason most respondents perceive that SMH has more upside potential. Low-valuation respondents tend to argue that AI is at risk of rendering many software companies obsolete: “I expect AI to commoditize software to the point it is free. For this reason, I do not believe IGV has a bright future”; “Software needs are threatened by AI’s ability to produce software.” High-valuation respondents, however, aren’t convinced the threat is so dire. As one writes, “Companies like Oracle and Microsoft have broad capabilities and deep commercial relationships and they’ve turned previous technological threats, such as cloud computing, into very profitable opportunities. Agents may well be the future of software but if the likes of Microsoft and Salesforce control access to those agents for most large corporates then that will be a profitable business model for them well into the future.”
• Poor 2026 YTD returns: A frequent high-valuation argument treats IGV’s recent underperformance as a reason to expect a better forward return: “Software is in the oversold territory.”; “I expect IGV to modestly outperform because software…valuations have already compressed, leaving more room for recovery.” Low-valuation respondents tend to read the recent poor returns as confirmation that the industry is likely “to keep underperforming” because the shift of value to frontier labs “has further to run rather than being fully priced.”
• Underlying companies: Respondents who look inside the IGV basket often point out that the companies are not all equally exposed. “Roughly 40% of the fund is seat-priced application SaaS [software as a service] that loses in high-AI worlds,” argues one low-valuation respondent, while several high-valuation respondents argue the fund “has drifted toward being substantially a cybersecurity vehicle,” and “cyber security may do even better in the coming months and years as concerns over AI enabled threats grow.” Another high-valuation respondent, addressing the risk to SaaS, writes, “whilst I think AI is revolutionising apps, it will never fully replace enterprise ready apps like Oracle (honestly who’d want to use a vibe coded database), Microsoft etc.”

Data Center Investment

Question. What will be the total annual US private fixed investment (in 2025 USD) in (a) data center structures, (b) electrical and communication structures, and (c) Information Technology equipment for each of the below resolution years?⁴

Question (I). (a) Data center structures

Results. Forecasts are indexed so that 100 equals 2025 investment in data center structures, a base of $40.8 billion. All three groups expect rapid growth by 2026, continuing through 2028, when experts give the highest median forecast at 176, followed by superforecasters at 170 and the public at 148. Investment keeps rising through 2035, but at a slower annual rate, and superforecasters overtake experts to give the highest median, at 250. Superforecasters are the most uncertain group at that horizon: their median 10th and 90th percentile forecasts for 2035 span 132 to 400. At every horizon, experts' and superforecasters' median forecasts are statistically indistinguishable, while the public is significantly lower at the 50th percentile throughout.

Question (II). (b) Electrical and communication structures

Results. Forecasts are indexed so that 100 equals 2025 investment in electrical and communication structures, a base of $146.4 billion. All three groups expect steady growth and broadly agree on its pace at each horizon, with median forecasts of around 105 by 2026, between 112 and 120 by 2028, and 132 to 138 by 2035 — roughly a third above the 2025 level. This is far slower than the growth they expect for the data centers this infrastructure powers, which experts see rising 76% by 2028 alone. By 2035 every group's median 10th percentile sits above the base period.

Question (III). (c) Information Technology equipment

Results. Forecasts are indexed so that 100 equals 2025 investment in IT equipment, a base of $514.8 billion. All three groups expect substantial growth: median forecasts of around 120 by 2026, between 140 and 148 by 2028, and 180 to 200 by 2035. This places IT equipment between the two other categories, growing faster than the electrical and communication structures that power these facilities but slower than the data center shells themselves. Experts and superforecasters are statistically indistinguishable at every horizon, while experts sit significantly above the public at the 50th percentile for 2026, 2028 and 2035. Superforecasters are the most uncertain about the upper tail, giving a median 90th percentile of 291 for 2035 against 242 for experts and 230 for the public.

Rationale analysis

General Drivers

• AI buildout sustainability: A central cross-category consideration is whether current spending on the AI buildout is sustainable. High-investment respondents typically conclude that it is: “I now think most of this investment is backed by real, growing demand rather than hype: demand looks like it's outrunning supply, with power and grid connections the main bottleneck rather than any lack of appetite. So I expect the buildout to keep climbing.”; “Enterprises have barely begun moving substantial production workloads into persistent AI systems.” Low-investment respondents often express skepticism: “In all cases there is an overcapacity possibility that would need to be accounted for.”; “The historical trend line is unsustainable.”
• Offshoring: Several respondents note a buildout that migrates offshore wouldn’t be captured by a measure of US private investment. As one writes, “China’s push toward domestic chip supply chains decoupled from the HBM[high bandwidth memory]/memory crunch, plus Gulf-state sovereign buildouts, means a growing share of new AI infrastructure investment increasingly lands outside the US NIPA [national income and product] accounts entirely—this doesn't shrink global AI capex, but it directly caps the US-specific structures and equipment lines [this] question asks about.” Another respondent notes that “nationalization moves spend out of private accounts,” as well.

