Content is user-generated and unverified.

The AI "Picks and Shovels" Playbook: Infrastructure Companies Built to Win No Matter Which AI Model Wins

TL;DR

  • Whether ChatGPT, Gemini, Claude, or some other model "wins," they all run on the same physical stuff — chips, memory, data centers, power, and networking — so the companies that sell those building blocks are the safest way for beginners to invest in the AI boom.
  • The four biggest cloud companies (Microsoft, Google/Alphabet, Amazon, Meta) are on track to spend about $725 billion on capital expenditures in 2026 — up a staggering 77% from 2025's record $410 billion, according to Goldman Sachs — and that money flows straight to the hardware and infrastructure names below.
  • The highest-quality picks (very profitable, strong balance sheets) include NVIDIA, Broadcom, TSMC, Arista, Vertiv, Eaton, and Constellation Energy; the riskier or less-profitable names include Intel, Samsung, and newer/heavily indebted players. ETFs like SMH, SOXX, DTCR, and GRID let you own a basket instead of betting on one stock.

Key Findings

1. Hyperscaler spending is exploding and accelerating. The "hyperscalers" are the giant cloud companies that buy most AI hardware. Per Goldman Sachs (June 2026), Meta, Microsoft, Amazon and Alphabet "collectively plan to allocate $725 billion to capital expenditures in 2026 — up a staggering 77% from last year's already record-breaking $410 billion." Looking further out, Goldman now expects a combined $5.3 trillion of capex for those same four hyperscalers from fiscal 2025 to fiscal 2030 (raised from a $4.5 trillion estimate before Q1 2026 earnings). Individually for 2026, Amazon guided to about $200 billion in capex, Alphabet to $175–190 billion, Meta to $115–135 billion, and Microsoft toward roughly $145–190 billion (depending on calendar vs. fiscal-year framing). The key point for investors: this is a multi-year buildout, and the companies repeatedly say they are "supply-constrained," not "demand-constrained."

2. The "picks and shovels" idea. In the 1849 Gold Rush, the people who reliably made money weren't the miners — they were the ones selling picks, shovels, and jeans. Same idea here: instead of guessing which AI model wins, you can own the companies that sell the tools every AI company must buy.

Details

Category 1: AI Hardware (Chips/GPUs)

NVIDIA (NVDA) — The undisputed king. NVIDIA makes the GPUs (graphics chips) that train and run most AI. Per Silicon Analysts (April 2026), "NVIDIA commands approximately 80-90% of the AI accelerator market by revenue as of 2025," with its share peaking near 87% in 2024 and projected to ease toward 75% by 2026 as rivals scale. In its fiscal year 2026 (ended January 2026), revenue hit about $216 billion with net income of about $120 billion, and data center revenue alone was $197 billion. Gross margins are around 75% — extraordinary. Quality: Top-tier. Hugely profitable, tons of cash, dominant moat (its CUDA software locks in developers).

AMD (Advanced Micro Devices, AMD) — The clear #2, though with only about 5–8% of the merchant AI accelerator market. Makes EPYC server CPUs and Instinct AI GPUs (the MI300/MI350 line). Full-year 2025 revenue was a record $34.6 billion, up 34%, with GAAP net income of $4.3 billion. Landed a big deal to supply OpenAI with 6 gigawatts of GPUs. Quality: High. Profitable and growing fast, though far smaller than NVIDIA in AI chips and with lower margins.

Intel (INTC) — The struggling giant. Once dominant in chips, Intel now holds under 1% of the AI accelerator market. It lost money in 2024 and has had a rough stretch, though it returned to a quarterly profit in Q3 2025 (helped by U.S. government funding and a deal where NVIDIA invested $5 billion in Intel stock at $23.28 per share). Quality: Turnaround story / lower quality. Weakest balance sheet and profitability of the big three; more speculative.

Also worth mentioning — TSMC (Taiwan Semiconductor, TSM) — The company that actually manufactures the chips. Almost every advanced AI chip (NVIDIA, AMD, Broadcom, even Apple) is physically made by TSMC. In 2025 it earned about $55 billion in net income on $122 billion in revenue, with gross margins near 60%. Quality: Top-tier. The ultimate "picks and shovels" name — it wins regardless of which chip designer wins. Main risk is its location in Taiwan (geopolitical).

Category 2: AI Memory (HBM)

AI chips need special high-speed memory called HBM (High Bandwidth Memory). Only three companies make it at scale, and it's sold out well into the future. Bank of America (cited by SK hynix) estimates the 2026 HBM market will reach $54.6 billion, a 58% increase over 2025, and calls 2026 "a supercycle similar to the boom of the 1990s."

