How our numbers are made

Five rules behind every valuation

Value the business, not the narrative. Every valuation rests on cash a shareholder can eventually claim. Where a company's story and its arithmetic disagree, we publish the arithmetic and let the story argue for itself in the bull case.

The base case is the most likely path, not the prudent one. Each base-case input is the single most likely value on the evidence: contracted revenue, guidance, disclosed pricing and the company's own record. Caution belongs in the bear case, where it is priced explicitly and given a weight; pipeline and ambition that the evidence does not yet support belong in the bull case, priced the same way. A base case built from a cautious choice at every step compounds a discount it never declares.

Every risk is priced once. A risk sits in the cash flows, in the scenario weights or in the discount rate — never in two of them. Each report prints a map of where every material risk is priced, and where it is deliberately not.

The capital stack comes before the shareholder. Debt, preferred, accreting project equity and dilution all rank ahead of the common. Any multiple applied to revenue or assets that ignores those claims is measuring the wrong thing.

The model is published, not summarised. Every report carries its working model with the assumptions exposed as controls. A fair value you cannot interrogate is an assertion, not an analysis.

Different businesses need different arithmetic

There is no single valuation model that works everywhere, and pretending otherwise is how research goes wrong. We select the framework from the economics of the business, state which one we used in the header of every report, and version it here.

The standing commitments

What every report publishes, and which number the rating rests on

The modal base case is the value of the most likely path. It is the number the model on the page shows on arrival.

The fair-value band is the same base case run across a grid of discount rates and terminal growth rates around the central setting. Each report prints the grid and its ranges. The band, not any single cell, is the output, and the band anchors the rating.

The probability-weighted value averages the scenarios at the analyst’s own weights. When it sits away from the base case, the report names the skew and explains it rather than averaging it into the rating.

A reverse DCF runs the model backwards from the market price, one input at a time, and states what the price requires — a growth path, a margin, a discount rate — so the distance between our view and the market’s can be read in the business’s own terms.

A twelve-month trading value, where the market’s multiple regime matters, estimates where the shares might trade on the multiples the market applies today. It is labelled as a market-regime estimate, published beside the fair value rather than blended into it, and it is never the anchor of the rating.

How the index multiples are built

The charts on Index multiples give a trailing P/E and P/S for five index funds, every trading day: QQQ and SPY from mid-2010, IGV from the start of 2011, SOXX from August 2011 and DRAM from its launch in April 2026. A sixth chart is an index of our own, the quantum pure plays, measured on a different yardstick for reasons set out below. Nothing in them is taken from a data vendor's screen. The holdings are the ones each fund reports to the SEC, with their filed weights: Form N-PORT every quarter since 2019, and before that the quarterly N-Q and N-CSR schedules for the iShares funds and the shareholder reports of the two trusts, which published no quarterly holdings: QQQ's basket before 2019 is a September snapshot held for a year, SPY's before September 2019 a March or September snapshot held for six months. A fund that holds a share through a total-return swap rather than outright is counted at the swap's notional, which is the exposure, not at the swap's value, which is only its unrealised gain: DRAM keeps two fifths of its assets in swaps on Micron, SK hynix and Samsung against a pile of T-bills, and Micron is its largest position. Between two filings a weight drifts with the share price, and a change the fund makes inside a quarter appears with the next filing, which is published about sixty days after the quarter ends. The earnings and revenue of every constituent come from its own 10-K, 10-Q, 20-F or 6-K, taken on the day the filing was made and never from a later restatement, and summed over the trailing four quarters. Market value is the share price times the company's own filed share count; where that count misleads, it comes from elsewhere in the same filings — Berkshire Hathaway's from the Class A and Class B figures on each report's cover, Visa's from the as-converted total in its equity note, and the weighted average for companies that file their shares issued, treasury stock included, as the basic count.

The ratio is the aggregate form: the market value of every name the fund holds, at that day's weight, divided by the sum of their trailing twelve-month earnings. A loss-making constituent reduces the earnings pool rather than being dropped, so a group carrying heavy loss-makers reads higher here than on a screen that quietly excludes them — software in particular. P/S is built the same way on revenue. The line breaks on any day the covered basket as a whole lost money, because a negative P/E is not a number worth printing.

Quantum computing is charted on an index we built rather than on a fund. The listed quantum funds hold mostly other things: the Defiance Quantum ETF (QTUM) is five sixths chipmakers, software makers and other technology companies, and the WisdomTree Quantum Computing Fund (WQTM) keeps its pure plays at a quarter of its weight beside IBM, Microsoft and the chipmakers, so their multiples describe those companies, not quantum. The Sonde Quantum Index holds only companies whose business is quantum technology: quantum computers and the parts and software built for them (IonQ, Quantinuum, D-Wave, Rigetti, Infleqtion, IQM, Pasqal, Xanadu, Quantum Computing Inc and Horizon Quantum), and quantum-safe security, meaning quantum key distribution, quantum random numbers and post-quantum cryptography (QuantumCTek, SEALSQ, BTQ, Quantum eMotion and Arqit). A company with a quantum unit inside a larger business, such as IBM or Honeywell, is left out, and so is WISeKey, whose accounts consolidate SEALSQ's. Each member counts from its first day of trading as a public company — for one that listed by merging with a SPAC, the day the merger closed, not the shell's years at ten dollars a share — and members are weighted by market value, so the index reads the sector the way the market sizes it rather than counting its smallest names as heavily as its largest. Its level, 1,000 on 14 December 2022, is kept the way the S&P 500's is, with a divisor that absorbs a listing or a share issue so that only share prices move it, and is charted with the members on its own page.

