On Investors’ Minds

AI data center ROI, construction risks, and regulatory headwinds discussed

12 min Listen to the original Episode page
AI InfrastructureData CentersCloud ComputingSemiconductorsRegulatory RiskPower GridInvestment ReturnsConstruction

Summary

Speaker_0 analyzes the profitability of AI data centers, citing estimates that owning infrastructure yields over 40% returns compared to 25% for renting. The episode highlights that while inference activity generates high margins, overall company profitability depends on covering R&D and training costs. Construction capacity in the US has risen significantly, but risks include electricity shortages, community pushback, and regulatory delays. Despite a major cloud provider lowering its 2026 guidance due to supply delays, the speaker forecasts 2027 spending could reach $1.3-$1.4T. The discussion concludes that stricter requirements may filter out speculative projects, favoring large cloud providers and shifting demand to regions like Singapore and Malaysia.

Key takeaways

  1. 1
    US Daily Center capacity under construction rose from 25.3 gigawatts at the end of 2025 to 37.7 GW gigawatts in June 2026. an increase of around 49%.

    Unidentified speaker (speaker_0) · 5:48

Episode sections

  1. 0:00 – 1:26

    Introduction and Profit Potential

    Speaker_0 discusses recent volatility in US and APAC tech stocks due to ROI concerns and US political changes. The episode outlines three key questions: whether investments will pay off, what might slow construction, and the implications for investors.

  2. 1:26 – 3:09

    Data Center Returns and Ownership Models

    Estimates suggest a 1GW data center could yield ~30% annual returns, with a hypothetical example showing $39B capex for $23B revenue. Sales-side research estimates owning infrastructure yields >40% returns vs ~25% for renting, with 75% of incremental revenue becoming operating profit for owners versus 30% for renters.

  3. 3:09 – 4:11

    Inference Profitability and Spending Outlook

    Inference activity is estimated to generate ~80% profit before other costs, though company-wide profitability depends on covering R&D and training costs. Infrastructure spending is expected to grow through Q3 2026, with one major cloud provider lowering its 2026 guidance from $175B to $170B due to supply delays, not weak demand.

  4. 4:11 – 5:06

    2027 Spending and Construction Risks

    The speaker believes the market underestimates 2027 cloud provider spending, which could reach $1.3-$1.4T. Risks include electricity shortages, building approvals, and construction issues that may delay projects and reduce returns, even if demand remains strong.

  5. 5:06 – 5:48

    Community Pushback and Regulatory Actions

    Post-midterm election concerns in the US include community opposition to data centers due to electricity, power reliability, and water use. There were >400 state/local actions to pause developments since Jan 2026, compared to 45 in 2023-2025. Most are temporary pauses, but rules may increase costs and timelines.

  6. 5:48 – 7:00

    Current Construction Status and Solutions

    US data center capacity under construction rose from 25.3GW (end of 2025) to 37.7GW (June 2026), a ~49% increase. Solutions include hyperscalers funding grid infrastructure, community investment programs (~$1B each by two large tech companies), and exploring self-generation of power to reduce grid reliance.

  7. 7:00 – 8:03

    Future Project Risks and Capacity Filters

    Projects planned for 2028+ face greater risks from restrictions, while underway projects are less exposed. Tougher requirements may filter out speculative capacity; >430GW proposed in the US, but only 59% is in site selection/approval/construction phases. This favors large cloud providers with strong finances.

  8. 8:03 – 11:43

    Geographic Shifts and Semiconductor Demand

    Demand may shift to more open states/countries or involve self-generation. This is positive for Asian semiconductor/memory chipmakers and economies like Singapore and Malaysia. The key challenge is turning investment into working centers and lasting profits, requiring active stock picking.

Claims worth verifying

A one gigawatt data center could earn an annual return of roughly 30% on the money invested.

Unidentified speaker (UNKNOWN) ·1:26

How to verify: Check JP Morgan Asset Management sales-side research or internal white papers cited in the podcast; verify against industry benchmarks for data center IRR. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

A hypothetical 1GW data center assumes $39 billion in upfront spending, $23 billion in yearly revenue, and $12 billion in yearly operating profit after tax.

