As generative AI demand surges, access to GPUs and high-bandwidth memory (HBM) is quickly becoming the defining battleground in tech infrastructure - not just for startups, but for nations and governments too.
Why it Matters in June 2025
By early June, the industry is grappling with persistent GPU shortages, long lead times on HBM3, and structural bottlenecks in global supply chains:
TSMC foundry capacity is booked out, while Samsung and Intel race to expand under new national incentives like the CHIPS Act and Europe’s counterpart programs.
GPU demand completely outpaces supply: Jensen Huang confirmed that even with production doubling in 2025, NVIDIA still coundn’t keep up with orders for H100 and Blackwell series chips.
Memory suppliers such as SK Hynix, Samsung and Micron are experiencing 6-12 month lead times on HBM3 packaging and 15-20% price increases on enterprise-grade NVMe SSDs.
This infrastructure crunch has prompted strategic moves from major players and these investments are not long-term playbooks, rather active capacity builds, driven by enterprise AI demand in Europe and North America.
AWS earmarked $5.3B for AI data centers in North Carolina with Graviton X processors and H140/HBM chips.
Microsoft committed $400M to double its Swiss data center capacity for enterprise AI workloads.
Infra = The New Capital Moat
In this market, equity in a startup means little if it can’t access infrastructure to serve users, train models, or scale. Companies that act like infrastructure providers are the new capital moats:
Crusoe is locking its supply with a $400M purchase of ~13K AMD MI355X chips to compete with Nvidia-dominated clouds. They are emerging as a modular AI provider building under the “neocloud” model.
GridFree AI is deploying off-grid, energy-efficient GPU centers designed to reduce capex and power strain, an unorthodox but scalable infra alternative in the Southeast US, UK, and beyond.
From my vantage - both in high-capex infra ops and venture sourcing, I’ve seen founder teams with great interfaces fail because they couldn’t get GPUs.
Infrastructure isn’t just a behind-the-scenes concern anymore - it’s front and center.
Founders must architect for supply certainty, not just API fidelity.
VCs need to evaluate not only product-market fit, but infra-market fit, i.e., does the startup control its compute destiny?
And, strategists and boards must treat procurement and physical capacity as strategic assets, not op-ex line items.
This isn’t temporary volatility, it’s structural realignment.
As the countries are launching AI infrastructure initiatives at national scale, Founders like Crusoe and GridFree are pioneering alternative models to hyperscale dependency, and chipmakers, AI labs and cloud infrastructure firms are rewriting compute delivery protocols from the ground up.
That’s why if you are still thinking AI without thinking hardware flow, cost curves, and supply dynamics - you are playing catchup. 👀 Upcoming Topics in All That Noise
Look out for:
A deep dive into Agentic AI infrastructure and autonomous orchestration
Case studies on invisible open-source founders building infra-first companies
Explorations of why analog obsessions - craft, slow tools - reveal deeper empathy in engineering culture
The views expressed are those of the author and do not necessarily reflect the views of any investment firm or portfolio company.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.