Seven companies passed our screen. In three of them the sum itself was wrong.

Prices from the 21 September 2026 pre-open panel unless stated. Estimates from the 18-21 September 2026 panel. Every accounting figure comes from the companies' own SEC filings. Prices and multiples move in real time.

We run a two-condition screen before anything else gets read: pay less times earnings than the business grows, in each of the next two years. It is deliberately crude. It is meant to throw away ninety-five per cent of the market in ten seconds so the remaining five per cent can be read properly.

These seven passed it. Super Micro Computer (SMCI), Century Aluminum (CENX), ACM Research (ACMR), Kulicke & Soffa (KLIC), Liquidia (LQDA), Strategic Education (STRA) and Excelerate Energy (EE). Seven for seven. Six clear both years outright; Liquidia (LQDA) clears year two and its year one is not computable at all, because the company went from a loss to a profit and there is no percentage to take from a negative base.

Then we read the accounts, and the seven split cleanly in two. In three of them the arithmetic itself is wrong — the sum that produced the pass does not say what it appears to say. In the other four the arithmetic is correct and insufficient: the screen computes the right number and the number is not what decides the outcome. This article is that split. It is more useful than any of the seven verdicts, because the failure modes repeat and the companies do not.

1 · ACM Research (ACMR): the multiple and the growth are not measuring the same profit

$71.85 · ~$5,000M market value · next results 7 November 2026

ACM makes the machines that wash silicon wafers between steps of chipmaking, plus a newer copper-plating and packaging line. It is American on paper — California headquarters — and Chinese in practice: almost everything is built, researched and sold in China through a separately listed Shanghai subsidiary of which the parent owns only 73.2%.

The screen shows 34.6 times earnings against growth of 39% and 42%. Comfortable pass, both years.

The 34.6 is computed from one profit figure and the 39% from a different one. The multiple uses official accounting: $2.11 a share over twelve months. The growth is measured from the panel's $1.56, which strips out stock compensation and the gains on a portfolio of listed Chinese shares that the Shanghai subsidiary happens to own. Those gains were large: 41 of every 100 dollars of pre-tax profit in the last twelve months did not come from selling machines. $84M of revaluation on Chinese listed equities, $70M of it in a single quarter, plus $21M from an affiliate selling securities at a gain.

Put the same profit on both sides and you are not paying 34.6 times. You are paying 46 times — more than 39 and more than 42. It fails both years.

Underneath, the business is genuinely growing: $389M of revenue in 2022 to $1,038M over the last twelve months, with a political tailwind that does not turn quarter to quarter, because Washington has forbidden the sale of advanced equipment to China and Chinese fabs have no choice but to buy at home. But in five and a half years it has declared roughly $590M of profit while the business consumed $71M of cash. The $1,000M of net cash on its balance sheet came from selling shares, not machines.

The $61 is the honest line, and it is not a fantasy number: the stock closed below it on 8 May, four and a half months ago, and in this name 15% moves happen in two sessions.

2 · Kulicke & Soffa (KLIC): the growth is division by almost zero

$85.54 · next results 5 November 2026

KLIC makes the machines that bond a finished chip to its package, with gold or copper wire thinner than a hair. Of every 100 dollars billed last quarter, 68 came from those traditional wire bonders and only 9 from the newer advanced-packaging line, which loses money — $12.6M in the quarter.

The screen approves year one on growth of +1,738%.

That is not growth. It is a rebound off a base of almost nothing. Fiscal 2025 earned $0.21 a share on the panel's basis and $0.004 on official accounting. Against a base of virtually zero, any ordinary profit produces a gigantic percentage. Every cyclical emerging from a bad year passes year one of this screen automatically, and that tells you nothing whatsoever about whether it is cheap.

Year two is the one that means something, and it does pass: +52.8% with 13.6 points of headroom. The business is real too — revenue in the April-July quarter went from $148M to $330M, and business profit from minus $6M to plus $68M. It has no financial debt at all, $516.6M in cash, and has bought back one share in four since 2016.

