By Block Analytics Ltd X Merkle 3s Capital
A model deduction suggests that by 2030, the economy may have a third more output than the path without AI, while the total income of all workers is only half a percentage point higher. Another outcome lies in the same deduction: AI has changed almost nothing, unemployment and today there is no difference. Both outcomes share the same set of data that has already occurred, and the disagreement won't materialize until after 2027, before you have to decide what to do. This article does not talk about its equations, but only about the results it calculates - the four shocks that fall on employment, education, investment, and society.
Report not intended to comfort anyone
This article is based on The Anthropic Institute, Economic Scenarios for Transformative AI (September 2026). This is a working paper number 2026-02 by five authors, including Charles I. Jones - representative of the semi-endogenous growth theory. The dissertation itself states that the opinion does not represent Anthropic's position and notes the use of Claude as a research and writing assistant.
What it does can be squeezed into one sentence: disassemble the impact of AI on cognitive work into the granularity of “tasks,” making it take time for the person being replaced to find a new position, and then extrapolate all the way to the beginning of 2030. There are only two mechanisms to remember - AI will not replace a profession as a whole, it will take away the tasks in this profession piece by piece; the people who are taken away will not disappear out of nowhere, they will take someone else's job. These two bars determine the shape of all the numbers that follow.
The difference between the three scenarios is just how big the AI ended up eating.
Moderate scenario: The task of 4% of the whole economy is done by AI, and the GDP in 2030 is 1.6% higher than the path without AI, in which AI is only an ordinary technology.
Significant scenario: 12% of tasks are taken over, which is about one-fifth of cognitive tasks, and GDP is 8.3% higher, which the report describes as "a hub technology more important than the Internet".
Extreme scenarios: 30% of tasks, about half of cognitive tasks, and 32.4% higher GDP.
The limitations must be said first: these three scenarios are exemplary, without probabilities, and the report repeats them in the summary, introduction, and conclusion. They are not predictions. The only conclusion it gives is that today's data cannot rule out any of them. Thinking of extreme scenarios as "Anthropic predictions" is a complete misreading.
Its value lies not in predicting 2030, but in turning a vague argument into an arithmetic question: if you believe in X, then by standard economics you must accept Y!
There is another detail that is easily overlooked: it is not what the model can do that determines the outcome, but the extent to which it is actually used. The gap in AI capabilities in the three scenarios is less than three times, but after adding "how many people are actually using it and how deeply they are using it", the share of tasks actually performed by AI increases from 4% to 30%, which becomes 7.5 times. There is a dispersion curve beyond the technology curve, which is the amplifier.
The report is accompanied by a survey for comparison: 10,980 US adults, and the median implied scenario falls around the salient scenario - GDP +8.6%, perceived employment -4.2%, and unemployment rate 4.6%. But reading this as "society has reached a consensus" is a statistical misreading. Regarding productivity gains, about 30% of respondents believe that AI will save no time at all, and 49% believe that it can save at least half; regarding Nobel-level discoveries, 40% believe that it will never be achieved, and 39% believe that it can be achieved before 2030. The median falls in the salient scenario, not because most people believe in the salient scenario, but because nearly as many people believe in the moderate as in the extreme.
Also make it clear what it doesn't look at. Robots and physical automation are completely ruled out, as are financial market turmoil, business cycles, and political economy factors. The report itself admits: "In large part for this reason, we do not extend our analysis beyond 2030." This self-statement will appear again later, and is the report's most honest and most disturbing sentence.
A report that does not give probabilities, does not give advice, and does not intend to comfort anyone is more difficult to refute than any prediction!
Employment shocks: who's jobs will disappear and in what form
Let’s first make it clear who this impact is aimed at. The report cuts U.S. employment in half: the cognitive group includes 12 major categories: management, business finance, computer mathematics, construction engineering, science, community social services, law, education, art media, medical practice, sales, and office administration, accounting for 62.4% of employment in 2025; the rest include medical assistance, security, catering, cleaning, personal care, agriculture, forestry, fishery, construction, installation and maintenance, production, and transportation, accounting for 37.6%.
The model only allows AI to impact the cognitive group, and the efficiency of the other 37.6% of people is not affected by AI at all. All changes in their wages come from an indirect effect: the economy has become richer, and their type of labor has become relatively scarce. Therefore, all the differentiations seen later are first of all the direct consequences of this setting, rather than empirical findings. If you don't accept the premise that "AI only affects cognitive tasks", the entire deduction is invalid - this must be before reading any numbers.
