Where will the next AI bottleneck emerge, and who can retain the profits?

Research note · 6 October 2026

In Essex, an AI data center’s opening depends on electricity arriving on time. The Telegraph reported in July that Nscale had been told supply would miss its anticipated 2027 opening. That is a concrete example of the next possible AI constraint: getting a particular site into service. [1]

Elsewhere, capacity is arriving. CoreWeave added nearly 500 MW of company-defined active power during the second quarter of 2026. Microsoft separately reported adding another gigawatt of capacity in its fiscal fourth quarter. Expansion and local shortages are happening together. [2][3]

Our view is that deliverable power is a plausible leading expansion gate in constrained locations over 2027–2029. Advanced packaging and high-bandwidth memory remain material competing constraints. The evidence points to an uneven deployment problem, with different stages binding at different sites. [4][5][6]

For investors, that leaves three questions: what limits deployment, who earns profits today, and who can sustain returns after competition and reinvestment. NVIDIA’s approximately 66.2% consolidated operating margin gives a striking answer to the second. Our answer to the third remains unresolved. Current scarcity and current profitability are insufficient grounds for a durable-profit ranking. [7]

Power can be available on paper before a site can use it

A cluster needs firm supply, a usable connection and installed electrical equipment at the same location. PJM’s July decision illustrates the distinction. From June 2027, specified new large loads lacking sufficient capacity, and unable otherwise to be served at its reliability standard, face an interim resource-adequacy service requirement. A connection and adequately firm service are separate hurdles. [4]

Equipment adds another clock. Reuters reported that generator step-up transformer lead times exceeded 160 weeks in early 2026, compared with a 143-week average in 2024. Those figures apply to that particular transformer category, but they show why the electrical side of expansion can take years to resolve. [8]

The Essex report puts a project behind the mechanism. Its evidentiary weight is narrower than a utility commissioning record, yet the reported consequence is clear: electricity would arrive too late for the planned opening. Together with PJM’s service conditions and equipment lead times, it makes power delivery a credible constraint in affected locations. [1][4][8]

CoreWeave’s portfolio shows why the stage of development matters. At the end of the second quarter, it reported 1.5 GW of active power against approximately 3.7 GW contracted. Dividing active by contracted power gives approximately 41%; subtracting the two leaves a roughly 2.2-GW difference between portfolio stages. The disclosures leave the causes of that difference unallocated. [2]

Source: CoreWeave Q2 2026 results. Note: Company-defined categories; contracted power includes capacity at different development stages. [2]

We therefore see constraints on the pace and location of growth, while substantial additions continue. Supply responses also deserve attention: Vertiv targeted a doubling of regional chiller production capacity by the end of 2026. That is prospective factory capacity; its contribution to operating sites depends on subsequent delivery and commissioning. [2][3][9]

Packaging and memory keep the power thesis conditional

Solving the electrical problem leaves the question of which components can actually be delivered. The strongest challenge to our power thesis comes from TSMC itself: it said packaging capacity was so tight that it limited customer growth. That is direct evidence of a semiconductor constraint deserving substantial weight. [5]

Micron supplies a second challenge. It reported agreements covering the vast majority of its calendar-2027 high-bandwidth-memory bit supply, with significant year-on-year price increases. Together, these disclosures support a local interpretation: power delivery may limit one site while qualified semiconductor supply limits another. We reject a universal handoff in which electricity simply replaces chips as the bottleneck. [6][5]

The supply response cautions against projecting scarcity indefinitely. TrendForce estimated that HBM’s share of leading suppliers’ DRAM wafer input would rise from 18% at the end of 2025 to 30% at the end of 2027. Those figures describe estimated and forecast wafer allocation; the amount of qualified HBM ultimately delivered remains a separate question. [10]

The observation that would discriminate between the competing explanations is straightforward: are energized, cooled facilities waiting for qualified components? Repeated project-level evidence of that condition would strengthen the semiconductor explanation relative to power. Faster firm connections and equipment deliveries, followed by actual commissioning, would weaken the local power thesis.

Profits are visible today; their durability is unresolved

Physical constraints tell us where deployment can stall. To assess investment returns, we need to follow a different set of observations: operating profit, capital obligations and cash generation.

NVIDIA’s current profit capture is substantial. In fiscal 2027’s second quarter, it reported approximately $89 billion of Data Center revenue. Its consolidated operating margin was approximately 66.2%, calculated by dividing reported operating income of 63,734 by revenue of 96,221 in the same financial-table column. The scope of that margin is the whole company. [7]

TSMC pairs scarcity with a large investment obligation. Alongside its statement that packaging constrained customer growth, it planned a $60–64 billion capital budget for 2026. That combination is central to the investment question: sustaining a scarce capability can require substantial reinvestment. [5]

Broadcom also shows rapid supplier growth, with $16.7 billion of AI semiconductor revenue in its fiscal third quarter, up 221% year on year. The disclosed category is broad; it supplies a revenue observation rather than the product-level economics needed to judge durable returns. [11]

Infrastructure buyers present a different picture. Microsoft reported $19.6 billion of company-wide free cash flow despite $41 billion of company-wide capital expenditures in its fiscal fourth quarter. Heavy investment can coexist with positive corporate cash generation. These are corporate figures, with AI-project returns still an open question. [3]

CoreWeave’s approximately $104 billion backlog at the end of June provides conditional future-revenue visibility, subject to delivery and service availability. Converting those commitments into operating revenue is one step; determining the cash ultimately available after investment is another. [2]

Our judgment is deliberately limited. These observations show current profitability, scarcity, growth and financing capacity. They leave the comparative persistence of returns unresolved. We cannot rank NVIDIA, TSMC, power owners or compute-capacity financiers by their ability to retain profits through the expansion cycle.

Demand is growing, while the economic test remains open

A supply-focused thesis must also survive the possibility that useful demand becomes the tighter limit. The retained observations weigh against broadly exhausted demand. Google’s disclosed API throughput rose from approximately 16 billion to 22 billion tokens per minute in a quarter. Dividing 22 by 16 and subtracting one gives roughly 38% growth in the reported rate. [12]

Microsoft reported more than 30 million paid Microsoft 365 Copilot seats. That adds a payment observation to the usage evidence. Customer returns and fully loaded inference profitability require more than either measure alone. [3]

Our view is that useful demand has not been established as the aggregate binding constraint. Nor can we quantify the profits ultimately available from it. The key change would be deteriorating utilization, renewals and realized pricing despite available energized capacity. That would shift the explanation from the difficulty of delivering supply toward the economics of consuming it.