Data Center Structures

• Announced capex: Many high-investment respondents argue published capex guidance and announced project lists is highly likely to lead to exceptional increases in investment in 2026 and 2028: “Federal Reserve research using project-level data estimated that broader US data center investment could reach roughly $370 billion at an annualized rate by the second quarter of 2026, while announced plans exceed $1.5 trillion, only a minority of which has yet been realized.” Low-investment respondents tend to point to such high levels of capex as a reason it can’t continue indefinitely: “The investments are becoming such a large fraction of the economy that they cannot grow at such a sustained pace in future, unless there is a economic explosion ahead of us which I am deeply sceptical about.”; “This will peak soon given the capex of the mag 7.”
• Physical constraints: Most respondents cite factors such as power interconnection, permitting, availability of appropriate sites, and labor as key considerations, with low-investment respondents tending to view these as setting a ceiling on potential growth: “Structures have very long useful lives, and annual construction eventually encounters the inconvenient physical characteristics of the United States. Suitable land with access to enormous quantities of power, fiber, cooling, water, transmission, construction labor, transformers…is finite.”; High-investment respondents tend to emphasize sharply rising demand for compute, and the pace of the recent buildout, above this consideration. One writes, “While planning delays, power constraints, permitting, and financing conditions may temporarily slow construction, I expect the overall trend to remain strongly positive through 2035.”
• Political opposition: Low-investment respondents frequently cite moratoria and zoning fights, with one pointing to “the recent New York state moratorium on new data centres” as evidence of “deeper public opposition in the US than in other comparable countries.” Several high-investment respondents acknowledge the opposition but doubt it will really slow down construction: “It may [see] success in some locations, but unlikely to succeed is slowing down overall US buildup.”; “The same [opposition] might also have been true with the huge Amazon warehouses—ultimately it takes some time for people to get used to new things but ultimately there is acceptance.”; “Political opposition will be a headwind throughout this time but not enough to slow growth.”

Electrical and communication structures

Rationale analysis

• AI’s limited impact: Many low-investment respondents stress that this category (along with the information technology equipment category below) is economy-wide: “Data centers are under 5% of [electrical and communication structures]’s level—the key fact about [this] category, since it makes [it] insensitive to a data center boom.”; “Even a dramatic AI-driven surge arrives diluted.” High-investment respondents tend to acknowledge this factor, but emphasize that the scale of new load is large enough to show up regardless: “The more data centers that are actually completed, the more obvious it becomes that generation, transmission, distribution, substations, fiber routes, and communications infrastructure need to be rebuilt around a load-growth regime that the United States has not experienced for decades.” Other high-investment respondents point to factors unrelated to AI: “The electrical grid is in desperate need not only for data centers, but also electric cars and tools. As climate change gets worse, there may be more of a push…to move away from gas power to electricity.”
• The flat history: Low-investment respondents often note that the real level of investment has barely moved in recent years: “[It] has been flat for five years despite constant talk of grid investment.”; “Fifteen years essentially flat, through the shale revolution, the renewables boom, and the first years of the AI surge.” Several high-investment respondents read the same flatness as having led to a backlog that will lead to a spike in investment in the years to come: “My 2028 figure has the bottleneck forcing real acceleration, and by 2035 I expect a construction boom of a scale not seen since the 1970s.”
• Bureaucracy and regulation: Some low-investment forecasters argue that utilities will struggle to expand quickly and overcome supply chain constraints given that “utilities are slow companies and permitting/construction is slow even if there is demand.” One adds that investment in this domain “is prohibitively expensive, politically tricky for governments to fund, and presents all the same permitting, design, construction labor, and other costs that create bottlenecks for the data centers.” High-investment respondents tend to acknowledge this element, but believe demand—along with “money, political influence, and geopolitical expediency”—will “force the issue toward the end of the decade.”

Information technology equipment

Rationale analysis

• The replacement cycle: High-investment respondents often cite the short life of hardware as a key reason investment in this sector will remain robust: “Structures [like data centers] last twenty-five years and servers last three or four. So once the installed base is large, replacement demand drives the equipment number even when construction stops growing.”; “Even if the growth of data centers slows, there will still be a demand for replacement of these systems.” Low-investment respondents don’t dispute that hardware has an existing shelf life, but instead point to reasons why this factor might not be enough. As one writes, “Open and smaller models could shift more inference onto existing or edge hardware, foreign supply constraints could limit quantities, and the late-2020s buildout could leave the economy digesting a mountain of underutilized equipment.”
• Falling prices: Respondents from both poles broadly agree hardware is getting cheaper per unit of performance; low-investment respondents typically argue that this will suppress total investment. One notes, “Real dollar spending can stagnate even as unit compute rises sharply, since the [cost of compute]...historically falls,” and another that “prices and efficiency improvements offset some of the increase in real investment.” High-investment respondents instead tend to focus on recent confirmed growth rather than projections: “We have already seen unprecedented growth in computers, communication equipment, and other apparatus in 2025, a nearly one third jump driven by the server and GPU surge”; “BEA printed +45.9% annualized real growth in Q1 2026.”
• Breadth of AI diffusion: Several high-investment respondents point to the likelihood that AI will diffuse broadly across the whole economy, and that as “AI becomes embedded throughout enterprise computing” it will begin “producing persistently higher equipment investment than before the AI boom.” One notes that we will also “have more robots interacting with the physical world and they will need IT support.” Low-investment respondents tend to emphasize factors like “memory/transformer scarcity” as reasons IT investment may remain muted, regardless of demand.

Footnotes

1. “Exists as an independent company” means that it has not failed, been acquired by another entity, or been nationalized. For more details, see the resolution criteria for Frontier AI Company Revenue below. ↩

2. In some cases, the "aggregate" refers to the mean; in others, the median is used, depending on which is more appropriate for the distribution of responses. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

3. We occasionally elicit participants' quantile forecasts (estimates of specific percentiles of a continuous outcome) to illustrate the range and uncertainty of their predictions. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

4. The historical data accompanying Question 3 initially displayed nominal investment values rather than the constant-2025-dollar values the question resolves on. These values were corrected to inflation-adjusted values on July 27th after the survey had been open for 2 weeks from the 13th July. Expert and superforecaster participants were notified of the correction on July 27th and given until August 3 to revise their responses; public participants saw the corrected figures from July 27 but were not separately notified. ↩