Wave Report
Wave 13: Short-Term Cruxes
In September, we released an editorial describing how we're improving the accuracy and relevance of our AI progress forecasts. In Wave 13 of LEAP we started to put some of these ideas into practice.
Published: November 19, 2026
Potential improvements we implemented in this wave, described in our September editorial, include:
- The inclusion of short-term resolving questions to enable faster identification of more accurate forecasters
- Improving the information we give to forecasters while they are forecasting
- Adding new forecasters with a track record of accuracy on other forecasting platforms
- Separating out forecasts from a group of panelists who expect faster AI progress than most LEAP forecasters
- Re-asking questions that appeared in earlier waves to understand how forecasters are updating their views
These experiments have already yielded some promising results. One question from this wave has already resolved—the quickest turnaround between forecasting and resolution across the entire LEAP project. We're also seeing some interesting differences in forecasts between our existing and new forecaster groups, although we need more data from upcoming waves before we can draw firm conclusions in this area.
The following report summarizes responses from 143 experts, in addition to 49 superforecasters, 133 AI risk experts, and 494 members of the public collected between Oct 06, 2026 and Oct 27, 2026. Expert respondents include 23 computer scientists, 31 industry professionals, 37 economists, and 52 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
-
Opposition to data centers among US voters ran higher than forecasters expected.
As of October 1, Ballotpedia was tracking 39 data-center-related ballot measures. We asked forecasters what percentage of the 39 would produce an outcome limiting data center development. The median expert forecast 29, (75%) the median superforecaster forecast 31 (80%).In actuality, all 39 of the ballot measures listed (100%) were decided in favor of limiting data center development. This exceeded even the median 90th percentile forecasts of experts and superforecasters who forecast 90% (IQR: 75–95) and 95% (IQR: 90–99), respectively. Of the 143 experts who completed this question, 18 gave 90th percentile forecasts of 100% (12.6% of experts) while 12 out of 49 superforecasters (24.5%) gave 90th percentile forecasts of 100%. Among AI risk experts, 10 out of 133 respondents (7.5%) gave a 90th percentile forecast of 100%.
In their written rationales, forecasters highlighted that opposition to data centers was broad and intense across the political spectrum. Forecasters also pointed to the concentration of ballot measures in Ohio—where 21 of the 39 measures were being voted on—referring to polls in Ohio suggesting that high numbers of residents opposed data center construction, and that the results across the state were likely to be correlated.
Each group’s median forecast and median 10th–90th percentile range for the share of the 39 measures limiting data center development, against the result of 100%. -
Forecasters updated toward faster AI progress and greater industry revenue.
In this wave we asked updated versions of two questions that forecasters had answered in earlier waves: the probability that AI will help solve another Millennium Prize Problem, and the combined future annualized revenue run-rate of OpenAI and Anthropic. In previous LEAP waves, forecasters have tended to underestimate AI progress. For example, in summer 2025 the median expert assigned just a 10% probability to AI helping solve a Millennium Prize Problem before the end of 2027—something that was likely achieved by OpenAI in September 2026.When re-asked versions of these questions, forecasters significantly updated toward expecting more AI progress. The median expert forecasts an 80% probability that AI helps solve at least one Millennium Prize Problem (other than Navier–Stokes) before the end of 2030—roughly a 4x increase over the equivalent forecast from summer 2025. The median superforecaster prediction for the same question rose from 20% in summer 2025 to 85% in fall 2026.
Forecasts for the combined OpenAI and Anthropic annualized revenue run rate also increased sharply. When asked between July and August, 2026 superforecasters predicted that combined revenue would reach $90 billion by the end of 2026. When asked the same question in October 2026, superforecasters gave a median of $175 billion—almost 2x their original forecasts. The median superforecaster's update for the end-of-2030 revenue was even sharper, going from $300 billion (forecasted in July–August 2026) to $700 billion (forecasted in October 2026), a 2.3x increase.
We can see the same trend in AI risk experts, who also gave revenue forecasts in July–August, 2026, although at slightly later date than superforecasters. In July–August, the median AI risk expert forecast a combined annualized revenue run rate of $100 billion in 2026 and $393 billion in 2030. When asked the same question in October 2026, the median AI risk expert gave updated forecasts of $175 billion (a 1.75x increase) in 2026 and $680 billion (a 1.73x increase) in 2030.
As well as median forecasts increasing across groups, the within-forecaster predictions also increased between waves. Among respondents who answered both waves, 91% of experts and 98% of superforecasters raised their probability of AI solving at least one additional Millennium Prize Problem by 2030. For the revenue question, 96% of experts and 90% of superforecasters who answered both questions increased their forecasts at the later asking.
Despite updating toward more significant AI progress on the Millennium Prize question, the median LEAP expert still forecasts a lower chance that another Millennium Prize Problem will be solved by the end of 2026, when compared to some prediction markets forecasts. For example, mid-way through our survey window, a Manifold market assigned this outcome a 39% probability, and an equivalent question on Polymarket assigned this a 23% probability, compared with a median LEAP expert forecast of 20%.
