Forward-looking commentary for a general audience, not personalised advice. Market data are StockAnalysis figures as of the US close on 25 Sep 2026; forward multiples are consensus. Company and industry figures are from filings, calls and forecasts dated below. Horizon: 2–5 years. All five report in the next six weeks, so this earnings run will refresh most of these numbers.
Everyone owns the AI chip story through the GPU makers. The companies that print the wafers, build the tools and license the designs have rerated too, but the market prices them very unevenly. Line them up and the strange part jumps out: the purest choke point in the stack is the cheapest stock in it.
TSMC, where high-performance computing is now two-thirds of revenue, runs a 60% operating margin and trades under 20x forward earnings. Arm trades at 130x. In between sit ASML, a genuine monopoly at a full price, Lam, a superb business at peak-cycle margins, and Marvell, the asymmetric, messy one with a catalyst eight days away.
The cycle is real, but it is not one cycle
Demand is not in question this year. SEMI's mid-year forecast (14 Jul 2026) has wafer fab equipment up 23.1% to $143.9B in 2026 and another 21.8% in 2027, and WSTS (Spring 2026) sees the chip market up 90% to $1.51T.
What matters for this basket is the shape. SEMI has DRAM equipment up 39% this year against foundry and logic up 19%, with total WFE growth slowing to about 14% in 2028. Memory is the hot end of this cycle, and memory capex is always the part that turns first. The logic side, meanwhile, is bottlenecked somewhere else entirely: advanced packaging. TSMC's C.C. Wei said on the 2Q26 call that CoWoS tightness is "limiting customer growth," and TrendForce (15 Jun 2026) expects the supply gap to narrow from roughly 20% to roughly 10% by year-end as TSMC grows packaging capacity by more than 60% into 2027.
The consumer end is ugly. IDC (26 Aug 2026) expects smartphone units to fall a record 16.7% this year as memory costs push prices up. That barely touches TSMC. It hits Arm, where phones are still 43% of royalties, squarely.
What the prices assume
A screener shows wildly different earnings multiples. The more useful observation is that four of the five sit at roughly 11x to 13x sales on our arithmetic; what separates them is how much of each sales dollar survives as cash, and for how long.
So we ran a deliberately simple reverse-DCF: 9% discount rate, 3% terminal growth, ten years at a constant growth rate, and a stated steady-state free-cash-flow margin. It asks what decade-long growth the enterprise value requires. It is a sanity check, not a model, and the outputs are approximate.
Read the last column against what each business is actually doing and the ranking mostly writes itself.
TSMC is the choke point, priced as if it is about to slow down
TSMC's July print was the kind that usually earns a premium. Second-quarter revenue was US$40.2B, up 33.7% and above guidance, at a 67.7% gross margin (58.6% a year earlier) and a 60.3% operating margin. HPC rose to 66% of revenue from 60% a year ago, and August sales were up 53%. Management guided full-year dollar growth "slightly above 40%," which on 2025 revenue of $122.4B (TSMC 4Q25 management report, 15 Jan 2026) means more than $171B this year, and Wei says demand stays strong to "probably 2029, 2030." Net cash is NT$2.49T, roughly US$79B.
Yet the ADR trades at 19.8x forward earnings and 19.5x EV/EBITDA, the lowest in the group on both. Our reverse-DCF says the $1.94T enterprise value needs roughly 13% annual growth for a decade at a 30% free-cash-flow margin, or about 16% if free cash flow stays near today's 23% of revenue while capex runs at about 39% of sales. Six years of 20% growth, then a slide to 3%, is enough. That prices a company growing 40%, with packaging capping its customers and a demand runway into the next decade, for a sharp and early deceleration. We think that gap is the most interesting mispricing in the stack.
The bear case deserves its due. The advanced capacity sits in Taiwan, and no spreadsheet prices a blockade. Tariffs are a truce with terms, not a settlement: January's Section 232 action put 25% on certain AI chips with broad data-center exemptions, and the US-Taiwan deal cut the general rate to 15% in exchange for $250B of investment and $250B of credit guarantees. Margins face two headwinds at once, with 2nm guided to dilute gross margin by three to four points as it ramps and Arizona and other overseas fabs structurally lower-margin. Two customers were 36% of 2025 revenue (press names Nvidia and Apple; TSMC does not). And $60B to $64B of 2026 capex means far less free cash than earnings suggest.
