Lynne Kiesling-Knowledge Problem

The Institute for Regulatory Law & Economics Five Prisms:

As @tylercowen is fond of saying, consider the equilibrium.

When PJM neared its peak demand record, they declared an energy emergency. Texas has blown past its pre-2026 record 7 times this month, incl. today. No emergencies, not even close. The difference? Solar. New record in Texas today: 1st time >35,000 megawatts #txlege #txenergy

As I read it, Dwarkesh is arguing that demand pressure will outrun capacity growth (a supply shift) for a while, and that supply is inelastic enough (a movement along the supply curve) for the changes to appear mainly in price. As always, I think Hicks-Marshall is helpful here, since compute is an input, and what matters is derived demand vs. supply (with some help from AI to get Greek for Twitter). And I get all spun around without writing it out. I see Dwarkesh as working through the basic decomposition R = pαC/(1 − m), where α is the share of compute devoted to inference and m is the inference gross margin. The real puzzle over the long run is what determines p, so let's focus on that. Let C be GPU-hours, A is technology that captures the amount of useful services (AI services) per GPU-hour, and p is the rental price. Output is X=AC, and q=p/A. Locally, write demand for effective AI as dln Xᴰ = dln B − E dln q. B is the extensive-margin demand shifter. Think new tasks or products. I am putting productivity on already-counted services in A and newly feasible uses in B so that we do not count them twice. Under competitive output pricing or a fixed markup, Hicks–Marshall gives the derived-demand elasticity of E = (1 − s)σ + sη. s is share, in this case, it is AI’s share of production cost, and our usual σ elasticity of substitution between AI and human labor, and η is the absolute elasticity of demand for final output, a scale effect. If cost pass-through varies, the scale term is closer to sκη, where κ is pass-through. The standard intuition is that when s ≈ 0, E ≈ σ: substitution away from labor dominates. When s ≈ 1, E ≈ η: final demand dominates. So we have demand for Since C = X/A, demand for GPU-hours is dln Cᴰ = dln B + (E − 1)dln A − E dln p. It depends on tasks done, elasticity, productivity, and price. It isn’t even necessarily true that better capabilities raise demand for raw GPU-hours. Holding B and p fixed, that takes E > 1. If E < 1, we need B to increase enough to offset the compute saved by greater efficiency. For price, we need to pin down supply too. Locally, dln Cˢ = dln C̄ + ε dln p. Think of C̄ as the committed or price-independent compute-supply shifter at a given price. Here, the underrated one, ε is the elasticity of compute supply over the relevant horizon. Price-induced fab construction can go into ε or into an endogenous C̄, but we should not count it in both. With supply and demand together, you get dln p = [dln B + (E − 1)dln A − dln C̄]/(E + ε). So for p to go up, we need dln B + (E − 1)dln A > dln C̄. That inequality is enough for the price to rise. A small ε is not necessary for the sign, but it is necessary if the increase is supposed to be enormous. For a pure outward demand shift g, dln p = g/(E + ε) and dln C = εg/(E + ε). The relative demand and supply elasticities E and ε determine how the adjustment is divided between price and quantity. In the constant-elasticity benchmark (so local is global), a 10x price increase that he mentions requires Δln B + (E − 1)Δln A − Δln C̄ = (E + ε)ln 10. Dwarkesh’s $250,000 calculation is about the height of the demand curve, or you could say about the current equilibrium. It is not the long run equilibrium. That distinction matters because E governs what happens as we move along the demand curve. The market price is set by the marginal H100 after millions of AI workers have entered, not by the value of the first use. A large E means compute demand is elastic, so a large increase in compute produces a relatively small decline in willingness to pay. A small E means the marginal value falls quickly. His high-skilled-immigration analogy is making one of two claims. It could be saying that E is large because final demand η is elastic enough to absorb much more AI-produced output. Or it could be saying that specialization, innovation, and new products keep shifting B outward as more AI workers enter. Those are different channels, and E plays a different role in each. The equilibrium equation gives three clean comparative statics: ∂ln p/∂ln B = 1/(E + ε), ∂ln p/∂ln A = (E − 1)/(E + ε), and ∂ln p/∂ln C̄ = −1/(E + ε). If the story is mainly about new tasks shifting B, a larger E actually dampens the price response to any given demand shift. A smaller price movement produces a larger movement along the demand curve, offsetting more of the outward shift. Dwarkesh then needs the B shift itself to be very large. If the story is mainly about better capabilities A, the key threshold is E > 1. Better models increase raw-compute demand only when the substitution and scale responses more than offset the GPU-hours saved by greater efficiency. Above that threshold, a larger E strengthens the capability-driven price effect. On the supply side, slower growth in C̄ raises price, while a small ε means that a given demand-supply gap produces a larger price increase and a smaller quantity response. I think the key is that Dwarkesh’s 3x number describes total compute growth. That alone doesn't pin down how much came from a shift in C̄ or how elastic supply ε is. So the horse race is not simply “demand outstrips supply.” It is between two demand channels, dln B + (E − 1)dln A, and the supply shift, dln C̄, with E and ε determining how strongly those shifts translate into price. The $250,000 calculation gives us a reason to think the demand curve may currently be high. The immigration analogy is doing the work of claiming either a large E or continued outward shifts in B. The fab, power, and data-center discussion is doing the work of claiming slow C̄ growth and a small ε. For a 10x price increase, all of those claims must add up to the much stronger magnitude condition above. Circling back to Hicks-Marshall: 1. σ and η determine E 2. E determines whether capability growth raises raw-compute demand and how fast marginal value falls 3. ε determines how much of a demand-supply gap appears in price 4. A, B and C̄ determine which curve shifts faster. B is obviously shifting fast. But A C̄ could outstrip it.

