Investing notes — the two buckets

The research from the ai.koglabs.com deep-research run, laid out so it's usable. The framework is universal; the concrete allocation is still blocked on three answers at the bottom.

source: ai.koglabs.com chat · 2026-08-04 · educational research, not personalized financial advice

00The flow

Two buckets. They never mix. Everything else is detail.

Bucket A — get rich slowly

AI stocks · compounding

Boring, diversified, time-in-market. This is where the real money actually compounds. Researchable, ownable edge.

~90% of capital

Bucket B — double up

Speculation · entertainment

Treat every euro as already spent. Upside with entertainment attached, not a strategy. Hard cap, chosen before you're excited.

5–10% of capital, max
The single biggest predictor of whether you come out ahead is not stock-picking skill — it's how you split those two buckets and never let B bleed into A. Bucket B bleeding into Bucket A is how people blow up.

01What to watch before buying anything

Score a stock against all of these. No single metric decides — the pattern does.

Valuation

  • P/E — context is everything. 30 is cheap for a 40% grower, insane for a utility. Use forward P/E, but remember forward estimates are optimistic and get cut.
  • PEG — P/E ÷ growth. ~1 reasonable, <1 potentially cheap, >2 you're paying up. Only as good as the growth estimate, which for AI names is a guess.
  • P/S — for names without real earnings. Above 10–15 prices in years of flawless execution; 20+ is euphoria.
  • Trailing vs forward — "cheap on forward" means analysts assume a big jump. Ask why.

Business quality

  • Growth AND margins together — growth with expanding margins is pricing power; growth with collapsing margins is buying revenue.
  • Moat — switching costs, network effects, scale, proprietary data. Ask: durable advantage, or a feature a hyperscaler replicates in a quarter?
  • Customer concentration — extremely common in the AI-chip chain. If 40% of revenue is 2–3 customers, one lost contract is a collapse.

Red flags

  • Insider selling — scheduled 10b5-1 sales are noise. Cluster selling, several executives dumping large amounts at once, is signal.
  • Dilution & stock comp — many AI firms pay staff in stock, inflating share count 2–5%/yr. Your slice shrinks even as the business grows. Check shares-outstanding trend and SBC as % of revenue.

AI-specific risk

  • The capex cycle — $500B+ projected AI infrastructure spend in 2026 (Goldman). Hyperscaler capex guidance is the best leading indicator for the whole chain.
  • Circular financing — chipmakers invest in AI labs, labs buy their chips, that revenue justifies more investment. Compared openly to dot-com vendor financing (CNBC, Jul 2026). If paying end-demand doesn't show up, the loop unwinds violently.
  • What's priced in — the whole game. After a stock triples on AI hope you're not betting on whether AI works, but on it working better than an already-euphoric market expects.

02The stack — where the names live

Map by layer. Each behaves differently. Verify live price and multiples before buying — these move fast.

L1Chips / semiconductors the picks & shovels
NVDATSMAVGOAMDASMLMU

Bull. Highest growth in the chain. NVDA is still the default compute layer with the CUDA moat and insane margins; TSM is a neutral toll-booth that profits whoever wins; ASML is the only maker of EUV litho — extreme moat; HBM memory is a genuine bottleneck.

Bear. Most cyclical layer. Peak-margin risk, customer concentration in a handful of hyperscalers, and NVDA sits at the epicenter of the circular-financing loop, priced for perfection. TSM carries Taiwan geopolitical risk.

TSM is often the lower-drama way to own the whole trend.

L2Hyperscalers / cloud the landlords
MSFTGOOGLAMZNMETA

Bull. They own the datacenters and rent the AI out. Real cash flows today fund the AI bet, diversified, hard to displace.

Bear. Enormous capex compressing free cash flow; the market punishes any quarter where AI spend doesn't show ROI.

Lowest-risk AI exposure for a beginner — profitable megacaps, not pure bets.

L3AI software / applications the frothiest layer
PLTRNOWCRWD

Bull. If AI creates real enterprise value, the application layer is where it gets captured.

Bear. Highest valuations, thinnest moats. Many trade at P/S of 15–30+, and "AI feature" is easy to copy.

Tread lightly here.

L4Infrastructure — power, cooling, networking the underrated pick
VRTGEVANET

Bull. They profit no matter which AI model wins, and the demand is physics-driven. AI datacenters are gluttons for electricity; liquid cooling becomes mandatory at AI rack density; the networking and optical interconnect layer scales with every cluster.

Bear. Also cyclical, and several of these — especially cooling and power — have already re-rated hard on exactly this thesis.

Where a lot of smart money looks, precisely because it's less obvious.

L5ETFs the "I don't want to pick" answer — often the smartest one

Broad market first. A global or S&P 500 index fund is already ~35% big-tech/AI. You may have far more AI exposure than you think without picking a single stock.