This dichotomy has three blind spots, and the report itself admits that "any dichotomy is a simplification." The two major categories of medical practice and education contain a large amount of physical labor and interpersonal labor that cannot be completed remotely, and the degree of exposure is overestimated; the salary and academic qualification thresholds of sales and office administration are much lower than those of management and professional positions, but they are regarded as the same kind of people in the model, and intra-group differentiation is completely covered up; there are no differences in seniority, skills, and regions within the same group. These numbers describe two statistical aggregates, not two specific categories of people.
Next comes the results. In the extreme scenario, the unemployment rate is 17.9% for cognitive jobs and 11.9% for the economy as a whole—these two figures must be distinguished, since mixing them would overestimate the unemployment rate by about 6 percentage points. The report uses the adjective "stunning" for that figure in cognitive positions. For reference, the annual peak in the United States after the financial crisis was just under 10%, and during the COVID-19 period, it was about 8%. In the extreme scenario, even the overall unemployment rate exceeded any post-war recession.
The difference between moderate and extreme is not a gap in degree, but a gap in magnitude. There is a six-fold change in cognitive unemployment from 2.9% to 17.9% and a sign reversal in cognitive wages from +0.4% to −11.5%, with no indicator in between that partially offsets the difference at either end. Considering that wages should have increased by about 2% per year without AI, the absolute level of cognitive wages in the extreme scenario will be lower than mid-2026 - four years later, it is not a slow increase, but a regression.
Employment losses are more concentrated than wages. The number of people in cognitive positions under the extreme scenario is 21.5% lower than in mid-2026, and the total wages received by these positions are 31.0% lower. Wages in other groups rose by 33.6%, the largest single increase in the entire report. The original words of the report are that the increase in the total wages of other groups is "almost the same" as the decrease in the cognitive group - the total amount is conserved, and a group of people has changed.
The shape and intuition of the impact are also different. AI is advanced task by task, so the profession will not disappear as a whole one day, it will be hollowed out first. What really triggers replacement is neither the most difficult nor the simplest tasks, but those links where AI can reduce the cost the most and take up the most payroll. The remaining part is left to people, the signboard is still up, the positions are still there, but there are fewer people to recruit.
There is another colder detail: whether AI replaces you or helps you, in this model, it is decided by the technology itself, and companies can only passively accept it. In the three scenarios, the proportion of "looks like a substitute" is half, three-quarters and 90% respectively - the corresponding space for "helping you" is reduced from 50% to 10%. The calibration is based on Anthropic's own observed usage: chatting uses Rio half of the image automatically, and calling through the interface uses Rio three-quarters of the image. Enhancement is a space that will continue to be squeezed, not a stable haven.
Your career will not suddenly disappear one day. It will be half-empty first, and then you find yourself competing for a position with three people who have also been hollowed out!
There is a sequence in which pain comes. When a company encounters "people are more expensive than AI", it will not cut salaries first, but directly lay off employees - wages are adjusted downward very slowly, and only half of the gap is covered in a year, so layoffs precede salary cuts. The first group of people who suffered were those who were laid off; the second group were those who stayed in their jobs, and their wages were pushed down to −11.5% year by year, and because they were supposed to increase by 2% every year, the physical experience was worse than the numbers; the third group were those who tried to cross borders. On the one hand, the laid off people flocked to other industries at once, and on the other hand, those industries could only absorb half of the gap at most every month. This mismatch itself is the source of the unemployment rate. Notably, by the end of the simulation in 2030, the perceived unemployment rate is still stuck at 17.9%—the adjustment is far from complete.
Finally, it must be said: these numbers should be considered lower bounds, not upper bounds. The model assumes that people who change industries immediately receive full wages in the new industry, with no seniority discounts, no skill depreciation, no income interruption during training periods, no moving costs, and no licenses and apprenticeships. The report cites evidence in its calibration section that in reality the wage loss of reemployment is usually 10% to 20%, and is concentrated among the group of people who change occupations. It named this batch of documents and then stated that they were not included. Even if one were optimistic to the point that "it only takes time to go from software engineer to electrician", the extreme scenario still gives a perceived unemployment rate of 17.9%.
