US AI Datacenter Boom Faces Chip Packaging Bottleneck
Satellite data reveals construction can't keep pace with projected 2028 AI chip demand
1Accelerating but insufficient growth
Satellite imagery shows US datacenter construction accelerating, with capacity projected to grow from 11 GW in 2025 to 16-18 GW in 2026. However, land clearing activity has plateaued, falling short of the 80+ GW needed by 2028. SynMax's analysis distinguishes actual construction progress from announced projects, revealing only half of planned 2026 capacity is underway. The gap suggests many projects may face delays or cancellations due to infrastructure constraints.
2Advanced packaging emerges as bottleneck
Beyond chip fabrication, advanced packaging combining multiple dies with high-bandwidth memory creates a critical constraint. Current packaging capacity supports only ~13 GW of accelerator power (17.5 GW total datacenter capacity). Two new packaging facilities in 2027 may add 6 GW capacity, but this still falls short of demand. Many US-made chips must be sent to Taiwan for packaging, creating logistical choke points between semiconductor production and operational datacenters.
3Texas and Virginia concentration risks
Development concentrates in Texas and Virginia, making national capacity vulnerable to regional delays. Texas faces unique grid challenges - datacenters may accept occasional power curtailments for faster connection. ERCOT's system allows 99% uptime if curtailments are predictable, but sudden 5-minute notices under the SCED system can disrupt operations. This creates financial viability challenges for operators needing stable power for AI workloads.
4Global chip allocation pressure
A 2023 study warned fulfilling all US datacenter plans through 2030 would require 90% of global AI chip supply growth. Current projections show packaging capacity growing slower than accelerator demand. The Jefferies report highlights how power access, permitting, and component supplies collectively constrain growth. Even with accelerated construction, physical and logistical bottlenecks may prevent matching ambitious demand forecasts.
5Satellite vs. announced projects
SynMax's weekly satellite monitoring tracks three key phases: land clearing, initial structures, and construction progress. This data reveals only 20% of 2028 pipeline projects show visible activity. The disconnect between announcements and ground progress suggests many projects face unaccounted delays. Satellite analytics provide more reliable capacity forecasts than corporate announcements or chip demand models alone.
6Practical limits for 2027
SynMax estimates a practical upper limit of low-20s GW by 2027, considering packaging and power constraints. This falls 60 GW short of some 2028 projections. The gap stems from compounding limitations: each finished accelerator requires multiple packaged components, and each datacenter needs reliable power infrastructure. These interdependent constraints create nonlinear scaling challenges.
7Financial viability calculations
Former ERCOT developer Jesse Gossett notes predictable curtailments can be financially modeled into service agreements. However, Texas' SCED system sometimes imposes unexpected 5-minute reductions - untenable for AI operations needing stable power. This creates a mismatch between grid reliability needs and market structures, potentially deterring investment in affected regions despite available land and demand.
8Reevaluating growth projections
The data suggests current AI infrastructure growth models may overestimate achievable capacity by not fully accounting for physical constraints. While chip production garners attention, packaging and power infrastructure emerge as equally critical bottlenecks. Industry forecasts may need recalibration to reflect these practical limitations, potentially impacting AI service rollout timelines and investment strategies across sectors.