Fine Eye

Our model builds in one pattern. Every theme becomes easier along an S-curve, so hard themes cross the ease line late and the Earn the right quadrant empties. The points below are the ones the model does not decide for us.

Impact on society and value to a company are different measures. A trend can reshape a decade and leave its builders with thin margins. The likelihood that a trend happens and the ease of building in it also differ. A trend can be certain and still need a fab, a licence or ten years of patient capital before the first customer pays. We invest before that first customer pays, so we use the two questions we would ask a founder at that point.

Take quantum computing. It may matter a great deal. For a team starting in 2026 it is hard to build in: the hardware is immature, few engineers know the field, and first revenue is years away. Humanoid robotics has the same profile. We read those themes as bets on timing.

A map is useful if it changes what we fund. This one orders themes. It does not allocate our capital, and every rule below is our judgment. Each card shows how often a theme in that quadrant stays there when we perturb the inputs. Where a quadrant is shaky, we ask for evidence before applying its rule, and we treat the Execute and Commodity rules as resting on our assumptions. If you are a founder, here is the 30-minute version of this paper.

1. Find your theme with the search box above the chart, or in the data table. Pick the closest one. A broad theme is your starting point, and your wedge is where the value sits.
2. Read its position in 2026 and 2032, and the year its ease crosses 50.
3. Write your wedge in one sentence: a narrow product at the easy edge of the theme.
4. Name your moat in terms the model uses: data, a licence, distribution, or a regulatory edge. If your honest answer is speed, our model assumes value falls, and that assumption is untested.
5. Answer the six questions further down, then tell us how you would survive to the year your theme crosses the ease line. Send us the answers.

Here is how we read each quadrant.

The companies are illustrative and the figures come from the model. The last one is a pass.

We built this model, so we tested it. Ease rises along an S-curve for every theme, so hard themes always cross the ease line late and the patient quadrant always empties. The claim we most want to test is the one the S-curve does not decide: that moat determines who keeps value as ease rises. We tested three things: random errors in our inputs, errors that all lean one way, and different choices about the model's structure.

Random errors. We ran 400 simulations. In each, every theme's profit pool moved by up to 10 points and its technology readiness and moat each moved by one notch with a one-in-four chance either way, and each of an archetype's six scores moved independently by one notch, with the same moves applied to every theme of that archetype. Earn the right held 50 themes in 2026 (80% of runs between 39 and 60) and Commodity held 62 in 2035 (between 54 and 68). These errors are independent from theme to theme, which flatters stability. The outside rater's inputs lean one way and move the map much more.

Errors that lean. Each row of the first table below shifts one input for every theme at once, or swaps in the outside rater's inputs. Across both tables, Commodity grows and Earn the right shrinks in 21 of 21 shifted scenarios, but that mostly tests the mechanics. The size moves a lot. Earn the right in 2026 runs from 25 to 104 themes across scenarios, and Commodity in 2035 from 12 to 88. Input shifts move the 2035 Commodity count by up to 28 themes and the outside rater's inputs by up to 48. Moat is the clearest case: with every moat set to zero, 67 themes are in Execute in 2035 against 63 on our inputs, so the total barely moves, but 12 themes change quadrant. Moat reshuffles who the winners are more than how many there are, and we scored every moat ourselves. Read every count in this paper as a range.

Our least secure claim. The idea that ease erodes value rests on our moat assumptions and on the shape of the value penalty. AI falls below 50 on value in 10 of 21 shifted scenarios and digital payments in 16. Neither falls below 50 on the outside rater's theme inputs (amber lines). A test that does not use our model, in the public-company section, was inconclusive.

This table keeps the inputs and changes the model's own parameters: how hard value is cut once a theme becomes easy, how fast ease grows, how sharp the S-curve is, and where the quadrant lines sit. These choices are ours and have no data behind them, and the results depend on them. With no value penalty, or a halved one, AI never falls below 50, and with slower ease growth many more themes stay in Earn the right in 2035.

Which calls are stable. 11 of the 16 main themes keep their 2035 quadrant in at least 90% of the random-error runs. We chose which themes count as main. Across all themes the average is 84%, and the quadrant cards above use the 2026 quadrant. Percentages carry about three points of simulation noise, and the 90% cut-off is sharp: re-running with another random seed gives 10 or 11 of the 16, with Hydrogen on the line. A crossing range that starts in 2026 means the theme is already easy in at least one run in ten. The themes that flip most often sit near a line:

A second opinion. One person scored our inputs, so we asked an outside rater: TypeSafe's Jev model. For each of the 124 themes we sent only the name and area, with none of our scores or rubric. We asked it to rate technology readiness, moat and profit pool on the same scales, and two yes/no questions: could a small seed-funded team reach first paying customers within 18 months, and could a company here reach a venture-scale outcome by 2035. It learned from the same public writing we did, so it is a second rater and not ground truth. TypeSafe AI is also an Agilefin technology partner, so a partner's tool ran this check and it is not a neutral audit. We changed no input because of it. The intervals in the tables treat themes as independent, which makes them too narrow, and the 400-draw bootstrap moves tail values by up to about 0.1 between runs. Its average confidence on moat was 0.46, which is low.

We read three things in the comparison. Technology readiness agrees strongly (0.85). Moat (0.54) and profit pool (0.42) agree only moderately, and with the archetype scores they are the inputs we expect founders to challenge. The rater's talent-and-distribution scores run far below ours even though the ordering agrees (see the archetype table), so we may read that question differently. The rater is also more optimistic than we are: it says nearly every theme could reach venture scale, so its yes/no answers do not separate themes at 50%. Its ranking carries signal and its level carries bias. Where it disagrees most we plan to look again. It rates 3D bioprinting, genetic engineering and advanced nuclear as more ready than we do, sees digital payments, sharing economy and metaverse as more durable, and sees a larger pool in green molecules, vehicle-grid integration and digital twins.

Second opinion on the archetype scores. The same rater scored all 13 archetypes on the six rubric dimensions. It saw each archetype's name and a few typical themes. Across the 78 scores the rank correlation with ours is 0.68 (0.52 to 0.80), and the rater is harsher on average, with a mean of 1.74 against our 2.31. On expansion, one third of our value rubric, the rater shows no agreement with us (0.04 (−0.69 to 0.69)). That part of our value score has the least support.

Where we think the model is wrong. Cybersecurity is the clearest case, and it sits on the quadrant line. Demand for security is adversarial because threats grow with the technology, so value probably holds better than our model says. Advanced nuclear is a second case: a policy decision can move ease in one year, and an S-curve cannot jump. A third gap is that we score moat for a theme, while a real moat belongs to a company. The rules and worked examples assume company-level moats the model never scores, so they rest on our assumptions. Finally, the 124 themes contain only 115 distinct input combinations, because many share every input (subscription economy, food-as-a-service and digital fitness, for example). Statistics that treat the themes as independent overstate our evidence, and the stability percentages are among them.

Two of our rubric scores make claims that company accounts can test: margin, and how little capital a company needs. We mapped each archetype to groups of industry (SIC) codes and read 3,588 annual reports (10-K filings) from US public companies for fiscal years 2017 to 2020, from the SEC's public financial-statement dataset. For each archetype we took the median gross margin and the median capital spending as a share of revenue.

Across archetypes with at least 50 filings, our margin score and the observed gross margin have a rank correlation of 0.66 (0.00 to 0.95) (10 archetypes), and our capital score and inverted capital intensity have 0.78 (0.39 to 0.97) (11 archetypes). That is moderate support with wide intervals. One end of the margin interval sits at zero, so we call it inconclusive. Defensibility against research-and-development intensity, a weak proxy for a science moat, gives 0.18 (−0.56 to 0.71) across 13 archetypes, which shows no relationship. We do not claim it validates defensibility.

A test that does not use our model. If easy-to-enter themes had thinner margins, archetypes with low capital intensity would show lower gross margins. Across 9 archetypes the rank correlation between ease (inverted capital intensity) and margin is −0.03 (−0.88 to 0.68): no relationship, with an interval wide enough to include both signs. With nine archetypes it has little power, and it compares archetypes with each other, not how one theme changes over time, so it neither supports nor contradicts the claim that ease erodes value. That is why the claim stays an assumption. Software, the easiest archetype, has among the highest margins, but public software companies are the ones that built moats.

Limits. Public companies are large survivors, not seed-stage start-ups. Capital spending over revenue measures the capital intensity of mature operations, not the capital needed before first revenue. Nuclear is proxied by electric utilities, which explains its high margin. The industry-code mapping is ours. The data stops at 2020.

We forecast patent filings for 31 themes to see whether invention momentum tracks our growth ceilings. It does not (rank correlation 0.02). On a 2022 to 2023 holdout, TimesFM's median error was 10% and 14%, against 9% and 10% for assuming no change, and it called the direction right for only 18 and 18 of 31 series. For 26 of 31 themes the 95% band includes no change. We use none of this anywhere else in the paper.

Some trends on a trend map are not markets. Ageing, urbanisation, deglobalisation, civil unrest and similar trends are forces. Ease of execution does not apply to them, so we leave them off the chart and ask a different question: which investable themes does each force push? We use a force to choose a theme and to check that a founder's story has a tailwind.

The map holds 26 forces, which sit in their own section, and 124 investable themes, three of them emerging, which are on the chart. Each is scored on separate inputs (technology readiness, profit pool and moat for the theme; capital, regulation, talent and distribution, margin, defensibility and expansion from its archetype) and the rest is arithmetic, set out under the chart. The dates in this paper are computed from that model, not written by hand, and shown as ranges where several themes share a story. A patent-filing cross-check was tried and set aside, as described above.

Fine Eye is not investment advice. Agilefin may hold interests in companies in these themes.

We plan to update this paper on the first of each month. Each version note says what changed and why. The next update is due on 1 November 2026.