No such titling-and-numbering chicanery will be tolerated on this blog. This is just straight-up Part Two, the sequel to Part One (which you can find here).
Quick Recap of Part One
Many, many market observers are projecting rapid, sustained growth in electricity demand through the next decade, driven by the growth of AI data centers. In order for these projections to come true:
Paying demand for AI products and services must skyrocket.
The compute efficiency of AI training and inference must only increase slowly, if at all.
The energy efficiency of compute hardware must only increase slowly, if at all.
There are several pretty good reasons to suspect that one or more of the three prerequisites above will not actually occur. Even if all three do occur, it’s not clear that the resulting AI-driven electricity demand growth would form an outsize share of total electricity demand growth.
In This Part Two We'll:
Hear from an industry expert on a few additional reasons to doubt that aggressive AI electricity demand growth will occur.
Understand what we do (and don't) know about how Goldman Sachs created their AI electricity demand forecast.
Discuss how AI-driven demand, if it does spike, figures into topline electricity demand growth through 2030.
Explore what energy project developers can do to limit the risks of building assets to meet AI demand.
Let’s get into it.
Someone Way Smarter and More Sciencey Than Me Lays Out the AI Bear Case
Rocky Mountain Institute cofounder and Stanford lecturer Amory Lovins, a near-mythical figure in the world of energy efficiency, renewables, and environmental policy, recently published a very well-written and comprehensive piece of analysis with a Hall-of-Fame title: “Artificial Intelligence Meets Natural Stupidity.” In the piece, which I encourage you to read in full, Lovins makes several key observations that cast doubt on hyper-bullish AI electricity demand growth forecasts. My takeaways from his writing:
A. Future electricity demand growth tied to AI is highly uncertain, driven by countervailing trends of increasing AI training and inference usage, and increasing per-query energy efficiency, both of which have rates of change that may prove to be exponential, linear, or something else (asymptotic)?
B. The revenue model to drive continued growth and eventual profitability for many large AI players is still unclear, as consumer and business willingness-to-pay for AI solutions will eventually depend on the extent to which AI applications meaningfully enhance productivity beyond the existing novelties of LLM text and image editing/generation services.
C. The hype around AI for energy markets is reminiscent of similar demand growth optimism during the dot-com boom, which drove the construction of many coal and gas plants that ended up being underutilized when these rosy growth forecasts did not materialize.
D. Firm/dispatchable power for data centers is less critical than popularly held, as many data centers have significant untapped load flexibility potential that can be optimized to better match the generation profile of renewables.
E. AI and data centers more broadly are just one piece of the electricity demand picture: AI only accounted for 8% of global data center electricity use in 2023 and is projected to reach 11% in 2025, and data centers' share of total electricity use was just 4% in the US and 1.5% globally in those years.
Items A and B map pretty well onto points that I made in Part One regarding the future trajectory of the three key unit conversions required to calculate AI electricity demand:
AI Tasks/Year * Compute/AI Task * Energy/Compute = Annual AI Electricity Consumed.
(See Part One Recap above or click back to the full Part One blog here for more on this.)
But there's far more to explore here, as Items A and B also provide a good jumping-off point to dig into how market observers like Goldman Sachs develop their load growth forecasts. Then, in Items C and D, Lovins raises additional points that I didn't consider at all in Part One. Finally, Item E returns to the idea that all this AI data center demand growth could just be a drop in the bucket of overall electricity demand – a topic that I alluded to in Part One but didn’t have space to cover in detail. Let’s lump Items A and B together, then take on Items C-E one-by-one.
Items A and B: Growth Assumptions and Forecasting Uncertainties
Lovins has this to say about upward bias in AI power demand forecasting:
Even as the world’s largest tech companies invest hundreds of billions in new data centers amid an AI-led arms race, there is broad consensus that most AI projects requesting power will never come to fruition:
Developers often consider multiple sites for a single project.
Utilities are bombarded with speculative power requests from landowners hoping to flip properties to data center firms. (Requesting power is usually free.)
Big tech firms “spray” power requests across regions to hedge power-supply lags and their own planning uncertainties.
Developers may reserve more capacity for initial model training than they’ll need later for inference, to reduce competitively damaging delays in time-to-market.
