Power shortages will delay more US AI data-center capacity than chip shortages through 2028.

Assessment

The evidence supports the claim more strongly for US data-center deployment than for global AI capacity, but not strongly enough for high conviction. The key distinction is between a chip being manufactured and a US facility being able to bring that chip online.

Why power appears more likely to dominate US delays

Morgan Stanley identifies power availability as the main current bottleneck to US data-center development. It also argues that, if US power is insufficient, globally demanded chips may be deployed in regions with excess power after a one-to-two-year delay. [6] This is directly aligned with the claim: power scarcity can delay or redirect US capacity even when the underlying chips exist.

TD Securities reports that its channel checks found utility-power procurement lead times of 2.5–7 years across 22 US markets. Its electricity-runway analysis forecasts reliability-rated power shortfalls in Northern Virginia in winter 2027 and New Albany, Ohio, in summer 2028. [7] Those lead times overlap the requested horizon and are structurally difficult to solve through short-term equipment purchases alone.

The deployment bottleneck is also visible in current industry reporting. A September 2026 report says Nvidia views power availability as a key barrier to getting AI server chips online as soon as they ship. [4] That is important because it describes a failure point after chip shipment: available hardware can remain unused if the data-center site lacks power.

Goldman Sachs forecasts that only about 50%–60% of data-center capacity scheduled for the next one to two years will come online on time amid delays and cancellations. [3] This does not attribute all slippage to power, so it should not be read as a power-only estimate. It does, however, indicate that a large share of the near-term buildout is exposed to execution constraints during the period leading into 2028.

EPRI's 2026 scenarios project US data centers will consume 9%–17% of national electricity by 2030, compared with 4%–5% today. [2] This is a demand projection rather than a direct delay estimate, but it indicates the scale of the power-system expansion required if the buildout proceeds.

Why chip shortages remain a serious counterargument

The IEA reports that high-end chip packaging constrained production in 2025 and that global high-end memory production became a binding constraint on AI-server production in the second half of 2025 and the beginning of 2026. [1] It reports that memory prices rose by an order of magnitude between late 2024 and early 2026 while inventory fell from almost four months to around three weeks. [1]

The IEA expects the AI-memory shortage to last at least until late 2027 and says current HBM production can support a maximum of around 25 GW of AI-ready servers per year through 2027. [1] This is a material chip-side ceiling that could delay or cap AI-server deployment independently of grid availability.

A separate manufacturing-industry analysis projects that 30%–50% of planned 2026 data-center capacity could slip to 2028. [5] This is a meaningful challenge to the thesis, but it is a projection rather than an observed outcome and does not provide a clean US-wide decomposition between chip shortages, power, permitting, financing, or other causes.

Bottom line for the 2028 horizon

On the available evidence, power is the more persuasive explanation for delays to US site commissioning and utilization, while chips remain a comparable constraint on the amount of AI-server hardware that can be produced. The claim should therefore be treated as directionally supported but conditional, not as an established quantitative fact. The main unresolved issue is attribution: public sources identify both bottlenecks but do not yet show, across the same US project sample, which one causes more delayed megawatts by 2028.