What is happening with AI-infrastructure capacity?
The constraint has not disappeared—it has migrated from “not enough GPUs” to “not enough complete, energized AI systems.” Hyperscalers can increasingly secure accelerators, but they still need HBM, advanced packaging, networking, power-conversion equipment, liquid cooling, transformers, grid connections and a completed data-center shell. The marginal bottleneck now varies by project, but power availability and memory are becoming more important than the GPU alone.
1. Memory is arguably the tightest semiconductor constraint
HBM consumes substantially more wafer capacity and cleanroom space than conventional memory, while each new accelerator generation requires more HBM per system. Micron says DRAM and NAND demand is materially exceeding supply and expects tight conditions to persist beyond calendar 2027. It also argues that HBM growth is displacing conventional DRAM capacity, while new fabs face long construction timelines, labor shortages, permitting complexity and energy-infrastructure requirements.
This is why memory pricing has become a much larger component of hyperscaler capex inflation. The constraint is no longer limited to HBM: allocating more cleanroom space to HBM can tighten server DRAM, enterprise SSD and even consumer memory availability. In investment terms, this supports unusually strong pricing and contract visibility for memory manufacturers, but also creates a risk that customers begin redesigning systems around greater memory efficiency.
2. Leading-edge foundry capacity remains tight, with relief weighted to 2027–28
TSMC is increasing N3 capacity specifically because of AI demand, including HBM base dies and HPC processors. However, its incremental capacity is arriving gradually: additional Taiwan production is scheduled for the first half of 2027, Arizona for the second half of 2027 and Japan in 2028. TSMC explicitly described capacity as tight while converting N5 tools and optimizing capacity between nodes.
Therefore, CoWoS and advanced packaging may be less catastrophically constrained than in the earliest Blackwell ramp, but the broader advanced-manufacturing chain remains heavily allocated. Every new custom accelerator from Google, Amazon, Meta, OpenAI or another hyperscaler competes for broadly the same foundry, packaging, substrate and HBM ecosystem.
3. Power and grid access are becoming the longest-duration bottleneck
AI-focused data-center electricity consumption increased approximately 50% in 2025. The IEA estimates that overall data-center electricity demand will roughly double from 485 TWh in 2025 to 950 TWh in 2030, while AI-focused consumption triples. More importantly, AI-server power density increased elevenfold between 2020 and 2025 and could rise another fourfold by 2027. That stresses transformers, switchgear, power electronics, backup generation, storage and cooling—not merely electricity generation.
This changes the industry from a chip-procurement problem into a utility-scale infrastructure problem. Having land and GPUs is insufficient if the site lacks an approved grid connection, substations, transformers, cooling water or on-site generation. Hence the increased interest in natural gas, nuclear, batteries, microgrids and dedicated power arrangements.
4. Cooling and electrical infrastructure suppliers are monetizing the constraint
Vertiv’s Q226 sales increased 24%, and the company raised its full-year outlook while acknowledging temporary supply-chain congestion and greater project complexity. It is expanding chiller and liquid-cooling capacity because conventional air-cooled facilities cannot efficiently support the rack densities required by next-generation clusters.
This is important for the equity debate: the AI-infrastructure opportunity is spreading beyond NVIDIA into power distribution, liquid cooling, heat rejection, electrical equipment and integrated data-center systems. The constraint is beneficial for suppliers’ backlog, pricing and revenue visibility, although valuations increasingly assume several years of near-perfect execution.
5. Networking is scaling almost as quickly as compute
Large clusters require scale-up networking inside the rack and scale-out networking across thousands of accelerators. Broadcom reported Q226 AI semiconductor revenue of $10.8bn, up 143% yoy, driven by custom accelerators and AI networking, and guided Q3 AI revenue to $16bn—more than 200% growth.
The architectural shift is toward a more diversified accelerator market—NVIDIA GPUs alongside Google TPUs, Amazon Trainium, Meta ASICs and other custom silicon—but almost every architecture requires more switching, optical connectivity and memory. Consequently, custom ASICs do not necessarily reduce overall infrastructure spending; they redistribute spending from merchant GPUs toward foundry, memory and networking suppliers.
6. Hyperscalers are still capacity-constrained despite enormous capex
Alphabet spent $44.9bn in Q226 capex, with approximately 60% directed to servers and 40% to data centers and networking. It raised 2026 capex guidance to $195–205bn and is temporarily using third-party capacity because internal supply cannot be delivered quickly enough—accepting some near-term margin pressure to capture demand.
Microsoft has similarly said customer demand exceeds available capacity and expected constraints to persist through at least 2026. Its latest quarter still showed strong cloud growth, with Microsoft Cloud revenue reaching $59.3bn, demonstrating that the constraint is suppressing potential revenue rather than indicating weakening demand.
Investment conclusion
The current narrative is not “AI infrastructure is becoming unconstrained.” It is:
Compute supply is improving, but the bottleneck is broadening into memory, networking, power, cooling and site delivery.
That remains supportive for TSMC, Micron, Broadcom, Vertiv and the broader electrical-equipment ecosystem. For hyperscalers, however, it creates a more complicated outcome: very strong demand and backlog, but weaker near-term free cash flow, higher depreciation, some gross-margin pressure and growing dependence on third-party capacity and external financing.
The key metric to watch is no longer announced capex or GPUs ordered. It is energized megawatts actually delivered, together with HBM availability, networking deployment and the speed at which those megawatts convert into billable cloud consumption. The principal risk moves into 2027–28: substantial capacity is being built simultaneously, so investors will eventually need to