We still land in the same place. The bear's margin path, a guided 65% to 67% next quarter, is a level most of the industry never touches. The stock is up 61% in a year and only about 6% off its high, so this is not a contrarian call. It is the observation that the most essential company in the chain is priced for the least.
ASML is the monopoly you pay full price for
ASML is the only company that ships EUV lithography, and 2026 has shown what that is worth. The full-year guide began at EUR 34B to 39B at a 51% to 53% gross margin and has been raised twice, to EUR 43B to 45B at 54% to 56%. Management says 2027 low-NA EUV capacity, itself rising 30%, is "close to being fully covered with orders."
The problem is that the price already knows. At 31.2x forward, our reverse-DCF needs about 15% annual growth for ten years at a 30% free-cash-flow margin, roughly second-quarter net margin. That is faster than SEMI's 2028 forecast for the whole WFE market, sustained for a decade, with no air pocket.
And the mix argues for one eventually. Memory was 49% of system sales last quarter, with memory systems guided up more than 75% this year: a cycle-peak mix. China should be about 20% of 2026 sales (CFO, CNBC, 17 Jul 2026), and the MATCH Act would push allies toward servicing bans, which matters when installed-base management brought in EUR 2.76B last quarter. High-NA is slower than bulls hoped, since TSMC skips it through 2029. And ASML stopped reporting bookings this year, removing the lead indicator investors used to trade.
None of that makes ASML a short. It makes it a name to accumulate on a bad week, like the 6% to 7% one-day drop in mid-September, rather than chase at 31x.
Marvell is the asymmetric one, and it is messy
Marvell's case rests on guidance. Fiscal second-quarter revenue rose 37% to $2.74B, with data center 79% of the total, and the third-quarter guide implies more than 50% growth. Management has put fiscal 2027 at about $12B and fiscal 2028 at about $18B, with custom silicon more than doubling, and targets a 38% to 40% non-GAAP operating margin by the fourth quarter of this fiscal year.
Believe the $18B and the stock looks reasonable: the $231B enterprise value is about 12.8x fiscal 2028 sales and roughly 33x the operating profit implied. Taking that step as given, our reverse-DCF needs about 17% a year for the following nine years at a 30% free-cash-flow margin, demanding but not heroic for custom silicon. The asymmetry is that the stock is 21% off its high, fell about 8% in a day in mid-September, and has a date when the $18B is reaffirmed or not: the 6 October Investor Day.
Dilution is a cost and we charge it. NVIDIA's $2.0B convertible preferred converts into up to 21.8M shares at about $91.84, and the Google warrant covers up to 59.0M shares at $206.58. Together that is about 9.2% of the share count, lifting the implied growth to about 18%. The twist: beyond a small time-vested tranche, the Google shares vest per $500M of Google custom revenue through fiscal 2033, and full vesting needs roughly $120B of it on our arithmetic. The worst dilution arrives only in the best business outcome. That softens the cost; it does not remove it.
The rest of the mess is real. There is about $1.35B of net debt, one distributor was 37% of fiscal 2026 revenue and the top ten 82%, GAAP EPS is about a third of non-GAAP, and a new CFO, Dan Durn, arrives just before the biggest disclosure event of the year. This is the high-beta sleeve, and it should be sized like one.
Lam is a great business at a peak-cycle price
Lam is doing everything right. June-quarter revenue rose 30% to $6.72B at a 52.0% non-GAAP gross margin, a 20-year high, and the September-quarter guide of $8.10B comes with a 39.5% operating margin. The trouble is that these are the margins of a cycle at full stretch. Memory was 46% of system sales, four customers each take 12% to 16% of sales (Micron, Samsung, SK hynix and TSMC are named as key), and China was 26% of the quarter and 33.8% of fiscal 2026. The BIS Affiliates Rule returns on 9 November; Lam's last estimate of its impact, about $600M for calendar 2026, dates from October 2025 and has not been updated.
At 33x forward, our reverse-DCF needs about 14% a year for a decade from the September-quarter run-rate at a 30% free-cash-flow margin (June-quarter free cash flow was about 19% of sales), or about 19% from fiscal 2026 revenue. That asks a cyclical toolmaker to post SEMI's 2028 growth rate every year for ten years without a downturn. The stock is already 28% below its high, so this is not a short. It is simply where you are paid least for the cycle risk.