In their rationales for the Millennium Prize Problem question, forecasters pointed to the fast progress in AI math on the whole as evidence that future problems may also be solved, with some noting that the difficulty of remaining problems and the slow verification process would limit new solutions before the end of 2026. For the revenue question, forecasters pinpointed the rate at which hypergrowth might slow as a key factor influencing their forecasts, with respondents often citing competition from open-weight and Chinese models as possible threats to revenue generation for OpenAI and Anthropic.
Median forecasts in Wave 13 compared with near-equivalent questions asked in Wave 2 (Millennium Prize) and Wave 11 (revenue run-rate). -
When respondents saw the group's views while forecasting, their forecasts did not change.
In order to assess whether exposure to the panel's aggregate forecasts changed participant behavior beyond receiving LLM-generated forecasts, half of participants were randomly selected to be shown the running median of participant forecasts. Experts and superforecasters within this group also received a mid-wave email summarizing the rationales submitted.We did not observe meaningful differences in the median forecasts between participants who saw the running median (treatment) and those who did not (control). Out of 184 comparisons (i.e., 46 forecasts across 4 groups), the groups differed significantly (p < 0.05) in only 7 comparisons, or what we would expect by chance. We also examined spread of forecasts, as aggregate forecasts become more accurate when individual judgments differ from one another so that their errors offset, and exposure to a running median could reduce the independence of errors (Davis-Stober et al., 2014). For participants in the treatment group, being provided the median did not appear to significantly reduce the spread across all forecasts. For experts who saw the running median, spreads were 8% narrower on average across the 46 forecasts (narrower in 24 of 46), but this was not statistically significant (p = 0.24).
In this wave, all participants were shown LLM-generated forecasts after submitting their initial forecast on each question. It is possible that respondents in both groups anchored to the LLM forecast and that this effect washed out any potential effects of shared aggregates. Wave 14 will randomize who sees the LLM forecasts to partially test this hypothesis.
-
The subset of experts who expect fast AI progress forecast similarly to other experts on short-run questions
In a recent post, we shared that much of the LEAP panel has underestimated AI progress on benchmarks and other capabilities indicators, while a subset of experts who expect faster AI progress has been more accurate. Beginning in this wave, we share forecasts of this “fast AI progress” subset of experts to understand how their forecasts compare with the group as a whole. We defined this “fast AI progress” subset as experts and AI risk experts who met the following criteria from their forecasts on previous waves:- a combined probability of more than 50% that AI reaches level 9 or 10 on the Technological Richter Scale by 2040 (Wave 8)
- a probability of at least 90% that, by 2100, more than half of LEAP panelists will agree that AGI exists (Wave 8)
- a median forecast of 2040 or earlier for when AGI arrives (Wave 8)
In this wave, 40 of the 276 experts and AI risk experts meet these criteria. Their forecasts are now available on this site for every LEAP question since Wave 1: select “Fast AI Progress Experts” in the Groups menu above any forecast’s chart. On short-term questions, the group is broadly similar to other experts. For questions with resolution dates by the end of 2026, the median forecast of fast AI progress experts is similar to the overall expert and AI risk expert median. For example, the median fast AI progress expert forecast a combined revenue run rate of OpenAI and Anthropic of $179 billion compared to $175 billion for both the median general expert and AI risk expert. For later resolution dates, the median forecasts separate. For example, in 2030 the median fast AI progress expert forecasts a revenue run-rate of $964 billion which is nearly double the median expert forecast ($523 billion) and above the median AI risk expert forecast ($680 billion). This pattern holds across several questions: the fast progress sample mirrors other experts on short-term questions and diverges only on later-resolving predictions. Where this pattern breaks, it is because the later forecasts do not diverge. On short-term questions, the fast-progress group's median stays close to others' on most forecasts, though it runs somewhat higher on a few (e.g., Democrats discussing AGI risk: 95 vs. 85 for experts and 80 for AI risk experts).
One implication of this is that these short-term resolving questions will not substantially differentiate these groups' medians in terms of accuracy. They may still be useful, though, in differentiating individuals' accuracy.
Median Wave 13 forecasts from experts, AI risk experts and fast AI progress experts on three questions asked about more than one resolution date.
