This question is motivated by the following hypothetical. Suppose a person such as Elon Musk derives little utility from ordinary consumption beyond a subsistence level but substantial utility from building and launching spacecraft. He would therefore save and invest much of his income. If he could persistently earn unusually high returns, he might eventually accumulate a large share of the economy’s wealth—although increases in asset values themselves would not count as GDP.

The production of his spacecraft would contribute to GDP through the value added in their construction. A spacecraft owned by a business and used repeatedly to produce services would normally be classified as investment; one acquired purely for personal enjoyment might instead be classified as consumption. Intermediate components would not be counted again separately. The more difficult question concerns the services produced when the spacecraft are used only by their owner. Home production is typically not included in GDP calculations. If comparable services are sold, their output can be valued at market prices. But Musk might refuse to sell them, or his fleet might become so abundant that access is not metered and no meaningful transaction price exists. National accountants might then value recognized own-use output using comparable prices or, lacking those, production costs and the user cost of capital. Some purely personal, self-provided services might instead fall outside GDP’s production boundary.

Price chaining creates a further complication. Real GDP growth is calculated by comparing quantities across adjacent periods using their respective price weights. If spacecraft services are novel, not traded, or eventually supplied at a zero price, there may be no usable sequence of prices with which to value and chain their growing quantity. Statisticians would have to impute prices or use cost-based measures of output. Moreover, as the service becomes abundant and its price approaches zero, its weight in chained GDP also approaches zero, so even an enormous increase in spacecraft services may add little to measured real GDP. Price chaining can therefore measure the resources devoted to producing the fleet more readily than the services or welfare the fleet ultimately generates.

A zero price for access would not mean that the spacecraft were costless to produce. It would mean that GDP—whether nominal or chain-weighted—could fail to capture the potentially enormous consumer surplus Musk receives from them. The example thus illustrates a broader limitation: GDP measures production using observed or imputed prices, not the welfare generated by goods with unusual demand, missing markets, or near-zero marginal cost.

This is a complicated question. At a high level, we can think of the rate of accumulation of wealth as a function of the rate of return on wealth (the rate of interest), the savings rate, taxation, and bequests. These questions were famously explored in Thomas Piketty’s book “Capital in the Twenty-First Century”, where he argued that a sufficient condition for wealth inequality to increase over time is that the rate of return on wealth is greater than the rate of growth of the economy (e.g., r > g). A high level intution is that average income growth tends to follow the growth rate of the economy, while the income of capitalists tends to follow the rate of return on wealth. If r > g, then the income of capitalists will grow faster than the income of the average worker, leading to increasing wealth inequality.

Piketty’s arguments have been challenged in many ways. Herein we highlight two. First, the rate of accumulation of wealth is not solely determined by r. A simple example when this would affect our predictions is if the government decided to increase taxes on capital in a transformative AI world. This would reduce the rate of accumulation of wealth, all else equal. Similarly, estate and capital gains taxes also affect the rate of accumulation of wealth, although tax avoidance and evasion mitigate their effects.

Second, r may not be greater than g in the future. This is because the interest rate is determined by the marginal product of capital, which is the additional output produced by an additional unit of capital. The growth rate of the economy is determined by average product of capital and labor, which is the total output produced by the economy divided by the total capital and labor employed. This situation may happen if there is an abundance of capital in the world, say from self-replicating robot factories producting goods, which humans are satiated with. If humans are satiated with goods (say cars), then their marginal utility of goods is close to zero, so the interest rate from building more factories will be close to zero.

A data center is mostly classified as private nonresidential fixed investment which is valued at cost. This is part of the ‘I’ in the formula GDP = C + I + G + (X - M). Additionally, GPUs and other imported components are also subtracted in the M term. This explains how the data center construction is included in GDP. Furthermore, once the data center is constructed, it offers services which other companies pay for. This often happens in the years after the data center is constructed. Since data center services are an intermediate good, they contribute to GDP via their contribution to final consumption (of AI services, ads, services from firms who use AI, etc…).

This is a broad question, but here’s an illustrative framework. In the short run, the stock of compute is roughly fixed. Such a fixed factor earns scarcity rents—payments above its cost of production. So compute owners will earn rents as long as demand is large relative to capacity. How compute is allocated, and who captures the surplus, depends on two factors: competition among model providers and the elasticity of demand for model outputs (intelligence).

Consider a 2×2 matrix combining lab-market structure with demand elasticity.

Labs produce similar quality/intelligence models + elastic demand. Labs bid for compute up to the value of the additional output it lets them sell. Competition prevents them from having much markup unless they have cost advantages, while customers readily reduce their purchases when prices rise. Compute flows to the labs with cost advantages, but demand elasticity limits the overall price levels. Most of the surplus is captured by the compute owners, labs only earn their cost advantage over the runner-up.

Labs produce similar quality/intelligence models + inelastic demand. Customers keep buying even as prices rise, so their willingness to pay for the fixed supply of model output is high. Competition among labs passes most of that revenue through into their bids for compute. Labs earn little above their other costs, compute owners capture most of the rents, and capacity limits output. Compute flows to the labs with cost advantages and overall prices for AI services are high. Most of the surplus is captured by the compute owners, labs only earn their cost advantage over the runner-up.

One lab produces much higher quality/intelligence models + elastic demand. A single lab of the most intelligent model can mark up its output and may restrict quantity, but customers’ willingness to substitute away disciplines that markup. Compute owners still receive the scarcity value of their capacity; the lab keeps a rent as well. Compute owners still capture most of the surplus.

One lab produces much higher quality/intelligence models + inelastic demand. The lab producing the most intelligent model can raise prices and capture surplus. It may restrict output below capacity to sustain those prices. How the resulting rents are divided between the lab and compute owners depends on bargaining, contracts, and compute owners’ alternative customers. The lab captures much of the surplus.

A toy example. Suppose 100 units of fixed compute are sold competitively. Lab A can turn each unit into $10 of value; lab B can turn each into $9. Both labs face competitive output markets, so each is willing to bid up to the full value it can produce. Lab A wins all 100 units because it creates more value—but needs to bid only slightly more than B’s $9 per unit. Of the $1,000 created, compute owners collect about $900 in scarcity rents and A keeps about $100: its productivity advantage over the runner-up. No compute monopoly is needed for this result; fixed capacity and competing users are enough. More elastic customer demand would reduce both labs’ willingness to pay, while model-market power could cause a lab to restrict output rather than bid according to the full value its compute could create. Compute owners capture most of the surplus, labs only earn their cost advantage over the runner-up.