The grid doesn’t have a demand problem; it has a supply problem.
The dream customer for a utility would consume a massive amount of power at a constant rate, 24 hours a day, with no seasonal variation. The data center is about as close to that as there is, with an enormous appetite for power and a typical load factor of 86%. The load factor is the utilization rate, defined as the average consumption divided by the peak. The load factor for a large electricity grid, such as ERCOT or MISO, is ~64%. A higher load factor results in lower costs for consumers by maximizing the use of all the equipment.
The data center requires significant resources to satisfy its insatiable appetite for electricity, but it also generates enormous revenue that can help a utility maintain economies of scale. Due to its high load factor, it can raise the grid’s load factor, improving the total capital efficiency. Customers like this, in principle, should reduce everyone else’s electricity bill.
But the reality is that utilities are struggling to serve new data centers. Electricity prices have risen in recent years, and the blame is often placed on the growing demand from data centers. But this doesn’t seem to make sense.
Electricity demand in the US grew historically, even as the inflation-adjusted rate for electricity fell. See Figure 1. Rising electricity demand was not a cause of past rate increases; utilities simply built new generators and additional transmission as needed. Demand rose during the Industrial Revolution, the introduction of air conditioning, and the advent of the internet, all while electricity prices fell. Virginia is home to the largest number of data centers in the world, which consume more than 25% of the state’s electricity, yet their electricity rates are below the national average.
Electricity prices started to rise around 2020, and suddenly, we are having difficulty accommodating new demand. Data centers are beginning to abandon the grid to generate their own electricity behind-the-meter (BTM). It’s estimated that 25% to 33% of new data centers will pursue this route by 2030, most of them being powered by natural gas. Why gas, isn’t solar the “cheapest” way to generate power? In the past, it would have been unthinkable for a company to generate its own electricity. It had to focus on its core expertise and leave power generation to the pros.
This article is not about data centers; they are the canary in the coal mine. What has changed?
The four drivers of instability
Over the last several years, most of the new generation has come from intermittent renewable energy sources (solar and wind), while dispatchable generators like gas and coal are being retired. This adversely affected electricity generation in four distinct ways: the missing money problem, the intermittent effect, the dispatchable disinvestment syndrome, and grid queue fouling.
The missing money problem.
Most electricity is sold in the day-ahead wholesale market in 1-hour blocks. Plant owners either bid to supply an hour of power at the minimum price they will accept or self-schedule their plant to run regardless of price. Bids are stacked from lowest to highest until the expected demand for the hour is met. The highest bid is the clearing price for which all bidders are paid, regardless of their bid.
This system worked well historically, but solar and wind power have changed the dynamics, making it hard for dispatchable plants to survive. Wind and solar plants are usually self-scheduled into the market without a bid. This has the effect of bidding zero. Self-scheduling is done for the following reasons:
In many states, it is policy to run solar and wind plants as much as possible to meet emission reduction goals.
Nearly all solar and wind plants have a Power Purchase Agreement (PPA) with a utility or a large company. When the plant is running, the PPA guarantees the plant owner a fixed payment regardless of the clearing price. The plant owner’s goal is thus to run the plant as much as possible to maximize the payments it receives.
Solar and wind farms have zero marginal costs, which means they can run when the clearing price is zero without losing revenue to pay for fuel. Gas, coal, and nuclear plants cannot afford to do this.
Solar and wind farms may benefit from substantial production subsidies. This allows the offtaker of a PPA to receive significant revenue even if the clearing price is negative.
When the weather is sunny or windy, the clearing price can be zero, sometimes even negative. When solar or wind plants bid, it is often negative. The effect is to suppress wholesale energy prices. This is frequently cited as evidence of renewable energy’s low cost. However, the actual cost to society is determined by the PPA price plus all subsidies, not by the wholesale market price. PPA prices are proprietary; hence, public data on them is not generally available. Figure 2 shows wholesale prices for Texas (ERCOT).
Prices for most of the years are at or below $30/MWh. No generator of any kind that depends solely on wholesale prices can be reliably profitable at these levels. The spike in 2021 was due to high gas prices that year. Ironically, the beneficiaries of high gas prices are the offtakers of PPAs with solar and wind plants. They receive the clearing price with zero fuel cost.
