Published evidence log
Where will the next AI bottleneck emerge?
Current position
Working answer
The next AI bottleneck is most likely to emerge in physical infrastructure: power, advanced AI-chip capacity, and high-bandwidth memory, followed by networking, cooling, electrical systems, and data-center buildout. Demand remains strong, so the limiting factor is increasingly the ability to manufacture, connect, cool, and energize complete AI systems rather than a lack of customers or training data.
The strongest evidence points to a physical-infrastructure bottleneck rather than weak demand or immediate data scarcity. Epoch AI’s forward-looking analysis identifies power availability and chip-manufacturing capacity as the principal constraints on continued training-scale growth. The industry sources add system-level detail: Broadcom reports rapidly rising demand for custom AI accelerators and networking; Micron describes tight memory conditions and sharply expanding DRAM, NAND, SSD, and HBM demand; and Vertiv is accelerating capacity expansion to serve data-center demand. Nvidia’s system-level framing reinforces that usable compute now depends on integrating chips with interconnect, cooling, power, and other data-center components. Microsoft’s strong cloud growth and contracted obligations suggest customers are still demanding AI capacity. The likely next bottleneck is therefore the ability to manufacture and deploy complete, powered AI systems—especially advanced chips and memory, followed by networking, electrical infrastructure, cooling, and data-center construction. The supplied Weixin source could not be substantively verified and does not change that conclusion.
Chronology
Evidence timeline
April 4, 2024
The Data Center is the New Compute Unit: Nvidia's Vision for System-Level Scaling[1]
Published April 4, 2024, the article presents Nvidia’s view that the data center—not the individual chip—is becoming the new unit of compute. It discusses system-level scaling and emphasizes copper, cooling, and compute density as key pillars of that approach.
How this bears on the question
This article supplies the system-level framework for understanding why the bottleneck may move beyond individual accelerator chips. If the data center becomes the effective compute unit, copper, cooling, compute density, and system optimization become part of the scaling problem. It is an analytical framing rather than direct evidence of a measured shortage.
2025
Can AI scaling continue through 2030? | Epoch AI[2]
Epoch AI examines whether AI training can continue scaling through 2030. It evaluates four constraints—power availability, chip-manufacturing capacity, data scarcity, and the latency wall—and predicts that training runs of 2×10^29 FLOP will be feasible by 2030. The source is identified as a 2025 publication; the searchable excerpt does not expose a more precise date.
How this bears on the question
Epoch AI provides the clearest comparative assessment of candidate bottlenecks. It examines power availability, chip manufacturing, data scarcity, and latency, and projects that very large training runs could remain feasible through 2030. Its framework makes power and chip capacity the leading constraints, ahead of data scarcity and latency, although the relative severity can change as infrastructure expands.
June 3, 2026
Broadcom Inc. Announces Second Quarter Fiscal Year 2026 Financial Results and Quarterly Dividend | Broadcom Inc.[3]
Published June 3, 2026, Broadcom reported second-quarter fiscal 2026 revenue of $22.187 billion, up 48% year over year. Its CEO cited increasing demand for custom AI accelerators and AI networking, while the company projected third-quarter AI semiconductor revenue of $16.0 billion, more than 200% above the prior year.
How this bears on the question
Broadcom’s results provide current evidence of accelerating demand for the specialized silicon and interconnect needed to assemble large AI systems. The release demonstrates strong demand and guidance rather than proving a shortage, but it supports the view that custom accelerators and AI networking are likely pressure points in the supply chain.
June 24, 2026
Micron Technology Fiscal Q3 2026 Earnings Call Prepared Remarks[4]
Micron’s fiscal Q3 2026 prepared remarks describe AI-driven growth in data-center DRAM, NAND, SSDs, and HBM4. The company says calendar-2026 data-center DRAM and NAND bit shipments are expected to more than double from two years earlier, while supply conditions for both DRAM and NAND are expected to remain tight beyond calendar 2027. The source is dated to fiscal Q3 2026; no more precise publication date is exposed in the searchable text.
How this bears on the question
Micron’s remarks identify memory and storage as important constraints within AI infrastructure. The source says industry data-center DRAM and NAND shipments are expanding rapidly while supply conditions for both remain tight beyond calendar 2027. That suggests accelerator availability alone will not determine deployable AI capacity.
July 29, 2026
FY26 Q4 - Press Releases - Investor Relations - Microsoft[5]
Released July 29, 2026, Microsoft reported Microsoft Cloud revenue of $59.3 billion, up 27% year over year, and commercial remaining performance obligations of $678 billion, up 84%. The release describes continued customer confidence in Microsoft’s AI and cloud offerings.
How this bears on the question
Microsoft’s results are indirect but important evidence against customer demand being the immediate bottleneck. Strong cloud growth and a large increase in contracted future obligations imply that demand for AI and cloud capacity remains substantial, placing more pressure on physical infrastructure and suppliers.
July 29, 2026
Vertiv Holdings Co. - Vertiv Reports Strong Second Quarter 2026 with Diluted EPS Growth of 53% (Adjusted Diluted EPS Growth of +60%); Raises Full Year 2026 Guidance Across All Key Metrics[6]
Released July 29, 2026, Vertiv said data-center demand remained strong and that its growth fundamentals were durable. The company said it was accelerating capacity expansion and strategic investment to meet demand and capture market share, while raising its full-year 2026 guidance across key metrics.
How this bears on the question
Vertiv’s response is a market signal that power and thermal infrastructure are becoming strategically important parts of AI scaling. Accelerating capacity expansion does not by itself prove a shortage, but it indicates that suppliers of the equipment required to operate dense data centers are preparing for sustained demand.
August 1, 2026
Weixin Official Accounts Platform[7]
The source is identified only as a Weixin Official Accounts Platform page. No substantive searchable excerpt or verifiable publication date was available, so its claims and relevance to the research question cannot be reliably summarized.
How this bears on the question
The supplied Weixin snapshot could not be substantively verified through searchable source text. Because no reliable passage or publication date was exposed, it provides no dependable evidence for ranking the next AI bottleneck.
Bibliography
Sources
- 1.
- 2.
- 3.
- 4.
- 5.
- 6.
- 7.