Goldman Sachs published a research note in May 2026 that landed differently from most AI infrastructure commentary. Not because the bank suddenly turned bearish on AI demand, hyperscaler spending, or long-term compute growth. The note stood out because Goldman reframed the discussion around a different question entirely. Instead of asking how much infrastructure is being built, Goldman focused on what determines whether those assets will actually generate durable returns.
That distinction matters because much of the AI infrastructure market is still being valued through scarcity-era assumptions. Over the last three years, investors have rewarded almost any business tied to GPUs, power capacity, or data center expansion. Goldman’s argument was that the next stage of the cycle will look very different. The winners will not necessarily be the companies building the most infrastructure. They will be the operators building the highest-quality infrastructure.
Power Access Matters More Than GPU Count
Goldman identified power quality as the first major variable separating durable infrastructure assets from speculative capacity expansion. Specifically, the bank argued that markets are still treating very different power arrangements as though they carry equivalent risk profiles. A campus with long-term grid-connected power agreements has fundamentally different economics from one dependent on behind-the-meter gas generation, temporary power structures, or volatile wholesale pricing exposure. The headline megawatt figure may look identical. The operating risk does not.
This distinction is becoming increasingly important as AI workloads shift toward persistent inference demand rather than episodic training cycles. Recent projections from the International Energy Agency suggest AI-related electricity demand could become one of the fastest-growing sources of industrial power consumption over the coming decade. Inference infrastructure requires predictable uptime, stable operating economics, and long-term reliability. Grid pressure is already emerging as a major bottleneck across multiple global markets, including Northern Virginia and parts of Europe. Dominion Energy has repeatedly warned that accelerating data center expansion is increasing strain on regional transmission and interconnection infrastructure. Goldman’s framework suggests that the market still has not fully priced the difference between infrastructure that merely exists and infrastructure that can operate economically under long-term AI demand conditions. Relevant reporting from Data Center Dynamics, Utility Dive, and the International Energy Agency all point toward the same conclusion. Broader projections around AI infrastructure investment and compute demand have also been explored by McKinsey and BlackRock in recent infrastructure analyses.
Interconnect Infrastructure Is Being Systematically Underpriced
The second variable Goldman emphasized was interconnect infrastructure. This is arguably the most underweighted component of current AI infrastructure valuation models because the market still tends to focus on raw compute deployment rather than network quality. During the first phase of the AI cycle, training workloads dominated demand. Training is relatively tolerant of latency because jobs run over extended periods and can operate at large geographic scale. Inference changes the economics completely.
Inference workloads increasingly require low-latency access to enterprise systems, applications, and end users. That makes carrier density, fiber access, and cloud interconnection structurally more important. A facility with abundant power but weak network positioning can still support training clusters, but it becomes materially less competitive for high-value inference workloads. Goldman’s analysis effectively argued that the market is still undervaluing network gravity and interconnection density as long-term infrastructure moats. That helps explain why companies such as Equinix and Digital Realty continue to maintain strategic advantages tied to carrier ecosystems and enterprise connectivity. That aligns closely with arguments already explored in The New Due Diligence: How Investors Are Rethinking Data Center Valuation, which examined how institutional investors are increasingly reevaluating AI infrastructure through operational-quality metrics rather than raw capacity metrics alone.
Contract Quality Is Becoming a Defining Variable
Goldman also highlighted customer contract quality as a central determinant of infrastructure durability. This is particularly important because many AI infrastructure businesses emerged during an extraordinary scarcity environment where almost any available GPU capacity could command premium pricing. That environment allowed transactional compute businesses and enterprise-grade infrastructure platforms to appear more similar than they actually were.
As supply expands, however, revenue quality becomes increasingly important. The difference between month-to-month GPU rental demand and multi-year enterprise infrastructure contracts is substantial, especially in a higher-rate financing environment. Long-duration agreements improve utilization visibility, stabilize refinancing risk, and create more predictable operating economics. Recent infrastructure analysis from BlackRock and McKinsey has similarly argued that AI infrastructure investment is increasingly becoming a long-duration capital allocation story rather than a short-term technology cycle. Goldman’s research strongly implied that markets are still underpricing the difference between volatile spot compute demand and durable enterprise infrastructure relationships. That framing also supports the broader argument made in The Neocloud Sector Is About to Find Out Which Business Models Actually Work, which argued that many neocloud operators will eventually face a separation between sustainable infrastructure businesses and commodity compute providers.
The Market Is Transitioning From Scale to Quality
The broader implication of Goldman’s note is that the AI infrastructure market is entering a new phase. During the early buildout cycle, expansion speed itself created value because demand consistently exceeded supply. Investors rewarded larger GPU deployments, larger campuses, and larger financing rounds almost automatically. Goldman’s analysis suggests that this framework is beginning to break down as infrastructure differentiation becomes more visible.
