Hyperscalers or Neoclouds, Where AI Infrastructure Value Will Stick

The AI infrastructure debate has sharpened quickly. Meta expanded its CoreWeave commitment by about $21 billion through 2032, CoreWeave added Anthropic, OpenAI deepened its AWS partnership by another $100 billion over eight years, and Anthropic separately signed for multiple gigawatts of next-generation TPU capacity from Google and Broadcom starting in 2027.

Scarcity Built the Trade

Neoclouds win because they solve today’s problem faster than hyperscalers can solve tomorrow’s problem. Bernstein mentioned that self-build offers the best long-term ROI, but neocloud offers low commitment, fast deployment, and a simple way to bridge a capacity gap when customers cannot wait for owned campuses or 15-year lease decisions.

That logic is visible in Meta itself. Morgan Stanley estimates Meta’s recent Nebius and CoreWeave agreements raise 2027 cost of revenue and lower 2027 EPS by about 6%, which is a reminder that outsourced scarce compute can protect growth, but usually at a higher economic cost. Bernstein does not expect easing in data center capacity until 2028, and says that could slip to 2029 if AI demand stays especially strong. In that environment, neoclouds are still the fastest relief valve for customers that need capacity now, do not yet want to commit to self-build, or want to test workloads before scaling more permanently.

Why Hyperscalers Keep the Moat

For equity investors, however, the better question is not who rents capacity this quarter, but who keeps the economics once scarcity normalizes.

Here hyperscalers still have the stronger hand. Morgan Stanley estimates the big four are headed toward about $834.7 billion of 2027 capex, 14% above consensus. Amazon is now modeled for 31% and 33% AWS growth in 2026 and 2027, with backlog near $350 billion after adding the enlarged OpenAI relationship. Microsoft reported Azure growth of 39% and commercial RPO of $625 billion, Google Cloud grew 48% with a $240 billion backlog and 30.1% margin, and Oracle’s RPO reached $553 billion while management said AI cloud demand still exceeds supply. Those numbers suggest hyperscalers are not only spending more, they are increasingly monetizing that spend inside broader, already profitable ecosystems.

How to Express the Theme in Stocks

$Amazon (AMZN.US)$

Morgan Stanley now sees AWS backlog reaching about $350 billion after incorporating the expanded OpenAI relationship, and models AWS revenue growth accelerating to 31% in 2026 and 33% in 2027 while margins stay in the low-to-mid 30s. In other words, Amazon is not just selling AI capacity, it is turning scarcity into long-duration contracted revenue. The main risk is that non-cloud headwinds such as fuel, fulfillment, and retail reinvestment can dilute how cleanly that cloud upside shows through to consolidated EPS.

$Microsoft (MSFT.US)$

Microsoft is still the highest-quality AI infrastructure compounder because its demand base is the most diversified. Azure and other cloud services revenue grew 39% in Microsoft’s fiscal second quarter of 2026, and Bernstein argues that Microsoft is better protected than most even in a potential AI bubble because demand is supported by both OpenAI-related workloads and Microsoft’s own software stack, rather than by one narrow customer cohort. That matters because it reduces the risk that AI demand proves to be a short-lived training spike. Microsoft may not have the sharpest near-term upside narrative, but it probably has the lowest strategic fragility among the hyperscalers.

$Oracle (ORCL.US)$

Oracle is the highest-beta hyperscaler rerating story. Oracle’s fiscal Q3 2026 RPO reached $553 billion, up 325% year over year, while OCI IaaS revenue grew 84%, and management raised FY2027 revenue guidance to $90 billion. Bernstein’s core argument is that Oracle has moved from being a secondary cloud player to a major AI training and inference platform, with OpenAI-related contracts playing a central role in that jump. If Oracle executes, the upside is that the market may still be materially underestimating how big OCI can become over the next several years. The risk is that this story is far more balance-sheet intensive than the market is used to, so any financing, deployment, or timing wobble can create sharp volatility.

$CoreWeave (CRWV.US)$

CoreWeave is still the best pure scarcity trade, but might not be the best long-duration compounding story. The bull case is obvious: Meta just expanded its agreement with CoreWeave by about $21 billion through 2032, and CoreWeave finished 2025 with $5.13 billion of revenue and $66.8 billion of backlog. The bear case is just as obvious: net interest expense was $1.229 billion in 2025, the business remains capital hungry, and Bernstein still rates the stock Underperform because it believes neocloud contracts are the easiest to churn once data center supply starts easing, likely in 2028 or later. So CRWV works best as a high-beta expression of ongoing capacity shortage, not as the safest way to own the full AI infrastructure stack.

$NEBIUS (NBIS.US)$

Nebius is the smaller-cap, earlier-stage version of the neocloud trade. Its advantage is that it gives investors direct exposure to AI cloud build-out before the model is fully mature. Nebius reported 2025 revenue of $529.8 million, fourth-quarter revenue of $227.7 million, and 2025 purchases of property and equipment of $4.066 billion, which tells you how aggressively it is building ahead of demand. That operating leverage can be powerful if capacity shortages persist, but it also means execution, funding, and utilization matter much more than they do for the hyperscalers. In this framework, NBIS is not the “best business,” but it may remain one of the more explosive trading vehicles while the market is still in a land-grab phase.

Hyperscalers still look like the safer long-term compounders, but in a world where AI demand keeps outrunning physical supply, Neoclouds may be the better growth trade for longer.

118K Views