Unidentified speaker (UNKNOWN) ·1:37

How to verify: Verify against the specific JP Morgan research note or presentation slide referenced by the speaker for the 1GW model assumptions. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

Sales-side research estimates that AI developers owning their infrastructure could earn returns above 40% on their invested money.

Unidentified speaker (speaker_0) ·2:19

How to verify: Check JP Morgan Asset Management sales-side research reports on AI infrastructure ownership vs. renting models. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

Roughly 75% of each extra dollar of revenue could become operating profit for owners, compared to 30% for renters.

Unidentified speaker (speaker_0) ·2:19

How to verify: Verify against the specific margin assumptions in JP Morgan's sales-side research comparing owned vs. rented AI infrastructure. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

One major cloud provider lowered planned spending for calendar year 2026 from $175 billion to $170 billion.

Unidentified speaker (UNKNOWN) ·3:49

How to verify: Check the 2026 capital expenditure guidance or earnings call transcripts of major cloud providers (e.g., Microsoft, Amazon, Google) for Q3 2026 or late 2026 updates. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

The planned spending for 2027 seems to be around $1 trillion, but this could reach $1.3 to $1.4 trillion in total.

Unidentified speaker (UNKNOWN) ·4:26

How to verify: Compare JP Morgan's 2026 infrastructure spending forecast against consensus estimates from other investment banks or the aggregated guidance of major cloud providers. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

More than 400 state and local actions to pause developments of data centers since January 2026, compared with only 45 across all of 2023 to 2025.

Unidentified speaker (speaker_0) ·5:23

How to verify: Search public records, news archives, or regulatory filings for local government resolutions, zoning pauses, or community opposition actions in the US between Jan 2026 and Nov 2026. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

Primary source

US data center capacity under construction rose from 25.3 gigawatts at the end of 2025 to 37.7 gigawatts in June 2026.

Unidentified speaker (UNKNOWN) ·5:48

How to verify: Check industry reports from firms like JLL, CBRE, or specialized data center tracking agencies (e.g., Synergy Research Group) for Q2 2026 construction pipelines. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

Two large tech companies have each announced roughly $1 billion in community investment programs.

Unidentified speaker (speaker_0) ·6:45

How to verify: Check press releases or corporate sustainability reports from major hyperscalers (e.g., Microsoft, Amazon, Google, Meta) for community investment announcements around 2026. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

More than 430 gigawatts of US data center capacity has been proposed, but only 59% is moving through site selection, approvals, or construction.

