Mind

Does Jim Chanos Get it Right - A different Perspective

June 20, 2026

The Bottleneck Isn't the Cost Line

The Setup

Watch the video above. Chanos is making a bear case on SpaceX's $75B IPO at roughly $2 trillion valuation, 90 times revenues, negative free cash flow, Starlink slowing - I have no gripe on the financial logic. The "hopes and dreams IPO" framing is to the point. But I keep coming back on data centres and power. Maybe because I know a little here. You know what they say, a little knowledge is dangerous - so I decided to dig in with my handy dandy AI assistant to help me with my research.

The point that I do agree with is that companies trading at 50 -70 times earnings or 30 - 40 times EBITDA because they'll power AI data centres have priced in perfection - this is a very traditional view in my opinion. I am not saying that it is wrong, I am just saying that it misses a critical point that changes the discussion. Maybe this was intentional as he knows his business - I can't nor could I compete with that. Instead, I offer a different perspective of the problem.

I think he ignores the Walmart problem - ok bare with me and you hopefully will see my point.

Walmart Didn't Win on Price

Everyone remembers Walmart as the low-price retailer. That's the headline. What they did was fundamentally different. Walmart won by controlling where they viewed a bottleneck - logistics, supply chain visibility, real-time inventory, the infrastructure that nobody sees but everything depends on. While everyone else was managing storefronts, Walmart was managing the flow. This was a very different approach.

I believe the AI power situation has the same shape.

The financial world is looking at power as a cost line. The people building and operating the infrastructure know it's an availability problem. Those are not the same problem. One you manage with procurement. The other you manage with integration - or you don't build at all.

What the Data Actually Says

The numbers on power demand growth are incredible regardless of what you believe or who you believe.

IEA Base Case: Global data centre electricity consumption roughly doubles to ~945 TWh by 2030, growing ~15% annually — more than four times faster than total electricity demand growth. (IEA, Energy and AI, 2025)

Goldman Sachs: US data centre power demand more than doubles from 31 GW in 2025 to 66 GW by 2027. Data centres' share of total US peak summer power demand jumps from 4.1% in 2025 to 8.5% in 2027. (Goldman Sachs Commodities Research, 2026)

Department of Energy (Lawrence Berkeley National Lab): US data centres consumed ~4.4% of total electricity in 2023, projected to reach between 6.7% and 12% by 2028. (LBNL, 2024)

McKinsey: AI data centre demand grows 3.5x from 2025 to 2030, reaching 156 GW worldwide. (McKinsey, 2025)

The spread in the stated ranges by experts (EPRI historical) and high case (BCG) for US consumption by 2030 is enormous - approximately 206 TWh to 970 TWh. That range exists because nobody knows how fast AI adoption compounds, how quickly chip efficiency improves, or how many projects actually get built on schedule. Goldman itself estimates only 50 to 60% of data centre capacity scheduled for the next one to two years will come online on time.

The forecast range is wild, but the direction is the same.

The Supply Side Has a Physics Problem

Here is where the 2 to 3 year resolution timeline runs into reality.

Data centres take 2 to 3 years to build. Power infrastructure takes something else entirely as you will see. Transmission lines, 10 - 20 years, Grid interconnection - 5+ years, then there's upgrades, using alternative mechanisms for generation.

Transmission lines: 10 to 20 years to permit and build. The US constructed an average of 1,700 miles of new high-voltage transmission annually from 2010 to 2014. By 2020 to 2023, that dropped to 350 miles per year. Over the past two years: 180 miles. (Grid Strategies, 2024; Camus Energy, 2025)

Grid interconnection queues: The median time from interconnection request to commercial operation now exceeds five years. Nearly 2,300 GW of generation and storage capacity was waiting for grid connection as of end of 2024 — more than the entire installed generating capacity of the US power grid sitting in line. (Lawrence Berkeley National Laboratory)

Power upgrades vs. build cycles: Power infrastructure upgrades take 8 or more years versus 2 to 3 years for data centre construction. (BCG, 2024)

Gas turbines: Manufacturers have a five-year backlog. (ITIF, 2025)

High-voltage transformers: Lead times of 2 to 4 years. (Enki Research, 2026)

The regulatory fix assumes the bottleneck is paperwork. Some of it is. But a significant portion is physical. You cannot permit a transformer into existence faster than it takes to build one. You cannot wish a transmission corridor through terrain and local opposition on a two-year schedule when the historical build rate is under 200 miles a year.

Goldman flagged elevated reliability risks specifically in Mid-Atlantic, Mid-Continent, and Northwest markets because generation capacity additions are limited relative to incoming demand. Some regions will simply turn data centres away.