The speculation and duplication issues Lovins identifies here indicate savvy forecasters should understand that when they see 10 GW (let’s say) of data center load requests in utility queues, they should expect only perhaps 1-2 GW of those requests to become reality:
“Conservatively, you’re seeing five to 10 times more interconnection requests than data centers actually being built,” said Astrid Atkinson, a former Google senior director of software engineering and now co-founder and CEO of grid optimization software provider Camus Energy.
(Quick aside: I worked with Astrid and the Camus team as a pre-MBA intern in Summer 2021 – they are doing great work to help utilities and project developers build a cleaner, more resilient, and more flexible grid.)
With this in mind, let’s revisit Goldman’s projections. They see total US electricity demand (all sectors) growing at a 2.4% CAGR through 2030, and they expect 90 bps of that growth to come from data centers. They also say that data centers will account for 8% of US power demand by 2030, vs. 3% in 2022. How are they arriving at these projections? First off, they helpfully note:
A key question impacting compute demand is whether that demand is:
Pent-up (i.e., available new servers will be bought regardless of budget),
Not pent-up and constrained by demand itself, or
Constrained by customer budgets.
In other words, [as hardware becomes more powerful and energy-efficient per unit cost,] will customers buy equal amounts of the more powerful servers as they would the less powerful ones? (PDF p. 15)
The answer to this question has huge implications for forecasting resulting electricity demand. If AI players continue to buy GPUs in quantities that represent the minimum of
(as much compute as is available to buy from GPU manufacturers)
and
(as much compute as the AI companies can afford to buy),
and they operate all of it at very high utilization rates over a long period of time, then the impact on electricity consumption of the sheer growth in compute deployed and used will far outweigh any per-unit-compute energy efficiency gains. On the other hand, if the big AI players decide they only need to grow their compute capability by a fixed amount in a given year, and they can do this by buying fewer total units of a more efficient new GPU model that uses, per unit, twice the energy to achieve eight times the compute speed of the previous model (i.e. a 75% reduction in energy use per unit compute), then the marginal energy demand growth from each new unit of compute installed would be one-quarter that of the prior hardware, with electricity demand growth correspondingly weaker as a result.
Which side of this does Goldman take? Well, they hedge a bit:
We have applied both a server supply-driven forecast and a compute speed demand-driven forecast, with a heavier weight applied towards the supply-driven methodology. (PDF p. 16)
I think this is fair, although I would like more specifics on the actual weighting applied. In the near term, it’s hard to see any meaningful incentive for AI companies to show capex discipline, so a heavier weight on supply-driven forecasting makes sense. People are shouting from the rooftops that “capex is the moat,” and that whatever Nvidia GPUs someone like Google doesn’t buy, someone like OpenAI will gladly snap up instead to gain an advantage. But in the long run, when we consider the $600B AI revenue gap that David Cahn wrote about last year (as discussed in Part One), there is no reason for all of these players to be shelling out to build additional AI data centers whose combined capacity far exceeds the actual compute needs of their AI products. When I say “actual” needs, I mean both short-term ability to serve every ChatGPT and Claude and Gemini query that comes in and long-term ability to provide a higher tier of service (whether via faster speed, higher-quality outputs, or both) to businesses and consumers who are actually willing to pay for the service. This is in line with the commentary of Lovins, Atkinson, and others suggesting that future AI electricity demand is exaggerated, based on the relatively low share of proposed data center capacity that will actually be built, the extent to which capacity built is actually utilized by inference after the initial model training rush, or both.
Consider this thought experiment: If OpenAI had a bigger-than-DeepSeek breakthrough and released a suite of models tomorrow with exactly the same performance as the existing GPT-3/4/5 variants, except that each of these new models used one-billionth (1E-12) as much compute per query as the old versions, would they still be trying to build, buy, or contract with new data centers faster than every competitor? For me, the obvious answer is no – unless you think Sam Altman just wants to make everybody else blow all their cash so that the rest of the competition goes out of business, like the Reagan Administration funding bullshit space laser projects so the Soviets would bankrupt themselves trying to match.
That’s certainly possible, but I view it as unlikely. Sure, it’s convenient for the big players that there’s a high cost of entry for smaller competitors, hence some of the panic that followed the DeepSeek announcement in January 2025, when it looked like a much smaller AI player had found a way to cut by 90% the capex required to train an equally good model. However, if your goal is to create a sustainable business model and you’ve just destroyed the competition on cost per outcome such that you can charge your subscribers orders of magnitude less and still make gargantuan profits that justify your stratospheric valuation, why keep lighting all that money on fire for compute you don’t need?