Arm is a great franchise priced for perfection
Arm is the one we would let come to us. At 130x forward earnings, 317x trailing and 64x sales, our reverse-DCF needs about 29% annual growth for ten years even at a generous 50% free-cash-flow margin, or twelve straight years of 25%. Trailing free cash flow of $1.4B is a yield of about 0.4%.
The bull case deserves a fair hearing. Data-center royalties more than doubled, and the new AGI CPU, co-developed with Meta and with OpenAI and Cloudflare named as customers, has more than $2B of demand across fiscal 2027 and 2028. But capacity is secured for only $1B of that, and Arm's own 20-F flags the risk that selling chips puts it in competition with licensees. Mobile is still 43% of royalties while IDC expects phone units down 16.7% this year. Stock comp of $343M was about a quarter of revenue, leaving a GAAP operating margin of 7.1%. Arm has stopped disclosing remaining performance obligations, related parties including Arm China were 30% of quarterly revenue, and SoftBank controls the company with a float of roughly 13%. That is a lot of opacity at 130x.
China and policy cut across all five
The toolmakers carry the most China risk, and the MATCH Act is the one to watch. It sits in a 150 to 240 day window, may ride the FY27 defense bill, and the Dutch and Japanese governments oppose it (Export Compliance Daily, 21 Sep 2026).
What would change our mind
A reversal in WFE or memory capex would hurt Lam and ASML first and pull the whole basket down with them. The MATCH Act passing with servicing bans would cut into ASML's installed-base business and Lam's support revenue. Any escalation around Taiwan breaks the TSMC thesis outright, and a TSMC gross-margin guide below the mid-60s would tell us 2nm and overseas dilution are running worse than management expects. For Marvell, a cut to the fiscal 2028 outlook on 6 October would remove the one number the valuation leans on.
How we rank them
Our criterion is simple: how much essential, durable cash flow you get for the growth the price already assumes. On that basis TSMC ranks first, because it is the choke point priced for the slowest growth. ASML is second, a monopoly whose quality is fully priced, so the entry point matters. Marvell is third, the most asymmetric setup with the most moving parts, where the Investor Day decides whether the $18B holds. Lam is fourth, a superb operator where peak margins and China exposure leave the least cushion. Arm is last, because even the bull case has to run for more than a decade to earn today's multiple.
What we are watching
Marvell's Investor Day on 6 October, ASML's third quarter on 14 October, TSMC's on 15 October, Lam around 21 October (our estimate; not yet confirmed), Arm on 4 November, and the Affiliates Rule on 9 November. Figures here are as of the 25 September close, and this run of reports will refresh nearly all of them.
Bottom line
Own the choke point that is cheapest, which is TSMC. Accumulate ASML on weakness rather than strength. Keep Marvell a small, high-beta sleeve into 6 October, knowing the dilution and concentration come with it. Be patient on Lam until the cycle gives a better entry, and let Arm's multiple come to you. The shared risk is that this is one AI capex cycle, and when memory capex turns, the correlation between these names will go to one. This is research, not personal advice.
Primary sources
- StockAnalysis statistics for ASML, TSM, LRCX, ARM, MRVL (close 25 Sep 2026)
- TSMC 2Q26 management report and 6-K (16 Jul 2026); 4Q25 management report (15 Jan 2026); 2025 annual report; August 2026 revenue release
- ASML Q2 2026 release and 6-K (15 Jul 2026); CFO interview, CNBC (17 Jul 2026); 2025 20-F
- Lam Research June-quarter release, SEC Ex. 99.1 (29 Jul 2026) and FY26 10-K
- Arm Q1 FY27 release and 6-K (29 Jul 2026); AGI CPU announcement (24 Mar 2026); FY26 20-F
- Marvell Q2 FY27 release (27 Aug 2026); FY26 10-K; NVIDIA preferred (31 Mar 2026) and Google warrant (18 Aug 2026) filings
- SEMI mid-year equipment forecast (14 Jul 2026); WSTS Spring 2026; TrendForce on CoWoS (15 Jun 2026); IDC smartphones (26 Aug 2026); Tom's Hardware and Reuters on TSMC roadmap and High-NA (22 Apr 2026); Reuters on US-Taiwan deal (15 Jan 2026); Export Compliance Daily on the MATCH Act (21 Sep 2026)