what are the biggest ai bottlenecks currently and beyond in the US?

The biggest bottleneck depends on whose output matters

A frontier lab needs large, coordinated clusters. A cloud operator needs to commission capacity and sell it. An enterprise needs a process that works after the model encounters its own data, permissions and exceptions. A single ranking conceals those differences: Nvidia describes a constrained supply chain, while surveyed business leaders identify worker skills as their largest barrier to integrating AI into existing workflows. [5] [3]

The ranking changes with the buyer and the stage of deployment.

Sources: utility and reliability disclosures, semiconductor company commentary, enterprise surveys and operator financial statements. Rankings are our judgments. [1] [6] [2] [3] [7]

We place cooling, construction labor, water and local permission below power in the national infrastructure ranking, while recognizing that any can stop an individual project. Prince William County’s rejection of the Dulles Cloud South proposal shows how land-use permission can become the first obstacle. Congressional researchers, meanwhile, identify substantial gaps in public information about data-center water sources. The evidence supports close scrutiny of individual sites rather than a precise national ranking of water-constrained capacity. [8] [9]

Electricity has to arrive at the campus, on time

A power purchase agreement starts a commercial relationship. Operating a campus also requires a connection, suitable substations and transformers, transmission access, and permission to build. Those steps determine whether a purchased accelerator can be switched on. In November 2025, Microsoft chief executive Satya Nadella described the consequence: “you may actually have a bunch of chips sitting in inventory that I can’t plug in.” [10]

The equipment has its own clock. Reuters reported that lead times for generator step-up transformers exceeded 160 weeks in the first quarter of 2026, compared with an average of 143 weeks in 2024. That particular category affects the connection of new generation, illustrating how a campus’s supply problem can extend well upstream of its own electrical equipment. [11]

CoreWeave shows both the constraint and the progress being made. It added nearly 500 MW of active power during Q2, reaching 1.5 GW, against approximately 3.7 GW contracted across its portfolio. Dividing active by contracted power gives about 41%. The remaining pipeline includes future deliveries across a portfolio extending beyond the US; its conversion into operating capacity is the key execution task. [7]

This is why we favor power over chips as the broad US campus bottleneck. The research points in both directions: Nvidia says its supply chain is running flat out, while Microsoft has described difficulty energizing equipment. Those observations fit different sites and stages. Our ranking reflects the long, location-specific path to delivered electricity; it is a judgment rather than a measured national total of output lost to each constraint. [5] [10] [11]

The reference case requires roughly 52 GW more average facility power by 2030 than in 2024.

Source: DOE/LBNL. Annual TWh divided by 8.76 gives average GW; 2024 is a historical estimate, and subsequent values are scenarios. [4]

The reference trajectory rises from 192 TWh in 2024 to 649 TWh in 2030, equivalent to approximately 22.5% annual growth. LBNL attributes 55% of total data-center electricity consumption in its 2030 reference case to AI servers. AI therefore drives a large part of the challenge, but the grid must serve the entire facility and the surrounding economy. [4]

After connections, dependable generation becomes harder to ignore

A connected campus still needs electricity during the hours when the regional system is under stress. That is where the distinction between annual energy and dependable capacity becomes decisive. PJM’s forward auction for 2028–2029 procured 138.3 GW of reliability-adjusted resources, yet finished 6.83 GW below its reliability requirement after the applicable accounting. The result signals a future regional resource squeeze. [12]

NERC places PJM in its high-risk category beginning in 2029, as rising demand, planned generator retirements and uncertainty about replacement resources converge. Our view is that power remains the leading infrastructure constraint beyond 2028, but the problem broadens: more sites need both a connection and confidence that dependable supply will arrive behind it. [6]

Meta’s Louisiana development makes the scale concrete. Entergy’s plan pairs a campus potentially drawing up to 5 GW with seven planned combined-cycle plants exceeding 5,200 MW and approximately 240 miles of proposed 500-kV transmission. The campus requires an accompanying expansion of the power system. Plants and lines have to reach service together. [13]

Flexibility can relieve part of that burden. Google has incorporated 1 GW of demand-response capability into long-term utility contracts. Its advanced-energy executive Michael Terrell explains that the company can “limit or shift a portion of machine learning (ML) workloads.” Where peak demand binds, moving work away from stressed hours can help a utility accommodate more computing. The capacity is contracted flexibility, with its value dependent on when and how it can be called. [14]