SK Hynix (Korea; 000660.KS, also trades over-the-counter as HXSCL) — The HBM leader. Per Counterpoint Research, "SK hynix now has 62% of the HBM chip market supply, Samsung's share fell to 17% in Q2 2025, while US-based Micron holds 21%." It is NVIDIA's main HBM supplier. In 2025 it posted record full-year revenue of ₩97.15 trillion (about $68 billion) and record profit of ₩42.95 trillion (about $30 billion), driven by HBM revenue that "more than doubled year-on-year" — overtaking Samsung in annual operating profit for the first time. Quality: High. The clear winner of the memory supercycle.

Micron (MU) — The only U.S.-based maker, with about 21% HBM share. Fiscal 2025 (ended August 2025) revenue was $37.4 billion with $8.5 billion net income; momentum is huge, with fiscal Q1 2026 revenue of $13.6 billion and $5.2 billion net income. It exited consumer memory (shutting its Crucial brand) to focus on AI. Quality: High and improving. Easiest way for U.S. investors to play HBM directly.

Samsung (Korea; 005930.KS, OTC SSNLF) — The biggest memory maker overall but the laggard in HBM (about 17% share) after quality issues qualifying chips with NVIDIA. Quality: Solid but mixed. Huge, diversified, profitable company, but trailing in the highest-value AI memory.

Category 3: Data Center REITs

REITs (Real Estate Investment Trusts) own the actual buildings. They benefit from AI demand and usually pay dividends.

Equinix (EQIX) — The world's interconnection leader with 270+ data centers and 500,000+ connections. Q4 2025 bookings jumped 42% year-over-year; 60% of its largest Q4 deals were AI-driven. Plans to spend $4–5 billion a year through 2029. Quality: High. Profitable, premium franchise, dividend-paying.

Digital Realty (DLR) — Owns the big campuses used to train AI; 300+ data centers. Signed its largest-ever hyperscale lease and grew Q1 2026 revenue 16%. Quality: High. Strong, though carries more hyperscaler concentration.

Iron Mountain (IRM) — Best known for records storage, but its data center business is the growth engine — revenue grew about 30% in 2025 with a 52% profit (EBITDA) margin. Quality: Good, but higher debt. A diversified way in; leverage is on the higher side (about 5x).

Category 4: Data Center Power/Energy

The real bottleneck for AI is electricity. Robert Schein, Chief Investment Officer at Blanke Schein Wealth Management, estimates "$1.4 trillion is needed just for AI Data Center Electrification by 2030." (Separately, an April 2026 PowerLines analysis of 51 U.S. investor-owned utilities found planned capex of "at least $1.4 trillion on capital projects through 2030.")

Vertiv (VRT) — Makes the cooling and power systems inside data centers. 2025 revenue about $10.2 billion (up 28%), net income $1.3 billion (up 169%), with orders up 81%. Quality: High. Strong profits, manageable debt, huge backlog.

Eaton (ETN) — Makes the electrical equipment (switchgear, power management) for data centers. CEO Paulo Ruiz Sternadt said on the Q4 2025 call that the backlog "is also over 200% up, and it equates to eleven years of what was built in 2025," with data center orders up about 200% in Q4. 2025 net profit margin was near 15% with strong free cash flow. Quality: Top-tier. Big, stable, dividend-paying industrial.

Constellation Energy (CEG) — The largest U.S. nuclear operator, signing long-term power deals with Meta (a 20-year deal for the full output of its Clinton plant), Microsoft, and CyrusOne. Bought Calpine to become the largest private-sector power producer in the world. Quality: High. Strong cash flows from long-term contracts, though earnings can swing.

Vistra (VST) — Texas-based power producer (gas, nuclear, renewables) signing data center deals. Q3 2025 revenue about $5 billion with 21% operating margins. Quality: Good but volatile. Strong growth, but a higher valuation and more execution risk.

NRG Energy (NRG) — A power producer that was the S&P 500's biggest gainer in 2025. Partnering with GE Vernova and Kiewit to build over 5 gigawatts of gas-fired power for data centers. Quality: Improving. Turned strongly profitable; more of a momentum/turnaround story.

Also worth mentioning — GE Vernova (GEV) — Makes the gas turbines and grid equipment to power data centers, with multi-gigawatt deals (Chevron, NRG, NextEra) and a fast-growing backlog. Quality: High and improving.

Category 5: Networking

AI data centers connect tens of thousands of chips, which requires massive high-speed networking.

Arista Networks (ANET) — The leader in high-speed Ethernet switches for AI data centers; it has overtaken Cisco in data center switching. Full-year 2025 revenue hit $9 billion (up 29%), and it raised its 2026 AI networking target to $3.25 billion. Quality: Top-tier. Very profitable, no debt, strong growth. Note: about 48% of revenue comes from cloud/AI customers, so it's concentrated.

