Key Takeaways
- The capital-utility flip: Tech giants’ 2026 infrastructure spend will quadruple the entire US electric utility industry’s annual investment.
- Physics as the bottleneck: Growth is no longer gated by chip supply, but by a three-to-five-year timeline for grid expansion.
- The neocloud arbitrage: Specialized providers like CoreWeave offer 60% cost advantages and immediate GPU access over hyperscalers.
- The end of elasticity: The “pay-as-you-go” era is being replaced by rigid, multi-year “take-or-pay” capacity contracts.
- The governance shift: Securing compute has evolved from a procurement task into a high-stakes vendor risk management strategy.
For most of the past decade, enterprise cloud strategy operated on a simple premise: pick a hyperscaler, migrate your workloads, and let the provider handle the infrastructure. That model is under real pressure. The AI capacity crunch has exposed constraints that the traditional cloud buying motion was never designed to navigate. Organizations that fail to adapt their sourcing approach will find themselves locked out of the compute they need to compete.
A capital war without precedent
The scale of investment flowing into AI infrastructure is unlike anything the technology industry has seen. Amazon has projected $200 billion in capital expenditure for 2026, Alphabet is targeting $175 to $185 billion, Meta $115 to $135 billion, Microsoft tracking toward $120 billion or more, and Oracle targeting $50 billion.
Combined, these five companies alone plan to spend roughly $660 to $690 billion on infrastructure in 2026, the vast majority directed at AI compute, data centers, and networking.
To put that in perspective: the entire US electric utility industry invested approximately $160 billion in infrastructure in 2024. The technology sector is now outspending the utility industry on energy-adjacent infrastructure by a factor of four.
Despite that spend, demand is still outrunning supply. Microsoft has disclosed an $80 billion backlog of Azure orders that cannot be fulfilled due to power constraints, while Alphabet’s cloud backlog surged 55% sequentially to over $240 billion.
Get ready for what’s next with insights and breakthrough topics in cloud, AI, and innovation. Join our newsletter for curated topics delivered straight to your inbox.
By signing up, you agree to Cloud Latitude’s Privacy Policy and Terms of Use.
The real bottleneck is power, not capital
What makes this infrastructure cycle structurally different from previous cloud buildouts is the constraint that is actually limiting growth. Power constraints, not capital limitations, create the main bottleneck for building data centers. The physics of AI infrastructure have fundamentally changed the equation.
Microsoft CEO Satya Nadella stated plainly that the company is no longer chip-constrained, but power-constrained. An AI query consumes ten times more energy than a traditional web search, and an AI server rack demands fifty times the power density of a traditional server rack.
As of mid-2025, more than 36 projects representing $162 billion in investment were either blocked or significantly delayed due to power, permitting, and community opposition. The timeline problem compounds the pressure: growth is no longer gated by a six-month chip production cycle, but by a three-to-five-year timeline required to expand grid infrastructure.
The rise of the neocloud
The capacity crunch has opened the door for a new class of provider. Specialized “neocloud” providers—led by CoreWeave, Lambda Labs, Crusoe, and Nebius—are pure-play vendors competing on price, availability, and technological specialization.
The cost differential is significant: renting an NVIDIA A100 40GB GPU on CoreWeave costs approximately $1.39 per hour, versus $3.67 per hour on Azure or Google Cloud—a 62% cost advantage. Beyond price, neoclouds offer immediate availability. Teams can launch large multi-GPU jobs within minutes, rather than sitting in hyperscaler queues for days.
| Provider | Primary strength | Ideal workload | Key constraint |
|---|---|---|---|
| Hyperscalers | Ecosystem, security, compliance | Integrated enterprise apps | Capacity queues, higher cost |
| Neoclouds | Speed, GPU availability, price | Large-scale training/inference | Fragmentation, younger orgs |
The evolution of GPU pricing models
As compute becomes a scarce commodity, the “pay-as-you-go” elasticity that defined the early cloud era is being replaced by more rigid, industrial-scale procurement. Enterprises now face a fragmented pricing landscape:
On-demand pricing: provides the most flexibility but the lowest priority. In a capacity crunch, on-demand instances are the first to be throttled or reclaimed by providers.
Reserved and committed use: hyperscalers are increasingly requiring one-to-three-year commitments to guarantee GPU availability. This shifts cloud spend from OpEx back toward a CapEx-like profile.
Spot instances: while theoretically cheaper, spot instances are nearly non-existent for high-end H100 or B200 GPUs due to baseline demand.
Neocloud contracts: these providers often offer significant discounts (30% to 50%) but require multi-year “take-or-pay” contracts, further emphasizing the need for accurate capacity planning.
FinOps as a service: navigating the cost of AI
The complexity of these pricing models, combined with the sheer cost of AI training, has elevated FinOps from a back-office accounting function to a strategic necessity. Traditional cloud cost management focused on “turning off” unused virtual machines. AI FinOps is about maximizing the “return on compute.”
Because AI workloads run at sustained 90%+ utilization, a 10% inefficiency in code or data orchestration doesn’t just waste money—it consumes power and capacity that cannot be easily replaced. Organizations are moving toward “FinOps as a service” models to gain real-time visibility into GPU utilization, ensuring that high-cost clusters are not sitting idle during data ingestion phases. In this new era, cost transparency is the only way to prevent AI initiatives from cannibalizing the rest of the IT budget.
Enterprise strategy: a more complex provider mix
The emergence of neoclouds makes enterprise cloud strategy more consequential. By integrating neoclouds into their portfolios, organizations can avoid vendor lock-in and ensure access to scarce resources. However, this also introduces a “middleman” risk.
When hyperscalers fulfill GPU quotas by subcontracting to neoclouds, enterprises may have signed a contract for the stability of a Tier 1 brand, but their critical AI workloads are actually running on a much younger company’s infrastructure. Organizations now carry an indirect dependency on a neocloud’s performance, security, and financial stability.
The advisory imperative
The decision of where to place AI workloads is fundamentally a governance decision. Hybrid architectures, distributed computing, and on-premise modernization are not regressions—they reflect a mature understanding of AI’s operational implications.
Hyperscalers continue to anchor the market through core services, while neoclouds function as elastic capacity layers. The question for IT and procurement leaders is how to position themselves across both. This means moving away from single-vendor dependency and building procurement strategies that account for GPU availability as a variable.
Cloud Latitude works with organizations to evaluate cloud strategy without the bias of a single vendor relationship. If AI demand is rewriting your roadmap, let’s talk. No cost, no commitment.
