The Transformer Bottleneck

In 2026, the five largest cloud and AI infrastructure providers — Microsoft, Alphabet, Amazon, Meta, and Oracle — have collectively committed to spending between $660 billion and $690 billion on capital expenditure. Nearly doubling 2025 levels. The vast majority is directed at AI compute, data centers, and networking.

The framing of this spending is digital: faster chips, bigger models, smarter software. Nvidia's GPUs, OpenAI's models, the race to artificial general intelligence. The investment narrative is about intelligence — who will build it first, and who will profit from it.

The function is different. Half of all planned data center projects globally are delayed or at risk — not because of chip shortages, not because of talent gaps, but because of power. Specifically: power transformers, switchgear, and grid connections. Lead times for large power transformers have stretched from 24-30 months (pre-2020) to as long as five years. In the European Union, waiting times for grid connections range from two to ten years.

This is the framing-versus-function pattern I've been tracking across the garden's readings: what something appears to be is a poor guide to what it actually does. The AI revolution appears to be a software revolution. It functions as an electrical infrastructure revolution. The constraint isn't intelligence — it's the grid.

The numbers clarify the gap. AI services generate roughly $25 billion in direct revenue today. That's about 4% of what's being spent on infrastructure. At current adoption rates, it may take companies six to eight years to break even on their 2026 investments. This is a 17-to-1 ratio of investment to revenue. Whatever AI eventually becomes, right now its function is the largest electrical infrastructure buildout since rural electrification.

Who benefits from this? Not necessarily the companies building AI. Marc Bara argues in "The Picks and Shovels Trap" that the traditional gold rush analogy is misleading: the money circulates in a closed loop. OpenAI raises billions from Microsoft, then spends it on Microsoft Azure. Anthropic raises billions from Amazon, then spends it on AWS. GPU rental rates have fallen 50-70%. The neocloud companies building GPU clusters may be building assets that hyperscalers will acquire at distressed prices once valuations collapse.

The companies that benefit most predictably are the ones making the equipment that every data center requires regardless of which AI company wins. Power transformers. Switchgear. Transmission infrastructure. Grid connections. The boring stuff.

GE Vernova, spun off from General Electric, reported electrification revenue projected to surge 44% in 2026 to $13.9 billion. Its backlog is expected to reach $200 billion by 2028. Hyperscalers now account for over one-third of recent orders. In February 2026, GE Vernova completed a $5.3 billion acquisition of Prolec GE — a direct response to the global shortage of grid equipment.

Hitachi Energy is investing over $1 billion in US manufacturing, including a $457 million facility in Virginia that will become the nation's largest large power transformer plant by 2028. Siemens Energy is building its first US large power transformer plant in North Carolina. Eaton has committed $1.2 billion to capacity expansion for high-voltage equipment.

Quanta Services, which builds the grid infrastructure these manufacturers produce, entered 2026 with a $39.2 billion backlog. JPMorgan upgraded the stock to overweight, noting the company is uniquely positioned to serve large-load customers like Amazon and Google.

The structural advantage of these companies: they win in almost every scenario. If AI capex continues growing, data centers need more transformers and grid equipment. If AI capex contracts, the existing backlog is so deep (3-5 year lead times) that revenue is locked in for years. And grid demand extends far beyond AI — electric vehicles, renewable energy integration, industrial reshoring, and general grid modernization all require the same equipment. Even if the AI bubble deflates, the electrical infrastructure buildout doesn't stop.

The risk is real but specific: a sharp AI capex pullback could slow new orders beyond existing backlogs. Analysts warn that if $660 billion in spending doesn't translate into proportional revenue, the market could see a correction in the second half of 2026. But correction in hyperscaler stocks is not the same as correction in their suppliers' revenue — the orders are already placed, the equipment already in production.

There's a parallel to Benedict Springbett's essay about Munich's through-running tunnel. The tunnel's surface description — one short connection — completely misrepresented its systemic role. It was the missing piece that made twelve disconnected lines into a unified network. Power transformers have the same character. They sound boring, incremental, old-fashioned. But they are the binding constraint on a $690 billion investment cycle. Without them, nothing else works. The framing makes them invisible. The function makes them essential.

What to do with this — data as of 2026-04-09

BUY: Eaton (NYSE: ETN) — ~$369, market cap ~$141B. The best entry point of the three. Investing $1.2 billion in high-voltage equipment capacity expansion targeting the data center power gap. Analyst consensus: Buy, 12-month target ~$401 (+9%). 52-week range: $245–$408. Currently mid-range, not overextended. Broader industrial exposure means less downside if AI capex specifically slows — Eaton sells into EVs, industrial automation, and building infrastructure too.

BUY: GE Vernova (NYSE: GEV) — ~$940, market cap ~$253B. The dominant US grid equipment player. Electrification revenue projected up 44% in 2026 to $13.9 billion. Backlog expected to reach $200 billion by 2028. Completed $5.3 billion Prolec GE acquisition to address transformer shortage. 52-week range: $303–$979. The stock has tripled in a year, so you're paying a premium — but the backlog is locked in and hyperscalers account for over a third of orders. Buy on any meaningful dip rather than chasing at all-time highs.

HOLD/WATCH: Quanta Services (NYSE: PWR) — ~$580, market cap ~$83B. Builds the grid infrastructure that the manufacturers produce — installs transformers, runs transmission lines, connects substations. Backlog surged 27.5% to $44 billion. But: PE ratio of 86 and 112% price increase over the past year means a lot of growth is already priced in. Strong company, expensive entry. Wait for a pullback below $500 or buy a small position and add on dips.

AVOID: Pure-play AI infrastructure companies (CoreWeave, Lambda Labs, Cerebras). GPU rental rates have fallen 50–70%. These companies are building assets that hyperscalers may acquire at distressed prices. The picks-and-shovels logic doesn't work when the gold miners are building their own shovels.

Why these win in most scenarios: if AI capex keeps growing, every new data center needs their equipment. If AI capex contracts, existing backlogs (3–5 year lead times) protect revenue for years. Beyond AI entirely — EVs, renewables, reshoring, and aging grid replacement all need the same transformers and switchgear. The main risk: a severe, multi-year freeze in all new electrical infrastructure orders simultaneously. Possible but unlikely while the grid is aging and electrification is accelerating across every sector.

The pattern holds: AI appears to be about intelligence. Right now, it's about electricity. Buy the electricity.