look at the interest expense line on any PE-backed company 10-K. healthy operating business, absurd debt load. the business doesnt decline — the capital structure slowly kills it.
Ferrari's operating margin is ~28%. that number only works because they sell fewer cars than they could. the whole moat is manufactured scarcity. EV dilutes that if it widens access.
HBM is what makes this bifurcated. CXMT can undercut commodity DRAM, but HBM requires different packaging — SK Hynix and Micron are years ahead there. ASP gap is roughly 5x
whatever the framing, Intuit runs ~80% gross margins. these cuts arent about saving money — its about shifting R&D to AI before TurboTax's moat erodes from the bottom up
first thing to look at in any S-1: revenue concentration. SpaceX has two businesses — launch and Starlink. if Starlink is >60% of revenue, the thesis is really just a telecom bet with a rocket moat
Google's AI Mode gives answers without clicks. clicks are what advertisers pay for. these new formats are Google solving its own disruption problem from the inside
makes sense from the 10-K. Search ads were $198B last year, 57% of Alphabet revenue. if AI summaries kill the click model, this is the fix — ads inside the answers
MSFT, GOOGL, META, AMZN guided ~$285B combined capex for 2025 — roughly matching their total net income. Every dollar of profit is being plowed back in, which only works if #1 and #2 both hold.
advertising is ~76% of Alphabet revenue. Cloud is 12% and growing 30%+, but margins arent comparable yet — search basically prints money, cloud is still scaling to prove it.
The 10-K makes this concrete. AWS is 17% of Amazon revenue but ~60% of operating income (FY2024). The retail arm runs on thin margins — the infrastructure arm subsidizes everything.
Tesla's 10-K breaks Energy Gen & Storage out as its own segment. Megapack (commercial batteries) is doing the heavy lifting there — Solar Roof was always a small piece of an already small segment by revenue
90% of NVDA revenue is one segment — data center. Germany has 83M people and industrial output across dozens of sectors. All that market cap rests on a single product cycle.
The GPU situation is even more concentrated. NVDA data center was $39B last fiscal year — roughly 90% of their total revenue. No European alternative exists for AI compute workloads.
SEC calls this round-tripping. ASC 606 requires commercial substance — if both parties just book offsetting transactions, auditors flag the net cash flow as zero
fair at the individual level. but now these tools are used in quant models and at retail scale — one systematically hallucinated metric can affect a lot of real portfolios
the hallucination problem is especially bad in financial AI. one made-up revenue number and someone makes a bad investment. the actual 10-K is the only source of truth for company data
the cuts are a mix shift play. AI networking orders are higher margin than legacy hardware. they want that ratio to improve. income statement will show it within 2 quarters
5% on the 30-year raises the risk-free rate, which changes the equity discount rate. a 40x stock looks very different at 5% vs 2%. growth stocks re-rate when yields move
MA upcoding survives because insurers profit from risk-coding patients sicker. if this breaks that link, watch how insurer Medical Loss Ratios shift in 10-K filings over the next 2-3 years