Evidence since mid‑2023 shows that the severe shortage of high‑end AI GPUs (Nvidia A100/H100‑class and equivalents) persisted well beyond a year and was still meaningfully constraining the industry into 2025, which matches Sacks’s “year or two” timeline.

Key points:

• In 2023, Nvidia’s AI accelerators were capped by TSMC’s advanced CoWoS packaging capacity; demand vastly exceeded supply. Analyses note that even by early 2024, this packaging bottleneck was only starting to ease, with TSMC planning to more than double CoWoS output by 2025 precisely because supply remained tight relative to AI demand.【(siliconanalysts.com)】
• High‑bandwidth memory (HBM), which is required for modern AI GPUs, has been a separate choke point. Industry analysts and Micron’s CEO have said HBM shortages are expected to weigh on GPU production into 2026, indicating that the underlying constraints did not disappear quickly in 2023–2024 but instead persisted for multiple years.【(luckboxmagazine.com)】
• A concrete example from a hyperscaler: Amazon’s internal documents show its retail division struggled with a lack of GPU capacity throughout 2024, to the point it launched “Project Greenland” in July 2024 to ration and centrally manage scarce GPUs. Reporting in April 2025 says Amazon’s retail business had only then resolved its internal GPU shortage and expected surplus capacity later in 2025—i.e., constraints lasted well past mid‑2024 and into early 2025.【(businessinsider.com)】
• Separate coverage of Nvidia H100 supply notes that through 2024 customers still faced multi‑month lead times; only by later in the year was it plausible that wait times might disappear, and even then large LLM developers like OpenAI were singled out as still hitting supply bottlenecks due to their scale.【(1ai.net)】
• Broad 2023–2025 retrospectives on the “GPU crisis” and AI boom describe a persistent global imbalance where big tech, specialized GPU clouds, and a handful of well‑funded players soaked up most of the Nvidia GPU supply, leaving many startups and smaller firms facing long waits and capacity constraints through 2024 and into 2025.(buzzbyte.org)}

Putting this together: the AI‑GPU shortage clearly did not materially ease within a few months of July 2023; instead, serious constraints persisted for roughly two years, with major players still short into early 2025 and structural bottlenecks (CoWoS, HBM) projected to remain tight even beyond that. That aligns well with Sacks’s claim that the shortage was “probably not going to get better for a year or two,” so the prediction is best judged as right.