Sector Report
The AI Compute Buildout: Who Wins the Next $1 Trillion
May 18, 2026 · 22 min read · Winvestor Analyst Team
The next phase of the AI compute buildout will require more than a trillion dollars of cumulative capex across chips, networking, power, and real estate. We map the value chain and identify the businesses with durable economics rather than fleeting demand spikes.
Why the spend is structural, not cyclical
Hyperscaler capital expenditure is no longer a discretionary investment cycle. Training frontier models has become a strategic asset for Microsoft, Alphabet, Meta, and Amazon, and inference workloads are scaling with every new consumer and enterprise deployment. Combined 2026 capex guidance from the top four U.S. hyperscalers exceeds $400 billion — roughly 3x the level seen five years ago. We expect cumulative AI-related capex to surpass $1.2 trillion by the end of 2028.
The four layers of value capture
We segment the buildout into four layers: (1) advanced logic and HBM memory, dominated by TSMC, Nvidia, and SK Hynix; (2) networking and optical interconnect, where the supply of high-end switches and transceivers remains structurally tight; (3) power and cooling, including gas turbines, transformers, and liquid cooling; and (4) data center real estate. The economics of each layer are very different — semis enjoy 60%+ gross margins while colo REITs operate closer to 20% — but each is supply-constrained for the next 24 months.
The chip layer: durable oligopoly economics
Advanced logic at 3nm and below is effectively a three-player market — TSMC, Samsung, and Intel Foundry — and only TSMC currently produces leading-edge dies at scale and yield. HBM is even more concentrated, with SK Hynix, Samsung, and Micron splitting more than 95% of supply. We expect HBM bit demand to grow 60%+ annually through 2027, well above stated capacity additions. Pricing has been firm for six consecutive quarters and contract structures have lengthened to 12-18 months — a meaningful regime change from the historical spot-driven cycle.
Networking: the second derivative the market still misses
Every doubling of accelerator count in a training cluster requires a more-than-doubling of optical interconnect bandwidth. 800G transceivers are sold out through mid-2027 and 1.6T qualification is pulling forward. The vendor list is short, and the optical components inside — DSP chips, EMLs, silicon photonics — sit with an even narrower set of suppliers. Gross margins in transceivers have expanded 600 bps over two years and we see room for another 200-300 bps as mix shifts to 1.6T.
Where we see the best risk-adjusted exposure
Rather than chase the obvious names at peak multiples, we focus on second-derivative beneficiaries. Power equipment OEMs trade at mid-teens earnings multiples despite multi-year backlogs. Specialty chemicals and substrates suppliers have pricing power that the market continues to underestimate. Two of our active coverage names sit in this category and screen attractively on free cash flow yield even after a 40% rally.
Power and cooling: the gating constraint
By 2027, U.S. data center power demand will exceed the entire current electricity consumption of Italy. Gas turbine OEMs have order books extending into 2029. Large-power transformer lead times have tripled. Liquid cooling, until recently a niche, is now specified in roughly 70% of new hyperscale builds. The companies that touch these workflows enjoy the rare combination of secular volume growth, pricing power, and limited competitive entry given the capital intensity and engineering complexity.
Sovereign AI and the next leg of demand
A new buyer cohort has emerged: governments and national champions in the Gulf, India, Japan, France, and Korea are committing to multi-billion-dollar sovereign AI infrastructure. Unlike hyperscaler capex, sovereign spending is policy-driven and far less price-sensitive. We estimate sovereign AI capex could add $150-200 billion of cumulative spend through 2028 on top of our hyperscaler base case. This is not yet reflected in consensus estimates for most equipment suppliers.
Key risks to the thesis
A breakthrough in model efficiency that materially reduces training compute would compress the timeline. A power grid bottleneck — already visible in northern Virginia and parts of Texas — could delay revenue recognition for builders. And concentration risk is real: roughly 70% of merchant AI capex flows through three hyperscalers, so any single budget cut sends shockwaves through the supply chain. We monitor hyperscaler capex commentary quarterly and would trim aggressively on any signs of a coordinated pause.