I ran 28 AI infrastructure companies through my two-stage filter this month, and only 5 made my list: TSMC, ASML, Nvidia, Broadcom and Microsoft. Today I'm walking through every test in that filter, the cutoff I set for each one, and why I picked that number.
The cutoffs matter more than the list. Lists go stale every quarter. A cutoff you understand works on any stock you look at for years.
The whole filter is free, as a one-page checklist and a copy-paste AI prompt. The link is at the bottom.
Why a filter at all
I think AI is a long-term shift. That's exactly why I'm picky about AI stocks. In a buildout this big, money flows through hundreds of companies, and it sticks in only a few.
My bar for any single stock is whether it's worth owning instead of an S&P 500 index fund. An index fund costs almost nothing and asks nothing of me, so an individual stock has to earn its spot.
The filter runs in two stages. Stage one is 3 checks, and failing any one ends it. Stage two is 8 questions worth a point each, and it takes 7 to make my list.
Stage one: 3 checks, no exceptions
1. Can it survive?
The cutoff: more cash than debt. Or, after subtracting its cash, debt under 1.5 times a year of EBITDA (earnings before interest, taxes, depreciation and amortization, a rough measure of operating cash earnings).
Why 1.5: that's debt the company could pay off with about 18 months of earnings. A company at that level doesn't need lenders to be in a good mood. AI spending can slow down fast. Debt payments don't slow down with it.
Why it allows some debt: plenty of good companies borrow on purpose. Broadcom carried debt from buying VMware, and still came in at about 0.69 times EBITDA (company filings, as of October 15). It passes easily.
2. Does it make cash?
Three numbers, all over the last twelve months.
Free cash flow margin of at least 15%. Free cash flow is the cash left after a company pays for everything it needs to keep running and growing. At 15%, it keeps 15 cents of every dollar of sales. That's a business that pays for its own growth instead of asking shareholders or lenders for money.
Return on invested capital above 15%. This asks how much profit each dollar put back into the business earns. Above 15% is comfortably more than money costs a big company to raise, so every reinvested dollar adds value. Under that, growth can quietly cost more than it earns.
Stock pay under 10% of revenue. Paying employees in stock is a real cost, but it doesn't show up in free cash flow. It shows up as new shares, which shrink your slice of the company. Below 10%, I can live with it.
That last one is where Meta landed. When I ran it on October 15, its free cash flow margin cleared 15%, but stock pay was about 11% of revenue, just over the line. Broadcom was at 9.5%, just under.
Meta's results last night took it further from the line. In the third quarter it spent $2.1 billion more on data centers and equipment than its business brought in, so free cash flow for the quarter was negative. Over the last twelve months, free cash flow fell to about 12% of sales, and stock pay stayed at about 11% of revenue (Meta filings, through September 30). It now misses two of the three.
3. Is the price fair?
Three conditions.
Forward P/E at or below its own 5-year average forward P/E. Forward P/E is the price divided by next year's expected earnings. I compare a company with its own history, because a chipmaker and a software company have never traded at the same multiple.
PEG at or below 1.5, with a review band up to 2.0. PEG is the P/E divided by the expected growth rate. A stock at 30 times earnings growing 20% a year has a PEG of 1.5. It's a check that I'm paying for growth that's actually expected. Between 1.5 and 2.0, a company doesn't fail automatically. I have to write one line on why the growth justifies the price, and the flag stays on the scorecard. Above 2.0, it fails.
I added that band this month because of Microsoft. When I first ran the numbers, its PEG was under 1.5. After the stock's run from its June low, it was about 1.7 by October 15. A huge company with slower, very durable earnings can deserve a little more room than a fast grower, but I want that call written down every time, so I can check it later.
No run-up of 50% or more in 3 months. After a run like that, I'd be paying for someone else's excitement. A great company can still be a bad price for a while.
Arista is the clearest example. It's an excellent business with $13 billion of net cash and no debt, and it failed here: about 42 times forward earnings against its own average of about 38 (as of October 15). KLA had fallen about 43% from its June high and still failed on price. A falling price and a fair price are two different things.
Stage two: 8 questions, 7 points to make the list
A clear yes gets a full point, a partial yes gets half.
The 4 number questions
Is it growing with strong margins? Revenue growth above 10%, an operating margin above 20%, and a gross margin above 50% for software or 35% for hardware. Hardware gets a lower bar because it has to pay for materials and factories that software doesn't.
Are few investors betting against it? Short interest under 5% of the shares that trade. This one is a positioning check. When a lot of professionals bet a stock will fall, I want to know what they've found before I go further.
Has it beaten the S&P 500 over 5 years without piling on new shares? This is the index fund test in plain numbers. If a company couldn't beat the index over 5 years, it has to show me why the next 5 will be different. And if the share count ballooned, existing owners paid for that growth. Microsoft's 5-year return was +85.0% against +85.1% for the S&P (as of October 15), a tie, so it got half a point.
Have I written down what would change my mind? One line on why I'm watching it, and one or two measurable lines that would take it off my list. I take this question most seriously, because it's the one I'll actually use when the stock drops 20% and I'm tempted to explain the drop away. I walked through how I build one of these in my October 8 issue.
The 4 judgment questions
A moat, and a CEO I'd trust with the cash? A moat is something a competitor can't copy with money and a few years. Micron got half here. Right now, demand for memory is so far ahead of supply that customers can't easily switch, and that gives Micron real pricing power. But that protection comes from the shortage. Once SK Hynix and Samsung add enough capacity, or a Chinese competitor like CXMT can offer something similar, a customer can move to whoever offers the better price. A moat still holds after supply catches up. If that day comes, I'd have to re-evaluate Micron's score.
Is the trend behind it years from done? The market has to keep growing for years, and the company has to be the one capturing that growth.
Is the culture strong? I look for years of doing what management said it would do. Broadcom got half a point here, after customer backlash following the VMware deal.
Does it keep its best people? For these companies, a few hundred engineers can be the whole advantage.
Why 7 out of 8
One weak spot is normal. Every real company has one. Two weak spots usually means I'm talking myself into it.
Micron scored 6.5, half a point short. That's the filter doing its job: it's a company I respect, and it still doesn't make the list.
What passing doesn't tell you
Passing means a company is worth my research time. It doesn't tell me when. Microsoft passed while sitting 2% off its high, well above its 200-day average. Broadcom passed while 31% below its high. For timing, I use the price lines from my October 15 issue.
Run it yourself
The filter is free: a one-page checklist with every cutoff above, plus a copy-paste prompt that works in ChatGPT, Claude or Gemini. Start with a stock you already own, so you'll spot it fast if a number looks wrong. Then check three of the numbers against the company's own filing before you trust any of it.
Stay disciplined,
Koh
Figures as of October 15, 2026, unless dated otherwise.
Positions as of publication: NVDA, MSFT, META, AMZN, GOOGL, MU, DRAM
Disclaimer: Nothing in this email, the Reset and Invest newsletter, the Make Your Own Alpha book, or any course or digital product from Starshine Media LLC constitutes investment advice or a recommendation to buy or sell any security. Numbers and observations are as of publication. I may hold positions in companies discussed. Always do your own research and consult a licensed financial advisor before making investment decisions.