The atlas presents ten companies in each of its 25 sublayers. Selecting them requires a definition of importance that can be measured, applied on the same terms to a uranium enricher and a model developer, and revised as the evidence changes. The 5LC Importance Index provides that definition. It scores companies within one sublayer at a time, using a single method and sourced inputs that carry their own uncertainty.
The index produces two measures. Importance today estimates how much of a sublayer’s AI-related output would be lost if a company disappeared, and for how long. Rising identifies the companies with the fastest documented growth in AI-related activity, measured in three ways: growth in activity, gains in share and contracted expansion. In each sublayer, eight places follow Importance and up to two follow Rising; a Rising place for which no company qualifies goes to the next company by Importance. A weekly recalculation maintains the selection, and companies left with no placement in any sublayer move to a public archive. Each sublayer page presents its measure card, the full ranked table of scored candidates and a downloadable ledger.
01.The meaning of importance
A company is important to the extent that the AI supply chain would lose output if it disappeared. Size alone does not establish importance in this sense. A large supplier whose output others could replace within weeks matters less to the continuity of the chain than a smaller supplier whose product has no qualified alternative. The index therefore combines two quantities: the share of a sublayer’s AI-related output that passes through a company, and the time the rest of the chain would need to replace it.
This definition draws on the criticality assessments used in European Union and United States reviews of critical supply chains, which judge a supply by the consequences of its loss together with the difficulty of substituting for it. The construction of the index, including its weighting, its treatment of uncertainty and its sensitivity analysis, draws on the Handbook on Constructing Composite Indicators published by the OECD and the European Commission’s Joint Research Centre.
Scores compare companies only within one sublayer. Importance in Memory and Importance in Fuel describe different outputs, and placing them on one scale would require a rate of exchange between them that the evidence does not supply. A company active in several sublayers is scored separately in each, and its placements cannot be added together.
Each sublayer has a measure card, which defines what AI-related output means there and how it is measured. It states the sublayer’s scope and boundaries, the basis on which shares are recorded, the functions into which the output divides with the weight and source of each, the preferred market-share data and the known gaps in the evidence. The card is reviewed and locked before any company in the sublayer is scored, and its fingerprint is recorded at that point: a cryptographic hash of its contents, which changes if any part of the card changes. A later change requires a revision with a stated reason and the editor’s approval. Fixing the specification before any results exist prevents weights or definitions from being adjusted to favour a particular company.
The unit of comparison is the economic supplier: the business that sells the output and controls its production, whatever its legal form. A controlled subsidiary is therefore consolidated with its parent unless the card treats the unit as the supplier, for instance where a business unit reports its own results, and one economic group can hold at most one place in a sublayer. Each sublayer’s candidate universe is established before scoring from earlier research, market-share reports, industry rankings and news of new entrants. It extends well beyond the companies already in the atlas and records every exclusion with its reason.
02.Reach across functions
A sublayer usually comprises several distinct functions. Memory, for instance, divides into high-bandwidth memory, server DRAM, data-centre solid-state drives, and the interface and controller chips that connect them. The measure card assigns each function f a weight wf equal to its share of the sublayer’s AI-related value or physical output, taken from a published source. In Memory the weights follow spending on memory and storage for AI servers; in Fuel they follow each fuel’s share of the electricity that data centres use. Every weight carries a low and a high value as well as a central one, the central weights sum to one, and no weight is set by preference.
Company i’s share of function f, sif, is a percentage taken from market-share data, production, capacity, shipments or revenue in that function, in the order of preference the card specifies. Reach is the weighted sum of these shares:
Reach is therefore the percentage of the sublayer’s AI-related output that passes through company i. Each share is recorded on the card’s single basis, such as revenue over a stated period, operating capacity in megawatts or an annualised revenue run-rate, so that all shares within a function describe the same quantity.
Where a component inside another company’s product has a near-sole supplier, the card gives the component a small function of its own, valued at the supplier’s sales and deducted from the function that contains it. Nothing is counted twice, and the component’s importance arises from its own replacement time.