Digital technology is turning the power grid into a more modular, layered system. But devices alone are not enough. Data access, interconnection, prices, contracts, and governance will determine whether the grid becomes a platform or a bottleneck. knowledgeproblem.com/p/a-new-power-…

Everyone from Weber to Mokyr puts culture at the center of the rise of the West. But nobody has had a long-run cultural series you could drop into a growth model. So I built one: LLMs read 23,000 books from the Western canon, scoring what each endorses. Year 0–1920, one chart:

In honor of his birthday, I have a piece out in today's The Daily Economy on Alexis de Tocqueville and why post liberals like Patrick Deneen are wrong to claim him as their own. Link below.

@ArushiSF @EmeraldAi_ @nvidia @JensenHuang Like you I thought first of data, mutually agreed upon telemetry standards, and third party metering. On the economics/market design side, having actual demand-side bidding instead of an administrative demand curve is essential.

@EmeraldAi_ nailing the Aurora @nvidia factory data-to-dispatch is step 1. We need commercial proof of possibility, and technical data streams to build on. Team @JensenHuang is going to help us get there. Making it possible could also involve PJM accepting third party metering for the site, and load telemetry over ICCP and a market standard undergirding that (protocols that control systems can integrate). Not that different from ERCOT accepting device level-to cloud -to grid ADER telemetry in addition to meter data from utilities/load serving meter entity collecting revenue grade. Some utilities stuff. IDRs for quality data, retail tariff accomodation of load participation in PJM Resource programs (tweaks, nothing brand new). Note- I said it is impossible *today*- not forever - to deliver C&M style dispatch solutions which don't overprocure /over signal emergency load shed at the cost of running a larger baseline operating system that does more with less drastic calls for curtailment. I'd be interested in what you think PJM needs too, @knowledgeprob !

@ArushiSF @bslotterback What would PJM have to change in order to deliver a dispatchable load program?

It is impossible for PJM to deliver a dispatchable load program today that puts loads and generators on the same or even similar footing in real time dispatch to solve grid constraints. The term was misappropriated last year around the time of the announcements from the White House, and absolutely does not apply to anything possible in PJM today.

Didn't PJM also promise a connect and manage proposal? I'm certain I heard that at the FERC-sponsored meeting.

@johnjhorton I don't have a good idea for how we deal with false positives, given that the LLMs are trained on our output.