The thing the screen cannot see sits in the analysts' own numbers: 2026 estimated $3.86, 2027 $5.90, and 2028 $5.36 — down 9%. When a consensus draws its own ceiling and its own decline, it is drawing a cycle. And in cyclicals the price historically moves opposite to the multiple. In 2022, with a record $7.45 a share, this stock traded at five times earnings, because everyone knew it would not last. They were right: a year later it earned $1.82.

Today it costs $85.54. The highest close of the entire previous wave was $74.06 (7 September 2021). You are paying more than at the top of the last cycle for a profit that has not yet reached that peak.

3 · Excelerate Energy (EE): the base year carries a one-off that inflates the growth

$36.02 on the 18 September panel; closed Monday 21st at $33.92, down 5.8%, with no SEC filing or company news explaining it · next results 5 November

Excelerate owns about twelve ships that are really floating gas ports, rented on long contracts to governments and public utilities — Bangladesh, Pakistan, Argentina, Brazil, the Emirates, Finland, Germany, Jordan and, from 2027, Iraq — where it gets paid whether the client uses the ship or not. At 31 December it had $3,315M of minimum rents already signed, averaging 5.8 years.

The screen sees 24.8 times earnings against +25.8% this year and +28.6% next. Passes by a point.

The +25.8% is almost entirely an artefact of the base. The 2025 figure it grows from — $1.28 a share — contains $34M of one-off costs from buying the Jamaica gas business: lawyers, banks, integration. The company itself publishes the clean number. Strip those costs and 2025 was $1.52, so growth to 2026's $1.61 is +5.9%, not +25.8%. At 24.8 times, +5.9% does not come close.

So the case rests entirely on year two, and year two rests on one thing: the first Iraqi gas terminal, a five-year contract with the Ministry of Electricity. Read the calendar in its own filings:

This is not an opinion about whether it will work. It is a calendar, and it has already moved nine months. The company writes that it continues "despite the ongoing conflict in the Middle East" and is "adapting execution plans". The port is in Basra. From the May downgrade you can back out what Iraq is worth: roughly $35M of cash per six months of delay.

And there is a second thing the screen cannot see, which is the shape of the company. There are two share classes. You can buy the 31.4M class A shares; the founding family holds 82M class B shares, which do not trade but convert one for one and are 2.6 times everything currently listed. They hold 72.3% of the operating partnership and the majority of the votes.

4 · Super Micro Computer (SMCI): the profit is real and the cash is negative

$40.49 (21-9-2026, pre-open) · Nasdaq · fiscal year ends 30 June

Supermicro assembles the racks that fill AI data centres. Of every 100 dollars it bills, 89 go straight back out to buy the parts. It is a giant assembler, not a patent business, and a single supplier — unnamed in the annual report — sold it 63 of every 100 dollars of everything it bought last year.

The screen passes it with room: 12.5 times earnings against +19.6% and +23.3%. And unusually, year two survives its own accounts without inventing a margin, which is rare on this list.

Here is what the profit calculation cannot see. Last year the company earned $2,230M. In the same year, $6,810M more cash went out of the business than came in. That is not a typo; it is in the cash flow statement of the annual report filed on 31 August.

Inventory and unpaid invoices are growing far faster than sales. To sell $17,000M more it had to tie up $12,100M more in parts and receivables. And it is not a freak year: across the last three fiscal years it earned $4,430M in total and burned $7,640M of business cash.

The gap is closed by shareholders. In June it raised $5,640M: 52.3M new shares at $27.50, $4,312M of preferreds costing 7% a year that convert by force in June 2029 into between 131 and 157 million more ordinary shares, plus a market programme of up to $1,250M. Two years ago there were 556M shares; by July there were 657M; counted fully, with preferreds and convertibles, the number approaches 890 million.

The binary is clean. Either it keeps growing as promised, in which case it needs more warehouse, more receivables and probably more money from investors, and your slice shrinks; or growth slows, cash comes back, and the 2028 profit does not arrive. Both at once, no. Today's price assumes both.