That is a meaningful counterweight to the power-first thesis. So are CoreWeave’s rapid additions. We give these developments substantial weight, particularly over the longer horizon. But the case for relief ultimately rests on commissioned equipment, available fuel and workable operating arrangements at specific campuses. [7] [14] [13]

More accelerators still need memory, packaging and networks

Power does not make the semiconductor constraint disappear. Nvidia chief financial officer Colette Kress told investors in August: “we expect supply to remain a bottleneck at least through the end of fiscal year 2028.” Jensen Huang described the entire supply chain as challenged. Our view is that the useful unit of analysis is the complete system: accelerator, advanced packaging, high-bandwidth memory, networking and rack. [5]

High-bandwidth memory, or HBM, is a particularly important shared input. Micron said on September 30 that it had completed agreements for the vast majority of its expected calendar-2027 HBM supply. It also anticipated industry DRAM bit shipments growing in the low-20s percentage range in both 2027 and 2028 while supply remained constrained. Strong volume growth and tight availability can coexist when demand grows faster. [2]

The manufacturing burden helps explain why. TrendForce’s corresponding end-2026 estimates assign HBM 22% of DRAM wafer input but 9% of bit supply. Comparing wafer input per bit with the remaining DRAM implies approximately 2.85 times the manufacturing intensity: (22/9) divided by (78/91). The estimate describes competition for fabrication resources; HBM’s value comes from the performance those bits deliver beside an accelerator. [15]

Alternative suppliers broaden the choices without removing every shared constraint. Broadcom reported $16.7 billion of AI semiconductor revenue in fiscal Q3, up 221% year on year. AMD’s Q2 Data Center revenue reached $6.7 billion, including both EPYC processors and Instinct GPUs. Its agreement with Meta envisages up to 6 GW of GPU deployments over multiple years. These businesses expand the routes to compute, while the systems still require memory, packaging and power. [16] [17] [18]

The constraint also changes with the workload. Large training runs require thousands of accelerators to work together efficiently. Inference can instead run into memory capacity, bandwidth and latency limits. A vLLM experiment using FP8 KV-cache delivered 14.9% higher output throughput in its specified serving workload. Software changes can therefore release useful capacity from existing hardware, although the gain depends on the model and requests being served. [19] [20]

We regard geographic semiconductor concentration as a strategic exposure rather than the leading explanation for current US deployment delays. Domestic manufacturing is progressing in stages: TSMC’s first Arizona fab entered high-volume N4 production in Q4 2024, while its second fab targets N3 volume production in the second half of 2027. Resilience improves as actual production and its supporting supply chain arrive. [21]

For enterprises, reliable work is scarcer than model access

Completing the infrastructure answers what can run. It leaves open what should be entrusted to it. Our view is that proprietary data access, workflow integration, evaluation and worker skills are more immediate constraints for many businesses than obtaining another increment of compute. Deloitte’s surveyed leaders identified insufficient worker skills as their largest barrier to integrating AI into existing workflows. [3]

Adoption remains uneven. The Census Bureau found recent AI use hovering between 17% and 20% of US businesses from December 2025 to May 2026. Among respondents to McKinsey’s international survey, 37% reported a positive AI contribution to organizational EBIT. The populations differ, but together they show the distance between availability, use and measured financial benefit. [22] [23]

Reliability creates an additional cost: someone must detect and repair mistakes. Stanford’s 2026 AI Index reports 66.3% accuracy on OSWorld for the cited 2025 agent performance, up sharply from roughly 12%. That is substantial progress on structured computer tasks, with a substantial residual failure rate. An enterprise deploying autonomous work must decide which mistakes are tolerable and how the others will be caught. [24]

The return depends on the whole process. In METR’s narrow randomized study, experienced open-source developers using early-2025 AI tools took 19% longer to complete their tasks. The lesson we draw is the importance of measuring completed work after review and correction. The study’s older tools and specialized participants make it an instructive case, rather than a verdict on today’s coding products. [25]

We are more skeptical of a universal, imminent training-data ceiling. Epoch’s 2024 estimate put the effective stock of quality- and repetition-adjusted human-generated public text at roughly 300 trillion tokens. That is a meaningful constraint on one route to scaling. The implications depend on progress in synthetic data, other modalities and alternative training methods. For enterprise adoption today, obtaining permission to use the right internal data and evaluating the resulting system is the more actionable problem. [26] [3]