Obviously, the “billion-times-better” hypothetical is extreme, but the point is to show that capex for compute is not an absolute good in and of itself – it’s only worthwhile if you utilize it fully enough in the future to achieve a satisfactory return on invested capital, or at least enough to meaningfully advance your R&D efforts toward helping you one day achieve said return, or if it’s truly a moat that keeps out other competitors by ensuring you can train bigger and more sophisticated models than they can (which I think is a pretty bad bet).
So, revisiting Goldman’s assessment, the question for me is not whether to assume either that GPU demand is:
A) Infinite, while supply is rate-limiting for deployment, or
B) Finite and simply has not yet been satisfied by existing supply.
Instead, it’s when we switch over from state of the world A to state B that matters – for the AI industry in general, and especially for projecting AI electricity demand growth. I expect that this switchover will occur sooner rather than later for the publicly traded companies, with heavyweights like Google and Microsoft under pressure to demonstrate that all this capex is creating real shareholder value in the form of revenue-printing AI products and services. Privately held players like Anthropic and OpenAI may have a longer leash, but having raised tens of billions to date, with current valuations in the hundreds of billions, they too will eventually need to start raking in some real cash to justify IPOs and deliver a return on the massive growth capital they’ve received.
Credit to Goldman where it’s due: While they don’t provide specifics about how exactly they’ve probability-weighted their forecasts with respect to GPU supply- vs. demand-driven projections, they are quite transparent about another key assumption, energy use per unit of compute, which is the final unit conversion in the AI energy consumption equation I cited above. They see chip energy efficiency growing at just 2-4% annually through 2030, above the trough value of 1% observed in 2022, but well below the high-single-digits to low-teens growth rates between 2015 and 2020.
This assumption looks questionable to me. Sure, there’s a consistent down-trend in the chart (below) from 2016 onward, but using a possible outlier 2022 value to re-base the forecast through 2030 is likely to introduce a downward bias. I haven’t been able to find 2023 and 2024 actuals, but if those actuals show a bounce-back in efficiency gains to, say, 5-10%, then we could probably adjust upward our expectations for efficiency gains through 2030, and correspondingly adjust downward our AI electricity demand growth forecasts.
https://www.goldmansachs.com/pdfs/insights/pages/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf, PDF page 18
The Goldman forecasters’ methodology is not quite so clear for the other two terms in the AI energy demand equation, which are:
Revenue-generating AI usage per year
As I discussed in Part One, it’s naïve to think that broad-based AI usage will continue to accelerate much beyond the next few years if no one can make any money from it, because the supply side has to get paid. It’s also hard to see usage growing as rapidly once everyone has to pay for the service, because right now we are functionally at what economists call the “giveaway quantity,” the amount demanded when a product is offered for free, and charging any nonzero price (Econ 101 claims) will always reduce quantity demanded relative to the giveaway quantity.
As revolutionary as AI products are or could become, they will still obey this economic law – I don’t see people declining to use AI tools for free because they’d rather pay for the same tool instead. Of course, the major AI players are starting to roll out subscription plans and limit non-paid access, but this points to the idea that total active users will soon decline, or at least level off. The big question: Will each paid user use the product enough to offset the reduction in usage from free users who now find their usage capped? It would have been helpful to understand how Goldman’s analysts see this playing out in their base case, but the report doesn’t mention the underlying assumptions around market adoption of AI products and services themselves, which makes it hard to understand how they are arriving at the energy demand forecast.
Compute per unit of AI output
The Goldman report doesn’t engage at all with the concept of AI compute efficiency – not energy consumed per unit compute, but units compute consumed per AI output produced. In fairness, the report I’m referencing came out in April 2024, ten months before DeepSeek’s apparent great leap forward (ha) in training and inference compute efficiency challenged the prevailing thesis that more compute means better models. In a post-DeepSeek world, it’s less clear that this should be the case, and therefore more important than before to be explicit about how much more (or less) compute we expect the AI models of tomorrow will require per unit of output produced, compared to today’s models.
Given the lack of detail on these two critical drivers of AI electricity demand, it’s hard to evaluate the rigor of Goldman’s forecast, as they’re only sharing detailed assumptions on the third driver (Energy usage per unit compute). More transparency would allow us to dig into the quantitative factors baked into their forecast and better understand what you need to believe in order to agree with their growth figures.
