AI Infrastructure Boom
◆ What's Happening
The largest infrastructure buildout in human history is accelerating. In Q4 2025 earnings, the five largest hyperscalers collectively committed to $660-690 billion in capital expenditure for 2026 — Amazon at $200B, Alphabet at $175-185B (more than doubling from $75B guidance just one quarter earlier), Meta at $115-135B, Microsoft at $120B+, and Oracle at $50B. Add the $500B Stargate project (OpenAI/SoftBank/Oracle joint venture, now with 5 sites under construction and 7 GW of planned capacity), and the total AI infrastructure investment pipeline approaches $1.2 trillion through 2029.
What's notable is the acceleration, not just the magnitude. Every hyperscaler reports being supply-constrained, not demand-constrained — GPU shortages, power bottlenecks, and construction timelines are the binding constraints, not customer demand. Alphabet's cloud backlog surged 55% sequentially to $240B, and Microsoft disclosed an $80B Azure backlog unfulfillable due to power constraints alone.
DeepSeek's efficiency breakthrough, rather than slowing the buildout, triggered a Jevons Paradox: as inference costs fell, demand exploded. Hyperscalers are pivoting infrastructure from training-focused to inference-optimized facilities while maintaining or increasing total spend. The supply chain bottleneck is shifting from chips to power and cooling — the key question is no longer whether this spending happens, but whether physical infrastructure (power, cooling, fiber) can be built fast enough to absorb it.
📈 Bull Case
AI compute demand exceeds supply through 2027. With $690B in hyperscaler capex and $500B in Stargate investment, the physical infrastructure buildout dwarfs any prior technology cycle. Revenue acceleration cascades across the entire supply chain — from AVGO's custom ASICs and CRDO's active electrical cables, to NVT's liquid cooling systems, VST's nuclear power plants, and CSCO/ANET's AI networking ($2.1B in CSCO AI orders in Q2 alone). The Jevons Paradox is playing out in real time: efficiency gains drive demand expansion, not spending cuts. Companies building the physical infrastructure of AI are where the most durable revenue growth lives.
📉 Bear Case
A second DeepSeek-style efficiency breakthrough genuinely reduces hardware requirements by 10x, breaking the Jevons Paradox and causing hyperscalers to slash spending commitments mid-year. Overcapacity emerges in 2027 as the near-doubling of capex from 2025 to 2026 proves to be a cyclical peak, crushing utilization rates and margins for infrastructure providers. Meanwhile, tariff disruptions on semiconductor equipment and rising power costs make some projects uneconomical.
◆ Supply Chain Coverage
59 stocks across 15 categories. Tap a category to expand.
📋 Also Impacted — scored for this event but uncategorized
On-device AI chips for mobile/edge. Not data center infra.
Solar can power data centers but DC operators prefer baseload (nuclear/gas). Utility-scale solar is supplemental.
Industrial automation and control systems used in data center cooling and power management. Aspen Technology acquisition adds AI/software.
AI infrastructure deal advisory (IPOs, M&A). Goldman advising on AI company transactions.
Builds electrical infrastructure for data center campuses. Indirect AI beneficiary through power delivery.
Renewable energy and battery storage for data center campuses. FPL utility serves Florida DC market.
Small modular reactors could power future data centers. Long-term optionality, not near-term revenue.
Advanced fission could power future AI data centers. Sam Altman connection to AI + nuclear is the narrative.
JPM invests heavily in AI/tech ($17B+/yr) but is a consumer, not an AI infrastructure provider.
Rare earth magnets in some DC cooling/power systems.
Utility serving Southeast US where data center demand is growing. Vogtle nuclear provides clean baseload.
FSD and Dojo AI compute. Not an infra supplier.
🛡️ At Risk — negative exposure to this event
◆ Why Some Stocks Are USER — Research Methodology▸ expand
The Framework: Supply Chain First, Entry Quality Second
Core picks are selected by mapping the complete AI data center supply chain — every layer from silicon to software — identifying the dominant or most differentiated player in each, and only then filtering by entry quality and valuation. This produces picks with both structural importance and investable timing.
