AI Infrastructure Bottleneck Shifts from GPUs to Power
Key Questions
What is the current main bottleneck in AI infrastructure buildout?
The bottleneck has shifted from GPU scarcity to power infrastructure limitations. Half of planned 2026 US data centers are delayed due to transformer lead times of up to 4 years and prices rising 77%.
Which companies are impacted by power supply constraints?
Siemens Energy is sold out through 2030. Space-based data centers are advancing with AI1 launching next year on Starship to address terrestrial power issues.
What new AI systems and products were recently launched?
AMD launched the Helios AI rack-scale system with Microsoft as a buyer. Nvidia released Spectrum-6 networking, while Huawei ships Attention-FFN Disaggregation for MoE serving.
Which startups raised funding in AI infrastructure and materials?
TYLsemi raised $43M for chiplet-based AI silicon. CuspAI secured $450M for AI materials discovery, and Infinity raised $15M for non-Nvidia inference.
What manufacturing expansions support Nvidia's roadmap?
Wistron is building a $700M Texas factory for Nvidia GB300 and Vera Rubin systems. Corridor Ventures invested in Carbice for advanced cooling solutions.
How is Google approaching AI hardware investments?
Google is prioritizing AGI development over TPU sales with planned capex of $195-205B, resulting in negative free cash flow.
What new research findings address training efficiency?
NVIDIA research indicates AdamW has a scale ceiling, pointing to potential training efficiency gains. Google also released JAXBench for TPU kernel optimization.
How does DeepSeek's work relate to alternative hardware?
DeepSeek-V4 was post-trained on Ascend SuperPOD, demonstrating viable training on non-Nvidia infrastructure amid ongoing hardware diversification.
AI buildout bottleneck moving from GPU scarcity to power and physical infrastructure. Half of planned 2026 US data centers delayed; transformer lead times up to 4 years, prices up 77%. Siemens Energy sold out through 2030. Space-based data centers moving to reality with AI1 launching next year on Starship. New massive rounds: Lumilens $900M for optical networking, Hadrian $1.37B for defense manufacturing, Valar Atomics $1B for nuclear reactors. Velaura AI $110M for power-efficient silicon. Fractile $6.5B for inference chips. Nvidia explores stake in Mercor for data supply. Nvidia and Wall Street firms partner on $500B AI financing venture. Intel raises $15B+ equity. Micron Ventures launches $250M AI fund. Cloudflare's AI push causing reliability problems. Stripe paying $7B for OpenRouter. SoftBank leads $200M Series A in Gravis Robotics. Chip giants pour record cash into AI startups ($250B+ rounds). Ed Zitron calls Nvidia's AI funding 'desperate market'. 'The Real AI Crash Will Start This Year' — bearish thesis. Higher interest rates reinforce rotation from growth to value.