Radar · Beta
Who supplies AI compute, where the physical chokepoints are, and what's shifting in the fab, datacentre and power layers beneath the model layer.
incumbents vs new entrants · ranked by mentions in compute/infrastructure coverage
from the compute/infrastructure-tagged set, most recent first
Co-packaged optics could eventually address bandwidth and interconnect bottlenecks in large AI and HPC systems, although the supplied report provides no demonstrated performance or deployment results.
fab capacity · HBM · packaging · grid · export controls
If sustained, the development could reduce Nexperia China's exposure to overseas wafer disruptions and strengthen domestic semiconductor manufacturing resilience.
Co-packaged optics could eventually address bandwidth and interconnect bottlenecks in large AI and HPC systems, although the supplied report provides no demonstrated performance or deployment results.
Coming soon
A timeline of announced compute capacity commitments — new fab, datacentre and cluster buildouts, tracked by announcement date.
Waiting for: A larger population of items carrying both a computeCategory and a verified announcement date; the fields exist but are not live yet (migration not applied).
Coming soon
Announced AI-related investment broken down by compute category — accelerators, datacentres, fabs, networking, energy, memory, cloud capacity.
Waiting for: The AI desk's own investment extraction (FR-613, shipped this sprint) to accumulate enough rows to chart — today's 155 investments rows are still almost entirely defence-desk.
Coming soon
The split between compute/infrastructure coverage that serves training workloads versus inference/deployment workloads, over time.
Waiting for: workloadPhase (FR-614, shipped this sprint) to be live and populated — its migration is written but not applied, so no item carries it yet.
Coming soon
Which countries are building, exporting, or restricting AI compute capacity — a measured pillar, not a hand-curated one.
Waiting for: Enough geo-tagged compute/infrastructure coverage to measure a posture per country rather than curate one — deliberately deferred rather than shipped as a guess (see the roadmap's excluded-scope note).
The approaches could shape future HBM performance, customisation, thermal behaviour, and scalability for advanced compute systems, although the report provides no demonstrated performance results.
The upgrades could preserve manufacturing capacity in China while export-control and supply-chain uncertainty increases around advanced chipmaking equipment.
More efficient use of AI chips could ease physical and economic scaling constraints, but the supplied text does not provide specific projects, technologies, performance data, or commitments.
Greater concentration of Nvidia-related AI memory demand at SK Hynix could affect supplier dependence, competitive positioning, and the resilience of the AI hardware supply chain.
A changing DRAM capacity balance could affect competitive positioning and the resilience of regional memory semiconductor supply chains.
Memory bandwidth and interconnect efficiency are important constraints on large-scale AI infrastructure, so progress in these areas could affect accelerator-system performance and the competitive position of major memory suppliers.
Greater on-site storage and self-sufficiency could improve the resilience of AI data-center power supplies and reduce their impact on Taiwan's electricity grid.
Backside power delivery could affect the performance, efficiency, and competitiveness of future advanced processors, although the supplied text does not establish comparative technical outcomes.
The development indicates growing demand for AI accelerator manufacturing and could affect the allocation, pricing, and availability of advanced-node capacity.
Advanced packaging and HBM availability are important constraints on AI accelerator production, so persistent TSMC bottlenecks could alter supply-chain dependencies and foundry competition.
The development may reshape the geographic dependencies and supply-chain exposure of AI-server manufacturing, although the source provides no investment figures, company names, or details on specific components.
The reported 1.33x average throughput improvement, rising to 1.54x with matched calibration data, suggests a potentially material way to run reasoning-oriented MoE models under constrained GPU memory.
Sustained AI infrastructure growth could constrain the hardware, electricity and cooling inputs needed to expand compute capacity.
The reported diversification may help preserve near-term AI-compute supply-chain activity and reduce disruption during NVIDIA's platform transition, although the article provides limited operational detail.