There is also the part that produces two-figure falls in a day: in March 2026 New York prosecutors charged three people linked to the company, including a co-founder and former director, with diverting servers containing restricted chips around export controls. The company is not charged, and says its own investigation with an outside law firm and forensic accountants found no evidence current management knew. It continues to receive requests from the Department of Justice and the export control office. The stock fell 33% in one day, from $30.79 to $20.53, on 20 March 2026. Separately, the auditor — new since late 2024 — gives an adverse opinion on internal controls: one unremediated failure over who can access the IT systems the accounts are made on. Of four material weaknesses, three are fixed.

5 · Century Aluminum (CENX): the cheap multiple is the warning, not the opportunity

$39.92 (21-9-2026, pre-open)

Century smelts aluminium in three plants — two American, one Icelandic. It does not set its selling price: it charges the world aluminium price in London plus a premium for delivering inside the United States, and that premium is set mostly by the American import tariff.

The numbers are spectacular. The US delivery premium went from $855 a tonne in the first half of 2025 to $2,518 in the second quarter of 2026, nearly three times, after the import tariff rose from 25% to 50% in June 2025. Each tonne sold in the US brought in $6,021 against $3,432 a year earlier: 75% more for exactly the same ingot. Operating profit went from $18M in Q2 2025 to $212M in Q2 2026 — twelve times, on sales only 20% higher.

The screen sees 6.9 times trailing earnings and 4 times 2026 estimates, against +307% and +8.6%. A normal company trades at 20.

In a smelter, that is exactly backwards. The multiple is tiny because the market does not believe today's profit lasts. Everything is at a record simultaneously — record margin, record premium, record profit — alongside a multiple at its lows. When every record in a cyclical coincides with a price that looks like a giveaway, that has historically been a sell signal, not a buy one.

And the market is already saying so. The stock touched $70.43 on 2 June 2026 and is at $39.92 — 43% lower — while analysts have not cut their estimate, still around $10 a share. Either the three analysts are wrong or the price is. Not both.

Then there is the check that actually matters. For 2027's $10.88 a share to come out of its accounts, redone in tonnes rather than percentages: with all three plants full, at exactly Q2 2026 prices, it would bill around $3,970M and earn about $12.6 a share. So 2027 works, with 16% of headroom. Each $100 a tonne the US premium falls removes roughly $0.34 a share a year. Which means the premium can fall to about $2,000 a tonne — 91 cents a pound — before 2027 stops working.

That is where it already is. On 20 August the United States announced it would cut the tariff on Canadian aluminium. The September delivery premium fell 8% in a day to 95 cents a pound, around $2,090 a tonne, and October and November contracts fell more than 12%. The premium is less than $100 from the line.

So the whole thing reduces to a political decision. Either the 50% tariff holds through the end of 2027 and the $10.88 arrives, or Canada's is cut — Canada supplies roughly 3 of the 4 million tonnes the US imports — the premium falls toward $1,400 and 2027 profit drops to about $9. The analysts assume the first. The price, down 43%, is discounting the second.

One thing has genuinely changed and deserves saying: the balance sheet is no longer the risk. It owes $480M against $343M of cash at 30 June plus $94M of federal production subsidy collected in July, and the company says that by end-July it had more cash than debt. For the first time in a decade it owes nothing net. If the cycle turns, this time it does not go under; it just earns very little.

6 · Strategic Education (STRA): the growth is real and its source has a ceiling

$79.11 (21-9-2026, pre-open, −0.91%) · next results 29 October

STRA owns Strayer and Capella, two American for-profit universities teaching working adults almost entirely online, plus Torrens in Australia and New Zealand, and a third leg that is the one actually growing: Sophia Learning, a cheap subscription whose courses convert into credits.

Its year two survives its accounts comfortably — as, to be fair, do Kulicke & Soffa's and Supermicro's. What is unique here is that we could not find a second-order problem behind it either. It pays 13.2 times, grows 16.7% then 17.2%, and we checked the second year from scratch against its own filings under three separate tests — standard, with the tax it actually pays, and stripping out the Australian dollar tailwind. It needs between 81% and 93% of its best-ever margin. All three pass. It has no bank debt, and what it generates yields 8.3% a year against a 5.25-6.25% cost of margin: you earn two to three points a year simply holding it.

