AI-driven automation process
The article below is a snippet of a book chapter. The chapter is on Organic Composition of Capital (OCC) and Technical Composition of Capital (TCC) and the section is talking about OCC and TCC using AI data centre build-out as an example of applied Real Economics/Classical analysis.
Introduction (and some clarity)
Why bring Organic Composition of Capital (OCC) and Technical Composition of Capital (TCC) into the analysis of AI?
The shift in software development has been from a fairly balanced process of automation and capital investment (growth in OCC) to a sudden increase in machine investment at the expense of employment (TCC). The explicit goal of these large AI tech firms companies is full automation of this workforce.
Large technology companies were already highly capital intensive companies, but relied on a significant pool of labour to engage with consumers and corporate enterprise. Writing and refining software, designing new hardware, and never mind assembling that hardware is very labour intensive.
Sitting on huge piles of cash-generating assets has meant a lot of capacity to invest in automation. However, not many options had existed before AI to automate the most labour intensive process of coding or "software engineering". Commercial software is essentially custom meaning someone has to interface with the consumer and build or customize parts of the system. This means upgrades are also custom and so more people to do that have to be employed. New clients means more workers.
The competitive advantage of automating this variable capital is obvious. However, it turns out it is very expensive to run physical assets at such a scale that it can process the data for even marginally useful AI models.
The tendency, then, was to make the models as general as possible to support all downstream firms in automation of their workers as well. The goal that all firms to must automate, driving that ever present tendency for the OCC to rise faster.
The history of AI thus far is to do this everywhere all at once. But, that is only possible if everyone is convinced it is going to be worth it. The burning of the cash on hand, layoffs of developers, is funding this demo.
The hope is that the process of increasing technology implementation in the workplace will yield cheaper products and we will land at a cost/price equilibrium (and thus the lowest quality viable product) that meets some minimum acceptable level.
Unlike the ad, none of this is guaranteed.
But, the transition is also an unstoppable force. Once we have started down the path of automation under capitalism, competition with the Regulating Capital makes it impossible to not follow.
Or, at least this is the case from a Classical/Real Economic Theory perspective.
Quantifying TCC and OCC in the AI-driven automation process
Let's start with some fun esoteric concepts from Classical economic theory.
TCC: Technical Composition of Capital
- The ratio of the total mass of constant capital to the amount of variable capital.
- That is, the quantity of physical capital (machines) to total number of workers' hours in a firm.
- TCC over time can act as proxy for the direction of travel for how many machines are used to replace a worker.
OCC: Organic Composition of Capital
- The ratio of the value of materials plus a firm's fixed costs to the value of labor power used in the production of a commodity.
- Over time, it is a measure of the replacement of workers with machines to produce output. There is a natural tendency for an increase in the OCC as workers are replaced with machines/technology.
The story we are interested in is the use of AI in the production of software and support for automation processes.
The reason that this is an interesting question to answer is that if the costs of replacing workers does not either increase productivity or decrease costs per worker in the production process, AI has no benefit to the profit rate. And, therefore, AI will not be successful as a generalist automation tool. Or, at least, it might put it in an interesting conflict with the natural tendency of rising OCC.
The output commodity we are looking at is a software program. Both TCC and OCC will increase through AI usage, either through direct automation/replacement or as a tool for augmentation of coding work. The reason both rise is we need to build data centres that you did not need before (a TCC increase) and you need to build capacity to operate them (increasing OCC).
If using AI creates a productivity gain, increases the output of the number of programs (a weird measure for software but it exists as software is custom), or betters the response to ongoing security issues as a unit measure of a program's value, then we should see adoption of AI increase. Again, this is the case if the end result is a per-unit decrease in cost of production.
Under Real Economics, there is a time element that needs to be incorporated in the analysis. The turbulent dynamics that result from the battle between firms means that which firm is winning will change as costs, prices and profit rates are in competition. The general tendency is for firms to adopt new technology from the firm that has the newest tech, and so the lowest unit prices, and highest profit rates. We call this firm the "Regulating Capital" as it regulates the price for that short period of time it is winning.
