Celestica Positioned as Critical Integrator in AI Cluster Buildout

We are witnessing the foundational buildout of a new technological S-curve. The current AI investment wave is not just about deploying chips; it is the construction of the physical and logical rails that will power the next paradigm. The winners in this race are defined by their position in this infrastructure layer, not merely in the compute layer itself.

The scale of this buildout is staggering. The five largest US cloud and AI infrastructure providers – Microsoft, Alphabet, Amazon, Meta, and Oracle – have collectively committed to spending between $660 billion and $690 billion on capital expenditure in 2026. This figure nearly doubles their combined 2025 levels, marking a sprint to secure the fundamental capacity for the AI era. This isn't a marginal upgrade; it's a full-stack industrial mobilization.

Crucially, this spending requires far more than just chips. The evidence is clear that AI deployment runs on a complex stack of physical and logical components. It demands data center capacity, high-speed networking, optical interconnects, and rack-scale integration. The hyperscalers are building not just compute clusters, but entire ecosystems of real estate, network fabrics, and integrated systems. This creates a multi-layered opportunity for winners beyond the GPU manufacturers.

An even more telling insight is the persistent gap between analyst estimates and actual spending. Consensus figures have consistently underestimated the true scale of AI capital expenditure. For two consecutive years, the consensus estimate for 2025 capital expenditure has been too low. This pattern of underestimation highlights the exponential, often unpredictable, natural of infrastructure buildouts during a technological inflection. It also explains the recent divergence in stock performance, as investors begin to reward companies where this massive capex is demonstrably translating into revenue, while rotating away from those where the spending is debt-funded and earnings growth is under pressure.

The bottom line is that we are in the early, steep part of the S-curve for AI infrastructure. The $690 billion commitment is the capital expenditure required to lay the rails. The companies that master the full stack of data centers, networks, and integration will be the true winners, as they provide the essential, non-negotiable capacity for the entire AI economy to scale.

The Infrastructure Layer: Where the "Non-Company" Winners Are Built

The true winners in the AI infrastructure S-curve are the builders of the physical and logical rails. While the world watches the GPU race, the foundational work is happening in real estate, networking, and system integration. These are the "non-company" winners-asset classes and business models that provide the essential, non-negotiable capacity for the entire AI economy to scale.

Real Estate Investment Trusts (REITs) are at the heart of this buildout, providing the physical land and shell for AI clusters. Companies like Digital Realty Trust and Equinix are locking in multi-year leases with hyperscalers who are pre-committing capacity well ahead of power-on. This creates a stable, dividend-generating income stream tied directly to the AI capex wave. Their REIT structure, which mandates paying out 90% of taxable income, aligns investor returns with the long-term, capital-intensive nature of data center ownership.

Simultaneously, networking suppliers are capturing the non-linear scaling that comes with massive AI clusters. As cluster sizes explode, the bandwidth requirements multiply, creating outsized demand for high-speed interconnects. Specialized firms like Arista Networks are deploying 400G and 800G platforms to build the Ethernet fabrics that connect thousands of GPUs within a single rack and across data centers. Their growth is tied not to chip sales, but to the sheer volume of data that must flow between them.

A more direct example of this systems-level buildout is the rise of electronics manufacturing services (EMS) firms. Celestica is a standout, gaining significant market share by building the physical systems that house AI compute. The company is providing networking solutions for custom AI chip leader Broadcom and counts major players like Marvell, AMD, and Intel among its partners. Its connectivity and cloud solutions segment, which serves the server and storage end markets, saw revenue jump 43% year-over-year in Q3 2025. Crucially, Celestica is cited as having 55% share of custom Ethernet switches, a key component in AI cluster fabrics. This positions it as a critical integrator, turning raw components into qualified, deployable racks.

This shift is accelerating a fundamental reset in enterprise IT strategy. For years, the narrative was to move everything to the cloud. But as organizations move from experimentation to production AI, many are finding that running these steady-state, compute-hungry workloads exclusively in the cloud is unsustainable. Cost, data gravity, compliance, and performance are pulling compute back on-premises. This creates new, immediate demand for integrated systems and services firms like Celestica, as enterprises need help building and deploying their own AI data centers. The infrastructure layer is no longer just about leasing space or buying switches; it's about the full stack of design, integration, and deployment services required to bring AI to life in any environment.

Goldman Sachs' Framework: The Next Phase of the AI Trade

The initial phase of the AI trade, dominated by infrastructure buildout, is now entering a new, more selective chapter. According to Goldman Sachs Research, the next wave will shift focus from the capital expenditure itself to the companies that can actually deploy AI to boost productivity. This represents a critical rotation in investor attention, separating the builders from the beneficiaries.

