Actionable insights — Can Arista Networks win the AI race?

1. Run a fixed numbered checklist with published thresholds, and score it

1. Fix the same 15 questions for every candidate — business model, management, moat, end market, risks, balance sheet, capital intensity, capital allocation, profitability, dilution, past growth, forecast growth, valuation, owner's earnings, shareholder returns.
2. Attach a numeric threshold to each so the answer is pass or fail, not an opinion: gross margin >40%, ROIC >15%, net debt/FCF <4x, goodwill/assets <20%, CAPEX/revenue <5%, ROE >20%, net margin >10%, FCF/net income >80%, SBC/net income <10%, revenue growth >5%, EPS growth >7%, owner's earnings CAGR >12%, since-IPO return >12%.
3. Score each step out of 10 and average to a Total Quality Score, so businesses are comparable across sectors.
4. Keep quality and price as separate outputs — the score answers "is this good", not "should I buy it".

• A high score created by one exceptional metric. Averaging fifteen scores hides which one is load-bearing.
• Thresholds that suit some sectors and not others — a 5% CAPEX/revenue bar excludes most industrials by construction.

2. For an asset-light business, read R&D as the real capital intensity

1. Start with the standard capital-intensity tests: CAPEX as a share of revenue and of operating cash flow.
2. When both come back trivially low, do not conclude the business is cheap to run — ask where the reinvestment actually goes.
3. For a technology or research business, substitute R&D as the reinvestment line and apply its own bands: R&D/revenue 10–20%, R&D/operating cash flow 25–50%.
4. Then decide whether the innovation spend is the moat or merely the cost of staying still.

• Capitalised software development moving spend from the income statement to the balance sheet and flattering both ratios.
• R&D that must be spent forever to hold position. That is maintenance capital wearing a different label.

3. Subtract stock-based compensation from free cash flow before you value it

1. Take the forward free cash flow estimate as reported.
2. Subtract the full stock-based compensation figure — it is a real transfer of ownership away from you, even though it never leaves the bank account.
3. Add back growth CAPEX, calculated as total CAPEX minus depreciation and amortisation, so that expansion spending is not treated as a cost of standing still.
4. Discount the adjusted number, not the headline one — and check whether the same adjustment was applied to the earnings multiple you quote elsewhere.

• Companies where the buyback exists only to offset dilution. That is compensation, not capital return.
• Your own inconsistency: adjusting the DCF but quoting the unadjusted multiple in the summary.

4. Convert the price into a required growth rate, then judge that rate against history

1. Fix the return you require — here 10% a year — rather than forecasting a price.
2. Solve backwards for the free-cash-flow growth rate the current price implies over your holding period.
3. Compare that required rate to the company's own realised growth over five and ten years, and to the base rate for businesses of that size.
4. Treat a required rate near or above the historical rate as a red flag, because you are being asked to pay for a repeat of the best decade.

• Anchoring on a spectacular past rate. Ten years of 46.7% growth makes 18.9% feel conservative; base rates for $180bn companies say otherwise.
• A ten-year horizon hiding the terminal assumption. Most of the value sits beyond the forecast.

5. When three valuation methods disagree, decide in advance which one binds

1. Always compute all three: forward multiple against its own ten-year history, an expected-return decomposition, and a reverse DCF.
2. Record each as an explicit pass, fail or unknown rather than blending them into an average.
3. Decide beforehand whether a single fail vetoes the idea, or whether a majority carries it.
4. Where the answer comes from judgement rather than the rules, say so.

• The Earnings Growth Model's exit multiple doing the work. Assuming a fall from 37.7x to only 32.0x is itself a forecast.
• Methods that share an input. All three here start from the same analyst earnings estimate.

6. Size customer concentration by who the customers are, not just how much they buy

1. Find the revenue share of the top one or two customers and treat anything above roughly a third as a structural risk, not a footnote.
2. Then ask the sharper question: could those customers replace you? Scale, engineering capability and vertical integration matter more than the percentage.
3. Check whether their spending is discretionary or contracted, and how visible the pipeline is.
4. Cross-check against inventory — a supplier holding stock for long lead times is exposed twice if orders pause.

• Hyperscalers designing their own networking hardware. These particular two customers have done exactly that in adjacent components.
• Concentration that is presented as a growth story while the cycle is up.

7. In a standards fight, back the cheaper open standard against the proprietary one

1. Identify where a supplier is simultaneously a partner and a competitor — that is where the standards fight is.
2. Ask which side of the fight the buyer's incentives sit on. Buyers generally prefer the open, cheaper, multi-vendor option once volumes get large.
3. Check that the open-standard vendor actually captures the economics — software, support and switching costs, not just the box.
4. Track the buyer's disclosed architecture decisions rather than either vendor's marketing.

• The proprietary vendor bundling the standard into the chip, which removes the choice entirely.
• "They almost never switch" claims that have not been tested by a genuine price war.

8. Be willing to finish the work and still say no

1. Complete the full quality assessment before looking hard at the price, so the verdict is not reverse-engineered from a view.
2. Separate the two conclusions explicitly: what the business is, and what it currently costs.
3. Record the pass as a watchlist entry with the condition that would change it, not as a rejection.
4. Publish the no. Work that only ever produces buys is not a process.

• A no without a number. This one does not say at what price Arista becomes a buy, unlike the FICO pass which names $901.
• Waiting indefinitely for a great business to get cheap. The cost of that patience is invisible and can be large.