Eight Chip Stocks, Two Questions: Is the Business Good, and Is the Price Cheap? #

4 September 2026 | Data as of 3 September 2026 close | Names: Arm, Nvidia, AMD, GlobalFoundries, Intel, Micron, TSMC, ASML | Verdict: three good businesses, zero cheap prices

Written as an informational analysis, not personalised financial advice. It is a fact-based look at eight semiconductor companies through one consistent ruler, and an honest framing of why "cheap" is the wrong word for every one of them today. Figures were pulled from company filings, Yahoo Finance and Finnhub on 3 September 2026 and will move.

The answer in one table #

"Value ÷ price" is a probability-weighted estimate of what each business is worth two years out, divided by what it costs today. Above 1.25 is cheap, 0.90 to 1.25 is fair, below 0.90 is expensive. Every input is shown below so you can disagree with a number rather than with a conclusion.

Three of the eight are excellent businesses. None is on sale. That is the whole post; the rest is the working.

Why chip stocks need their own ruler #

Three things make a generic value screen useless here.

The cycle dominates. Semiconductor revenue swings 20 to 50 percent from peak to trough and earnings swing far more, memory most of all. That flips the usual signal: a low P/E at the top of the cycle is a warning, not a bargain, because next year's earnings will be lower. Micron trades at 6 times next year's consensus earnings today. In May 2026 it traded at 7 times and the stock went on to rise another 35 percent, then fall 24 percent. The multiple told you nothing; the cycle told you everything.

Six different businesses share one label. An IP licensor with a 97 percent gross margin and no factories (Arm), a commodity memory maker that spends 40 percent of revenue on equipment (Micron), and a company with a literal monopoly on the machines that print the smallest chips (ASML) have nothing in common except the word "semiconductor." Each needs a different definition of "good" and a different gauge of "cheap."

Attention is the sector's steady state. AI spending is the loudest story in markets. Every one of these names reaches an ordinary investor through headlines, Reddit or YouTube. By then the easy money is gone. A framework exists so that a price decides, not a story.

The framework, in plain words #

Step 1. Which kind of chip company is it? IP licensor, fabless designer, foundry, integrated turnaround, memory maker, or equipment maker. This decides which numbers matter.

Step 2. Where is it in its own cycle? Five gauges, each read against the company's own five-year history: revenue against trend, gross margin against its range, inventory days, capital spending against history, and price-to-book against its range. Four or five gauges at extremes means Peak. A company at Peak can never be called cheap on this year's numbers; only on normalised ones.

Step 3. Is the business good? Eight tests, each scored 0, 1 or 2:

1. Does it generate free cash flow after paying staff in stock, even in the trough year of its last cycle?
2. Is the moat visible in the numbers: gross margin at the trough, market share, prices rising on flat volume?
3. Does it earn a real return on the capital it employs, averaged through a cycle?
4. Would the balance sheet survive a bad year without asking the market for money?
5. Does it fund itself by issuing shares? Rising share count, warrants, converts, stock pay above 5 percent of revenue for hardware or 15 percent for asset-light models.
6. Are insiders buying or selling? Cluster selling by the CEO and other officers within 90 days is a hard fail, planned or not.
7. Does the story match the financials, and has guidance been met four quarters running?
8. Concentration: any customer above 15 percent of revenue, China exposure, subsidy dependence.

A business is good at 12 or more out of 16, with no zero on tests 1, 5 or 6.

Step 4. Is the price cheap? Seven tests, of which three decide:

• Graham break-even growth. Benjamin Graham's rough formula says value is earnings times (8.5 plus twice the growth rate). Turn it around and today's price tells you what growth rate it assumes. Compare that with the growth the business has actually delivered over a full cycle. Cheap means the price assumes at least five points less than the business has shown, with a quarter left over.
• Reverse discounted cash flow. What ten-year free-cash-flow growth rate does today's price require, at a 10 percent discount rate and 3 percent terminal growth?
• Scenario value. Bear, base and bull earnings two years out, each with a probability and a multiple appropriate to the business model. Add them up, divide by price.

Cheap needs all three. Fair is a scenario value between 0.90 and 1.25 times price with nothing absurd in the other two. Everything else is expensive.

Where the cycle sits #

Memory prices, the sector's most sensitive dial, are still rising but slowing: DRAM contract prices rose 95% in the first quarter, 58 to 63% in the second, and are forecast to rise 13 to 18% in the third. Micron's own management now talks of "meaningful moderation in the rate of price increases."

Every gauge that has moved since the spring has moved further into late-cycle territory, not back.

