Picks, Shovels, and Denim: The Real Winners of the AI Infrastructure Build-Out
Picks, Shovels, and Denim: The Real Winners of the AI Infrastructure Build-Out
During the 1849 California Gold Rush, the people who reliably got rich weren't the prospectors panning for gold -- they were the merchants selling durable canvas trousers to the miners. Levi Strauss built a lasting business not by finding gold, but by solving a mundane, unglamorous problem every miner shared. Three years into the current AI boom, the same pattern is repeating in plain sight. The foundation-model companies burning cash to train ever-larger systems remain, for the most part, unprofitable. Meanwhile, the businesses solving AI's physical bottlenecks -- power, cooling, and compute financing -- are already generating tens of billions in backlog and revenue. This piece walks through four data-verified niches in that "picks and shovels" layer, their underlying success factors, and where founders and allocators without hyperscaler-sized checkbooks can realistically compete.
- Only 5GW of ~12-16GW -- the share of 2026-targeted US data center capacity that had actually broken ground as of April 2026, per Bloomberg reporting relayed by domestic Korean coverage
- 20-31% CAGR -- the range of forecast annual growth for the global data center liquid cooling market through the early 2030s, cross-verified across Grand View Research and MarketsandMarkets
- ~$23-25B -- 2025 neocloud (GPU-cloud rental) industry revenue, nearly tripling year over year, with Synergy Research projecting roughly $180B by 2030
| Niche | Core Bottleneck | Cross-Verified Market Signal |
|---|---|---|
| Power & Transformers | 3-4 year interconnection queues | Hyosung Heavy Industries backlog exceeds ~$11B equiv. (Jun 2026) |
| Liquid Cooling | 30-50kW racks exceed air-cooling limits | $6.65B (2025) → $29.46B (2033e), GVR |
| Neocloud / GPU Brokers | Capital intensity of owning GPU fleets | $23-25B revenue (2025) → ~$180B (2030e) |
| Skilled Labor & Assurance | Manual coil-winding can't be automated | Early-stage; directional estimate only |
1. Why the Bottleneck, Not the Model, Is the Business Opportunity
Map the AI value chain onto the gold-rush metaphor and the layers become clear. The "gold" is the AI model itself -- OpenAI, Anthropic, and other foundation-model labs. The "pickaxes" are GPUs, dominated by Nvidia. Both layers are already crowded and well understood by public markets. Two layers underneath receive far less attention despite being structurally non-optional: the physical constraints of power and cooling, and the capital markets infrastructure needed to finance it all. As of April 2026, reporting indicated that of roughly 12-16 gigawatts of US data center capacity targeted for 2026, only about 5 gigawatts had actually broken ground. Capital and intent to build are abundant; the transformers, transmission capacity, and cooling systems required to actually operate the facilities are the scarce resource. That scarcity is where the denim-selling opportunity lives.
2. Niche One -- Power & Transformers: The Least Glamorous, Most Durable Bottleneck
A single hyperscale AI data center now draws over 100MW, versus the 10-25MW a conventional facility historically consumed. The constraint isn't generation capacity in the abstract -- it's grid interconnection. In PJM Interconnection's territory, the largest US wholesale power market, the queue for new data center grid connections averages three to four years. The structural root cause: more than 70% of large power transformers in the US have already exceeded their 25-30 year design life, meaning replacement demand and new-build demand are colliding simultaneously.
Korean manufacturers are turning this bottleneck into hard numbers. Hyosung Heavy Industries operates the only US-based facility -- in Memphis, Tennessee -- capable of producing 765kV-class ultra-high-voltage transformers, the grade required for long-distance transmission and large-load interconnections. As of June 2026, the company's power-equipment order backlog, concentrated in ultra-high-voltage transformers, was reported to exceed KRW 15 trillion. In the same period, HD Hyundai Electric and LS ELECTRIC both raised their new-order guidance -- a signal, per Korean equity research, that this trend is being driven by actual backlog and capacity expansion rather than AI-keyword sentiment alone.