Unidentified speaker (UNKNOWN) ·7:39

How to verify: Verify against JP Morgan's or another analyst's data center pipeline report detailing proposed vs. active projects in the US as of late 2026. For recent events, verify 2026-08-10 through 2026-11-10. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify the $39B capex and $23B revenue assumptions for a 1GW data center against industry reports from firms like McKinsey or BCG.
  • Check the specific cloud provider's Q3 2026 earnings call transcript to confirm the guidance cut from $175B to $170B and the stated reason (supply delays vs demand).
  • Review the list of >400 state/local actions to pause developments cited for 2026, comparing them against public records or news archives for 2023-2025 to validate the volume increase.
  • Analyze the 59% conversion rate of proposed US data center capacity (430GW) to active projects using industry trackers like CBRE or JLL.
  • Assess the financial health of large cloud providers versus smaller players to evaluate the claim that stricter requirements favor well-capitalized firms.
  • Investigate the specific grid infrastructure funding commitments and community investment programs (~$1B each) mentioned by the two large tech companies.
  • Research the regulatory landscape in Singapore and Malaysia regarding data center incentives and power availability to assess the geographic shift claim.
  • Compare the claimed 40%+ return on owned infrastructure against historical real estate or infrastructure fund performance data to validate the yield differential.
Full transcriptexpand
0:00
Unidentified speaker (speaker_0): Over the past few weeks, we've seen increased volatilities in both US and APAC technology stocks, on the back of questions over a return on investment, and also the new political setup in the US that could impact on data center build out. So what comes next for AI investment? Demand for computing power remains strong and
0:19
Unidentified speaker (speaker_0): potential profits are attractive. The challenge is to build the data centers needed to meet this demand and make a lasting profit. Today we ask three questions. Could the investment pay off? What might slow the construction? And what does this mean for investors?
0:41
Unidentified speaker (speaker_0): Hi, and welcome to On Invest Minds. I'm Tai Huei, the chief market strategist for Asia Pacific at JP Morgan Asset Management.
0:47
Unidentified speaker (UNKNOWN): And
0:48
Unidentified speaker (speaker_0): thank you for giving us a few minutes of your time to understand what investment
0:50
Unidentified speaker (UNKNOWN): is.
0:50
Unidentified speaker (speaker_0): investors are thinking about right now and how that fits in with building the right portfolio. In this episode, we are going to look ahead to the next stage of AI investment: the potential profits, the spending outlook for 2027, and the challenges facing new projects from 2028 onwards. We will discuss how electricity
1:10
Unidentified speaker (speaker_0): supply, building approvals, and community support could affect where. when data centers open and when the investments start to pay
1:19
Unidentified speaker (UNKNOWN): off.
1:20
Unidentified speaker (UNKNOWN): So
1:20
Unidentified speaker (speaker_0): let's start with the potential profits. What could launch AI data center earn once it opens?
1:26
Unidentified speaker (UNKNOWN): It
1:26
Unidentified speaker (speaker_0): is estimated that a one gigawatt data center could earn an annual return of roughly 30% on the money invested.
1:34
Unidentified speaker (UNKNOWN): A
1:34
Unidentified speaker (speaker_0): gigawatt is a measure of power capacity.
1:37
Unidentified speaker (UNKNOWN): An
1:37
Unidentified speaker (speaker_0): example assume. assumes $39 billion in upfront spending, $23 billion in yearly revenue, and $12 billion in yearly operating profit after tax.
1:48
Unidentified speaker (UNKNOWN): These
1:49
Unidentified speaker (speaker_0): are estimates based on assumptions, not actual results.
1:52
Unidentified speaker (UNKNOWN): The
1:52
Unidentified speaker (speaker_0): question ahead is whether developers can deliver those returns while managing the huge upfront costs.
1:59
Unidentified speaker (UNKNOWN): The
1:59
Unidentified speaker (speaker_0): choice between owning and renting computing power could also affect future profits. Sales-side research estimates that AI developers owning their infrastructure could earn returns above 40% on their invested money. It also estimates that roughly 75% of each extra dollar of revenue could become
2:19
Unidentified speaker (speaker_0): operating profit. For developers renting computing capacity, their estimates are around 25% in after-tax after tax return. a bit lower, with roughly 30% of each extra revenue dollar becoming operating profit.
2:33
Unidentified speaker (UNKNOWN): The
2:33
Unidentified speaker (speaker_0): return measures are not directly comparable, and none of these estimates are guaranteed, but they do show why the choice between owning and renting could matter.
2:43
Unidentified speaker (UNKNOWN): Now,
2:43