Is the business good? The scorecard #

What the zeros mean:

• AMD, Micron and Arm all fail on insiders. At AMD, Lisa Su sold 335,000 shares (about $130 million) in March, May and June under a pre-set plan, and the CFO, CTO and two senior vice-presidents sold in the same window. At Micron, the CEO sells a fixed 40,000 shares every month (about $136 million since May) while the Chief People Officer sold $57 million. At Arm, the CEO and the CFO both sold, the CFO four times, most recently on 27 August. Not one officer at any of the three bought a share on the open market. Planned sales are still sales. When everyone who knows the business is selling, the unanimity is the signal.
• Arm does not produce cash for its shareholders. Free cash flow after stock compensation was negative in three of the last four fiscal years. Stock pay runs at 21 to 33 percent of revenue and there are no buybacks to offset it. The 97 percent gross margin is real; it flows to employees, not owners.
• Intel is a turnaround, not yet a business. Free cash flow after stock pay has been negative for four straight years, about minus $47 billion cumulative. The share count rose 20 percent in twelve months through government, SoftBank, Nvidia and public placements. Return on capital is negative. To its credit, the CEO bought $10 million of stock in August and the CFO bought in January; nobody senior is selling.
• GlobalFoundries fails on returns. Net cash, a flat share count, no insider cluster, but a return on equity of 6 to 7 percent, and it renegotiated its long-term customer agreements downward in 2025. A good balance sheet attached to a mediocre earning engine.
• Nvidia's zero is concentration. Three customers were 54 percent of revenue in the latest filing, every one of them designing its own AI chips, and the China business has gone to essentially zero on export controls.

Is the price cheap? The scorecard #

Two things stand out.

Forward P/E is the sector's favourite lie. Micron at 6 times next year and Nvidia at about 15 times the year after look like the cheapest large stocks in the market. Both sit at Peak on four of five cycle gauges. Micron's consensus for fiscal 2027 spans $107 to $221 per share across 39 analysts; even the low end would be a cyclical high, and what follows a cyclical high is a cyclical low. The cash side says the same thing: no name here yields more than 2.3 percent in free cash flow after stock pay, and every one needs more than 20 percent annual cash-flow growth for a decade to justify its price.

TSMC is the one that nearly clears. Its price assumes about 11 percent growth on normalised earnings against 22 percent delivered over five years and 16.5 percent over ten. Graham's formula at 15 percent growth gives a value around $520, a 20 percent margin against today's $415. The framework asks for 25 percent at a cycle peak. Closest of the eight, and still not there.

The scenarios behind "value ÷ price" #

Earnings two fiscal years out, times a multiple that fits the business, weighted by probability. Bear probabilities are set at 30 percent (35 for Micron) because the cycle gauges read late or peak everywhere.

Memory at peak earnings has historically cleared at 5 to 8 times, which is why Micron gets 7 to 9 on fiscal 2027; foundries 17 to 24; equipment 25 to 32; fabless designers 16 to 30 depending on the cycle read. Change a multiple and the ratio moves; none of the changes I can defend moves a name from fair to cheap.

Name by name #

TSMC. The best business of the eight and the benchmark for the rest. Gross margin of 67.7 percent at a 60 percent operating margin, $80 billion of net cash, a share count unchanged in five years, a dividend raised 17 percent in August, and wafer prices going up 5 to 10 percent in 2027 with a further surcharge on AI chips. Taiwanese insiders filed 151 purchases against two sales in six months. The risks are known: Nvidia and Apple are 36 percent of revenue, Taiwan is Taiwan, and capital spending at 45 percent of revenue is why a business earning 30 percent on equity yields only 1.7 percent in free cash. Fair at $415. Cheap starts near $316, roughly where it traded in January.

ASML. The only monopoly in the set, with a backlog that funds the next two years: the 2026 sales guide was raised three times to €43 to 45 billion, 2027 orders are "substantially secured," and capacity is being raised 30 percent for 2027. China fell from 42 to 14 percent of system sales in four quarters, so the export-control hole is mostly behind it, except that Washington is now pressing The Hague for a total ban including servicing of installed tools. Net cash of €5.6 billion, share count down 4.6 percent, negligible stock pay. Two soft spots: it stopped publishing quarterly bookings this year, and insider dealings were not checked (ASML reports to the Dutch regulator, not on US Form 4). Fair at €1,423. Cheap starts near €1,076.

Nvidia. On the checklist a good business: free cash flow after stock pay was positive even in the fiscal 2023 trough, $68 billion of net cash, share count down 4.5 percent in five years, an $80 billion buyback and a 25-fold dividend increase in May. On the cycle gauges the most stretched name here: inventory up 128 percent against revenue up 85 percent, purchase commitments doubled, gross margin guided down to 71 to 72 percent on memory costs, and three customers at 54 percent of revenue who are all building their own chips. No officer bought a share on the open market; one director sold $445 million. On next year's earnings it is the cheapest of the eight. On the cycle it cannot be called cheap. Fair at $230. Cheap starts near $182.

GlobalFoundries. Down 52 percent from a May high that coincided, to the day, with its controlling shareholder Mubadala selling a $1.9 billion block. Then a soft third-quarter guide, flat cash flow, and a sector sell-off. Net cash, a buyback, a first dividend, no insider cluster, utilisation in the high 80s. But phone-chip revenue has fallen four quarters running, automotive flipped from +24 percent to minus 10 in the latest quarter, the long-term contracts were repriced lower, and return on equity is 6 to 7 percent. In June this blog called it "a genuinely great foundry at a price that asks for perfection." At $44 the price no longer asks for perfection; it asks for a business that earns more than 7 percent on its capital, and that is the part still missing. Fair at $44. Cheap starts near $32.