Load Serving Entities (LSEs) are reluctant to build new gas plants or sign PPAs with gas merchant plants due to these market distortions and low-carbon policies. LSE is the regulatory term for any organization that provides electricity to the public, typically a utility. There are greater profits in renewables due to generous subsidies, while building new gas plants only reaps criticism from activists and politicians. Analysts sometimes refer to this as the missing money problem.
The intermittent effect
With enough subsidies, intermittent sources have produced a respectable amount of electricity, but the obvious problem is that you don’t necessarily get it when you need it. In most states, there are two major electricity markets: energy and capacity. In the energy market, plant owners sell their power at the highest price they can get. But the energy market cannot guarantee that power will always be available. The purpose of the capacity market is to ensure that power is always there.
In states with capacity markets, regulators require that LSEs have sufficient capacity to ensure demand is always met, even during the highest peak. LSEs can fulfill this obligation either with generators they own or by purchasing the right to have power available from the capacity market. Basically, if a plant owner has spare capacity, they can rent it to an LSE in need for a specific period. The plant doesn’t necessarily have to run, but it must be available to run. The price unit is typically $/MW-day or $/kW-month.
The type of capacity matters. Each generation type is rated by the regulators and assigned a reliability factor known as the Effective Load Carrying Capacity (ELCC). The ELCC is the percentage of the nameplate capacity that can be counted on to supply peak load. For instance, solar and wind power have very low ELCCs. The LSEs are not paying for nameplate capacity; they are paying for effective capacity.
The capacity market is a measure of the grid’s health. If the price of capacity is too high, it means there is a shortage of reliable plants, and if nothing is done, there will be blackouts. Capacity prices in PJM recently soared to a record $329/MW-day, sparking panic in energy markets. See figure 3.
According to Utility Dive: “Gas-fired generation accounted for 45% of the cleared capacity, followed by nuclear at 21%, coal at 22%, hydroelectric at 4%, wind at 3% and solar at 1%, according to PJM. Demand response offered in the auction was essentially flat at 8,010 MW, PJM said.” Note that wind and solar supplied negligible effective capacity.
In California, a similar situation has been brewing in the resource adequacy market. California’s resource adequacy market is synonymous with PJM’s capacity market. Here, the price units are $/kW-month. Since around 2020, resource adequacy prices have risen exponentially.
Texas takes a different approach. They are an energy-only market with no separate capacity or resource adequacy market. During peak demand periods, prices can occasionally spike to as much $5,000/MWh. Dispatchable plants are expected to make profits by selling at high prices during peak consumption.
Problems have materialized here as well. In recent years, due to the low average wholesale prices, dispatchable plants have been unable to remain profitable. To remedy this, in 2023, the state began subsidizing the construction of new gas plants, offering state-backed loans at 3% interest to cover 60% of construction costs. Starting in 2025, Texas can curtail large loads without advanced notice, except for designated “critical load industrial customers” and certain critical gas facilities. These customers now have to provide for their own resource adequacy.
The dispatchable disinvestment syndrome.
The surging capacity markets have created enormous new demand for gas plants. The markets are working as designed, but we have another serious problem. Nearly all gas turbines are made by three companies: GE, Siemens, and Mitsubishi. They have been told over the years that there is no future in gas plants. Solar, wind, and storage are the future, and they should get with the program.
Naturally, no company wants to invest in manufacturing obsolete products. Wanting to be good corporate citizens, they divested from gas turbine manufacturing in favor of windmills and other green technologies. They are now having difficulty supplying this surging demand.
It used to take 1.5 to 3 years to produce a new gas turbine; now it takes closer to 4 to 7 years. Until recently, CAPEX for a new gas plant was around $1,000/kW. There are now reports of prices more than double that. The tight supply paired with soaring demand is causing prices to explode. Figure 5 shows estimates of future prices for combined-cycle gas plants.
Many analysts, misreading the situation, cite the higher turbine prices as more proof that solar, wind, and batteries are inherently cheaper. They are doubling down on pushing for an accelerated deployment of intermittent energy sources, which could risk causing a negative feedback loop.
