Every week someone sends me a chart comparing NVIDIA to Cisco in 2000. "See? Same pattern. Bubble. Get out." And every week I look at NVIDIA's income statement — $215.9 billion in revenue, 71% gross margins, 53% net margins — and I wonder if these people have ever actually read a 10-K.
Let me be clear: parts of the AI trade are in a bubble. But the core of it — the infrastructure layer — is the most fundamentally sound tech rally I've seen in my career. The trick is knowing where the line is between real and hype.
Kyiv Rental Yield 2026: Data by District That Matter
Let's start with what the hyperscalers are actually spending. In 2026, the five largest cloud and AI infrastructure companies — Microsoft, Alphabet, Amazon, Meta, and Oracle — have committed to spending between $660 billion and $690 billion on capital expenditure. That's not a forecast. That's committed spend.
Break it down: Amazon at $200 billion, Alphabet at $175-185 billion, Meta at $115-135 billion, Microsoft at $120 billion+, Oracle at $50 billion. Approximately 75% of this goes to AI-related infrastructure — servers, GPUs, data centers, networking, power.
These are not startup companies burning VC money. These are the most profitable businesses on Earth, deploying their own cash flow (and increasingly, debt) into AI infrastructure because their customers are demanding it. Microsoft has an $80 billion Azure backlog they literally cannot fulfill because they don't have enough data center capacity. This is demand-pull spending, not supply-push speculation.
The Valuations: Cheaper Than You Think
Here's the part that surprises people:
NVIDIA (NVDA) — $199.55, ~22x forward earnings. When the AI rally started in early 2023, NVIDIA was trading at 60-70x forward earnings on speculative revenue. Today, after three years of explosive growth, the forward multiple has actually compressed because earnings grew faster than the stock price. At 22x forward, NVIDIA is trading below the S&P 500's average multiple. Read that again.
Microsoft (MSFT) — $415.75, ~24x forward earnings. Down 18.67% from its 12-month average P/E of 33x. The voluntary retirement program affecting 7% of its workforce spooked people, but this is classic Microsoft: cutting costs while investing in Azure. Next earnings report: April 29.
Alphabet (GOOGL) — $341.68, ~31x trailing earnings. This one's trickier. The P/E is 42% above its 4-quarter average, partly because Google Cloud is finally profitable but also because the market is pricing in AI search dominance. Q1 revenue expected at $92 billion. The 31x is rich, but if Gemini monetizes well, it's justifiable.
Meta (META) — $659.15. Zuckerberg's pivot from metaverse to AI infrastructure was one of the best strategic decisions I've seen from a public company CEO. Meta is now a top-3 AI infrastructure spender, and the advertising business continues to print money. I'm long.
AMD (AMD) — $274, target $310-320. The Meta custom chip deal was a game-changer. AMD is no longer just selling standard GPUs — they're designing custom silicon for hyperscaler workloads. Bank of America and Stifel both raised targets above $300. Earnings May 5.
Infrastructure vs. Applications: Where the Real Money Is
This is the critical distinction most investors miss. The AI value chain has two layers:
Layer 1: Infrastructure. Chips (NVIDIA, AMD, Broadcom), data centers (Equinix, Digital Realty), power (Vistra, Constellation Energy), networking (Arista Networks), and cloud platforms (AWS, Azure, GCP). This layer is profitable today. NVIDIA's margins are obscene. The hyperscalers are generating cash. The demand is real and measurable.
Layer 2: Applications. Software companies building AI products — chatbots, coding assistants, enterprise tools, autonomous agents. This layer is mostly unprofitable. OpenAI generated $13 billion in 2025 revenue but doesn't expect profitability until 2030, with projected cash burn of $17 billion in 2026, $35 billion in 2027, and $47 billion in 2028. Anthropic, Cohere, and most AI application startups are in the same boat.
My position is simple: own the infrastructure layer, be selective in the application layer. The picks-and-shovels strategy isn't original, but it's correct. Every gold rush makes money for the people selling shovels.
The Dot-Com Comparison: Why It's Mostly Wrong
People love the dot-com analogy because it's scary and simple. But let me walk through why this is fundamentally different:
1. Profitability. In 2000, Cisco — the NVIDIA of that era — traded at 200x trailing earnings. NVIDIA trades at less than 50x trailing earnings. More importantly, the top five dot-com companies generated combined revenues of about $50 billion. The top five AI companies generate over $1.5 trillion in combined revenue. This isn't comparable.
2. Revenue quality. Dot-com companies derived revenue from ad impressions and page views — metrics that collapsed overnight when spending dried up. AI infrastructure revenue comes from multi-year cloud contracts with Fortune 500 companies. Microsoft Azure doesn't lose 80% of its revenue in a quarter.
3. Enterprise adoption. The internet in 2000 was still mostly a consumer phenomenon. AI in 2026 is an enterprise transformation. Every major bank, pharma company, manufacturer, and retailer is deploying AI in production workloads. This is not speculative demand.
4. Concentration. This one cuts the other way. The top five companies now hold 30% of the S&P 500 — the highest concentration in half a century. In 2000, it was about 18%. If you're looking for a genuine bubble risk, it's here: the market's dependence on a handful of names.
Where I Think the Bubble Is Real
Let me be honest about where I see froth:
Private AI companies. The valuations being assigned to pre-revenue AI startups are insane. Companies with $10 million in revenue are raising at $5 billion valuations. Most of them will fail. This is the dot-com parallel that actually holds.
AI SaaS multiples. Some public AI software companies are trading at 30-50x revenue (not earnings — revenue) on the assumption that AI will justify premium pricing forever. It won't. Competition will compress margins, just like it did in cloud computing.
The "AI wrapper" companies. Startups that are essentially a ChatGPT API wrapper with a logo and a marketing budget. These have zero moat and will be wiped out as foundation models get cheaper and more capable.
Hyperscaler debt. The big five raised $108 billion in debt in 2025 alone, with projections suggesting $1.5 trillion in total debt issuance over the coming years. Aggregate capex now exceeds projected free cash flow. If AI revenue growth disappoints even modestly, the debt servicing becomes a real issue.
My Position and Thesis
I'm long the infrastructure layer with approximately 25% of my equity allocation in AI-related names. The specific positions:
- NVIDIA: Core holding. Adding below $200. The valuation doesn't support a sell thesis at these multiples.
- AMD: Building position. The Meta deal changes the narrative from "also-ran" to "legitimate alternative."
- Microsoft: Holding. Azure backlog is bullish, but I want to see April 29 earnings before adding.
- Meta: Holding. The advertising business funds the AI capex, which creates a virtuous cycle.
- Alphabet: Small position. The 31x P/E gives me less margin of safety than NVIDIA at 22x.
I'm hedging the position with put spreads on the QQQ (Nasdaq 100 ETF), which gives me defined-risk downside protection if the whole AI trade reverses.
The Bottom Line
Is AI a bubble? In spots, absolutely. Private market valuations are disconnected from reality. Some public AI software names will crash 60-80% when growth disappoints.
But the core infrastructure trade — the chips, the data centers, the cloud platforms — is backed by real revenue, real profits, and real enterprise demand. This isn't 1999 Pets.com. This is 1996 Cisco — expensive, yes, but in the early innings of a structural shift that will run for another decade.
My biggest risk isn't that AI is overhyped. It's that I'm not positioned aggressively enough. And that's a risk I can live with, because I'll have plenty of time to add if the thesis keeps playing out.