Unidentified speaker (speaker_0): leading AI model developers are using a mix of rented computing power and infrastructure of their own. Earning more could give them greater control and potentially improve returns, but it would also require more money up front and bring greater risks if the project runs late or over budget. Their future profits would depend on getting that
3:03
Unidentified speaker (speaker_0): balance right and earning enough revenue from a computing capacity they use.
3:09
Unidentified speaker (UNKNOWN): Now
3:09
Unidentified speaker (speaker_0): there's another important question. How much of the profit from using AI will reach the company's bottom line?
3:15
Unidentified speaker (speaker_0): When a trained AI model answers a request, the activity is called inference. Estimates suggest that roughly 80% of each extra dollar of revenue from this activity could become profit before other company costs. Now these costs include training new models, research and development, and building infrastructure.
3:34
Unidentified speaker (UNKNOWN): So
3:35
Unidentified speaker (speaker_0): AI services could be profitable, but that does not mean the whole company will be. The challenge ahead is to cover all those costs and still earning a lasting profit.
3:46
Unidentified speaker (UNKNOWN): So,
3:46
Unidentified speaker (speaker_0): what could happen to spending from here?
3:49
Unidentified speaker (UNKNOWN): Now
3:49
Unidentified speaker (speaker_0): we expect that spending on infrastructure continues to grow faster through the third quarter of 2026. One major cloud provider lowered this planned spending for calendar year 2026 from $175 billion to $170 billion.
4:03
Unidentified speaker (speaker_0): The reason behind that was a delay in capacity supplied by other companies, not due to weaker demand.
4:10
Unidentified speaker (UNKNOWN): Now
4:11
Unidentified speaker (speaker_0): further delays could push spending into later years without changing companies' appetite to invest. Looking into next year, 2027,
4:20
Unidentified speaker (UNKNOWN): we
4:20
Unidentified speaker (speaker_0): believe that the market markets is underestimating how much the largest cloud providers could spend on infrastructure.
4:26
Unidentified speaker (UNKNOWN): The
4:26
Unidentified speaker (speaker_0): planned spending for now seems to be around $1 trillion, but this could reach $1.3 to $1.4 trillion in total. Again, we have to face the reality that some areas such as electricity shortages, building approvals, and construction problems could push some of the spending into later years. And that's not the
4:46
Unidentified speaker (speaker_0): same as customers losing interest in AI. But these delays will still have a cost. If a data center takes longer to build or costs more than expected, revenue could arrive later and investment returns likely to fall. Strong demand does not remove these risks of getting projects built. After the midterm elections, there are great
5:06
Unidentified speaker (speaker_0): concerns that data center build-out in the US could be delayed. Building approvals and community support will help to determine how quickly. new data centers can open. Communities are concerned about electricity bill, reliable power supplies, and water use.
5:22
Unidentified speaker (UNKNOWN): A
5:23
Unidentified speaker (speaker_0): recent calculation shows that more than 400 state and local actions to pause developments of data centers since January 2026, compared with only 45 across all of 2023 to 2025.
5:37
Unidentified speaker (UNKNOWN): Many of
5:38
Unidentified speaker (speaker_0): these are temporary pulses, not permanent bands. The rules and agreements that follow could affect how long future projects to take and how much they will cost.
5:48
Unidentified speaker (UNKNOWN): We
5:48
Unidentified speaker (speaker_0): should not forget the fact that a large amount of capacity is already being built. US Daily Center capacity under construction rose from 25.3 gigawatts at the end of 2025 to 37.7 GW gigawatts in June 2026. an increase of around 49%.
6:05
Unidentified speaker (speaker_0): That shows construction has continued despite the challenges. Looking ahead, the question is how quickly these facilities can open and how tougher rules could affect the next round of projects. One possible solution to this deadlock is determining who pays for extra electricity infrastructure. Energy regulators want
6:25
Unidentified speaker (speaker_0): hyperscalers to make stronger. financial commitments so that households do not have to pay those added costs. Hyperscalers are also offering to fund infrastructure and support local communities. Two large tech companies have each announced roughly $1 billion in community investment programs. Some hyperscalers are also exploring
6:45
Unidentified speaker (speaker_0): their own power generation to reduce reliance on stretched electricity networks and grids. These steps could help projects to go ahead, but there will be added costs and also time spent negotiating impacting on potential returns.
7:00
Unidentified speaker (UNKNOWN): So,
7:00
Unidentified speaker (speaker_0): what does this mean for the next phase of AI investment and also for investors? First, projects planned for 2028 onwards do face greater risks from new restrictions. Projects that are already underway seem to be less exposed. For future developments, building approvals, power supplies, and access to funding could play