Success Factors
- The entry barrier is the moat. Ultra-high-voltage transformers require rigorous certification and manual coil winding by skilled labor -- capacity cannot be added quickly, which concentrates pricing power among a small number of qualified suppliers.
- Demand is a physics constraint, not a fashion cycle. Regardless of which AI model or lab wins, GPUs don't run without electricity -- this demand persists independent of any single AI company's fortunes.
3. Niche Two -- Liquid Cooling: Air Conditioning Has Hit a Physical Wall
Racks carrying the latest AI accelerators now dissipate 30-50kW of heat -- beyond what traditional air cooling can manage. Direct-to-chip and immersion cooling have consequently shifted from a nice-to-have to a core architectural requirement. Grand View Research puts the market at roughly $6.65B in 2025, growing to $29.46B by 2033 at a 20.1% CAGR. MarketsandMarkets, using a different methodology, estimates $4.07B in 2026 growing to $27.65B by 2033 at a 31.5% CAGR. The point-estimates diverge meaningfully by source -- readers should treat the underlying figure as a range rather than a precise number -- but every methodology we reviewed agrees on the direction: roughly a fourfold expansion over the period.
Success Factors
- Retrofit demand compounds new-build demand. Existing facilities converting from air to liquid cooling create a second demand stream layered on top of new construction.
- The opportunity fragments below the system level. Coolant fluids alone are forecast to grow from $0.28B (2025) to $2.01B (2032) at a 32.7% CAGR -- a sub-niche within the niche that a smaller, specialized company could credibly own.
4. Niche Three -- Neoclouds: Renting Out GPUs Without Building an AI Lab
The most conceptually interesting model here belongs to "neoclouds" -- CoreWeave, Lambda, Nebius, and Crusoe -- companies that build no foundation models of their own. Instead, they acquire GPU fleets at scale and rent compute to hyperscalers and AI labs, effectively functioning as a hidden subcontracting layer beneath Microsoft, Meta, and OpenAI. CoreWeave's 2025 revenue reached roughly $5.1B, up from $1.9B in 2024, with contracted backlog growing from $66.8B at year-end 2025 to $99.4B by March 2026. Industry-wide neocloud revenue exceeded $23B in 2025 -- more than triple the prior year -- and Synergy Research projects roughly $180B by 2030.
Success Factors
- Speed is the competitive edge. Building a data center from scratch takes 18-30 months; a neocloud can lease existing space and rack servers far faster.
- Take-or-pay contracts lock in cash flow. Customers commit to fixed GPU capacity for 2-5 years and pay regardless of utilization, giving these businesses unusually clear revenue visibility for infrastructure-stage companies.
The risks are equally clear, however. Microsoft alone accounted for 62% of CoreWeave's 2024 revenue and 67% in 2025 -- a customer-concentration risk investors should weigh carefully. Annual interest expense of roughly $1.2B is fixed regardless of utilization, and because financing is frequently structured against GPUs as collateral, a decline in GPU resale value directly threatens the collateral base. Nvidia's parallel move to establish a $500B+ compute-financing platform with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR -- treating GPUs as a financeable infrastructure asset akin to toll roads -- represents the capital-markets layer underneath this entire niche, and is a thread we examine in more depth in our companion piece on Fed policy and AI-backed credit, linked below.
| Company | Profile | Scale Indicator |
|---|---|---|
| CoreWeave | Public; largest independent; high Microsoft concentration | ~$5.1B 2025 revenue (from $1.9B) |
| Nebius | Public; major Microsoft and Meta contracts | $3.0B AI cloud ARR |
| Lambda | Private; heading toward IPO | Raised $1.5B+ in late 2025 |
| Crusoe | Built OpenAI's Stargate Abilene campus | Valued at roughly $10B |
5. An Emerging Model Worth Watching -- Skilled Labor, Insurance & Maintenance
One layer hasn't yet been consolidated by large public companies: skilled transmission and distribution labor (coil winding remains difficult to automate), specialized infrastructure insurance products for data centers, and maintenance services for aging grid equipment. Reliable market-size data for this layer isn't yet available at the depth we hold other figures in this piece to, so we present it directionally rather than with a specific forecast -- readers should treat this as the author's estimate rather than a sourced figure. Structurally, it scales with -- and lags -- growth in the power and cooling niches above it, which makes it a plausible entry point for smaller operators who cannot compete on the capital intensity of transformers or GPU fleets.