Me, in delivery room: "I think this baby is worth naming"

BTM and off-grid are not anti-grid. They’re exactly what the grid needs. High-renewables grids require a source of well-funded, dispatchable reserve capacity. Solar and wind are intermittent, and batteries handle average daily fluctuations well, but not extreme events. Something that has puzzled the grid policy community for ages is: who will pay for, build, and maintain that low-utilization capacity? Until recently, we had no idea. Today, we’re building gigawatts of “behind-the-meter” power for data centers. These are engines and turbines that typically wouldn’t be used for baseload operations. On the grid, big reciprocating engines and smaller aeroderivative gas turbines are used for peaking, essentially backup power, and renewables balancing. It’s only off-grid that they’re used for “prime power.” While these engines and turbines are running baseload today to serve their data center customers, the vast majority of these systems will connect to the grid over time. Their customers want the grid, and their own financing works better when there is a grid connection. There’s no reason to avoid the grid once it gets its act together and actually offers to connect. When they do connect, these BTM DCs will bring gigawatts of flexible backup power to the grid. And unlike the CCGTs many utilities are begging to build at astronomical costs, they won’t run all the time, polluting 24/7 and locking in long-term emissions. They’ll run only when they’re really needed because they’re low-capex, high-opex systems. That’s exactly how the grid is designed to work, and a perfect complement to high solar/wind/battery systems. The dispatchable reserve capacity we direly need to scale solar, wind, and batteries to high penetrations will have been paid for by high-margin compute. Not by ratepayers or taxpayers. And if AI is a bust (imagine it’s a huge bubble and it all deflates) then those costs will be borne entirely by the BTM and AI companies. Compare that with a scenario where we massively upgrade the grid immediately to serve AI, and then all that demand goes poof. That would be a disaster for ratepayers who are all ready exasperated with high prices. If you’ve been told BTM and off-grid power are anti-grid, ask what the alternative is and if its ever been shown to have momentum. So far, this is the most promising path to actually building gigawatts of dispatchable grid capacity without putting the cost, and the downside risk, on ratepayers.

@thomasforth @dieworkwear @PrivateWhiteVC Not menswear, but With Nothing Underneath makes excellent women's shirts using men's tailoring. I make a beeline for their Marylebone shop whenever I'm in London.

If @dieworkwear -- or a British equivalent -- on the back of that great @PrivateWhiteVC thread, could maybe share a big list of British clothes manufacturers doing similarly great things that would be amazing. Thanks in advance.

Don't know anything about Andy Burnham, but I recognize the room in the middle photo. That's Private White VC. Let me tell you about them. 🧵

@asymmetricinfo Last week at the @LawEconCenter retreat someone just meeting my husband referred to him (somewhat wittily) as Mr. Kiesling, and he took it in stride. He's very gracious and communicates well that he is proud of my work and reputation capital.

However, some of our more socially conservative friends have their kids call me Mrs. Suderman and that's fine, I am not mad about it. I might have taken his last name in other circumstances, if for no other reason than that it is easier to spell.

I was already professionally known under my maiden name and changing my name would have sacrificed a great deal of reputational capital.

I've now listened to all of the episodes of Everyday Abundance, something I rarely do. Highly recommend! @vpostrel and @CharlesCMann are a great duo—entertaining and informative. Many great stories that will help you appreciate the world around you. abundance.institute/EverydayAbunda…

California just made it much harder to use one of its most effective wildfire prevention tools: goats. New minimum wage and overtime rules will drive up costs (requiring goat herders to be paid ~$240,000/yr) and push herding companies in the state out of business. My latest with @JarrettDieterle In @reason reason.com/2026/07/25/cal…

“If AI is so powerful, why is its impact so small?” Look, *decades* into the Industrial Revolution, GDP per capita in the UK had only increased about 20%, total! Impact takes time. Diffusion is slow. Even if it is 10x faster than in 1800, it's still early.

The knowledge that makes AI useful is diffused. It lives with scientists, engineers, clinicians, firms. For AI to benefit from distributed knowledge, it must itself be distributed. Agree with Jensen that this is a future worth building.

Capitalism is just so beautiful. $8 trillion worth of market cap coming out in support of open-weight models because it's good for their business and makes them look good... and, totally coincidentally, it also happens to be better for us consumers and practically every business except for a couple of labs. we just get that benefit for free with MSFT and NVDA's self-interest. it brings a tear to the eye 🥲

Barry Weingast on Adam Smith and James Madison, new at Constitutional Political Economy