AI chip shortage eases as NVIDIA export controls reshape supply chain

The global AI chip shortage is easing in 2026 as NVIDIA H100 and Blackwell allocations stabilise, but export controls on advanced GPUs to Southeast Asia are redirecting demand toward alternative suppliers and custom silicon. The shift is creating new winners in the semiconductor supply chain while reducing the bottleneck that constrained enterprise AI expansion throughout 2024 and 2025. NVIDIA's production capacity for data centre GPUs has increased roughly 40 percent year-on-year, driven by new fab output from TSMC's advanced packaging lines and improved yields on the Blackwell architecture.

The easing is not uniform across geographies. Export controls imposed by the United States restrict sales of advanced AI accelerators to customers in China and several Southeast Asian markets, which has redirected demand toward AMD Instinct products, Google's TPU v5, and custom silicon developed by hyperscalers including Amazon, Microsoft, and Meta. The diversion is accelerating diversification of the AI hardware market, with enterprises and cloud providers evaluating multi-vendor procurement strategies that would have been impractical when NVIDIA dominated supply.

AMD and custom silicon gain share

AMD has increased its data centre GPU revenue by roughly 60 percent year-on-year, with the Instinct MI300 and MI325 products gaining traction in enterprise and cloud deployments where price-performance is the primary selection criterion. The MI325 offers roughly 80 percent of the training performance of an NVIDIA H100 at roughly 60 percent of the acquisition cost, a gap that narrows further when NVIDIA's software ecosystem premium is factored in. Enterprises that lack the engineering resources to optimise for CUDA are finding AMD's ROCm software stack mature enough for production workloads, particularly in inference and fine-tuning scenarios where memory bandwidth matters more than raw tensor throughput.

Custom silicon developed by hyperscalers is also entering the market. Amazon's Trainium2, Microsoft's Maia, and Google's TPU v5 are designed for specific workload patterns that each company encounters at scale, and they are being offered to cloud customers as cost-optimised alternatives to NVIDIA GPUs. The trade-off is portability: code written for CUDA requires adaptation to run on custom silicon, and the adaptation cost is only justified for workloads that run at sufficient volume to amortise the engineering effort. Enterprises with stable, high-volume inference pipelines are the natural candidates for custom silicon, while organisations that need flexibility across model architectures and research tasks remain on NVIDIA.

TSMC advanced packaging remains the bottleneck

Despite increased GPU production, advanced packaging capacity at TSMC remains the binding constraint on AI chip supply. The CoWoS and SoIC packaging technologies required for high-bandwidth memory stacking on GPU and custom silicon modules have lead times of 40 to 50 weeks, according to semiconductor supply chain data. That bottleneck affects NVIDIA, AMD, and hyperscaler custom chip programmes equally, which means the easing of the AI chip shortage is driven more by demand normalisation than by a structural increase in supply.

TSMC is expanding its advanced packaging capacity in Japan and the United States, with new fabs scheduled to come online in 2027 and 2028. The expansion will eventually relieve the packaging constraint, but it does not address the near-term allocation problem. Enterprises that need large GPU clusters in the next 12 months should expect competitive pricing and reasonable availability, but lead times for custom configurations and the largest cluster sizes will remain extended compared with pre-shortage norms.

Australian enterprise access and local infrastructure

Australian enterprises rely on cloud providers for AI compute rather than owning GPU infrastructure directly, which means the AI chip shortage easing is felt through cloud pricing and availability rather than through direct hardware procurement. AWS, Google Cloud, and Microsoft Azure have all improved GPU allocation for Australian region customers over the past six months, with spot pricing for H100 and Blackwell instances dropping by roughly 20 percent from the 2025 peak. The improvement reflects both increased global supply and the shift of some enterprise workloads to alternative chip architectures that are less constrained by NVIDIA allocation limits.

Nexus Century, an Australian AI infrastructure provider, has been testing AMD Instinct clusters in its Sydney data centre as an alternative to NVIDIA-based offerings for enterprise clients. The company reports that inference workloads on AMD hardware are achieving comparable latency to NVIDIA at lower cost, though training workloads still favour NVIDIA's software ecosystem. The divergence suggests that the AI chip market is segmenting by workload type rather than converging on a single dominant architecture. Explore more enterprise AI infrastructure analysis at the Tech & Ideas hub

For NVIDIA's latest supply and product information, see NVIDIA data centre. AMD Instinct product documentation is at AMD Instinct. TSMC advanced packaging updates are published at TSMC investor relations.