So what is wrong? Nothing in the numbers. Everything in where they come from.

The 17% growth comes from three places and none of them is new students: keeping more of each dollar through cuts and automation, buying back shares — last quarter half the EPS growth came simply from there being fewer shares — and charging more per student. For next quarter the panel itself expects sales up just 2% and earnings per share up 20%.

Both levers have ceilings. By 2027 the margin will already be very near its record. And the buybacks are spending more than the company generates: $207M returned over twelve months against $147M generated, with cash falling from $153M to $134M in six months. At a sustainable pace it can buy back one share in twenty a year, not the one in thirteen of the last year. The board's authorisation expires on 31 December.

And its largest supplier is the US government, which is changing the rules. The 90/10 rule removes federal aid from a for-profit university that collects more than 90 of every 100 dollars from the government two years running. Strayer was at 89.64 in 2024 — a third of a point from the limit. Unlimited federal graduate loans ended in July 2026, capped now at $20,500 a year and $100,000 in total; Capella has 30,218 of its 51,753 students in master's or doctoral programmes, precisely the ones who borrow most. None of this appears in the promised growth, because analysts project from the current trend.

The verdict, stated precisely: the screen decides correctly here, for the next two years, and today's price sits inside the rule with 21% to the limit. What it does not tell you is what happens in the third. Take it as a company that will earn somewhat more for two years, not one that will keep growing at 17%.

7 · Liquidia (LQDA): the accounts are clean, and a judge decides

$66.90 (21-9-2026, 4:00 ET)

Liquidia sells one drug: YUTREPIA, an inhaled treprostinil powder for pulmonary hypertension, approved on 23 May 2025 and in pharmacies since that June. Fifteen months.

What it has done in those fifteen months is the most impressive operating performance on this list:

Sales are rising in a straight line, roughly $39M more each quarter. It costs 6 dollars in every 100 to make. It no longer burns cash: it generated $133.2M in the first half of 2026, against burning $70.2M in the same period a year earlier. And the panel's numbers match official accounting in every published year: this one does not dress anything up.

The screen passes it easily — it trades 40% below the limit. And that margin is almost meaningless, for a reason no accounting test can reach.

United Therapeutics, which sells the same active ingredient with fifteen years' head start and bills $3,183M a year, has four suits open against it. One of them — patent '327, covering use in pulmonary fibrosis patients — went to trial in June 2025. The verdict has been pending for fifteen months. Liquidia's own lawyer said in August it "could come any day".

The company does not disclose in its filings what share of sales comes from each indication, so the loss cannot be calculated. What is known is that the fibrosis group is the larger of the two: more than 60,000 patients against 45,000. And if it loses, United Therapeutics can additionally claim damages on everything sold to date. A trade secrets trial follows in January 2027.

Year two is also at the edge on its own numbers — the $5.57 a share asked for 2027 requires 105% of its record margin — but that is the second-order problem. The first-order problem is that what moves this share 40% in a day is a judge, not a quarter.

The split: three wrong sums and four right ones that are not enough

Set the companies aside and the pattern is the point.

Group one — the arithmetic itself is wrong. In these three the sum that produced the pass does not say what it appears to say:

Group two — the arithmetic is right and it is not what decides. In these four the screen computes the correct number, and the correct number is not the thing that will determine the outcome:

Only the first group is a criticism of the instrument. The second is a criticism of what we ask the instrument to do. Arithmetic cannot see a tariff decision, a court date, a customer concentration, or a cash flow statement that contradicts the income statement above it — and it was never going to.

What we would do

Seven passed. On today's prices only one — Strategic Education (STRA) — sits inside the rule with room we would act on, and even there the two caveats are real. The other six we would either wait on or decline outright. That is not a failure of the screen. That is the screen doing its job: it got seven names down from a whole market, and then it stopped being useful, exactly where it was always going to stop.

In one sentence

"Seven companies passed a screen designed to find cheap growth; in three of them the sum itself was wrong, and in the other four it was right and not what decides."

← All articles