Our Take
Infrastructure moat is real; the companies with spare GPU capacity win the next 18 months of enterprise AI deals.
Why it matters
AI workloads are compute-constrained, not talent-constrained. Whoever controls supply controls the market for enterprise agents, RAG, and fine-tuning. Smaller cloud providers and startups are squeezed.
Do this week
Infrastructure teams: audit your cloud egress costs and multi-vendor contracts this quarter before GPU scarcity locks you into a single provider for three years.
Amazon and Google secure the supply edge
Amazon and Google have pulled ahead of Microsoft, Meta, and other AI companies in the race to acquire and deploy AI compute capacity, according to the Wall Street Journal. Both companies control vast data center infrastructure and have invested heavily in custom silicon (AWS Trainium and Inferentia chips; Google's TPU line). Their vertical integration from chip design through cloud services gives them first access to their own production capacity and the ability to offer sustained-use discounts that smaller competitors cannot match.
The advantage is not theoretical. Companies building AI applications face 6-to-18-month wait times for GPU slots from third-party vendors. Amazon and Google can serve internal demands first, then monetize excess capacity to external customers. This creates a two-tier market: priority access for those willing to commit spend to a single cloud provider, and scarcity pricing for everyone else.
Control of supply reshapes vendor relationships
Compute scarcity is an infrastructure problem disguised as a technology problem. Microsoft has significant GPU inventory (through partnerships with NVIDIA and its own chip efforts) but is committed to OpenAI's training and inference pipelines first. Smaller cloud providers (Oracle, IBM, regional players) have minimal AI chip allocation and cannot compete on price or availability. Startups and mid-market enterprises face a choice: accept multi-year lock-in to a dominant provider, or accept longer latency and higher costs for portable, multi-cloud infrastructure.
This matters because cost and latency drive adoption curves. If a startup can only get reliable GPU access by committing 50% of its cloud budget to one vendor, it will do so. That decision cascades: data residency, API choices, and model selection all follow the compute decision. Switching costs become structural.
Audit your cloud dependencies now
If your AI workload runs on AWS or Google Cloud, document exactly what services you depend on and whether equivalent alternatives exist on other clouds. Specifically: identify which models you use (OpenAI, Anthropic, open-source, fine-tuned), where inference happens (first-party service, self-hosted, hybrid), and whether you have negotiated volume discounts or committed use contracts.
If you do not yet have a cloud-committed AI workload, negotiate multi-cloud provisions in any long-term agreement you sign. Include explicit egress pricing caps, portability guarantees, or compute-quota swaps between providers. The cost of portability today is far cheaper than the cost of renegotiating under scarcity in 2026.
Finally: if you are evaluating open-source models as a hedge against vendor lock-in, remember that self-hosting requires spare GPU inventory of your own. For most mid-market teams, that inventory is rented, not owned, which means the infrastructure lock-in problem still applies. The only real hedge is contractual, not technical.