The driver of this rotation is a clear inflection in the adoption curve. While AI spending by the largest tech companies is expected to decelerate from the previous year, the adoption of AI tools across the broader corporate economy is accelerating. This divergence is already causing significant volatility in the stock market. The average stock price correlation among the largest US tech stocks has collapsed from 80% to just 20% in recent months. Investors are no longer rewarding all big spenders equally; they are rotating away from infrastructure companies where capex is debt-funded and earnings growth is under pressure, and toward those demonstrating a clear link between investment and revenue.

Goldman's framework identifies two primary beneficiaries in this next phase. First are the AI platform stocks that provide the operating systems and development tools for enterprise AI. Second are the productivity beneficiaries-companies across sectors that can leverage AI automation to cut costs and improve efficiency. The bank's analysis points to firms like software services provider EPAM Systems and buy-now-pay-later platform Affirm as potential leaders in this cohort, based on their exposure to AI-driven productivity gains.

Yet the pace and profitability of this entire buildout are now constrained by factors beyond compute power. The critical bottlenecks are shifting to power, land, and permits. As the initial wave of data center construction proceeds, these physical and regulatory constraints will determine which projects are viable and which are stranded. This means the winners will be those who have secured long-term power contracts, own strategic land positions, and have the operational moats to navigate complex permitting processes. The infrastructure layer is no longer just about building racks; it's about securing the fundamental inputs that make them run.

Valuation and the Long Run: Separating Froth from Fundamentals

The investment wave in AI infrastructure is a classic S-curve buildout, and like all such inflections, it carries both froth and fundamental justification. The froth is undeniable. When a single chipmaker commands 8% of the S&P 500, it's reasonable to question if an AI bubble is inflating. Yet the long-term demand for data center capacity appears justified by the sheer scale of commitments from the world's most advanced technology companies. This cycle is different from past hype, underpinned by long-term contracts with the world's most advanced technology companies.

The key differentiator in the coming shake out will be underwriting discipline. Winners will be those who model the profitability of individual projects after the cost of power and capital. As the initial wave of construction proceeds, the critical bottlenecks are shifting to power, land, and permits. This means the math must include the increasingly expensive cost of electricity and the significant capital required to secure it. Business models that fail to account for these hard constraints will not survive.

The outcome of this cleansing will be a market where weak business models are eliminated, but the underlying assets remain valuable. The hard assets being built-power grids, land, interconnects, and the operational excellence to manage them-will form the backbone of the next economy. Past infrastructure cycles show that while some asset prices become inflated and some companies fail, the capacity itself endures and achieves compounding returns. The shake-out will not destroy the rails; it will refine them, leaving only the most resilient operators and the most strategically positioned resources.

Catalysts and What to Watch: The Next Phase of the AI Trade

The thesis of a durable infrastructure buildout now hinges on a few critical catalysts. The next phase will be confirmed not by more spending announcements, but by the execution of massive joint ventures and the market's rotation toward productivity beneficiaries. Watch for these signals to separate the fundamental from the froth.

First, the execution of the Stargate project is a major test. This joint venture between OpenAI, SoftBank, and Oracle aims to mobilize up to $500 billion in AI infrastructure investment by 2029. Its success will demonstrate whether the capital being committed by the hyperscalers can be efficiently channeled into tangible, long-term projects. A slow or stalled rollout would challenge the sustainability narrative, while rapid progress would validate the scale of the opportunity.

Second, the market's rotation will be the clearest signal of the adoption curve accelerating. Goldman Sachs expects AI spending to decelerate from the previous year while company adoption increases. This divergence is already causing volatility, as seen in the collapse of correlation among large tech stocks. The next winners will be those demonstrating a clear link between investment and revenue, not just those with the biggest capex budgets. This rotation will likely favor the AI platform stocks and productivity beneficiaries identified by the bank, creating two-way risk for the broader index.

Finally, the critical constraint is shifting from compute to power, land, and permits. As the initial wave of construction proceeds, these physical and regulatory bottlenecks will determine which projects are viable and which are stranded. This means the math for profitability must now include the increasingly expensive cost of electricity and the significant capital required to secure it. The winners will be those who have secured long-term power contracts, own strategic land positions, and have the operational moats to navigate complex permitting processes. The infrastructure layer is no longer just about building racks; it's about securing the fundamental inputs that make them run.