The most acute supply-demand imbalance is now expected to persist through at least 2028 to 2029. (Hanwha Data Centers, 2026)

That is not 2 to 3 years. That is the floor — and it assumes current investment plans proceed on schedule. An estimated $64 billion worth of US data centre projects have been cancelled or delayed since 2023 due to local opposition and permitting battles alone. At least 142 grassroots advocacy groups across 24 states were actively mobilizing against data centre development by 2025. (MMCG, 2026)

This is actually a structural constraint running at multiple levels all at once.

Two Countries, Same Problem, Different Execution

The US and Canada are running parallel versions of this problem and the contrast is important enough to talk about.

The US challenge is scale and fragmentation. Interconnection timelines in PJM — the grid covering the Mid-Atlantic and parts of the Midwest — average 6.3 years from application to commercial operation. The state-by-state regulatory patchwork, federal NEPA reviews, local opposition, and aging infrastructure compound each other. CenterPoint Energy in Texas reported a 700% increase in large load interconnection requests between late 2023 and late 2024, growing from 1 GW to 8 GW in a year. (TD Economics; Camus Energy, 2025)

Data centres drove roughly half of all US electricity demand growth in 2025 — the single largest contributor to the country's power appetite. (IEA Global Energy Review, via Fortune, 2026)

Canada's situation is structurally different but not easier. Canada does have some advantages that are real: approximately 84% of electricity generation is non-emitting, hydro provides significant firm baseload capacity, and the provincial-led model enables more coordinated decision-making in some cases than the US state-federal split allows.

But there are challenges. Every ying has a yang.

Ontario's IESO forecasts data centre load rising to 13% of new provincial demand by 2035, with an additional 5,000 megawatts required to meet growing electricity needs. The IESO flags large "step-load" connections as requiring proactive transmission and interconnection planning — not a standard connection process. (IESO 2026 Annual Planning Outlook; Telehouse Canada, 2026)

Ontario introduced new regulations in 2025 requiring Ministerial approval before data centres can connect to the grid (this may seem insignificant but can actually stop a project - for another post). Alberta capped large load connections at 1,200 MW through 2028 under an interim framework. British Columbia introduced legislation limiting AI and data centre power allocations through a competitive call process. Quebec's Hydro-Québec paused new data centre procurement in 2024 pending its Action Plan 2035. (Osler, 2025; NES Fircroft, 2025)

Provincial governments are not obstructing AI development. They are managing a genuine tension between abundant clean electricity as a competitive advantage and the risk of committing that resource to large, certain, continuous loads while other electrification priorities compete for the same megawatts. That is a reasonable thing to manage carefully. I don't believe that this problem can be resolved within a two-year timeline.

The transmission constraint in the Greater Toronto Area is already material. Developers are being pushed toward behind-the-meter solutions — microgrids combining natural gas turbines, battery storage, and renewables — not because that is the preferred architecture, but because the alternative is waiting in a queue. (Mordor Intelligence, 2026)

And then there is Markham District Energy.

The Answer That Already Exists

Here is a thing that has been working for 25 years that almost nobody outside the industry knows about.

In 1998 an ice storm knocked out power across eastern Ontario and Quebec for weeks. The kind of event that makes people reconsider their relationship with centralized infrastructure. Markham's response was to build a district energy system. A municipal heat network that doesn't depend on the grid staying up. The first anchor customer in 2000 was IBM Canada. A data centre. The logic wasn't complicated: IBM generates heat continuously as a byproduct of computing, the surrounding community needs heat continuously in a Canadian winter, put a pipe between them.

The heat pump project was commissioned in late 2023. A facility that was previously dumping thermal energy into the atmosphere is now heating a city. The data centre's "waste" became a community asset. The community's energy resilience improved and the grid demand for heating went down.

Equinix's TR5 data centre in Markham now exports residual heat to the Markham District Energy network, supplying over 14 million square feet of mixed-use development — more than 9,000 homes, local businesses, two municipal swimming pools, York University's Markham campus, a hotel, and a hospital. (Equinix, Markham Review, 2025)

In January 2025, a delegation from Northern Virginia — home to the largest concentration of data centres in the world, over 265 facilities with approximately 100 more planned — flew to Markham specifically to study the model and assess how to replicate it. (Northern Virginia Regional Commission, 2025)

The world's biggest data centre market is looking at a Canadian city as the template.

If it was a pilot, it no longer is, and since 2000 we have had the answer. Now we just need to scale.

The Demand Side Is Where the Leverage Is

When supply is constrained and building more of it takes a decade, you need to focus on the only variable you can actually move in the short to medium term and that is demand - how much power you consume, when you consume it, and how intelligently you manage it.

Most buildings have never had a proper energy audit. The preliminary audit finds something, a report gets written, a proposal goes to procurement, and somewhere between month four and month fourteen the world has moved on and the data is already stale. (That's if the report isn't just shoved in a desk). If it is used, issues could have exacerbated, funds reallocated, or the interventions get implemented against old, manually acquired data. Ineffective.