7:20
Unidentified speaker (speaker_0): a bigger role in deciding which project goes ahead when they open and how much they earn. Ultimately, high-quality companies with stronger balance sheets are likely to be in a better position. Second, tougher requirements could reduce the risk of building too much capacity based on uncertain demand.
7:39
Unidentified speaker (UNKNOWN): It
7:39
Unidentified speaker (speaker_0): is estimated that more than 430 gigawatts of US data center capacity has been proposed, but only 59% is moving through site selection, approvals, or construction.
7:50
Unidentified speaker (UNKNOWN): A
7:51
Unidentified speaker (speaker_0): proposed project is not a completed data center. Tougher requirements could favor large cloud providers with stronger finances and hyperscalers whose projects have a clearer path to profits.
8:03
Unidentified speaker (speaker_0): Third and finally, investment could move elsewhere. Stronger demand for compute power could draw new projects to states or countries that are more open to developments. Hyperscalers could also generate their own electricity, although that would add costs. A change in where or when projects are built does not necessarily mean a
8:23
Unidentified speaker (speaker_0): change in demand for AI. This is actually good news. for semiconductors and memory chipmakers in Asia, as well as a number of Asian economies who are already participating in the data center build-out, such as Singapore and Malaysia.
8:37
Unidentified speaker (UNKNOWN): So
8:38
Unidentified speaker (speaker_0): looking ahead, the AI story will depend on more than how much companies promise to spend.
8:43
Unidentified speaker (UNKNOWN): The
8:43
Unidentified speaker (speaker_0): potential returns are attractive, but developers will need to finish projects, earn revenue, and manage the costs. Lower spending in the near term does not always mean weaker demand, and strong demand does not remove the risk of delays, rising costs, or difficulty in raising money.
9:00
Unidentified speaker (UNKNOWN): The
9:00
Unidentified speaker (speaker_0): key question is whether companies can turn their investment into working data centers and those data centers into lasting profits.
9:09
Unidentified speaker (UNKNOWN): And
9:09
Unidentified speaker (speaker_0): this requires require significant active selection and stock picking.
9:13
Unidentified speaker (UNKNOWN): So
9:13
Unidentified speaker (speaker_0): thanks for listening. I hope you're
9:14
Unidentified speaker (UNKNOWN): going to be able to do it.
9:14
Unidentified speaker (speaker_0): you found this useful. If you do, please share this with your friends and colleagues and of course consider subscribing. Any questions? Feel free to reach out to your JP Morgan Asset Management Client Advisor. I'm Taikohi. I'll see you next time to discuss what's on investments. Meanwhile, stay informed and stay invested.
9:34
Unidentified speaker (speaker_1): This content is a general communication being provided for informational purposes.
9:39
Unidentified speaker (speaker_1): only. It is educational in nature and not designed to be taken as advice or a recommendation for any specific investment product, strategy, plan, feature, or other purpose in any jurisdiction, nor is it a commitment from JP Morgan Asset Management or any of its subsidiaries to
9:59
Unidentified speaker (speaker_1): participate in any of the transactions mentioned herein. Any examples used are generic, hypothetical, and for illustration purposes only. This material does not contain sufficient information to support an investment decision, and it should not be relied upon by you in evaluating the merits of investing in any
10:19
Unidentified speaker (speaker_1): securities or products. In addition, users should make an independent assessment of the legal, regulatory, tax, credit, and accounting implications and determine, together with their own financial professional, if any investment mentioned herein is believed to be appropriate to their personal goals, investors
10:39
Unidentified speaker (speaker_1): should ensure that they obtain all available relevant information before making any investment. Any forecasts, figures, opinions, or investment techniques and strategies set out are for information purposes only based on certain assumptions and current market conditions and are subject to change without prior
10:58
Unidentified speaker (speaker_1): notice. All information presented herein is considered to be accurate at the time of production, but no warranty of accuracy is given and no liability in respect of any error or omission is accepted. It should be noted that investment involves risks. the value of investments and the income from them may
11:18
Unidentified speaker (speaker_1): fluctuate in accordance with market conditions and taxation agreements, and investors may not get back the full amount invested. Both past performance and yields are not reliable indicators of current and future results. JP Morgan Asset Management is the brand for the asset management business of JP Morgan.
11:39
Unidentified speaker (speaker_1): Chase and Company and its affiliates worldwide.
This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights