AI’s Achilles Heel

July 1, 2026

Currently, a topic of hot debate is whether AI’s physical infrastructure can support its growth.

Investors obsess over model capabilities while the physical systems that power AI face serious constraints.

The risks are real — but so are the reasons for optimism.

Consider the bottlenecks first:

Infrastructure Layer	Concentration	Current Status
Advanced chips	TSMC = 90% of the advanced chips	Single point of failure
AI accelerators	NVIDIA = 80-90%	Long wait times
HBM memory	3 suppliers = 95%+	Sold out through 2026
Power Demand	Mismatch in Timelines	Months vs Years

The Risks Are Real

Chip concentration: TSMC is estimated to manufacture 90% of the world’s advanced chips. Every major AI company — NVIDIA, AMD, Apple, Qualcomm — depends heavily on TSMC due to its advanced manufacturing capabilities. TSMC’s most advanced capacity remains heavily concentrated in Taiwan, creating a geopolitical and operational concentration risk.

Taiwan’s energy system is also exposed. Taiwan imports 97% of its energy, and its LNG reserves could be exhausted in just 11 days without foreign shipments. The island’s gas-fired power generation accounted for 43% of total power generated in 2024. This makes Taiwan highly vulnerable to LNG supply disruptions in the event of a maritime blockade.

In 2025, roughly one-third of Taiwan’s LNG imports came from the Gulf region, and the Hormuz crisis led to Taiwan receiving no LNG cargoes from Qatar or the UAE in April and May 2026. Currently, Taiwan has been forced to secure the US and Australian LNG supplies to make up for the shortage; an extended shortage could result in power rationing hitting the Taiwanese semiconductor industry, including TSMC, and with it the entire AI hardware supply chain that depends on it.

GPU scarcity: It is estimated that NVIDIA commands 80-90% of the AI accelerator market by revenue as of 2025

NVIDIA H100 GPU Rental Price Surge

NVIDIA’s H100 GPU rental prices have surged nearly 40% since October 2025, rising from $1.70/hour to $2.35/hour by March 2026. This is unusual because older hardware typically gets cheaper when newer models (like the Blackwell B200/GB200 series) arrive. Instead, the sharp increase indicates a mismatch between limited supply and high demand for AI compute power. This GPU shortage extends beyond the H100, with even the newer Blackwell chips facing long wait times that stretch into mid-2026. Despite the high demand, switching from NVIDIA is costly due to CUDA, the software ecosystem NVIDIA has built over 20 years. CUDA has 6 million developers, 3,000+ optimized applications, and deep integration into every major AI framework.

Memory bottleneck. High-bandwidth memory (HBM) is the new bottleneck. SK Hynix CFO Kim Jae-joon: “We have already sold out our entire 2026 HBM supply.” Micron CEO Sanjay Mehrotra confirmed the same: “Our HBM capacity for calendar 2025 and 2026 is fully booked.”Three companies — Samsung, SK Hynix, and Micron — control over 95% of global DRAM production. They’ve collectively shifted focus toward HBM for AI accelerators. Each HBM chip consumes 3x the wafer capacity of standard memory

Power constraints.

In today’s era of hyperscale data centers, campuses consuming 300–600 MW of electricity—comparable to the power needs of a mid-sized city—are increasingly becoming part of development discussions.

Currently, the entire U.S. data centre sector draws a total of less than 15 gigawatts (GW) of power, but the pipeline of new data centres under construction will, if all planned facilities right now are completed, add 140 GW of new load to the current total U.S. peak demand of 760 GW, an increase of almost 20%.

There are significant bottlenecks for meeting this power demand, the primary being a mismatch in timelines. Hyperscale data centres can be built in 18–36 months, but the transmission lines tations needed to power them take 5–10 years to plan and complete. This creates a mismatch in timelines. High-capacity power transformers used in hyperscale substations are experiencing procurement lead times exceeding two years

The Case for Optimism

The risks above are real. But markets are responding to bottlenecks — and early signs are encouraging.

Risk	Counterargument
TSMC concentration	Diversification underway
NVIDIA dominance	Markets are no longer a one-player story
HBM shortage	Memory makers are expanding capacity to ease supply constraints
Power constraints	New Approaches to solve constraints. SMR nuclear pipeline doubled from 25 GW to 45 GW.

TSMC is diversifying– For example, in 2020, TSMC chose Phoenix, Arizona, for its first advanced U.S. semiconductor manufacturing site — a milestone investment that has expanded from $12 Bn to $165Bn. The first fab at Arizona started high-volume N4 production in Q4 2024; the second fab’s volume production is targeted for 2H 2027; and the third fab’s volume production is targeted by the end of the decade. The second fab in Japan will be operational by the end of 2027, and the first fab started operations in 2024. European Semiconductor Manufacturing Company (ESMC) says total investments in Dresden are expected to exceed €10B; ESMC is a joint venture between TSMC, Bosch, Infineon, and NXP to establish an advanced semiconductor fab in Dresden, Saxony, Germany

NVIDIA dominance – NVIDIA’s moat is still strong, but the market is no longer a one-player story with AMD emerging as a challenger worth watching. AMD witnessed record AMD Instinct™ GPU revenue, driven by the ramp of its MI350 Series accelerators and expanding customer engagements for next-generation AI platforms. The company’s broader data centre segment revenue of $16.6Bn in 2025, representing 32% YoY growth, was driven by strong demand for both general-purpose and AI computing. Additionally, to reduce dependence on NVIDIA, hyperscalers are investing in custom accelerators tailored to their own training and inference workloads. These include Google’s TPU, Amazon’s Trainium, Microsoft’s Maia, and Meta’s MTIA, each optimized for internal scale, workload fit, and long-term cost control

Memory Makers expanding capacity: Micron’s new Singapore HBM packaging facility is expected to add meaningful advanced packaging capacity from 2027. Samsung anticipates that its HBM sales will more than triple in 2026 compared to 2025, and is proactively expanding its HBM4 production capacity. In Feb 2026, SK Hynix announced a KRW 21.6T investment for the first fab and five cleanrooms at its Yongin semiconductor cluster

New approaches to solve energy constraints: The pipeline of conditional offtake agreements between data centre operators and small modular reactor (SMR) nuclear projects has grown from 25 gigawatts at the end of 2024 to 45 gigawatts as of April 2026. In parallel, hyperscalers are locking up long-term power through nuclear PPAs and plant-restart deals, including Microsoft/Three Mile Island, Google/Duane Arnold, Amazon/Talen Energy, and Meta’s nuclear agreements.

Conclusion

The AI boom is real. But infrastructure constraints create risks. TSMC concentration, NVIDIA dominance, memory shortages, and power bottlenecks aren’t imaginary. At the same time, there are measures in the right direction that are likely to ease the AI’s infrastructural bottlenecks. The infrastructure race is already separating winners from laggards. The companies that navigate the next cycle won’t just have the best models — they’ll have secured the power, the chips, and the memory to run them. The 1,000x drop in inference costs is real and will unlock new markets. But it doesn’t reduce physical compute requirements — it creates demand for more of it. Infrastructure constraints won’t stop AI; they will determine who gets to participate.

FinStrat Insights