Questions
-
Members of Congress Discussing AGI Risk: According to the AI Policy Network (AIPN)'s Congress on Superintelligence tracker, how many current members of the U.S. Congress will be listed as having publicly discussed AGI, AI superintelligence, AI existential risk, AI loss of control, recursive self-improvement, or the Singularity, as of December 31, 2026, who are: a. Democrats (or independents caucusing with Democrats)? b. Republicans (or independents caucusing with Republicans)? ⬇️
-
Data Center Ballot Measures: Of the 39 data center-related ballot measures in 2026 tracked by Ballotpedia (as of October 1st 2026), what percentage will produce an outcome limiting data center development by December 31, 2026? ⬇️
-
AI Solves Millennium Prize Problems: (1) What is the probability that AI will solve or substantially assist in solving a Millennium Prize Problem in mathematics other than the Navier-Stokes Millennium Prize Problem by the end of the following calendar years? (2) What is the probability that AI will solve or substantially assist in solving at least three of the five remaining ‘unsolved’ Millennium Prize Problems (not counting the Navier-Stokes Millennium Prize Problem) by the end of the following calendar years? ⬇️
-
AI Spend Per Employee: What will December 2026 AI spend per employee be for firms in the following categories of the Ramp AI Index? ⬇️
-
Frontier AI Company Revenue: (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 (a), what will be their combined annualized revenue run-rate (in billions of 2026 USD) at the end of the following calendar years? ⬇️
-
Capabilities (ECI) of Closed and Open Models: What scores will the top-performing closed-weight system and the top-performing open-weight system have on the Epoch Capabilities Index (ECI) at the end of the following calendar 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.
Members of Congress Discussing AGI Risk
Question. According to the AI Policy Network (AIPN)'s Congress on Superintelligence tracker, how many current members of the U.S. Congress will be listed as having publicly discussed AGI, AI superintelligence, AI existential risk, AI loss of control, recursive self-improvement, or the Singularity, as of December 31, 2026, who are: a. Democrats (or independents caucusing with Democrats)? b. Republicans (or independents caucusing with Republicans)?
Results. At survey launch, the AIPN tracker listed 72 members of Congress: 46 Democrats, 2 independents who caucus with the Democrats, and 24 Republicans. All four groups expect the count to rise by December 31, 2026, and expect Democrats to keep outnumbering Republicans, but disagree on how large the rise will be. Experts give the highest median forecasts: 85 Democrats and 42 Republicans, 127 members in total, an increase of approximately 76%. AI risk experts' medians of 80 and 40 are slightly lower but statistically indistinguishable from experts', and the public's are 73 and 40. Superforecasters are the most conservative at 68 and 33, about 101 members in total, an increase of approximately 40%, significantly lower than both experts and AI risk experts.
Rationale analysis
- The chance of another shock: High-forecast respondents tend to give the occurrence of a new Hugging Face-like incident substantial weight: “Forecasters, and LLMs, are underrating the likelihood of large warning shots that change public sentiment.”; “If another major incident (misuse, accident, cybercrime/war) happens over the next 2 months I could see the topic and the discussion around it spike so that almost everyone has a public opinion / statement on it.” Many low-forecast respondents consider this angle, but deem it either unlikely to occur before December 31 or—if it does occur—unlikely to cause many more politicians to speak out. As one argues, “A shock the same size as September converts almost nobody new, because everyone with a threshold that low already spoke.”
- The midterms: Respondents at both poles cite the November 3 election, with high-forecast respondents often arguing that the campaign will force members to take positions on AI. One writes that “the discussion is in the zeitgeist…there will probably be some show votes and hearings on the matter.” Low-forecast respondents more often expect other campaign issues to crowd the subject out: “Affordability, the Iran war and fuel prices will compete for attention.”; “AI ranked 11th out of 16 in issues that are 'very important' to voters.” One also notes “Congress is out until after the Nov 3 election, and the lame duck is short and crowded.”
- AI progress: Some high-forecast respondents predict the rapid pace of AI progress will drive numbers higher: “I do expect the numbers to increase as AI capabilities continue to improve fairly dramatically in coming months, and we likely see more dramatic capability demonstrations and also alignment failures—especially as open source models near the Mythos-frontier become more capable.” Low-forecast respondents generally focus on the likelihood that talk of AI, when it occurs, will be more “about jobs, data centers, [and] electricity bills, which does not qualify.”
- Reading September's surge: Many high-forecast respondents treat the 29 members added between September 9 and October 4 as the start of a trend that is likely to persist and “not a one-off.” Low-forecast respondents more often link the surge directly to Jacob Coxon's resignation and the Sanders–Casar bill, with several noting that the October 14 update added only a single name: “September was a burst, not a trend.”
- Republican uptake: Most respondents expect Democrats to stay well ahead, with many arguing the issue allows Democrats to “stand in opposition to Trump.” Others point to the urban/rural divide: “Democrats in urban areas will appeal to their constituents with more specific topics of AI, while fewer Republicans will gain the same value doing the same in rural areas.” Another common sentiment is that, for Republicans, “if they criticize AI, they might run afoul of the president's position on AI and if they praise AI, they may be out of step with votes. Better to not raise it.” Those with higher Republican forecasts more often expect voter pressure, national-security issues, or the appeal of populist rhetoric to lead Republicans to speak out. Several also argue that “Trump's preference for the word 'superintelligence' may push some Republicans to be counted.”