Intel. The next node (18A) is showing up in the numbers: revenue up 18 percent in the latest quarter, gross margin at 40 percent from a 33 percent trough, data-centre revenue up 59 percent. The foundry still loses $2.1 billion a quarter, no external customer has been named for the node after that, and the company sold $20 billion of stock at $95 in August. The CEO and CFO are buyers and nobody senior is selling, which is the one thing Intel has that AMD, Micron and Arm do not. Six out of sixteen on the good test; 0.86 on price. Expensive, with a turnaround that may or may not fill the gap.

Micron. The textbook case for why this framework exists. Trailing P/E of 21, forward P/E of 6, and every cycle gauge except inventory at an all-time extreme: an 84.6 percent gross margin against a ten-year range of minus 9 to 59 percent, revenue that tripled on price rather than volume (DRAM shipments up low single digits in the quarter, prices up 60 percent), a price-to-book of 10.6 where no previous peak passed 2.5, capital spending guided from $27 billion to more than $45 billion, and China's CXMT at 10 percent of the DRAM market and shipping HBM. Over the last nine years Micron produced about $12 billion of free cash flow after stock pay in total. The market values it at $1.07 trillion. The CEO sells 40,000 shares a month.

AMD. Revenue up 47 percent and data-centre revenue up 107 percent, yet gross margin is stuck at 54 to 57 percent because its AI accelerators earn less than the company average. Free cash flow after stock pay yields 0.9 percent. The share count is up 35 percent in five years, and the company handed OpenAI a warrant for 160 million more shares at one cent, about 10 percent of the company, as part of a supply deal. Five senior officers including the CEO and CFO sold within 90 days. Expensive at 0.70 before the warrant, 0.64 after.

Arm. The best franchise in the set attached to the worst equity. Ninety-seven percent gross margin, 350 billion chips shipped, and yet free cash flow after stock pay was negative in three of the last four years. GAAP P/E of 251, and even on optimistic non-GAAP earnings two years out the price assumes 35 percent growth forever. In July it cut its royalty-growth outlook to "high teens," stopped disclosing the two metrics investors used to track it, and is moving into selling its own chips in competition with its licensees. The CEO and CFO both sold. SoftBank owns 87 percent with $18.5 billion of margin loans against the shares.

What this means in practice #

Watch, don't buy. For the three good businesses, the sensible action is to write down the price at which each turns cheap and set an alert there: TSMC below $316, ASML below €1,076, Nvidia below $182. Then stop looking. If those prices arrive it will be because something broke, and the first question at that moment is whether the break is a cycle turn (buy) or a thesis break (wait).

For the three insider-cluster names, no price is cheap enough yet. An open-market purchase by the CEO or CFO of AMD, Micron or Arm would reopen the question. Until then the tests are not met.

For GlobalFoundries, the number to watch is return on equity. Two quarters above 12 percent, the company's own 2028 target, would change the verdict at the right price.

For anyone already holding one of these (most index-fund investors hold all of them through a world tracker), nothing here says sell. It says: this is not the moment to add on top, and a forward P/E of 6 on a memory stock is a description of the peak, not a discount.

What would change the whole picture. Two quarters of falling memory prices, TSMC or ASML guiding below the prior year, or the large cloud companies guiding capital spending flat. That is when the cycle read flips, the good names approach their cheap prices, and this framework gets run again with very different inputs.

Bottom line #

Three of these eight are among the best businesses in the world: TSMC, ASML and Nvidia. All three are fairly priced at the top of their cycle, not cheap. GlobalFoundries is fairly priced for a business that does not earn enough. Intel is an expensive turnaround with honest insiders. Micron, AMD and Arm are expensive, and the people running them are selling. The cheapest-looking numbers in the sector, Micron at 6 times and Nvidia at 15 times forward earnings, are the ones the cycle says to trust least.

Good and cheap are two different questions. Right now, in semiconductors, the answers are "yes, three of them" and "no, none of them."

Sources: company 10-K, 10-Q, 6-K and 8-K filings and earnings releases (Nvidia Q2 FY27, AMD Q2 2026, Arm Q1 FY27, GlobalFoundries Q2 2026, Intel Q2 2026, Micron Q3 FY26, TSMC Q2 2026, ASML Q2 2026); SEC Form 4 insider filings; Yahoo Finance and Finnhub for prices, consensus estimates and analyst targets; TrendForce for DRAM and NAND contract pricing; Mercury Research for CPU share. Prices as of 3 September 2026 and will move. The framework itself is at /research/sectors/semiconductor-framework.md and the full working, including every scenario input, at /research/sectors/semiconductors-2026-09-03.md. Not investment advice.