Morgan Stanley has argued that the late-June selloff in AI-linked equities reflected a technical market reset rather than deteriorating demand for computing capacity, while identifying electricity-connected data-center sites as the most constrained asset in the sector. The bank’s July 27 report, “Playing the AI Infrastructure Dip,” estimates that the United States could face a 38-gigawatt gap between projected data-center demand and identified power supply for 2026 through 2028.
That assessment places Bitcoin miners and other operators with large, energized sites in a potentially valuable position. Morgan Stanley estimated that converting existing mining locations to AI data-center use could provide 10 to 19 GW of interconnection capacity within one to three years, far faster than many conventional grid-development schedules.
The report named TeraWulf, Cipher Mining, HUT 8, Riot Platforms, Applied Digital, and Galaxy Digital among companies it views as “Powered Shell Providers” — firms controlling land, buildings, and, most critically, electricity connections that can support high-density computing. Morgan Stanley said the group traded at enterprise values of roughly $2 to $4 per watt, compared with 20 to 25 times EV per watt for mature data-center operators. It applied a 15-times EV-per-watt target multiple to the category.
Power access reshapes the AI infrastructure debate
Morgan Stanley’s central argument is that AI infrastructure is increasingly limited by the time required to secure usable power rather than by the availability of capital for new projects. A “powered shell” refers to a site that has the physical structure and electrical connection needed to install computing equipment, avoiding the lengthy process of acquiring land, navigating grid interconnection, and building supporting infrastructure.
The bank put U.S. data-center electricity demand at about 68 GW over 2026–2028. Against that figure, it counted 15 GW of capacity under construction and another 15 GW of contracted grid capacity, leaving a 38 GW shortfall before additional solutions are considered.
Grid interconnection queues can run for five to seven years in some regions, according to the report. Those delays are compounded by shortages of electricians, welders, plumbers, and other skilled workers needed to build and connect facilities. Morgan Stanley grouped those obstacles under “3P”: people, power, and politics.
Political pressure is also becoming part of the construction equation. The report noted that several U.S. states have paused, amended, or reconsidered tax incentives for data centers as communities and regulators weigh the cost of new transmission infrastructure and rising local electricity demand.
At the federal level, Morgan Stanley cited House consideration of the Ratepayer Protection Act. The proposed legislation would require utilities to consider a “large load standard” designed to direct more grid-upgrade costs toward major electricity users such as data centers. The bank linked the proposal to a White House Ratepayer Protection Pledge signed in March by Amazon, Google, Meta, Microsoft, Oracle, and xAI.
Mining sites offer a faster route to capacity
Bitcoin miners built many of their operations around access to large power loads, often in regions with comparatively low-cost generation or underused industrial infrastructure. That history has made some mining campuses candidates for conversion or hybrid use as AI and high-performance computing sites.
Morgan Stanley estimated that fast-deploying gas turbines could contribute 15 to 20 GW of power within one to three years, while fuel cells could provide another 5 to 8 GW. Direct nuclear supply arrangements and mining-site conversions were included among the other potential sources of capacity.
Even after probability-weighting those options, Morgan Stanley’s base case retained a 1 GW shortfall, while its downside scenario produced an 11 GW gap. The analysis suggests that a data-center developer with an existing grid connection may be able to negotiate from a stronger position than a competitor starting with undeveloped land, even if the latter has greater financial resources.
The conversion process is not automatic. AI data centers generally require more sophisticated cooling, networking, redundancy, and building design than a conventional Bitcoin mining operation. A miner’s power contract, interconnection rights, substation configuration, and local permitting status would also shape whether an AI conversion is economically feasible. Yet securing the electricity itself can remove one of the longest steps in a development timetable.
Morgan Stanley sees demand surviving the stock pullback
The bank said the decline in global AI-related stocks since late June was driven chiefly by crowded positions, margin-financing deleveraging, and a reversal in momentum trading strategies. It did not interpret the drop as evidence that enterprise demand for AI services had broken down.
Concerns over corporate controls on employee AI token usage have contributed to the debate. Some companies have introduced spending limits, raising questions over revenue growth at providers of large language models and related services. Morgan Stanley’s enterprise-use-case analysis estimated average labor savings of about $55 for each AI-assisted call, against agent-based workflow costs of $2 to $5 per completed corporate task. The bank calculated that relationship as supporting returns of more than 10 times the cost.
Its infrastructure model also projected rising margins from newer generations of Nvidia systems. Morgan Stanley estimated net margins of about 58% for token sales from Blackwell-based deployments, rising to roughly 80% with Rubin GPUs and 90% with Feynman GPUs. In that scenario, token prices could fall by approximately 75% without eroding margins, allowing lower AI costs to expand use rather than necessarily weaken provider economics.
The report used the Jevons Paradox to address fears that cheaper Chinese AI training methods, including developments associated with Kimi K3, could diminish returns on large U.S. hyperscaler spending. The principle holds that greater efficiency can increase total consumption when lower costs open new use cases and encourage heavier usage.
Morgan Stanley cited Google executive comments that compute needs could double every six months, equivalent to a 1,000-fold increase over five years. It compared that demand trajectory with an estimated 140% compound annual growth rate in Nvidia AI-chip sales between 2025 and 2028, concluding that even an extension of that supply growth for five years would meet less than 10% of Google’s projected demand by itself.
For Bitcoin mining companies, the implication is less about a direct link between Bitcoin’s price and AI spending than about the value of infrastructure assembled for mining. Companies that can document durable power rights, suitable sites, and credible conversion plans may find their assets assessed increasingly alongside data-center developers rather than solely through hash rate, mining margins, and cryptocurrency market cycles.
Power constraints reshaping AI infrastructure? Learn how tokenized energy and crypto rails intersect in this deep dive.
Disclaimer: The content on this page is provided for general informational purposes only and does not represent the views or financial advice of Toobit. We make no guarantees regarding the accuracy or completeness of this information and shall not be held liable for any errors, omissions, or outcomes resulting from its use. Investing in digital assets involves risk; users should independently evaluate their financial situation and the risks involved. For further details, please consult our Terms of Service and Risk Disclosure.