Broadcom (AVGO) — A powerhouse. It makes custom AI chips (XPUs) for Google, Meta, and others, plus the Ethernet switch chips (Tomahawk/Jericho) that nearly every AI network uses. Q1 FY2026 AI revenue guided to $8.2 billion (doubling year-over-year); total AI order backlog about $73 billion. EBITDA margins around 67%. Quality: Top-tier. Hugely profitable, though it carries more debt from acquisitions (like VMware).

Marvell (MRVL) — Provides custom-chip building blocks (high-speed connections) for AWS and Microsoft's in-house chips. Data center grew to about 75% of revenue. Quality: Good. Growing fast on AI, but more dependent on a few customers.

Category 6: ETFs (Diversified Baskets)

ETFs let beginners own many of these names at once. These focus on hardware/infrastructure, not AI software:

  • SMH — VanEck Semiconductor ETF. Expense ratio 0.35%. Owns the 25 biggest chip stocks — NVIDIA, TSMC, Broadcom, AMD, Micron. The most popular semiconductor fund (roughly $47 billion in assets). Very concentrated in NVIDIA.
  • SOXX — iShares Semiconductor ETF. Expense ratio 0.35%. Similar idea, broader (about 30+ holdings), with names like Micron, AMD, Marvell, Broadcom, Intel.
  • DTCR — Global X Data Center & Digital Infrastructure ETF. Expense ratio 0.50%. Owns data center REITs and digital infrastructure — Equinix, Digital Realty, American Tower, Crown Castle.
  • GRID — First Trust NASDAQ Clean Edge Smart Grid Infrastructure ETF. Expense ratio about 0.56%. Targets the electric grid/power-equipment theme — Eaton, GE Vernova, Quanta, Schneider Electric, ABB.
  • AIPO — Defiance AI & Power Infrastructure ETF. Expense ratio 0.69%. A newer (launched July 2025), dedicated "AI + power" fund holding Quanta, Vertiv, Broadcom, NVIDIA, AMD. Smaller and less established.
  • NUKZ — Range Nuclear Renaissance ETF. Expense ratio 0.85%. For the nuclear-power-for-AI angle — Constellation, Talen, Cameco, GE Vernova.

(Tip for the newsletter: ETF assets and top holdings shift over time, so it's worth a quick same-day check on each issuer's official fund page before publishing exact figures.)

Recommendations

For a beginner-friendly newsletter, I'd structure picks in three tiers:

  1. Core, highest-quality holds (own these first): NVIDIA, Broadcom, TSMC (chips); Arista, Eaton, Vertiv (networking/power); Equinix or Digital Realty (REITs). These are all very profitable with strong balance sheets and clear AI tailwinds.
  2. Diversify with ETFs to lower single-stock risk: SMH or SOXX for chips, DTCR for data centers, and GRID for the power/grid theme. This is the simplest "set it and forget it" approach for beginners who don't want to pick individual winners — and it directly captures the "picks and shovels" logic.
  3. Higher-risk / situational: Micron and SK Hynix (memory is cyclical — prices boom and bust); Vistra and NRG (power producers with great momentum but higher valuations); Intel (turnaround bet); AIPO and NUKZ (newer, pricier, more concentrated ETFs).

Benchmarks that would change the thesis: Watch hyperscaler capex guidance each quarter — if Microsoft, Amazon, Alphabet, and Meta start cutting capex, the whole "picks and shovels" trade weakens. The 77% jump to $725 billion for 2026 is the green light; a flat or falling number would be the warning sign. Also watch memory prices (a sign of the cycle turning) and any shift in company language from "supply-constrained" to "demand-constrained."

Caveats

  • Valuations are high. Many of these stocks have run up a lot. Strong companies can still be poor investments if you overpay; consider buying gradually over time.
  • Concentration risk. Much of this demand comes from just four or five customers. If they slow spending, the whole chain feels it. Some sellers are themselves concentrated (e.g., Arista at ~48% cloud/AI revenue; SK Hynix supplies a huge share of HBM to NVIDIA).
  • Cyclicality. Memory (Micron, SK Hynix, Samsung) and power markets are historically boom-and-bust.
  • Some figures are projections. The 2026 capex numbers ($725 billion) and the $5.3 trillion 2025–2030 figure are Goldman Sachs estimates based on company guidance, not final results, and "supply-constrained" claims come from the companies themselves. The HBM market size and electrification spending are analyst forecasts.
  • Geopolitics. TSMC, SK Hynix, and Samsung are based in Asia; tensions involving Taiwan or trade restrictions are real risks.
  • This is educational information, not personalized investment advice.
Content is user-generated and unverified.
    AI Infrastructure Stocks: The "Picks & Shovels" Investment Guide | Claude