Legal requirements, local permission and organizational accountability can all slow individual deployments. We do not rank regulation as the leading national constraint on the evidence available. The more immediate enterprise test is whether a useful workflow remains useful once its exceptions and oversight costs are included. [8] [3] [25]

Capital can become the next bottleneck even while demand grows

Useful deployment must eventually support its capital cost. Our view is that financing already matters sharply for specialized clouds and smaller entrants, while cash-rich incumbents retain more room. The buildout’s growing burden makes investment returns a plausible dominant constraint later in the decade.

Across Microsoft, Amazon, Alphabet and Meta, calendar-Q2 cash purchases of property and equipment totaled approximately $165 billion, against $172 billion of operating cash flow. Summing the comparable cash figures and dividing spending by operating cash flow gives 96%, leaving roughly $6.7 billion before other cash uses. These are consolidated global businesses, including activities beyond AI; the comparison measures the funding burden borne by the operators. [27] [28] [29] [30]

Equipment spending absorbed most of the four operators’ combined quarterly operating cash flow, with substantial differences between companies.

Sources: company disclosures. Calendar Q2 2026, global consolidated cash flows; equipment purchases exclude noncash lease additions. Microsoft’s inputs are rounded. [27] [28] [29] [30]

CoreWeave exposes the economics more directly. In Q2 it reported $2.58 billion of revenue, $1.51 billion of adjusted EBITDA and a $49 million operating loss. Its reported cash equipment purchases of approximately $6.42 billion exceeded operating cash flow of $679 million by roughly $5.74 billion. The gap between EBITDA and cash available after investment is where financing becomes indispensable. [7]

There is considerable operating sensitivity beneath those figures. Holding Q2 operating costs fixed, a 20% revenue reduction would take CoreWeave’s operating loss to approximately $564 million: the existing $49 million loss plus $515 million of lost revenue. That fixed-cost stress assumes no cost response and abstracts from contractual protections; it shows why realized utilization and pricing matter so much to capital-intensive providers. [7]

The strongest financial counterargument is the incumbents’ earning power. AWS’s Q2 operating income of $16.6 billion on $42.2 billion of revenue implies a 39.3% operating margin. Google Cloud’s $8.81 billion on $24.8 billion implies 35.6%. Those broader cloud businesses provide substantial profits, and Alphabet still generated $53.3 billion of trailing-twelve-month free cash flow. We therefore see mounting discipline on investment, rather than an imminent universal funding stop. [28] [29]

Financing relationships also run backward through the supply chain. Nvidia’s OpenAI partnership included an intention to invest up to $100 billion progressively as capacity is deployed. Its separate $6.3 billion CoreWeave capacity agreement supports demand for a customer’s infrastructure. These arrangements can facilitate expansion while linking supplier returns more closely to downstream success. [31] [32]

Scarcity rewards suppliers, while operators carry the utilization risk

The clearest current profit capture sits in selected scarce inputs. Nvidia reported $89 billion of global Data Center revenue in fiscal Q2 2027 and a 75% company-wide gross margin. Broadcom’s rapidly growing AI semiconductor business supplies another route to accelerators and networking. These sales occur upstream of the operators’ eventual return on installed capacity. [33] [16]

Networking and optics participate as clusters expand. Coherent’s Data Center and Communications revenue increased from approximately $1.02 billion to $1.62 billion between fiscal Q4 2025 and Q4 2026. Its broader segment scope includes communications, but the growth illustrates the hardware surrounding the accelerator. Investors need to follow the whole system to see where spending goes. [34]

Dominion and Entergy sit on another scarce input: the path to delivered electricity. The utility opportunity comes with capital requirements and cost-allocation questions. Meta says its work with Entergy aims to ensure that other consumers are not paying its costs. Whether that holds depends on the executed tariffs, collateral and protections against abandoned investment. Operators and their financiers pay first; poorly allocated residual costs can reach other utility customers. [1] [13]

Our investment conclusion is deliberately narrower than a stock recommendation. Control of scarce inputs supports orders and pricing. The eventual shareholder return also depends on the price paid for that exposure, the capital required to supply it, and the capacity that competitors add. We have established bottleneck exposure, not completed a valuation case.