Item C: AI as Dot-Com Bubble 2.0
I mentioned in Part One that all trends continue until they can no longer, citing examples like Dutch tulips, mortgage-backed securities, and skinny jeans. Lovins’s piece draws a direct parallel between another such example, the dot-com bubble, and today’s AI boom:
Prof. Jason Bordoff said at Davos, “We could fill this room, and many like it, with reports from the late 1990s about how electricity demand was going to go through the roof because of the internet revolution. None of that ever happened because the technology and the chip productivity kept improving.” Here’s how: …
In 1997–2003, independent and utility generators were deceived by an urgent-demand-growth mirage (including IT) into ordering 225+ GW of new US capacity. (Gas-fired capacity added during 2000–04 totaled ~242 GW—140 GW just in 2002–03.) Most proved unnecessary, and their 12-figure investment disappointed… Power demand took a decade or two to grow slowly into the far-overbuilt supply.
Internet usage doubled in each of many years in the 1990s, but in a few 1995–96 spikes, it doubled in just 100 days. Those anomalies got extrapolated into a media meme, triggering a stampede to overbuild optical fiber networks, with 10× growth just in 1996–2000, much of it debt-financed. In 2002, 97% of the fiber was still dark. It took a decade to start making money, vaporizing billions.
A new RMI review finds that in 2006–23, US utility planners “on average overforecasted electricity demand by 8% in 5-year forecasts and by 17% in 10-year forecasts….The forecast error is even higher for more recent years: data from 2012 to 2023 shows that forecasts were, on average, 23% higher than actuals.”
Pardon the long pull quote, but Lovins says it far better than I could, with the added advantage of having actually lived through that time as an adult, whereas I was still in preschool playing with Tonka trucks in 2000 when the dot-com bubble burst. He continues (emphasis mine):
The biggest source of… load inflation seems to be projections of extremely rapid growth of giant data centers. Those assets combine high ambition to build, high competitive urgency, flexible siting, low load flexibility for AI inference (though bigger for AI training and for crypto), high “flight risk” that the load may later flee or default, highly concentrated risk (e.g. five firms made up 80% of projected data-center demand growth for Dominion Energy Virginia, whose prospects could be upended by just one firm’s strategic shift or business reversal), and uncertain business prospects. For the ponderous, long-lead-time, high-capital-intensity electricity industry, that’s a tough customer risk profile.
This quote expands on another topic that I touched upon briefly in Part One: Stranded asset risk. If energy project developers build projects specifically to serve data center load, they expose themselves to substantial risks with respect to how long this load lasts, what happens if it never materializes at all, and how to recover the sunk costs of energy infrastructure that the target customer no longer needs. More on this later.
Item D: Load Flexibility
Lovins makes the case that rather than assuming data centers’ load profile throughout the day is constant and building tons of new dispatchable generation assets to serve them, we should encourage data centers to shift their compute usage to better align with times when existing electricity supply is cleaner and more abundant:
First and easiest is optimizing when data centers are used, often splitting tasks into more-flexible parts and deferring compute-heavy tasks such as AI training, batch inference, data analysis, or model distillation. Right-timing [of demand] can avoid much new power-supply capacity, and can move parts of AI’s computing to times or places where power is cleaner, cheaper, and perhaps otherwise wasted—thus reducing curtailment and prioritizing marginal renewables. Making data centers a flexible, dispatchable load… could enable utilities to power many new data centers “tomorrow” from existing assets, not in five-plus years from costly new ones.
This actually runs counter to a statement I made in Part One regarding data center demand profiles. I wrote:
For data centers, whose utilization rate is fairly constant, i.e. their instantaneous power draw is relatively steady over time, we can make some basic assumptions that load translates fairly linearly to energy consumption over a certain period of time.
Lovins is asking, “What if we didn’t accept that utilization rate would be fairly constant?” The examples he provides of training and batch inference are indeed areas in which compute can theoretically be re-allocated to times at which excess existing grid supply is sufficient to meet the facility’s electricity needs. It’s a good idea in principle. However, given that AI companies are rushing to lock in unthrottled 24/7 compute to meet future inference demand, it’s not clear that the social benefit from shifting compute activity to use cheaper, cleaner energy from existing assets actually translates to private benefit for thedata center offtaker.
Meta, Google, OpenAI, and Anthropic are sparing no expense to train and deploy their models as quickly as possible to stay ahead of the competition; they would rather have compute available on demand even if they pay more to power it. If and when the AI industry pivots to focus on profitability, with a mandate to reduce costs per model trained and per inference query served, time-shifting data center demand may help limit the amount of new-build generation required to serve AI electricity loads. Until then, I’m less optimistic about this than Lovins is.