6. Market Outlook: Two Scenarios
AI data center investment continues through 2026-2030, gradually but not fully resolving the power and cooling bottleneck. Under this path, power-equipment and liquid-cooling order backlogs keep building, and neoclouds continue expanding their role as hyperscalers' hidden subcontracting layer. Multiple independent forecasters -- the EIA and Gartner among them -- converge on continued power-demand growth through 2030, which supports this reading.
If actual AI training and inference demand fails to keep pace with the current build-out rate -- echoing the "circular financing" concern where contracted revenue may overstate genuine end demand -- neocloud take-or-pay contracts could go unrenewed, pressuring GPU collateral values sharply. Separately, if governments constrain data center power and water use more aggressively (as Ireland and the Netherlands have already done), the pace of new infrastructure investment itself could slow.
7. Basic Investment Principles for This Special Report
This section is deliberately general and not a recommendation to fund, invest in, or found any specific company.
- Diversify across the value chain layer, not within a single niche. Spreading exposure across power, cooling, and compute-brokerage reduces single-bottleneck risk relative to concentrating in one.
- Favor niches with genuine entry barriers. Certification, skilled labor, and capital requirements (as with transformers) tend to produce more structurally durable competitive positions than software-only plays in the same space.
- Scrutinize contract structure, not just headline revenue. A neocloud's backlog looks impressive until customer concentration, debt load, and renewal risk are examined together.
- For founders, target the niche within the niche. Maintenance, specialized parts, and skilled-labor staffing around power and cooling infrastructure are more realistic entry points for smaller operators than competing directly with transformer manufacturers or GPU-fleet operators.
Related Reading
For the capital-markets layer underneath the neocloud financing model discussed above, see Jackson Hole Meets the Circular Financing Question: What Warsh's First Keynote Means for AI-Backed Credit.
For more on the power-versus-AI-demand tension underlying the transformer story, see AI Boom vs Energy Crisis: The Battle Shaping Global Markets in H2 2026, and for the memory-chip side of the same hyperscaler capex cycle, see SK Hynix Hits All-Time High: Is the HBM Super-Cycle Just Getting Started?
This article is published for informational and analytical purposes only. Nothing contained herein constitutes financial, investment, legal, or tax advice. All investment decisions involve risk, including the possible loss of principal. Past performance is not indicative of future results. Readers should conduct their own research and consult a qualified financial advisor before making any investment decisions. The views expressed are solely those of the author based on publicly available information and do not represent the official position of SEJONG DnC.
Sources
1. G-enews, "AI Data Center 'Infrastructure Shock'... 2026 Volume Halved," Apr 12, 2026 (Bloomberg report cited)
2. Electimes, "(2026 New Year Special) The 'AI Power War' Has Begun," Jan 15, 2026
3. Etnews, "Global Data Center Power Consumption Up 26% in One Year Amid AI Expansion" (citing Gartner), Jun 12, 2026
4. NewDaily, "US Transformer Lead Times Reach 143 Weeks... Hyosung Heavy Industries Expands Orders," Jan 2, 2026
5. CBC News, "Hyosung Heavy Industries: Ultra-High-Voltage Transformer Backlog Reaches KRW 15 Trillion," Jun 1, 2026
6. finance.thesmileinfo.com, "2026 Power Equipment Stocks TOP4 Analysis," Aug 2026
7. Grand View Research, "Data Center Liquid Cooling Market Report," 2026
8. MarketsandMarkets, "Data Center Liquid Cooling Market Report 2026-2033"
9. ABI Research, "The State of Neocloud: Four Trends for 2026," Dec 2025
10. ModulEdge, "Neocloud: How the GPU Cloud Business Works," Jun 2026
11. Signisys, "GPU Cloud Providers: The $20B Neocloud Era," Apr 2026
12. American Compute, "Neocloud Business Model and Unit Economics," Mar 2026
Comments
Post a Comment