None of these companies earns a profit and most sell little: on 9 October 2026 the fifteen were worth $52 billion against $485 million of trailing revenue, half of it IonQ's. What each of them does have is a research budget and a large pile of cash raised from shareholders. So the index is measured on EV/R&D: the members' combined enterprise value (market value, less cash and short- and long-term investments, plus debt) over their combined trailing twelve-month research and development expense, $1.4 billion that day. EV/sales is shown beside it. Enterprise value rather than market value, because cash is nearly a fifth of these companies' worth today and was half of some of them in 2023. The line starts in December 2022, the first month in which members worth at least 70% of the index had a year of results on file, the same floor as the funds', and has run from 12 times R&D in August 2024 to 184 times in October 2025.

The figures come from each company's filings, first-filed and dated by filing, like the rest; QuantumCTek's from its reports to the Shanghai exchange. Several members listed in 2026, and their last full year sits in a registration statement rather than an annual report; those years are read from the prospectus, and their cash from the last balance sheet before the listing until the first one after it, so the money a listing raised is missing from enterprise value for a quarter or two. The Canadian, European and Singaporean members report in their own currencies and largely outside the SEC's machine-readable format, so their half-year figures are typed from the filed reports and converted to dollars at each period's average rate. Foreign filers state their share count once or twice a year while selling stock every month, so for them the count is the month-end figure our price vendor publishes for the whole company. Quantinuum and Xanadu are valued on all their share classes: Quantinuum's Class B, held by Honeywell and the other pre-listing owners, is exchangeable one for one and carries most of the company, and Xanadu's unlisted Class A carries six sevenths of it.

Three cautions. Arqit reports no research costs at all, and Pasqal no R&D line comparable to the others' — it presents its costs by nature, and the R&D it does disclose is a fraction of its spending — so both count toward the index's value and its EV/sales but not its EV/R&D. Quantinuum's trailing R&D includes $294.9 million of stock compensation booked when its June 2026 listing vested older awards, which holds the reading down until that quarter leaves the trailing year in mid-2027. And the ratio is not comparable across the page: a profitable semiconductor company's R&D is small beside its earnings, so SOXX would read far higher than the quantum index on it without being any dearer. It is a measure of this group against its own past.

Coverage is the share of the fund's weight the reading actually measures (for the quantum index, the share of its members' value). A name is left out on a day when it has no price, no trailing year of results on file, or only one that ended more than eighteen months before. A name that files nothing with the SEC is normally never in; DRAM's Asian memory makers and the quantum index's QuantumCTek are the exception, read from the results each publishes at home (SK hynix's newsroom, Samsung's earnings release, Kioxia's Tokyo filing, the Taiwan regulator's quarterly table for Nanya, Winbond, Macronix and Phison, and a mainland data service for GigaDevice and QuantumCTek), converted to dollars at the period's rate. Where a market reports the year to date, quarters are differenced from the previous release; where it publishes no announcement date, the statutory deadline stands in. The table below is as at 9 October 2026 and is restated whenever this section is.

Earlier years are thinner where a constituent has since been acquired and delisted: no free price feed keeps a delisted name, so those are priced at each quarter's filed value until the next filing, and their earnings are fetched from the filings they made while listed. Filings before 2019 name their holdings without identifiers, so each name is matched to the company's SEC record by hand or by EDGAR's entity search. From 2012 coverage stays above 80% for every fund and above 90% for most years; a series starts on the first day it measures at least 70% of the fund, which is why none begins before mid-2010, when companies were still being phased into machine-readable XBRL filings.

What our conclusions mean

The test is applied to the reference price printed on the report, on the day of publication. Ratings published under version 1.2 were set by a different test — modelled fair value within 20% of the price either way — and they stand, and are measured, as published.

We deliberately avoid buy, sell and hold. Those words describe an action a specific person should take, and we do not know anything about any specific person.

A modelled fair value is an estimate, not a fact and not a price target. It is what our assumptions produce, and our assumptions are judgement calls made by people who can be wrong. Every one of them is printed on the report and exposed as a control precisely so that you can replace ours with yours.

What changed in version 1.3

Against version 1.2, version 1.3 changes six things.

Earlier reports are not restated under the new version: each keeps the version printed in its header, and its rating is read against the rules of that version. Version 1.2, published 27 July 2026, is kept unchanged.