The assets most likely to hold long-term value are increasingly those where power certainty, interconnect quality, and customer durability combine into a stable infrastructure platform. Commodity GPU rental businesses and weakly differentiated capacity expansion projects may struggle once scarcity premiums normalize. Goldman effectively told investors that they are still looking at AI infrastructure through the wrong lens. The next stage of the market will not be defined by who built the most infrastructure. It will be defined by who built infrastructure capable of remaining economically valuable after the scarcity phase ends.
Enterprise AI Demand Is Becoming More Persistent
What makes this shift particularly important is that the market is simultaneously moving from experimentation toward operational deployment. During the first phase of enterprise AI adoption, companies were primarily testing models, running pilots, and experimenting with internal workflows. Infrastructure demand was therefore concentrated around training bursts and temporary compute needs. That type of demand rewards whoever can deliver GPUs the fastest. It does not necessarily reward operators building durable infrastructure businesses.
The next phase looks materially different. Enterprises are now embedding AI directly into operational workflows, customer service systems, internal productivity stacks, software products, and data environments. That creates persistent inference demand rather than temporary experimentation cycles. Persistent demand changes infrastructure economics because reliability becomes more important than availability alone. The operators positioned to benefit most are not necessarily those with the largest clusters today. They are the operators capable of delivering stable latency, predictable pricing, long-duration uptime, and integration into enterprise ecosystems over multiple years.
Scarcity Markets Hide Weak Infrastructure
That distinction also changes how investors should think about risk. Scarcity markets tend to compress perceived differences between good assets and weak assets because nearly every available resource monetizes. Mature infrastructure markets do the opposite. Weaknesses become visible very quickly once supply normalizes. A facility with poor network density, unstable power economics, or weak customer durability can suddenly look structurally disadvantaged even if it appeared valuable during the expansion phase.
This is one reason why infrastructure investors are increasingly paying closer attention to operational metrics that previously received far less scrutiny. Questions around utility relationships, transmission access, refinancing exposure, fiber density, and customer concentration are becoming central to investment analysis rather than secondary considerations. Goldman’s research note effectively formalized this transition into a mainstream institutional framework.
Capital Markets Are Becoming More Selective
The implications for private capital are also significant. Infrastructure funds, sovereign investors, and pension-backed capital vehicles have poured enormous amounts of money into AI-related assets over the last several years. Much of that investment activity assumed that demand growth alone would protect long-term valuations. Goldman’s framework suggests that assumption may no longer be sufficient. Capital is likely to become more selective as investors increasingly separate infrastructure assets into quality tiers rather than treating all AI exposure as structurally equivalent.
That repricing could happen gradually in some areas of the market and very abruptly in others. Commodity-style GPU rental businesses remain particularly exposed because their differentiation is often limited once supply constraints ease. Operators with stronger enterprise integration, superior interconnect ecosystems, and durable power agreements are more likely to maintain pricing strength over time. In practical terms, that means the gap between premium infrastructure assets and weaker assets may widen substantially during the next stage of the cycle.
The Repricing Extends Beyond Data Centers
The broader market implications extend beyond neoclouds and data center operators themselves. Hardware suppliers, utilities, networking providers, and infrastructure financiers all sit downstream from these same economic realities. If infrastructure quality becomes the defining valuation variable, then capital allocation across the broader AI ecosystem also changes. More money flows toward power-secure campuses, network-dense facilities, and enterprise-oriented operators. Less money flows toward speculative capacity expansion without durable economic positioning.
Goldman’s Core Argument Is About Asset Quality
Goldman Sachs did not argue that AI infrastructure demand is slowing. If anything, long-term demand projections across the industry continue to rise. What the bank challenged was the assumption that demand growth alone guarantees durable value creation for every infrastructure participant. That is a far more important distinction than most market commentary currently acknowledges.
The AI infrastructure market is therefore becoming more comparable to traditional infrastructure sectors where asset quality determines long-term returns more than expansion speed alone. Airports, ports, telecom towers, and energy networks all eventually moved toward quality-based valuation frameworks once their initial growth phases matured. Goldman’s analysis strongly suggests AI infrastructure is beginning the same transition now.
For investors, operators, and enterprise customers, the message is ultimately the same. The market is moving beyond the phase where simply participating in AI infrastructure creation guarantees strategic value. Power quality matters. Network positioning matters. Contract durability matters. Integration depth matters. Those variables increasingly define whether infrastructure assets behave like durable long-term platforms or temporary scarcity trades.
That is the real significance of Goldman’s note. The bank was not merely updating investors on another phase of AI growth. It was signaling that the analytical framework used to evaluate the entire sector is starting to change. The infrastructure businesses that survive the next phase of the cycle are unlikely to be the ones that simply built aggressively. They are far more likely to be the ones that built intelligently.