Stocks marked USER were requested by you but don't pass this framework as core recommendations. They're still tracked because they're legitimate AI infrastructure companies — but there's a specific reason each one didn't make the cut.
✅ 5 Stocks Promoted to Core — My Misses
These were originally user-added but pass the framework. Honest accounting of why I missed them:
TSMC is the single most critical company in the AI supply chain — it manufactures every leading-edge chip. I excluded it because it was only 5% below ATH, which meant poor entry quality. But you can't analyze AI infrastructure without the company that physically makes every chip. The framework now says: map the monopoly first, flag entry timing second.
Super Micro is the #1 AI server assembler and the deepest value play (71% below ATH, 14x PE). I avoided it because of the auditor resignation, delayed SEC filings, and short seller reports. The framework says: if a company is the dominant player in its layer, include it and flag the risks — don't exclude it entirely.
Arista is the dominant Ethernet switching vendor for AI clusters. I skipped it for the same ATH-proximity reason as TSM. With the Feb selloff pushing it 24% below ATH, the entry has actually improved — which proves the original exclusion was wrong.
Arm is the CPU architecture monopoly with 96% gross margins. I excluded it because 75x PE seemed unjustifiable. But it's a royalty-collecting IP monopoly — every chip shipped pays Arm a toll. The framework catches irreplaceable IP companies regardless of PE multiples.
Nebius is CoreWeave's most direct competitor. I had CRWV but didn't know about Nebius (ex-Yandex restructure was recent). The framework would have caught this: once you map the gpu-cloud layer, you look for all players, not just the first one you find.
📋 9 Stocks That Remain USER — Why They Don't Make the Cut
These are real AI companies, but each has a specific disqualifier under the framework:
Layer saturated + poor entry. Power layer already has AEIS, VST, CEG. VRT is 7% below ATH at 30xx PE. Great company, terrible timing. Wait for a 20%+ pullback.
Layer saturated + poor entry. Same as VRT — power layer is covered. ETN is 4% below ATH at 25xx PE. Only ~25% DC revenue makes it a diluted play.
Unprofitable + speculative. Interesting fuel cell thesis, but negative earnings and high volatility. VST and CEG provide profitable power exposure with less risk.
Poor entry + layer covered. Connectivity layer already covered by CRDO (75-88% AEC share) and ALAB (86% retimer share). GLW is 4% below ATH at 35xx PE — no edge vs the pure-plays already in the list.
Losing to the layer leader. AI score 58. Arista is eating Cisco's lunch in high-end DC switching. The 17xx PE looks cheap because growth is low single digits.
Layer covered + poor entry. Connectivity layer has CRDO and ALAB as pure-play picks. APH is 14% below ATH at 28xx PE with only ~25% DC revenue. Reliable compounder, but not differentiated enough.
Indirect AI exposure. AI score 45 — analog chips are in DC power supplies but that's a stretch as a real AI play. Better positioned for China trade event (score 65) than AI capex.
Commodity storage. AI score 55. NAND flash is a commodity business. Memory layer already has MU (HBM, the real AI memory play). Currently 13% below ATH at 8xx PE.
Commodity storage. AI score 40 — lowest of all AI stocks. HDDs losing share to SSDs long-term. At 21xx PE, 8% below ATH. Memory layer covered by MU which has the real AI play (HBM).
🎯 The Lesson
The original research optimized for room-to-run — stocks with big discounts from ATH and asymmetric upside. This is a valid filter, but applying it before mapping the supply chain caused blind spots: the most important companies (TSM, ANET) were excluded because they looked expensive. The fix is simple: map first, filter second. You need to know what the full playing field looks like before deciding where to place bets.
◆ Catalyst Calendar
💡 Cross-Event Note
If you're also watching the China Trade War, note that AI infrastructure stocks with minimal China exposure (AVGO, CRDO, ALAB) are better positioned than those with significant China revenue (MU, QCOM). The China-AI intersection is where cross-event risk is highest — if export restrictions tighten while AI capex booms, domestic-focused names win biggest.