Utilities, on the other hand, have to deal with reality. They don’t care how the power is generated. They’re happy to use the technology that most pleases their customers as long as it’s reliable. But when prices rise, utilities are the convenient fall guy, and no one wants to hear their excuses.
Grid queue fouling.
Gas turbines remain essential for producing reliable, affordable power; thus, they will play a crucial role in electricity production for many years to come. However, successfully ordering a new gas turbine doesn’t mean you’re out of the woods. We have yet one more serious new development. It used to take 2 to 3 years to navigate the process of getting a new plant connected to the grid. According to Berkeley Labs, the median timeline as of 2023 is 5 years. Today, it may even be higher. This affects all types of plants, regardless of the source.
The grid operator will not allow a plant owner to simply connect their new plant to the grid. The grid operator must first conduct a study to ensure it can handle the physics of the new plant. There are several considerations, for example:
Will the new plant cause grid congestion?
Will the new plant introduce frequency fluctuations?
What kinds of grid upgrades will be required?
Solar and wind farms are more difficult to study because they can introduce frequency fluctuations and power spikes, plus they are located further from load centers. But the main problem is the sheer number of requests for new interconnections. Developers interested in building a new power plant submit an application, which is then entered into the queue.
The number of applications has been growing exponentially over the years. The capacity in the interconnection queue is now double the size of the entire U.S. grid. Some see this as proof that the energy transition is imminent. However, the vast majority of these applications are speculative and will never be built. See Figure 6.
The applications are considered on a first-come, first-served basis. This is a problem because the queue is filled with a massive number of tiny solar farms, each of which requires a study. For example, a 1 MW solar farm in front of a 1000 MW combined-cycle gas plant is processed first, even though the latter could improve grid stability and has 1000 times the nameplate capacity.
In summary, each state has a unique story behind its rising electricity prices. Still, these four factors go a long way toward explaining the recent rate hikes and difficulties of satisfying new demand.
Flexible load
The remedies renewable energy advocates offer to these supply problems are more storage, Virtual Power Plants (VPPs), Distributed Energy Resources (DERs), and new transmission corridors to distant regions with strong solar and wind potential. They claim these solutions will finally lead to lower prices. You can be forgiven for being skeptical of this, as batteries and long transmission lines add significant costs, and DERs lack economies of scale. These solutions will be discussed in a future post.
But another proposal, “flexible load”, has been getting massive attention lately. Flexible load is synonymous with demand response. The idea is that if a large consumer, like a data center, agrees to curtail or shift its power consumption during peak demand periods, its load could be served 99% or more of the time without requiring any new generation or transmission. The consumer would presumably shift this curtailed consumption to periods when power is less constrained. A popular paper co-authored by Tyler Norris explains the proposal.
Norris and his team reviewed years of hourly demand data published by the EIA and determined how much new load could be accommodated, based on curtailment rates ranging from 0.25% to 1% of total demand. If all consumers agree to curtail on the request of the grid operator, here are some examples of how much new load could be accommodated with existing electrical infrastructure:
0.25% curtailment on request can accommodate 76 GW of new load, which is 10% of the US peak demand.
0.5% curtailment on request can accommodate 98 GW of new load, which is 12.6% of the US peak demand.
1.0% curtailment on request can accommodate 126 GW of new load, which is 16.2% of the US peak demand.
In summary, if all consumers could curtail or shift up to 1% of their usage at critical times, this would be equivalent to adding enough new dispatchable generation equal to 16.2% of peak load, along with any required transmission upgrades. However, curtailing on demand is easier said than done. Industrial processes generally require constant energy input. The paper suggests three ways data centers can be flexible:
Temporal flexibility. AI training should be performed only when adequate power is available, leaving AI inferencing unconstrained.
Spatial flexibility. Data center in region A, where power is constrained, could offload some of its tasks to a data center in region B, where more power is available.
Temperature flexibility. Use a temperature storage system to make cold water or ice when power is available, then use it to cool the processors when power is constrained.
Unsurprisingly, the data center operators have not responded favorably to the proposal. They have enough on their hands running a data management enterprise; now, they’re being asked to master the art of electricity management as well. Recall that the load factor of a data center is already very high at 86%. This implies there is not much room for flexibility, and some of these remedies are quite expensive.