Data Center Ballot Measures
Question. Of the 39 data center-related ballot measures in 2026 tracked by Ballotpedia (as of October 1st 2026), what percentage will produce an outcome limiting data center development by December 31, 2026?
Results. As of October 1, 2026, Ballotpedia listed 39 data center-related ballot measures. Four had already produced outcomes limiting data center development, and the remaining 35 were due to go to voters in November. Three groups expect roughly three-quarters or more of the measures to limit development, while the public expects close to an even split. Superforecasters give the highest median forecast at 80%, significantly higher than every other group. They also agree most closely with each other: half forecast between 70% and 87%. Experts and AI risk experts give almost identical forecasts, with medians of 75% and 74% respectively. The public's median, 54%, is significantly lower than every other group's. It falls below superforecasters' median 10th-percentile forecast (62%) and only just above those of experts (51%) and AI risk experts (50%).
Rationale analysis
- Public opposition: Many respondents describe the opposition as broad and intense: “People really hate data centers right now. Rarely do issues receive so much agreement that people from all sides of the political spectrum attend rallies, protests, and even city council meetings in order to oppose the issue.” Several cite polling—“polls show really strong support for limiting data centers across the board”—and one argues that hard-to-poll, low-trust voters “are likely to be even more anti data centers.”
- Local costs versus local benefits: High-forecast respondents often point to rising electricity bills, water use, and noise as tangible local costs that lack a commensurate benefit: “Data centers take up land and strain local water and power, while creating few lasting jobs.”; “Benefits are diffuse in the society and costs are concentrated locally.” One describes data centers as an “easy target” given that they have become “a big symbol” of cost-of-living frustration. Low-forecast respondents push back on the narrative that local benefits will be meager: “I expect most restrictive measures to fail because of the strong economic incentives associated with data center investment, including tax revenues, job creation, and infrastructure development.”
- Prior votes on data centers: All four measures already decided produced outcomes limiting development (two by approving restrictions, two by rejecting pro-development proposals), with the limiting side taking between 66% and 90% of the vote. Many high-forecast respondents suggest this lends considerable insight into how the remaining measures will fare: “4 out of 39 have been decided, and they all led to limitations, so base rate 100%...With increased media reporting about the risks and costs of data centers, most other cases will probably be similar.” Others, however, stress the small sample: “The 100% base rate out of 4 definitely pushes my estimate up, although these probably passed early because they were clearer-cut than the others.” A few also note that the early contests drew low turnout and predict “a midterm electorate will be less lopsided.”
- Ohio's weight: With 21 of the 39 measures in Ohio, many respondents see the state as decisive. High-forecast respondents often cite state polling, such as “a BGSU/YouGov poll saying 75% of Ohio opposes a data center near them and 78% in favor of moratoriums.” Many also see Ohio as the main source of uncertainty given that “outcomes may be correlated across Ohio communities” and that depending on “how hard republicans and others campaign” there, the state “can actually go either way.”
- Citizen-initiated measures signal discontent: Many high-forecast respondents argue that citizen-initiated measures are, in and of themselves, evidence that local opposition is strong: “These measures are on the ballot in these places because of opposition to data centers in the first place”; “People that are happy with the status quo don't advocate for new policy measures.” Others note that this dynamic doesn't apply to all the measures listed, some of which are pro-development: “...industrial townships near St. Louis asking whether to permit data centers reads like the pro side picked the venue, so I lean favoring.”
- Influence money: Several low-forecast respondents expect industry money to prevail: “I think big tech has too much influence for there to be much change or limit to data center development.” Even some high-forecast respondents expect a few bans to fail in towns, arguing that “lobbying by strong interest parties…might mean that the current trend towards 100% will be curbed going forward.”
AI Solves Millennium Prize Problems
Question. (1) What is the probability that AI will solve or substantially assist in solving a Millennium Prize Problem in mathematics other than the Navier-Stokes Millennium Prize Problem by the end of the following calendar years? (2) What is the probability that AI will solve or substantially assist in solving at least three of the five remaining ‘unsolved’ Millennium Prize Problems (not counting the Navier-Stokes Millennium Prize Problem) by the end of the following calendar years?
Results. All groups consider it unlikely that AI solves, or substantially assists in solving, a remaining problem by the end of 2026, but their forecasts rise sharply for 2030. Experts and AI risk experts give identical median forecasts at both horizons: 20% by the end of 2026 and 80% by the end of 2030. Superforecasters and the public both give 10% for 2026 but diverge sharply for 2030. Superforecasters give the highest median of any group at 85%, statistically indistinguishable from both expert groups, while the public gives 44%, significantly lower than every other group.
A similar pattern holds for the probability that AI solves, or substantially assists in solving, at least three of the five remaining problems. Median forecasts for 2026 fall between 1.5% and 3% but rise substantially by 2030 and further by 2040. For 2030, superforecasters and AI risk experts give medians of 50% and experts 45%, which are statistically indistinguishable, while the public is significantly lower than all other groups at 20%. For 2040, superforecasters again give the highest median forecast at 88%. Experts and AI risk experts both give 80%, statistically indistinguishable from superforecasters, while the public is again significantly lower than every other group at 49%.