I have seen this enough times that it stopped being surprising. The building was running badly and nobody knew, because the only feedback loop was the electric bill — and by the time that lands, the problem is sixty days old.

The version of this that actually moves the needle runs differently: audit to deep audit to design and implementation to real-time monitoring to ongoing optimization as a recurring service. The data intelligence layer - sensors, integrations, live dashboards. This means the model of how a building performs updates continuously, not annually. Something changes in building behaviour and you see it in hours, not on the next electric bill. The gap between an event happening and someone knowing about it is where energy waste, equipment failures, and carbon exposure live. Reducing that gap is not an abstract benefit when a grid constraint can directly affect whether a facility stays operational.

The integration piece is where most operators leave significant value on the table. Building systems are not a single thing. Vertical transport, fire systems, parking, lighting, HVAC, booking platforms, visitor management, all affect energy behaviour and most of them do not talk to each other. When they do, and when that data feeds a common intelligence layer, the picture of how a site actually operates becomes legible in a way that changes what decisions are even possible. This layer sees patterns that are invisible when you are looking at each system in isolation. The Markham model is the same principle at the city scale — systems that were not designed to communicate, made to communicate, and the integration layer is what makes it work.

The dashboard is not a reporting tool. It is a re-engineering of how a site gets operated. Maintenance becomes directed rather than scheduled, equipment problems surface before failure rather than after. The technology watches the systems continuously, no sick days, no vacation, no shift changes, no forgetting to check something, no setting and forgetting. The standard of site management becomes consistent in a way human operations cannot match at scale.

For me this is what changes the conversation, not the percentage that power represents in a revenue model. The ability to compress consumption, extend equipment life, reduce failure risk, and capture the value of what the building is already generating, and doing it as an ongoing managed service rather than a capital project with a five-year payback.

Most ESCOs stop at project delivery. The monitoring, the ongoing optimization, the recurring service layer is where operational intelligence compounds. A building that was audited three years ago and retrofitted is not the same as a building under active management today - one has a fixed intervention, the other a continuously improving one.

The Walmart Argument, Properly Made

I agree with Chanos that a supplier of a 5 to 7% cost line probably should not trade at 50 times earnings.

And as much as I have admired him over the years, I think he's measuring the wrong thing. Or the right thing I guess depending on how you invest :)

Walmart did not win by being cheap. They won by controlling the layer that everything else depended on - logistics, supply chain, real-time visibility into inventory and demand. The companies building stores and filling them with product were price takers. If you dig into this you will see the retail industry specifically started to understand the logistics problem back in the late 90's and early 2000's. Walmart just put their money where their mouth was and built the infrastructure.

In the AI energy context, the question is not who supplies the power. It is who controls the integration between where power comes from and how intelligently it gets used. The utility provides the commodity. The integration layer of building systems connected and communicating with energy intelligence aggregated to a common platform with demand response capability, continuous optimization, waste heat capture and redirected - this is the infrastructure that everything else runs on.

The hyperscalers know this and are not waiting for the grid to solve itself. They are signing direct PPAs, funding nuclear restarts, pursuing co-location with generation, investing in behind-the-meter solutions. They are building their own infrastructure layer because they understand that availability and reliability are not cost lines — they are operating conditions.

The same logic applies at the building scale. The commercial real estate portfolios, industrial facilities, and institutional buildings sitting on constrained grids have the exact same problem. Less capital, more urgency than they know and the same fundamental need to get smarter about what they have before the grid dictates what you can have next.

Markham figured this out in 2000. Northern Virginia is studying it in 2025. The answer has been running for 25 years. It just needs to scale.

What This Actually Is

This is not a bearish call on AI infrastructure as everyone knows the demand numbers which are as accurate as they can be right now, are real.

I just think we need to take a more pragmatic approach to where the leverage is.

Supply-side bets on constrained timelines are priced for outcomes that the physical and regulatory infrastructure cannot deliver on the advertised schedule. Transmission lines that take a decade to permit do not become faster because the market needs them to be. Transformer backlogs do not clear because a valuation model requires it. Provincial grid operators on both sides of the border are managing genuine constraints, not bureaucratic inconveniences.

The demand side - intelligence, integration, optimization, and the capture of what buildings already generate - is ready now. Not in 2028, not after the interconnection queue clears. Now. In buildings that already exist, on systems that already run, with data and heat that are already being generated and mostly wasted.

The organizations that combine supply awareness with demand intelligence and operational pragmatism will be better positioned than the ones waiting for the grid to catch up.

Markham proved it. The question is who builds it next.

Sources

Research for this article was conducted with the assistance of AI. All sources have been verified.