03.Irreplaceability and replacement time
Reach measures how much output passes through a company, but not how quickly that output could be replaced. For each function in which a company holds a material share, normally at least 1% unless the card states otherwise, two raters assess irreplaceability against three criteria, each on a level from 1 to 5:
| Level | Qualified alternatives at comparable capability | Time to switch supply (qualification and ramp) | Capability or capacity gap of the best alternative |
|---|---|---|---|
| 1 | Many (six or more) | Weeks to three months | None |
| 2 | Several (four or five) | Three to twelve months | Small |
| 3 | Few (two or three) | One to two years | Material |
| 4 | One, partial | Two to four years | Large |
| 5 | None at parity | More than four years | No alternative at scale |
Each criterion level cites its evidence: the source, its date and the passage that supports the judgment. A rater’s level for a company and function is the mean of its three criterion levels c1, c2 and c3, and the final level is the mean of the two raters, a and b, where r denotes either rater:
The raters are two independent AI passes of the same model family, and the second works without access to the ratings of the first. Where their levels differ by more than one level, a third pass reconciles them with both records in view and sets the final level, and the case is referred to the editor. Agreement is reported for every sublayer as the share of company functions on which the raters fell within half a level of each other, together with a weighted kappa: a standard measure of agreement between raters that discounts the agreement expected by chance and counts near misses as partial agreement. The AI models used for rating and for weekly maintenance are developed by Anthropic and OpenAI, both of which the index scores. Because of that conflict of interest, the ratings of Anthropic and OpenAI are applied exactly as computed, with no editorial change; the editor reviewed them personally in the Frontier labs sublayer, and every entry for either company carries a disclosure.
Levels convert to a replacement time in years, which doubles from one level to the next:
A level of 1 corresponds to three months, a level of 3 to one year and a level of 5 to four years. The progression is geometric because the switching times in the rubric lengthen roughly geometrically from one level to the next. The final level is an average and can fall between whole levels; the replacement time is computed from that average directly, so a level of 3.5 corresponds to about 1.4 years. A share below the material threshold that no rater assessed is held at level 1, with a range extending to level 2.
04.Importance today
Importance today combines the two quantities function by function:
where Tif is the replacement time of company i in function f. Its unit is percent-years of the sublayer’s AI-related output at risk: the percentage of output that would stop, multiplied by the years needed to replace it.
Multiplication, rather than addition, gives the measure its intended behaviour. A large share that could be replaced quickly is discounted, while the near-sole supplier of one function scores highly even when its total revenue is small. Consider a function with a weight of 0.5. A supplier holding 60% of it at level 1 contributes 0.5 × 60 × 0.25 = 7.5 percent-years, whereas a supplier holding 40% at level 4 contributes 0.5 × 40 × 2 = 40 percent-years. The second carries less of the output but considerably more of the risk to it.
Companies are ranked by Importance, with ties broken by Reach. The value, the rank and every component behind them, including each function’s weight, share, level and replacement time, are published in the sublayer’s ledger. On the site, Importance is shown relative to the sublayer’s leader, which scores 100.
05.Rising companies
Importance today describes the present structure of a sublayer. Rising captures the fastest documented growth in AI-related activity, measured in three ways from what has already happened or has already been contracted: growth in activity, gains in share and contracted expansion.
The first component is activity growth: the year-on-year growth g of the company’s AI-related activity in the sublayer, measured on the card’s basis from two observations in the same unit whose periods end nine to fifteen months apart. Growth is expressed as a continuously compounded rate, the natural logarithm of the ratio between the current and the earlier figure, under which a doubling and a halving are equal and opposite movements. The growth g is limited to the range from −50% to +300%:
The limits keep growth from a negligible base from dominating the component; activity that rises from zero counts at the upper limit. The second component is the gain in share, the change in Reach over roughly twelve months in percentage points, available where two dated sets of share observations exist. The third is contracted expansion: backlog (orders received but not yet delivered), remaining performance obligations (contracted revenue not yet recognised), signed offtake (agreements by buyers to take future output) or capacity under construction, expressed as a ratio to current annual activity.
Each component is converted into a mid-rank percentile among all eligible companies that have it, including those among the eight by Importance. The lowest value scores 0 and the highest 100, tied companies share the average of their positions, and a component held by only one company scores 50. Among five companies with distinct values, for instance, the fastest grower scores 100, the middle one 50 and the slowest 0. Where Ki is the set of components available for company i and Pik its percentile on component k, the Rising score is their mean:
A company needs at least two of the three components to receive a score, and a missing component is left out of the mean rather than counted as zero. Eligibility requires documented current commercial activity and a Reach of at least 1% of the sublayer. This floor prevents a very small company growing from a near-zero base from taking a place. In five highly concentrated sublayers, Chip design, Foundry, Memory, Frontier labs and Developer tools, the card sets the floor at 0.25% and states its reason.
Because Rising combines three components, a Rising company need not have the fastest revenue growth in its sublayer, and each Rising company’s entry names the components behind its score. Rising is not a forecast of which companies become important, and it is not investment advice.
06.Selection of the ten
Before full research, every candidate receives a screening estimate of Reach and of its irreplaceability levels, each with a range. A candidate is screened out only when, even on its most favourable estimate, it cannot overtake the tenth company’s least favourable one, and it is not a prospect for a Rising place, which requires documented commercial activity and a Reach that could meet the Rising floor. A screened-out candidate remains in the record with the estimate that excluded it. Full research goes to candidates that can reach the ten or a Rising place; the others are scored from a lighter update of their screening evidence, which passes the same checks.
The ten places are then filled by a fixed rule. Let I8 denote the eight companies with the highest Importance, and E the set of companies outside I8 that are eligible for Rising and score at least 50. With R2 the two members of E that have the highest Rising scores, or fewer where E has fewer than two members, the selection is:
Because the percentiles are taken over all eligible companies, including those among the eight, a score of at least 50 requires momentum at least typical of them, not merely the best of the few left outside the eight. When fewer than two companies qualify, the remaining places go to the next companies by Importance, from rank 9, and several sublayers have no qualifying Rising company. The display lists the Importance places in rank order, followed by any Rising companies, marked as Rising. The rule of eight places by Importance and up to two for Rising is the same in every sublayer, so that the composition of each ten is comparable across the atlas.
The editor can replace an individual computed pick. Each override is an editorial decision with a public reason: it records the reason, the date and the computed pick it replaces, and it appears as an editorial decision on the sublayer page and on the company’s entry. The weekly rule never removes an override unless the company leaves that sublayer’s business or exits altogether, the company it replaced does not compete for a place in that sublayer while the override stands, and the weekly report continues to show what the formula would select.
07.Evidence and unequal disclosure
The index is built entirely from public sources: filings, company statements, official statistics, reputable press and published research. Consolidating that evidence and reasoning over it consistently is the work the index performs. Where a market total is published only in a subscription report, the card uses the best public evidence of it, such as press coverage of the report, company disclosures or usage data, at the uncertainty its tier implies.
Every input that feeds a score records its low, central and high values, its unit, basis and period, and its source: address, title, publisher, publication and access dates, evidence tier and, where practical, the location of the figure. Aggregators such as encyclopaedias, data-broker pages and search summaries can point towards a source but cannot be the source of a scoring input, and a company’s own statement remains the company’s claim rather than independent confirmation. Inputs are gathered and checked in separate AI-assisted passes. Every scoring input, not a sample, is checked for support in its cited source and for consistency of periods, units, business boundaries and arithmetic, and a company is scored only when both checks have passed on its inputs as they stand.
Each observation carries an evidence tier, which sets its default uncertainty when the source gives a single figure:
| Tier | Source | Default half-width around the central value |
|---|---|---|
| T1 | Audited or regulatory filing | ±5% |
| T2 | Company’s own statement (executive, release, investor material) | ±10% |
| T3 | Reputable press citing sources, such as Reuters, Bloomberg or the Financial Times | ±25% |
| T4 | Analyst or market-research estimate | ±40% |
| T5 | Derived by 5LayerCake from usage or other indirect evidence, with the method stated | ±60% |
The half-widths are relative to the central value, so a T3 share of 20% becomes 15–25%, and a range given by the source itself replaces the tier default. The highest available tier is preferred. Conflicting observations are all recorded: the best supported provides the central value and the others inform the range. Two further adjustments apply on top of either range. An observation whose period ended more than twelve months before the calculation is labelled stale, and its range widens by a further 20% of the central value at each end, so a stale T3 share of 20% becomes 11–29%. An observation on a basis other than the card’s, such as a fiscal-year figure where the card specifies calendar years, is used only with a visible label and a further 25% of the central value at each end; it is never silently converted. For a single figure x from a tier with default half-width ht, the range used in scoring is therefore:
where σ and β equal one for a stale or an off-basis observation respectively and zero otherwise. No range falls below zero, and no share exceeds 100%.
Missing never means zero. Disclosure differs greatly between a listed manufacturer and a private model developer, and an absent figure is not evidence of absent activity. A company that does not perform a function is recorded as a non-participant, with evidence. A function in which a company participates but which has not been measured blocks the calculation until it receives an estimate with a tier and a range. Private companies therefore enter with their latest credible reported figures or usage shares, labelled and ranged, and OpenAI and Anthropic are treated exactly like any other company. A business unit never takes its parent’s totals. Units such as Google DeepMind or Amazon Web Services are measured through segment disclosures, unit-specific statements or usage shares, and the rest of the parent’s business remains outside the sublayer. Every figure shown on the site carries its source and date, and estimates are labelled as estimates.
08.Uncertainty and stability
The ten are selected on central values. Because every input also carries a range, the index examines how far that selection depends on the precise values chosen. Function weights carry their low and high values from the card. Irreplaceability levels span the two raters’ levels and the final level, extended by half a level at each end to reflect judgment error and limited to the scale from 1 to 5. Shares carry the ranges described above.
The calculation runs a seeded Monte Carlo simulation of 10,000 draws: it repeats the scoring 10,000 times, each time with inputs drawn at random from their ranges, and the seed fixes the sequence of random numbers so that the result can be reproduced exactly. In each draw, every weight, share and level is sampled from a triangular distribution, which makes values near the central estimate the most frequent and values towards either end of the range progressively rarer. The weights are rescaled to sum to one, and Importance, the ranking and the selection are recomputed, with Rising scores held at their central values. For each candidate, the simulation reports its median rank, its 80% rank interval between the tenth and ninetieth percentiles, the share of draws in which it is among the eight highest by Importance, and the share of draws in which it is selected.
These frequencies define three stability classes. A company selected in at least 90% of draws is Firm, one selected in 60–89% of draws is Likely, and one selected in fewer than 60% is Borderline. The classes and the simulation results appear only in the downloadable ledger for each sublayer; they are not shown in the company list, on sublayer pages or in dossiers. The frequencies describe the stated ranges, not true probabilities. A Firm selection is robust to the uncertainty the inputs record, but the ranges are themselves estimates, and the classification inherits their limits.
Weight sensitivity complements the simulation. Each function weight is set in turn to its low and its high value, with the other weights rescaled proportionally, and the ledger records which companies enter or leave the ten as a result. A reader can therefore see which selections depend on a particular weight.
09.Validation against known bottlenecks
A measure of importance should recognise suppliers documented to have constrained AI output. Before any sublayer was scored, seven such cases were registered with their sources: TSMC’s advanced-packaging capacity, which constrained accelerator output from 2023 into 2025; SK hynix’s high-bandwidth memory, sold out for 2024; ASML, the only manufacturer of extreme-ultraviolet lithography systems; NVIDIA’s data-centre accelerators, for which demand exceeded supply in 2023 and 2024; the heavy-duty gas turbines of GE Vernova and of Siemens Energy, committed years ahead since 2025; and Dominion Energy’s 2022 constraint on new data-centre connections in Northern Virginia. These cases cover six sublayers; the other 19 have no registered case. The list and its fingerprint were recorded before scoring began, and cases cannot be added or removed without the editor’s approval and a recorded reason.
A sublayer passes when each of its case companies ranks among the top three by central Importance. A failure leads to a written diagnosis for the editor. Resolving it may require a change to the method, while editing a company’s inputs, weights or ratings to force a pass is prohibited. When the index was first computed, in October 2026, every case met the criterion: five of the seven companies ranked first in their sublayers and two ranked second.
The index also computes a reference ranking by the latest annual revenue of each company as scored, using a business unit’s own revenue where the unit is the supplier. The ledger reports the rank correlation between this baseline and the Importance ranking, and compares the selected ten with the ten largest companies by revenue. The baseline plays no part in selection. It shows what the index adds to a ranking by size, since any difference between the two arises from the functions a company supplies and the time needed to replace it.
10.Weekly maintenance and the archive
The index is recalculated every Sunday at 18:00 New York time. There is at most one run in any Sunday-to-Saturday week, and one catch-up pass follows any interruption. Each run monitors company sources, in full for selected companies and through filings and results for archived ones. A new figure that affects a scoring input counts only after passing the same source-support and consistency checks as the original inputs. Fixed, logged discovery searches look for new entrants, acquisitions, large financings, major contracts and new market-share data in every sublayer, and new names enter the candidate universe for screening. Between quarterly reviews, irreplaceability is reassessed only after a material event, such as the qualification of a new competitor. All 25 sublayers are then recomputed with the simulation and the validation cases. An unreachable source is a gap in coverage, not evidence that nothing has changed, and blocked sources are respected rather than circumvented.
A buffer rule limits turnover caused by small differences in measurement. A company selected by Importance keeps its place while it ranks tenth or better, and leaves after two consecutive runs ranked eleventh or lower; a member ranked eleventh one week and ninth the next keeps its place. A challenger takes a free Importance place after ranking eighth or better in two consecutive runs; when no place is free, it displaces the lowest-ranked member only after ranking fifth or better in two consecutive runs. A Rising company keeps its place while its score remains among the four highest of at least 50 held by eligible companies without an Importance place, and leaves after two consecutive runs outside them. A Rising challenger needs one of the two highest such scores in two consecutive runs.
An exit event requires a dated primary source. A company-wide exit, meaning a completed acquisition, a bankruptcy or departure from the business entirely, removes every placement at once and moves the company to the archive; leaving the business of one sublayer removes only that placement. Any place left empty, whether by an exit event or by a member leaving under the buffer rule, is filled at once: a Rising place by the highest eligible Rising score of at least 50, and otherwise by the next company by Importance. A Rising place filled by Importance in this way returns to Rising when a Rising challenger qualifies by holding one of the two highest Rising scores in two consecutive runs.
A company entering the ten without a current dossier that has passed an independent AI fact-check is shown as pending research until its dossier meets the site’s research standard and passes that check. Until then, the company it replaces remains visible. For the first eight weekly runs, the editor approves the exact proposal of each run before anything is published; after eight approved runs, the editor may allow changes to publish automatically, with a full report for every run. Every entry, exit, archive move and input change is recorded in a permanent change log.
On the first Sunday of January, April, July and October, the review also compares the weights on each measure card with new market data, searches the candidate universes beyond the fixed queries and repeats every irreplaceability rating with a fresh independent pair of raters. A change to a card follows the rule that applies before scoring.
A company with a dossier is either Selected, when it holds a place in at least one sublayer, or Archived, when it holds none. Companies with dossiers that the index does not select on its first calculation move to the archive. Thereafter, a company is archived when it loses its last placement under the buffer rule, or at once on a company-wide exit event; losing one placement while keeping another archives nothing. The archive record states the date, the sublayers concerned, the company’s last rank and Importance, the reason, and the fingerprint of the dossier as it then stood. Archived dossiers remain public, labelled “Archived on date; not updated since”. They leave full weekly research, but their scoring inputs continue to be refreshed from filings and results so that they can re-enter. A returning company follows the ordinary entry rule, and its dossier is brought up to date and passes the fact-check again before it is shown as Selected. Nothing is deleted.
11.Accuracy of the company dossiers
The index selects companies; the dossiers describe them. Their accuracy was measured in October 2026. Fifty dossiers were drawn at random with a recorded seed, two from each sublayer’s ten as selected before the index. For each dossier, an independent AI fact-check, without access to earlier review notes, examined about fifteen to twenty material claims, from leadership and financial figures to ownership, products, business relationships, recent news and risks. Each claim was compared with its cited source and, where necessary, a current independent source, and classified as correct, outdated, wrong, unsupported by its cited source, or impossible to check because the source was blocked. An error was material if it would mislead a reader, as an incorrect revenue figure, an incorrect chief executive or an invented relationship would, and minor otherwise.
Of the 968 claims that could be checked, 50 contained an error of some kind and 2 contained a material error. The material error rate is therefore 0.2%, with a Wilson 95% interval of 0.1–0.8%; the Wilson interval is a standard confidence interval for a proportion that remains reliable when errors are rare. Seventeen claims could not be checked and are excluded from these figures, and because claims about the same company are not independent, the interval is approximate. Both material errors concerned recent events: an acquisition still described as pending after it had completed, and a chief financial officer replaced shortly before the check. Most of the minor errors were citations that did not themselves support the statement they accompanied. The decision rule was set before the audit: a material error rate of up to 2% leads to correction of the errors found, a rate between 2% and 5% to a re-check of that type of claim across all dossiers, and a higher rate to a full re-check of every dossier. The result fell in the first band, and every error found was corrected.