Rationale analysis
- AI's trajectory in math: Most respondents predicting fast progress point to the pace of general AI progress in math, the rise of AI agents working in parallel, the September Navier–Stokes proof and OpenAI's release of other solved frontier-math problems, or all of the above, as a sign that Millennium-level mathematics is within AI's reach: “I don't think these math problems need fully new ways of thinking. They can be solved by brute force using a tree of agents.”; “The Navier-Stokes result says that solving Millennium prizes is a good deal easier than I used to think.” Others acknowledge that general progress has been fast but tend to downplay the Navier-Stokes breakthrough: “OpenAI's result is impressive, but it exploits a loophole in the official problem statement…and it builds on ground humans had already prepared extensively.”; “Creativity is not the same as brainpower/compute power.”
- Difficulty of the remaining five: Many skeptics argue that the remaining problems are quite different from and likely more difficult than Navier–Stokes: “These five problems are not interchangeable [in terms] of difficulty, and success on one does not necessarily imply that the others become straightforward.”; “The remaining five are mostly 'prove a universal statement' problems with no known viable program (P vs NP faces relativization/natural-proofs/algebrization barriers; RH has no credible attack; Hodge/BSD need new theory).” One adds, “There's no guarantee that all of them can be solved.” Fast-progress respondents tend to single out problems that are rumored to be within reach: “Hodge and BSD [Birch–Swinnerton-Dyer] seem like gimmes (both saw significant progress in the OpenAI drop).”
- The 2026 window and verification: Slow-progress respondents often point out that a proof would need to be found, released, and vetted to count toward a 2026 resolution. “Three problems is an absurd number for 2.5 months,” writes one. Another calls confirmation by year-end “nigh impossible.” Fast-progress respondents, however, tend to see the expert-panel route as viable: “The panel criterion is the only realistic path in 2026, and FRI's ruling that the Navier-Stokes claim meets it means a public, credible, ideally Lean-verified proof is enough.” Some suspect results are already waiting: “I think there's a decent chance one or more companies is sitting on a solution already.”
- Three of five: Most respondents agree part b is a much higher bar, but differ over the degree to which success in solving one of the remaining five problems will presage success in solving the others. Low-progress respondents tend to stress the uniqueness of the individual problems: “The mathematical machinery needed to approach the Hodge Conjecture may have relatively little to do with resolving P versus NP.” Whereas fast-progress respondents typically expect jumps in general capability to result in a domino effect: “If it solves another, it will soon solve them all.”
- Bragging rights versus the backlash: Fast-progress respondents often stress that labs have clear reasons to keep trying: “There's a literal prize for solving them, but I think the prestige, publicity, and ability to demonstrate their models' capabilities will be an even greater incentive.”; “Perfect as a showcase of AI prowess before an IPO.” Others emphasize that cost and mathematicians' hostile reaction could slow things down. “Given that Navier-Stokes already fell, the news value or propaganda value of another one is unclear. Might not be worth all the compute?” writes one. Another adds, “One factor against is the very negative reception from the mathematical community to Navier-Stokes.” Others speculate mathematicians might share less, “given the rumors of the AI Labs allegedly stealing ideas and methods from their users.”
- “Substantially assist”: Many fast-progress respondents argue that any future solution will involve AI, so the question comes down to whether the problems get solved at all: “I have a hard time imagining at this point that any solution will not substantially involve AI.” A few slow progress respondents question this assumption, with one writing, “Experts must judge that an original, important part of the solution likely could not have been produced without AI. Widespread use of AI in mathematical research would [still] leave this question open.”
AI Spend Per Employee
Question. What will December 2026 AI spend per employee be for firms in the following categories of the Ramp AI Index?
Results. The latest Ramp figures shown to respondents, for August 2026, put monthly AI spend per employee at $675.60 for the median firm among the top 10% of AI spenders and $12.50 for the median firm overall. Every group expects spending to rise by December 2026 for both kinds of firm, with groups agreeing more closely on the median firm than on the top spenders. For the median firm, all groups forecast $17–18, rises of roughly 35–45%, with no significant differences between them. For the top 10%, AI risk experts give a median of $925, experts $918 and superforecasters $875, rises of roughly 30–37%. These three groups are statistically indistinguishable, while the public is significantly lower than all three at $800. For both kinds of firm, forecasters see less than a 10% chance that spending will stay at or below its August level: every group's median 10th-percentile forecast is above the August figure.
Rationale analysis
- Extrapolating the trend: Most respondents start with extending Ramp's monthly series, but high-forecast respondents often extend the growth since late 2025—“If I compute the monthly growth rate from September 2025 to September 2026, I get about 12 percent for the top 10 percent and 10 percent for the median firm”—whereas low-forecast respondents tend to instead project the slower pace of recent months: “AI spend per employee grew explosively from November 2025 to May 2026, but since May both the top 10% and the median have grown only around 3% a month.”