Moving on to the final topic…
Item E: How AI electricity demand fits into the overall electricity demand picture
We’ve arrived back at the last what-if question I posed in Part One. If you’re still reading, maybe take a breath and have a glass of water, or crack open an adult beverage of your choice. We’re almost there.
What if AI-driven electricity demand grows as fast as is commonly forecasted, but is still just a drop in the bucket of overall electricity demand?
I’m going to quote myself from Part One, not because I’m self-absorbed but because I’m lazy. Here’s Past Coop:
It’s still highly possible – likely even – that said AI-driven demand growth is still a fairly small piece of the overall electricity demand picture, based on three things:
AI electricity demand as a share of overall data center electricity demand.
Data center electricity demand as a share of overall electricity demand.
The contributions to demand growth of other sectors, like EVs and industrial electrification.
Back to Present Coop: Goldman sees data centers contributing 90 bps (0.9%, for non-finance folks) to a topline of 2.4% total annual electricity demand growth through 2030. However, looking at their visualization, we can see that gross total demand growth split out by sector is actually 2.9%, with a negative 50 bps “other” category bringing us back down to the 2.4% net CAGR. They don’t specify what’s driving this “other” decrease – maybe energy efficiency improvements in agriculture? General utility demand response programs whose electricity savings aren’t attributable to a single sector? Who knows.
The point is that among sectors contributing to the increase in gross demand, data centers only account for 31% of growth, with another 21% each coming from transportation and residential demand. If data centers’ contribution to demand growth is only 1.5x that of two other key sectors, why do data centers attract almost all of the demand growth hype? According to these projections, there will also be significant new opportunities through 2030 to power homes, cars, trucks, etc, so why aren’t we talking about them? Where’s the enthusiasm for these sectors at RE+, at InfoCast, and in industry press and podcasts?
https://www.goldmansachs.com/pdfs/insights/pages/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf, PDF page 20
Drilling down further, Goldman’s base case has AI data centers specifically accounting for 2% of total US electricity demand by 2030, while non-AI data centers are 6%. Sure, AI data centers are a faster-growing segment, but they are also smaller today and will still be smaller in five years, while non-AI data center demand is already strong and continues to grow. So even if you like the story behind overall data center demand growth, why specifically obsess about serving the AI portion of this market? You can argue that bending over backward to serve AI players today sets you up for more growth in the long term – Deloitte expects AI to account for 70% of total data center electricity demand in 2035 – but it still seems strange for energy project developers to give so much attention to one specific kind of large load customer (AI) when there’s a very similar customer profile (non-AI data centers) that’s attracting far less attention but is also growing rapidly.
https://www.goldmansachs.com/pdfs/insights/pages/generational-growth-ai-data-centers-and-the-coming-us-power-surge/report.pdf, PDF page 20
What’s an energy developer to do?
If you’re an energy project developer, and if you think, based on what I’ve laid out here, that there’s even a 10% chance (or 1%, Snyder-verse fans?) that electricity demand growth from the AI infrastructure buildout is substantially lower than all of these very smart bankers and consultants are predicting... What do you do about it?
I see two straightforward answers for developers at risk of “putting all their eggs in one basket” by prioritizing AI demand:
1) Build a stronger, better-cushioned basket
Prioritize serving data center customers for as long as the current boom lasts, but make sure you get sufficient commercial protection built in. What does this mean?
If you’re an energy project developer who’s entered into a 20-year power purchase agreement to sell electricity to a data center, and all of a sudden that data center is underutilized such that it’s buying barely any electricity, or even shuts down entirely, you’re screwed – especially if you’ve built your asset behind the meter, without a grid connection that allows you the option to sell excess merchant energy into the wholesale markets as a backup plan. No matter how creditworthy OpenAI or Anthropic or even Google looks now, their massive spending to construct new data centers doesn’t guarantee that these facilities will be fully utilized, generate sufficient revenue, and therefore demand as much energy as anticipated. And if these companies own and finance their data centers through subsidiaries, it won't be their corporate balance sheet collateralizing your PPA with them. Plus, data center operators are obsessed with speed-to-power, meaning a behind-the-meter asset (or combination of assets) you’re building to serve them is likely only valuable in the very near term. Once they can get a full grid connection in year 4 or 5, they’ll likely be able to get cheaper 24/7 power from a front-of-the-meter PPA or from liquid wholesale markets than from your collocated generation assets.