Many data centers are already doing these to the extent that it makes sense economically. Additionally, these new costs will give large players like Google and Amazon an advantage over smaller companies, thereby reducing competition.
The sheer excitement of this proposal reveals an internal pessimism for the ability of intermittent energy to accommodate significant load growth. In this new world, instead of buying the electricity we need when we need it, we must adapt to using it when it is available. Instead of the utility bearing the burden of managing complex supply issues, this task is increasingly being offloaded to its customers, potentially resulting in lost economies of scale. And the situation will worsen over time as more dispatchable plants are replaced by intermittent sources.
Demand response programs have their place as they have historically reduced peak demand by ~6.5%. However, they are self-limiting. The price paid rises in proportion to the percentage of consumers able or willing to participate until it becomes more expensive than generation. US demand response growth has been relatively flat over the 5 years reported by FERC, as shown in Figure 7.
The nuclear renaissance
Concurrent with these problems, there is a new enthusiasm for nuclear energy. It’s not likely a coincidence. Big tech, early adopters of solar and wind power as the ideal path to zero carbon, is now desperate for nuclear fission. The illusion of achieving net zero using Renewable Energy Credits (RECs) while their facilities continue to rely on fossil fuels has been shattered.
They are locating their data centers near existing nuclear plants, making deals to restart shut-down plants, and building new plants. There is now even a proposal to complete the unfinished AP1000s at the V.C. Summer site, which began construction in 2013 and was abandoned in 2017.
Microsoft is reportedly paying $110/MWh to $115/MWh for power from the Three Mile Island reactor once it restarts. This is more than double what Big Tech was often paying for solar and wind. Microsoft will pay this much because nuclear power is the only way it knows how to honor its near-term pledge of zero carbon emissions. Solar and wind, even with substantial storage, still require fossil-fuel backup.
This emerging mania for nuclear energy is yet more strong evidence that we have a supply problem, not a demand problem. And the most acute issue is in generation rather than transmission. Yet there is still concern over whether nuclear energy is affordable and can be built on time at scale.
Historically, nuclear energy was among the lowest-cost electricity options. Before the Three Mile Island accident, the US consistently built reactors at prices of $1,500/kW to $3,000/kW over a 5- to 6-year time frame. See Figure 8.
A nuclear power plant is just a large building made of concrete and steel. The inflation-adjusted costs of concrete, steel, and labor are actually lower today than in 1970. The prices rose over the years primarily due to increasingly stringent regulations applied to plants during construction, leading to frequent work stoppages. The Three Mile Island accident was the final blow, turning the US against nuclear power, demanding near-perfect predictability.
If nuclear plants can be built at pre-Three Mile Island CAPEX prices and operate with modern capacity factors, the unsubsidized levelized cost is ~$40/MWh. China currently builds plants in this price range. Critics of nuclear energy overstate the difficulty of rebuilding a stable nuclear power industry. History and the experience of countries that have achieved low prices demonstrate that it’s simply a matter of implementing stable, sensible regulations and building plants based on standardized designs on a regular basis.
Conclusion
The attempt to replace dispatchable electricity with intermittent electricity is still in the early stages, and we are already seeing severe cracks in energy markets. The cost and difficulty of perfectly matching demand with intermittent sources are proving greater than expected. Things are going to get worse unless a new direction is taken.
Advocates of this approach need to explain why capacity prices are exploding, why new demand for gas generation is surging, why new load from data centers cannot be accommodated, and why Big Tech is willing to pay such high prices for nuclear energy. The markets are sending a message that policymakers need to understand.
If we can’t accommodate data centers, what does this imply about the ability to electrify the entire economy?
Our climate and energy policies have done more harm than climate itself has. We are shooting ourselves in the foot.
No more solar and wind. Move resources to natural gas, nuclear and grid improvements.
Let's hope policymakers see sense and require PV and wind to comply with the same rules as everyone else. No more "must run", no more subsidies of one sort or another, bid on the day-ahead market like everyone else, with appropriate penalties for failure to produce.
I'm not holding my breath, though.