- Falling token prices: Many low-forecast respondents note that token prices have fallen considerably. Some argue that this will offset heavier use: “Falling compute costs work against rising spend: token volumes are increasing, but prices are dropping just as fast, which is why token spend has been flat since August.” High-forecast respondents tend to expect cheaper AI will bring more use, with one noting, “So far, cheaper tokens have increased total spending rather than reducing it.”
- The top 10%: Many high-forecast respondents describe the top decile as AI-intensive firms that haven't come close to maxing out their productive use of AI: “Top 10% have already validated that AI spend is valuable to them; current numbers are still relatively low compared to salary expenses, so they have significant room to grow.” Some low-forecast respondents predict that the top decile is better positioned, relative to the median, to take advantage of falling token prices. As one notes, “The top 10% relies more on pay-per-use token spending, which is more exposed to price cuts, as the recent dip among the top 1% suggests.”
- The median firm: Nearly all respondents predict median-firm spending will keep lagging the top 10% by a large margin, but they differ on whether that margin will increase or decrease. Several respondents who foresee an increase point to the per-seat spending factor: “The median firm buys seats. Seats grow with headcount coverage at a few percent a month and do not swing with token prices, which puts December in the low 13s.” Others stress that median firms have ample room to expand the use of AI and narrow the gap with the top 10%: “At the median level, vibe coding has yet to hit the average humanities-based employee.” One respondent distinguishes between a median company in general and a median company in the Ramp system: “When we think of the median Ramp user, we have to think about the median company with enough employees and growth to make using Ramp worthwhile. The local HVAC company and flower shop, on average, are going to be very slow adopters of AI, if at all, and they aren't reflected here.”
- Cost versus benefit: High-forecast respondents tend to argue that leading firms will need to keep spending to stay competitive: “Start ups and existing orgs are incentivised to use AI as much as possible, so even if cost is falling, tokens will rise faster than that.” Several low-forecast respondents instead predict that as AI hype recedes, spending will be reined in: “The time period in question includes a frantic token-maxing period in which many organizations were very briefly completely cost insensitive when it comes to AI spend. That got reeled in very quickly.”
- Agents and new models: Respondents from both forecasting poles expect agent use and new models to lift spending, but many question whether another surge like the one in late 2025 will arrive by December. Some high-forecast respondents predict it will: “I expect October to show a marked increase in month-to-month percent change…people are going to start noticing that Astra 6, Opus 5.5, and subsequent models are extraordinarily capable.” Low-forecast respondents tend to express more skepticism: “They've already made a big move upwards early this year (likely from adoption of Claude Code and Codex) and I don't expect this to happen again by the end of the year.”
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 (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 groups are near-certain that at least one of OpenAI and Anthropic will still be an independent company at the end of 2026, and remain largely confident for 2030. For 2026, superforecasters give a median of 100%, experts and AI risk experts 99%, and the public 95%. By 2030, the medians fall to 90% for experts, 89% for superforecasters, 85% for AI risk experts and 75% for the public. Experts, superforecasters and AI risk experts are statistically indistinguishable for 2030, while the public is significantly lower than every other group at both horizons. The public's forecasts are also the most spread out: for 2026, a quarter forecast 55% or below.
Rationale analysis 5.a
Drivers of forecasts for question 5.a largely mirror the drivers cited when the question was originally posed in Wave 11. They include:
- Acquisition difficulty: 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: “A $2T company is close to un-acquirable (who's buying it and who's clearing that through antitrust?), and Chapter 11 doesn't even count as failure.” Pessimists tend to argue that this may not be an insurmountable barrier: “Loss of independence runs along one path: competitive displacement, then valuation collapse, then rescue acquisition.”
- Nationalization: Both poles point to state intervention as a potential threat to independence, but optimists tend to expect any state intervention instruments imposed will be of the lighter-touch variety that would leave the companies independent—“The tools Washington actually uses (export controls, release vetting, DPA, minority stakes) fall short of the bar, and a listed company with dispersed shareholders is hard to take over”—while pessimists don't discount the possibility that a harder version could prevail: “The scenario I take most seriously is a security-driven takeover, with frontier labs brought under conservatorship or similar state control in a race framing, which would likely hit both together.”
- Access to capital: Several optimists cite the companies' cash position and upcoming IPOs as virtually ensuring near-term independence: “Both companies are very large, well-funded and (for Anthropic) heading to an IPO, so at least one remaining independent is near certain through 2026 and still very likely through 2030.” Pessimists tend to look at the same balance sheets and instead emphasize the massive losses: “Standard startup business hazards apply but are amplified to an incredible extent by the size and scale of these companies, their borrowing, and the circular financial structure of it all.”
- High bar set by resolution criteria: Optimists typically focus on elements of the criteria that allow for restrictions on independence without triggering a resolution: “Only one company needs to remain independent, and the rules allow IPOs, reorganizations, and even a merger between the two.” Pessimists tend to focus on all the ways the question could resolve: “I noted that you are trying to rule out the variations of 'DPA orders' or Golden Shares in the resolution criteria, but that is not particularly clear—there are many versions of these concepts, and for all intents and purposes, these companies would not be an independent company in more stringent versions of these actions.”
- Base rate for large-company disappearances: Many forecasters consider the corporate cease-to-exist 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: “Once IPO'd, both would be the size of inclusion in the S&P500 and from a quick search, S&P 500 companies typically cease to exist as independent companies at [the] rate of 0.5% annually.” Pessimists tend to express skepticism that the base rate is relevant, given the nature of the business and reasons above.
Question (II). Given that at least one of these companies exists as defined in (a), what will be their combined annualized revenue run-rate (in billions of 2026 USD) at the end of the following calendar years?
Results. All groups expect continued rapid growth, broadly agreeing on the annualized revenue run-rate by the end of 2026 but diverging for 2030. For 2026, experts, superforecasters and AI risk experts all give a median of $175 billion and are statistically indistinguishable, while the public is significantly lower at $160 billion. For 2030, superforecasters give $700 billion and AI risk experts $680 billion, with the AI risk experts' median significantly higher than experts' at $523 billion (the superforecasters' difference is not significant). The public is significantly lower than every other group at $275 billion. Uncertainty widens sharply for 2030, with AI risk experts the most uncertain group: their median 10th and 90th percentile forecasts span $250 billion to $1.6 trillion.
Rationale analysis 5.b
- Whether hypergrowth slows: Most respondents expect growth will slow but disagree on how quickly. High-revenue respondents predict it will remain strong through 2030: “Combined AI company revenues have been growing at something like 700% over the last year. Even if the doubling time of AI revenues slows dramatically, revenues north of 2T by EOY 2030 seem very plausible.”; “Current adoption may represent only an early stage of a much broader transformation.”; “I expect substantial growth driven by broader AI adoption and expansion into new business areas, such as Anthropic exploring vaccine development.” Low-revenue respondents more often expect growth to fall toward rates seen at other large tech firms: “The recent growth is crazy…but this cannot be sustained. Comparison classes such as Google, Facebook, Amazon suggest 20-25% annual growth.”
- Competition: Low-revenue respondents often cite rapidly improving open-weight and Chinese models, along with proprietary-model competitors, as threats: “Not every employee needs a frontier reasoning model.”; “Other options certainly include Gemini and Grok. More importantly [the competition] includes open source models with their cheaper implementations. For most applications, the cheaper model will suffice.” High-revenue respondents tend to expect buyers to keep paying for the best: “For a lot of things you really do want to pay for the best possible performer (that's precisely why salaries are very high in talent-constrained areas).”; “On revenue, I believe in substitution of human labor rather than commoditization: companies won't pay for AI like software, but like labor, in a market measured in trillions of dollars of wages.”
- The disputed baseline: Many respondents note that the $135 billion baseline shifted mid-survey, when OpenAI's September run rate was reported at about $50 billion on its own accounting basis. Low-revenue respondents often start from roughly $115 billion and stress that the figure that resolves the question may be weeks old: “Resolution uses the latest figure reported by Dec 31, not the actual December number.” High-revenue respondents tend to build from year-end targets of about $70 billion for OpenAI and $100–120 billion for Anthropic, and emphasize a record of underestimates: “I deliberately put the median above trend-skeptic anchors given the history of under-forecasting here.”
- Transformative AI: A few high-revenue respondents tie their highest scenarios to recursive self-improvement and an “intelligence explosion,” after which the labs could “dominate a huge fraction of industries and economic activity.” Low-revenue respondents more often raise the political response such growth could provoke: “I fear continued exponential growth will lead to a backlash against AI due to automation of work.”; “It seems unlikely that an AI company with substantially >$1.5tn revenue would be stable for long in a democratic country.”
Capabilities (ECI) of Closed and Open Models
Question. What scores will the top-performing closed-weight system and the top-performing open-weight system have on the Epoch Capabilities Index (ECI) at the end of the following calendar years?
Results. At survey launch, the top closed-weight system scored 167 on the ECI and the top open-weight system 157. All groups expect both scores to keep rising, with open-weight systems continuing to trail. Experts' and the public's medians put the gap at 10 points or fewer at every horizon, while superforecasters and AI risk experts expect it to widen slightly. For 2026, medians range from 172 to 175 for closed-weight systems and from 162 to 167 for open-weight systems. Medians are similarly close for 2027, within 3 points of each other: 190 to 192 for closed-weight systems and 178 to 181 for open-weight systems. The groups diverge more for 2029. The public gives the lowest medians (212 and 204), significantly below every other group for both, while AI risk experts give the highest closed-weight median (227) and experts the highest open-weight median (210). Uncertainty also increases by 2029, with AI risk experts the most uncertain group: their median 10th and 90th percentile forecasts for closed-weight systems span 196 to 280.