How can energy projects mitigate these risks? A few possibilities:
Sign higher-priced PPAs for a shorter period of time, reducing the risk of PPA default or termination in the later years if data center utilization rate drops off over time as more compute capacity enters the market.
Insist on a take-or-pay contract structure in which the offtaker is obligated to pay for energy that the project generates or the capacity that the offtaker is entitled to call on command, even if the offtaker does not actually need or use this energy or capacity.
Use a two-tiered contract structure consisting of a PPA and a tolling agreement, under which the PPA provides behind-the-meter energy supply for a shorter period of time until the data center receives its large load interconnection rights, along with a longer-duration tolling agreement to provide callable backup power/peak-hour overage even after the data center gets its grid connection.
Push for high PPA termination fees and long notice periods, so that you have sufficient time to re-contract the asset if the offtaker does decide to exit the PPA early.
Insist on a ParentCo guarantee from your off-taker, so that your PPA is collateralized by their balance sheet (or at least the balance sheet of a reasonable large subsidiary/division) rather than by an AssetCo SPV that contains only the data center itself.
Ensure that your project obtains its own grid connection rather than staying fully behind-the-meter, even if this means filing a separate generator interconnection request after COD, to enable commercial optionality if the initial PPA is later terminated, downsized, or otherwise modified by the offtaker.
Ideally you’d pair two or more of the first five options along with the sixth to combine strong contractual protections from the off-taker with a backup plan for commercialization in the wholesale markets enabled by the physical configuration of the asset, should those contractual protections prove insufficient.
2) Put some of your eggs in a different basket
While it’s reasonable to pursue some projects aimed at serving data center customers, developers should diversify with projects that serve other customer profiles, like:
EV charging and fleet electrification
Heavy industry and advanced manufacturing
Fulfillment centers, warehouses, cold storage facilities, and other logistics real estate
Campuses and multi-building complexes that value energy resilience, like universities and hospitals
Here’s what I like about these four categories of electricity buyers: They all serve needs that already exist and must continue to exist for vast portions of society to function. Putting it in cavemannishly simple terms, we need to get where we’re going, and we also need to extract, make, handle, fix, and transport physical stuff – mine copper, can peaches, assemble airplanes, set broken bones, sort and deliver packages. While the new AI overlords are intent on taking as much of human lived experience out of the physical realm as possible, we are still creatures of the flesh, with all of the physical needs that this simple fact implies.
Even if we consider a hypothetical techno-utopia (dystopia)? in which universal multipurpose 3-D printer ownership allows you to make whatever food, furnishings, toiletries, clothes, medical supplies, and electronics you want or need using a single machine without ever leaving your home, the raw material to feed into such a miraculous omni-printer still needs to be sourced, packaged, and distributed through the physical world to arrive at your door. Whether the nutrients in your 3D-printed and auto-cooked meals come from strawberries, steak, or genetically engineered protein-and-vitamin slush, it’s just not very efficient for each household to try to grow or gather those things on its own. Ditto for durable goods and electronics: Good luck mining cobalt or copper in your backyard. Unironically: Yay for factory farms and open-pit mines! Please feel free to take this line out of context and cancel me.
Based on the idea that these material needs aren’t going anywhere, fully electrifying as much of the physical world’s energy consumption as possible strikes me as a good bet for investors and project developers – certainly less risky than betting that the AI behemoths manage to close the revenue gap for their products quickly enough to pay in full the massive hardware and energy bills they plan to rack up over the next few decades. Whatever you think of AI and its implications for energy use, GHG emissions, and the future of work, we already have a whole physical-world economy for which the transition to cheaper, cleaner power is equally urgent and likely a better investment.
Standard Disclaimers
This blog is written in my personal capacity and reflects only my own thinking, research, experience, and opinions.
This blog post is not sponsored by, endorsed by, or affiliated with my employer, although its content may be informed by some non-confidential aspects of my work.
This blog post is not investment advice and does not constitute any offer, solicitation to offer, or recommendation of any investment product or security.
This blog post is 100% human-drafted and edited. AI tools were not used in any way.
Want to weigh in with your thoughts? Propose a correction or improvement? Find me on LinkedIn or shoot me an email: cwetherbee at alumni dot gsb dot stanford dot edu.