Education shock: Why academic qualifications are becoming a depreciating asset
There is not a word about education in the report. However, under extreme scenarios, the two paths of wages for cognitive positions -11.5% and other positions +33.6%, taken together, are enough to pose structural doubts on the chain of "read good books, obtain academic qualifications, and enter white-collar jobs" - those who are damaged are precisely the 62.4% with the longest education and the highest wages. Over the past two centuries, the losers of technological progress have often been low-skilled manual workers, and education has been treated as a safe-haven asset. This report depicts an impact in exactly the opposite direction.
The first level is the evaporation of human capital rents. The reason why academic qualifications command a premium is because the tasks they correspond to are scarce, expensive, and cannot be performed by others. When the unit cost of these tasks drops to less than half of its original value in four years, the corresponding rent that has been spent for more than ten years to master them will evaporate simultaneously. Moreover, the cost decline curve in the report is still deepening by 2030, with no sign of peaking.
The second layer is signaling value. A degree is valuable partly because it is expensive—the cost of acquisition is lower for people with strong abilities, so it can be used as a sieve. After AI levels the output threshold of writing, analysis, and coding, the distinction between academic qualifications will inevitably decrease, and employers will turn to evidence that is more difficult to imitate: work samples, verifiable delivery records, and face-to-face judgment. This layer is our derivation, and the report does not discuss the signaling mechanism.
The third layer is the most dazzling. Historical experience is estimated that for every two tasks that are automated, technology will create a new one. This is the source of the comforting phrase "technology always creates new jobs" that has been quoted for two hundred years. In the three scenarios, this ratio is one new item for every two items and one new item for every four items, while in the extreme scenario it is zero - AI does not create any new tasks for humans. That comfort is clearly turned off in this scenario.
It’s not that reading is useless, it’s that the thing that sets the price for reading is getting cheaper!
So the retraining window becomes a very unfriendly arithmetic problem. The report said the divergence will occur after 2027, and "data over the next few years will tell us which scenario we are in." In reality, most licensed occupations—nursing, electrician, plumber, imaging technician—take 2 to 5 years, and advanced licenses such as registered nurses, professional engineers take 5 to 10 years. If it starts at the end of 2026, the earliest time to obtain the certificate is from 2029 to 2031, which will fall exactly in the range where the differentiation is a foregone conclusion and the conversion has not yet been completed. If we wait until the data in 2028 is "confirmed", the license will be postponed to 2031 to 2038, by which time the premiums of other groups will have already been diluted by the first movers.
This constitutes an asymmetric bet: if it turns out to be a moderate scenario, the loss of switching early is just learning another trade; if it turns out to be an extreme scenario, the cost of not switching is to find a job at a cognitive unemployment rate of 17.9%, and the efficiency of cross-line search at that time is only about a quarter of normal times. I would also like to add something that is not in the report: the model assumes that changing careers does not require any retraining, so the interruption in income during the training period is not included in those figures at all.
What's even more troublesome is that the more "successful" a cognitive career is, the harder it is to transfer out. The report's raw data showed that the probability of cross-group employment for unemployed sales and office workers was 28 percent, compared with only 10 percent for managers and professionals. The deeper the accumulation of professional-specific human capital, the higher the threshold for cross-border. This means that in this impact, the group of people who suffer the most damage are precisely the group of people who are most difficult to save themselves.
Whether English and communication skills are worth continuing to invest in, the answer should be looked at separately. The price-performance ratio of pure language conversion capabilities is declining significantly - translation, polishing, and cross-language writing are the types of tasks where AI costs are falling the fastest, and they are at the core of the impact. Investing in English as a "tool for information acquisition and text production" is investing in a competency curve that is rapidly decreasing in price.
However, the cost-effectiveness of communication skills as a carrier of trust and negotiation is increasing. The reason is very straightforward in the model: after the automated links become cheaper, the unautomated links become bottlenecks, and more money in the entire process will flow there. For those professions that rose 33.6% in the extreme scenario, the value lies precisely in the fact that people have to be present - the value of negotiations, relationships, team leadership is not in generating text, but in having someone sign off on the results.
Therefore, our judgment is: if the purpose is just to read English materials and write English emails, it is not worth investing a large amount of time; if the purpose is to negotiate in real time, host meetings, and build trust in a cross-border scenario, it is worth investing, but what is really scarce is the communication ability itself, and English is just the carrier. The most easily misled thing is to regard "fluent speaking" as the goal itself.
The academic qualification will not be invalidated, it is just returned from a ticket to a line of decoration on the resume!
Investment shock: Capital’s feast, and its bills
This is the strongest part of the report. Under the three scenarios, the total income of capital is +3.1%, +18.9%, and +81.4% respectively without AI; the total income of labor in the same period is +0.6%, +1.4%, and +0.5%. This is not a battle for shares, but two completely different curves.
How did the money go? The mechanism is so simple that it is a bit cruel: for every task taken over by AI, all the corresponding salary expenditures are directly transferred to capital, regardless of how much cost is saved. The labor share went from 60.0% to 59.4%, 56.1%, and 45.2%, and the capital share went from 40.0% to 40.6%, 43.9%, and 54.8%. The original words of the report in the extreme scenario are: "Full 15% of GDP shifts from being paid to labor to being captured by capital as a return on investment." But what is more worth remembering is actually the significant scenario-the labor share falls by 4 percentage points in four years, which is only slightly smaller than the entire decline in the United States for the full four decades after 1980.
Returns are also on the rise. Return on net capital rose from the baseline of 6.5% to 6.6%, 7.0%, and 8.3%, with an increase of approximately 28% in the extreme scenario. This number is extremely sensitive to one assumption: how quickly capital can be piled up. The report assumes that computing power can be financed globally and built within one or two years, so the supply elasticity is set on the high side; if the actual expansion is much slower than this, the net return will reach 10.3% in the extreme scenario, and the average salary will become −9.2%; if the computing power is as much as needed, the net return will remain at 6.5%, and the average salary will increase by 30.1%. The level of capital return directly depends on how hard the physical bottleneck of computing power expansion is.
Productivity did increase significantly, but in unexpected places. Total factor productivity was 0.7%, 3.1%, and 13.4% higher than the path without AI, and the annual growth rate rose from 1.0% to 1.2%, 2.3%, and 7.3%; the annual GDP growth rate rose from 2.0% to 2.4%, 5.4%, and 15.4%—the fastest year during the Internet boom in the 1990s was 1999 4.7%, the growth rate of the significant scenario has exceeded it. Capital stock is 2.3%, 13.8%, 56.3% higher. However, the "AI accelerates scientific research" line has almost no contribution: the stock of ideas is only 0.07%, 0.20%, and 0.61% higher than the path without AI, and the annual growth rate only moves from 1.67% to 1.69%, 1.76%, and 2.02%.
Almost all the growth in this report comes from making existing things cheaper, rather than discovering new things - "AI accelerates scientific research" as an investment argument before 2030, which is not supported in this model!
If we want to line up in order of benefit, the sorting should be based on physical bottlenecks rather than narrative popularity: the first is computing chips and accelerators, then power, transmission and distribution, and grid connection, advanced packaging and optical interconnection, then data centers and construction supporting facilities, then cloud and inference service layers, and finally downstream applications. The value realization of the downstream must be later than that of the upstream, because it relies on the slow variable "how many people actually use it", and that curve has just left the ground in mid-2026.
The report also proactively explained areas where it may have underestimated. It admitted that the model was missing a key mechanism linking capital returns to growth rates, and that adding it would "could significantly improve capital returns" - 8.3% would likely still be an underestimation. There are two other things listed as lower bounds: the innovation effect is low because the model cannot produce self-accelerating feedback, and unemployment and income losses are low because there is no wage discount for changing careers. A report that simultaneously admits to being conservative in three directions is a message in itself.
Then comes the bill. The report itself calculated: In extreme scenarios, to maintain the income of cognitive workers at the level without AI, a transfer payment of about 9% of GDP is needed, "roughly equivalent to the sum of Social Security and Federal Medicare"; after paying this money, the rest of the economy is still more than 20% ahead of the non-AI path. The total gain was about three times the cognitive worker's loss, and the report's original words were that "the gain more than compensated for the loss in a textbook sense."
The problem is that if this 9% is really raised through capital gains tax, corporate tax or some kind of computing power tax, the benefit chain above will be directly eroded. The report went on to say that technological transfer on this scale is unprecedented, and based on the experience of past trade shocks to the United States, it will basically not happen spontaneously. What’s worse is the worst parameter combination in the footnote of the report: capability and diffusion only reach the level of a significant scenario, but the substitution ratio is extremely high, cross-industry search is extremely difficult, AI does not create new tasks, and capital expansion is limited. At this time, GDP is only 7.2% higher, and total labor income is 4.3% lower than the non-AI path. Compensation requires 84% of the GDP gain. The logic of "make the cake bigger first and then divide it" basically fails in this corner.
The mild scenario also deserves its own mention: net returns there only go from 6.5% to 6.6%, and excess returns on the AI theme are essentially non-existent. And the report makes it clear that this scenario cannot be ruled out today. A position that is only established under extreme circumstances is actually placing a heavy bet on an example that does not even allow for probability.
The biggest risk on the investment side has never been whether AI will succeed, but who will get the benefits if it succeeds!
Social Impact: The cake gets bigger, but workers don’t get more
The most counter-intuitive sentence in the entire report is: the total income received by workers in the extreme scenario is less than that in the significant scenario. +0.5% versus +1.4%, while GDP growth jumped from 8.3% to 32.4% over the same period.
There is only one arithmetic explanation given in the report, and this is the only one retained in this article: the labor share fell from 60.0% to 45.2%, a drop of about one-quarter; GDP rose by one-third. 0.45 times 1.32 equals approximately 0.60, and the two forces exactly cancel out. The laborer cuts an almost as thin slice of a much larger cake.
There is no monotonic relationship between GDP and labor income - AI upgrades from "significant" to "extreme", the cake is one-third bigger, and workers don't get a penny more!
Average salary is the most dangerous metric here. It’s +0.7%, +2.1%, +9.7% in three scenarios, looking all the way up. But in the extreme scenario, the +9.7% includes −11.5% of cognitive jobs and +33.6% of other jobs, and it does not include the 17.9% of cognitive workers who have lost their jobs at all. Without them in the denominator, the average certainly looks good.
There is a more subtle trap: wage stability does not equal human security. The report has conducted a series of tests and flipped the switch on "whether wages can be adjusted downwards" from the most flexible to the most rigid. At the most flexible end, cognitive wages plummeted by 42.2%, but cognitive unemployment was only 2.6% and the overall rate was only 3.1%, almost the same as in normal times, but GDP was the highest at +36.6%. At the stiffest end, the cognitive wage is actually +2.8%, which is higher than the path without AI. The cost is that the cognitive unemployment rate soars to 24.0% and the overall rate is 15.2%. The same impact can be expressed as "everyone's salary will be cut by 40%", or it can be expressed as "one-quarter of the people are unemployed and the remaining people will have their salaries increased".
Protecting wages does not mean protecting people. The protected wages will only turn the impact into unemployment and pass it on to those who are no longer on the list!
This also explains why "the current macro data is still very mild" cannot be used as evidence. The report acknowledges that "the labor market effects of AI in macro data so far are relatively modest," but goes on to point out that this sentence can be read as either "we are in a moderate scenario" or "we are in the beginning of an extreme scenario." The three scenarios share the same set of readings for mid-2026, with previous GDP differences within a few tenths of a percentage point. Data for 2026 cannot distinguish between the three scenarios, while data for 2028 can. In the reality where wages adjust downward very slowly, early signals will appear first in hiring volume, not wages - by the time the wage scale becomes ugly, the adjustment has already been underway for a long time.
The pattern of wealth differentiation will also change. It is no longer mainly "high salary vs. low salary", but "assets vs. only salary". Cognitive workers are exactly the typical group with "high wages and low asset ratio" - the income side is exposed on the damaged side, and the asset side is not on the benefiting side. The risks are not spread, but superimposed. The situation within the labor force is equally dramatic: total wages in cognitive positions have dropped by 31.0%, while other positions have increased by almost the same amount. The winners and losers are not the same people.
A common retort is "Everyone has no money, who can sell things to?" In this model this problem does not hold true: output is fully converted into someone's income, but 15% of GDP is transferred from wages to capital gains, and the consumption and investment of capital owners fill the gap. However, the report admits in its own note that the model does not consider price rigidity and related demand effects, so it "cannot produce the kind of negative feedback in which an impact reduces demand and demand amplifies the impact." The risk of consumption interruption has not been falsified, but assumed. It cannot be used to negate the conclusion of the report, nor can it be considered that the report proves its non-existence.