Rationale analysis
- Historical ECI trajectory: Most respondents start by considering that the scores of closed frontier models have been rising about 14–16 points a year, with open-weight models trailing by about 8–10 points. High-forecast respondents tend to expect AI-driven research to accelerate this progress: “Beyond 2026, I don't expect capabilities to keep advancing linearly. My medians assume continued acceleration, with AI partly automating AI research and development but without full recursive self-improvement.”; “I expect the rate of capabilities gains to increase as we hit inflection points from automated researchers sometime between mid 2027 and mid 2028.” Others point to AI-driven research when setting their 90% quantile: “90% accounts for potential for recursive self improvement to pull trendline up, potentially substantially.” Low-forecast respondents tend to either anchor directly to the historical trend—“The trend looks almost ridiculously linear…No reason to deviate much from it.”—or cite some combination of the factors below when arguing that progress is likely to decelerate.
- Open-closed gap: Respondents who forecast that the gap will narrow often “expect the gap to close by 2029, whether by distillation, model weight theft, or more legitimate advances.” Others also “expect open models to gradually catch closed models over the next four years as China and others invest heavily in them.” Those who forecast that a wider gap will persist tend to stress the closed-model advantage in compute and other resources: “Proprietary labs invest massively in huge compute clusters, secret dataset pipelines, and early architectural breakthroughs”; “Chinese labs are compute-constrained by export controls.” Still others question whether open models, once they attain certain capabilities, will remain open: “Open models may very well be banned beyond a certain capability.”; “Since April 2026 [Alibaba's] flagship Qwen Max models have been available only through an API, while it keeps releasing smaller open models. Meta was reported to be planning the same move for its next frontier model.”
- Release delays: Some respondents argue that the ECI leaderboard will trail the true frontier, citing safety reviews and national-security concerns: “We're forecasting the highest score on Epoch's leaderboard rather than the best model someone has quietly running inside a laboratory.” Others treat “stronger models already known to exist but not on the public leaderboard” as evidence that the ECI has room to climb higher: “The upside [for 2026] comes mostly from OpenAI's internal model, which it says is stronger than Astra.”
- Compute resources: Several low-forecast respondents expect physical and data limits to slow scaling: “We're probably severely constrained by compute until 2029 for the most fantastic scenarios.” High-forecast respondents tend to view massive AI-buildout spending through the prism of national security and market-competition imperatives: “Closed-weight capabilities will continue to increase in terms of ECI given US geopolitical…and financial incentives on the part of the leading frontier labs, esp as they move towards IPO.”
- ECI as a measuring stick: High- and low-forecast respondents alike note that 2029 scores depend on Epoch adding harder benchmarks, and that how this is executed will impact the direction of scores: “ECI only stays meaningful if Epoch keeps adding harder benchmarks, so the wide [quantile] range…is as much about the measuring stick as about the models.”; “Numbers will depend heavily on the benchmark curation, not on model capabilities. Assuming the historical trend of the benchmark curation following a linear trend, I presume a linear trend too (it's basically predicting human behaviour rather than model capability projection).”
Notes
-
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
-
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
Cite this work
Please use one of the following citation formats.
APA
Murphy, C., Rosenberg, J., Canedy, J., Jacobs, Z., Flechner, N., Britt, R., Pan, A., Rogers-Smith, C., Mayland, D., Buffington, C., Kučinskas, S., Coston, A., Kerner, H., Pierson, E., Rabbany, R., Salganik, M., Seamans, R., Su, Y., Tramèr, F., Hashimoto, T., Narayanan, A., Tetlock, P. E., & Karger, E. (2025). The Longitudinal Expert AI Panel: Understanding Expert Views on AI Capabilities, Adoption, and Impact (Working paper No. 5). Forecasting Research Institute. Retrieved 2026-11-19, from https://leap.forecastingresearch.org/reports/wave13
BibTeX
@techreport{leap2025,
author = {Murphy, Connacher and Rosenberg, Josh and Canedy, Jordan and Jacobs, Zach and Flechner, Nadja and Britt, Rhiannon and Pan, Alexa and Rogers-Smith, Charlie and Mayland, Dan and Buffington, Cathy and Kučinskas, Simas and Coston, Amanda and Kerner, Hannah and Pierson, Emma and Rabbany, Reihaneh and Salganik, Matthew and Seamans, Robert and Su, Yu and Tramèr, Florian and Hashimoto, Tatsunori and Narayanan, Arvind and Tetlock, Philip E. and Karger, Ezra},
title = {The Longitudinal Expert AI Panel: Understanding Expert Views on AI Capabilities, Adoption, and Impact},
institution = {Forecasting Research Institute},
type = {Working paper},
number = {5},
url = {https://leap.forecastingresearch.org/reports/wave13},
urldate = {2026-11-19},
year = {2025}
}