Thematic AI/semi ETFs exist for a concentrated tilt — check the expense ratio and how much the holdings overlap what you already own.

EU point: as a European investor you'd buy UCITS ETFs, not US-domiciled ones, for tax and regulatory reasons.

03The bubble question

Read this twice.

Credible voices are flagging dot-com parallels: circular financing, capex outrunning proven end-demand, valuations pricing perfection. That doesn't mean don't invest — it means how you invest matters more than what.

Dollar-cost average

  • Buy fixed amounts on a schedule rather than lump-summing at a possible top. Removes timing stress and averages your entry.

Position sizing

  • No single stock above 5–10% of Bucket A. The AI theme shouldn't be your entire portfolio — pair it with boring diversification.

Both can be true

  • AI can be genuinely transformative and current prices can be a bubble that corrects 40% first. The internet was real; 2000 still wiped out people who bought the top on margin.

04Bucket B — the double-up game

Honesty first, math first.

House edge

Every venue takes a cut. Prediction markets have spreads, casinos have edges, memecoins have insider dumps and fees. You fight a headwind on every single bet.

Gambler's ruin

Bet a large fraction of your bankroll repeatedly and even a positive-edge strategy goes to zero — variance eventually deals a streak you can't recover from. Bet 50% per flip on a fair coin and you're mathematically doomed.

The retail scoreboard

The research is brutal: the vast majority of active day traders lose money, and the profitable minority barely beat costs. Memecoin traders face negative expected value from fees, MEV bots and insider allocations before they even pick wrong.

"Double or nothing"

To double reliably you need real edge (rare), or you're accepting a coin flip minus vig and calling it a strategy. One double feels great; the sequence ruins people — press winnings enough times and one loss zeroes everything.

The menu, ranked by how badly the odds are stacked against you

VenueOdds againstWhat it actually is
Prediction markets
Polymarket, Kalshi
▮▮▮▮ Buy Yes/No shares at a price equal to the market's implied probability — a share at $0.30 pays $1 if right. Your edge exists only where you estimate probability better than the crowd, on a question you know cold. Doubling means finding badly mispriced long-shots and being right more often than the market implies — hard and rare. Access is a live regulatory question in the EU.
Options ▮▮▮▮▮ Leveraged exposure for a fraction of the capital, cutting both ways brutally. LEAPS (1yr+) are the least-reckless version — more time for the thesis to play out. The trap: short-dated options routinely go to zero. Total loss is the base case, not the tail. A tool for a specific time-bound view, not a slot machine.
Leveraged ETFs
2x, 3x, inverse
▮▮▮▮▮ Fine for a single day's move, terrible to hold. They reset daily, so in choppy markets volatility decay grinds them down — you can be right on direction over a month and still lose money. Do not buy-and-hold a 3x ETF.
Crypto majors
BTC, ETH
▮▮▮▮▮ The acceptable-speculation tier. Volatile, but not engineered to extract you the way memecoins are. Can DCA a small amount.
Memecoins ▮▮▮▮▮ The closest thing to a lottery ticket with negative expected value. Insider allocations, rug pulls, sandwich bots and fees mean the deck is stacked before you click. People do hit 10–100x — the survivors are loud, the wiped-out majority silent. Pure entertainment money you've already written off.

05The rules that keep it fun instead of ruinous

The barbell: ~90% ultra-safe, ~10% ultra-aggressive, nothing in the mediocre middle.

  1. Cap Bucket B at 5–10% of total capital. Pick the number now, before you're excited.
  2. Per-bet limit: no single speculative position above 10–20% of Bucket B — roughly 1–2% of everything. The Kelly lesson: even with edge, bet small enough to survive variance.
  3. No re-loading. When Bucket B is spent, it's spent until the next scheduled top-up. This one rule prevents 90% of blowups.
  4. Never on borrowed money, never on money with a job to do — rent, emergency fund.
  5. Separate account or wallet so you can physically see the boundary. Blurred lines mean B eats A.
  6. Log every bet. Most people who think they have edge don't. The log tells the truth.

06What's still open

Three answers turn this map into a move list

The framework above is universal. The executable choices — specific account, starter allocation, which 3–5 tickers or ETFs, a sane Bucket-B setup — all depend on these.

1 · Region / tax residency The single biggest filter. Decides broker (IBKR vs local), which ETF wrapper you can buy (UCITS vs US-domiciled), your capital-gains treatment, and whether prediction markets are legally reachable at all.
2 · Ballpark amount + timeline Sets position sizes, and decides whether single stocks make sense at all versus just one broad ETF. Also: how much is genuinely losable on the Bucket B side.
3 · Risk appetite Steady compounding or swing-for-the-fences. Decides the A/B split — 95/5 versus 85/15.