The report explicitly excludes political economy, but it indirectly leaves behind three variables that determine fate: whether a country can build enough computing power and electricity, whether wage shocks are realized in the form of salary cuts or unemployment, and whether it has the ability to tax and redistribute. From this, we can deduce the differentiation of outcomes - places with votes, fiscal space and local computing power and energy industries are most likely to achieve high growth and high transfer; places with votes but no computing power industry will suffer from the outflow of capital gains and the loss of domestic labor; areas without both will suffer from a decline in the share of labor and a lack of capital gains. This paragraph is our derivation. The report itself only says that "what system can bring shared prosperity is beyond the scope of this framework."
Place your bets before the answer is revealed
The logical starting point of the whole thing is only one sentence: the difference will not be resolved until the data from 2027 to 2028, and it takes 2 to 5 years for most people to complete a career change-you must make a decision before you know the answer. The following are our deductions based on the results of the report. The report itself does not provide any personal advice.
For white-collar workers and professionals, these four are the most sharp.
First, break down your job into a task list and mark each item one by one to what extent AI can currently do it. In this model, the average exposure of the cognitive group is more than twenty times that of other groups. What you want to know is not "Is this occupation dangerous?" but "How much of my job falls on the side that has been taken away."
Second, take the initiative to move to the part where "AI is more like helping you, rather than doing it for you": external responsibility, cross-department coordination, on-site judgment, and reducing purely delivery-oriented text and analysis output.
Third, establish a verifiable record of delivery—published projects, verifiable performance, externally verifiable results, because the distinction between academic qualifications and titles is declining.
Fourth, do a stress test: assuming that jobs disappear within 18 months, cash reserves can cover several months; even in the significant scenario, the perceived unemployment rate rises from 2.9% to 4.5%, while in the extreme scenario we face 17.9%, and only a quarter of the inter-bank search efficiency is left.
For parents, the same applies.
First, do not hastily change your education path in the next one or two years in order to "avoid AI" - the mild scenario cannot be ruled out today. In that scenario, the cognitive unemployment rate is 2.9%, unchanged, and the expected losses from overreaction are real.
Second, shift the focus of investment from knowledge stock to the ability to have people present, take responsibility, and build trust. In the extreme scenario, the +33.6% increase in other groups is the largest single increase in the entire report.
Third, directions with licenses, apprenticeships, and on-site delivery attributes are preferred, but one must accept a limitation-the report excludes robots. The author personally acknowledged the shelf life of this conclusion by "not extending the analysis beyond 2030." The shift to blue-collar workers is a transitional strategy, not an end-game strategy.
Fourth, tell children frankly about the magnitude of uncertainty; teaching children to make decisions amid uncertainty is more consistent with the conclusions of this report than teaching any specific skill.
For those who have investable assets, there are also four rules.
First, write down which scenario your position corresponds to - under the mild scenario, the net return only goes from 6.5% to 6.6%, and the excess return on the AI theme is basically non-existent. This alone can filter out many narratives.
Second, hang your judgment on observable readings rather than news. The data from 2027 to 2028 will determine the winner. Any "consensus" before that is an illusion.
Third, understand that the order of benefits is determined by physical bottlenecks: computing power, power and grid connection, packaging and optical interconnection come first, data center and construction supporting facilities come second, cloud and inference services come third, and downstream applications come last.
Fourth, treat the political risk of redistribution as an independent risk factor - compensation in the extreme scenario requires 9% of GDP, and in the worst-case parameter combination, it needs to eat up 84% of the GDP gain, and both figures are from the report itself.
These three lists actually point to the same action. In extreme scenarios, labor is damaged and capital skyrockets. Holding capital itself is a hedge against one's own human capital. In turn, human capital is concentrated in the cognitive group, and financial assets have nothing to do with the AI benefit chain. It is the most vulnerable combination in which both the income side and the asset side are exposed to damage at the same time.
Buried in the report is a seldom-cited comparison: Respondents in cognitive occupations, a median, believed that only 27% of their working time would be covered by AI by 2030. However, when the same people evaluated eight general tasks, the median estimated that 71% of knowledge work would be covered. The public's judgment on the macro trend is not bad. The problem is that everyone exempts themselves.
Finally, two things not to do. Don’t use extreme scenarios as baseline assumptions to increase leverage - the report does not assign probabilities to any scenario, and the extreme scenario is the one with the highest political risk and the deepest fall in wages. The parts that are good for capital and the parts that are bad for capital are tied together. Don't stand still just because "nothing has happened so far." The report has made it very clear: this fact is compatible with both mild scenarios and the beginning of extreme scenarios.
Before you know the answer, you have to place a bet - and not placing a bet is itself a bet!