However, if there is a net increase in cost per unit of production, AI (in the general sense) will not be implemented everywhere all at once. In this situation, firms that do not implement AI may over time have a higher profit rate as they do not have to foot the bill for all the new highly capital-intensive production. However, this will only happen when real costs of production are imposed and the reality of costs flow through to prices and this may not be clear right away as it takes time to see the true costs because of financialization of the lending system.
At some point, there will be a fight around the real economics of capital intensive production versus the current model of employing people to implement code. And, eventually both of those versions will fight against some new version of cheaper TCC and less intensive OCC increases using ever newer AI or other technology mix.
This is as true for downstream tech-related production firms as it is for the upstream tech related production firms. The reality is that a firm cannot replace workers with more expensive and less productive machines and be assumed to survive. The debts paying for those large capital expenses have to be paid eventually.
"But," you say. "What about the fact that the output from AI is just from a rented machine and it produces free-to-use products as output? Surely that is different from an industrial OCC argument based on an old-style industrial production process?"
It might be different if a significant portion of the software that we used right now was not already free. So, such a comparison is probably not accounting for the change in real costs to a firm.
The comparison of the ease of creating new software is flawed at the level of the AI firm.
One example is that we have not captured the value of the free labour used to generate open source and the free software process. At the firm level, the software that was free and shared freely is now a new constant capital expense, renting the access to AI to produce code that is of significantly reduced quality. A quality that is so low you have to rent continued access to the models to sift for and correct errors. Which means you still need coders and you need to pay for AI tools rental.
It is not that the cost of wages went down, it is that constant capital costs went up. Significantly.
It is just as likely that using AI means the number of machines (and inputs) used for AI generation is much higher than the gain from reducing variable capital, if open source software was used to any degree. And, shared open source code is everywhere.
If proprietary and licenced software was used then you are still not avoiding this cost increase. The costs for the contracting firm (that is also definitely using AI) needs to be included. The cost for the total gain in output versus costs of value creation affects potential profitability of all parts of the production chain.
If variable capital (workers) costs are lower and output equivalent, then there will not be any continued move towards increased TCC. In some cases, a firm may shift and actually increase the number of workers, if the costs and output remains competitive.
An other example is the tendency of specific local models being designed that will undermine the profitability of large centralized models for the currently profitable enterprise-level productive industry applications.
Important to note here that we are not attempting to describe a complete deterministic model for AI adoption.
There are many actions a firm can take that can offset costs, models change in costs, outputs can get better while efficiencies are being found. But, the total costs of transforming the industry very rapidly to a very high TCC/OCC system from one that relied on workers has to be included in the analysis of where this is all going. And, that analysis should create a significant drag on expectations for AI driving significant growth in profits.
The financialized profit subsidy vehicle and AI
These days, one thing a firm can rely on to delay or offset the cost of introduction of new technology is through opaque and complex financial systems. Indeed, that is what these systems are for.
But, there are limits.
It should be clear that a firm that produces returns for shareholders cannot also cannibalize asset value without borrowing from the future somehow. For firms like the tech giants, moving towards very high OCC and increasing TCC to drive that shift takes massive investment in new capital and an increase in debt to fund it. This move reduces the cash flow available to shareholders as profits.
Fictional capital markets create a space where things can be obfuscated significantly enough to trick people that something magical is going on.
That something magical is usually underpinned by some partially hidden mass transfer of wealth.
Does large capital have access to different financial vehicles that allows it to better mask the transfer of wealth across time?
Maybe through public markets and opaque private capital lending. However, it still must pay its ongoing bills. Layoffs alone cannot pay for all new investments if productivity gains are not realized from investments.
Cannibalization of future costs is the only way this can happen in real terms. Eating into future costs is a form of borrowing or wealth transfer depending on who is paying. And, it is very clear that there is significant cannibalization of future costs going on in the AI data centre market.
Let's have a look at one version of this narrative playing out right now.
A common narrative is that even if we are in an AI build-out bubble, at least physical data centres will be left over even if poor investment decisions lead to over-construction, bankruptcy, and an economic crisis. Such a result is assumed to be net wealth transfer from shareholders to "the public good".
A transfer like this is talked about like it is a public sector investments into productive production that runs at a loss, but are essential to life. Water services, nuclear development, or medical advancements.
The assumptions for this outcome are:
- Data centres are a net positive when built.
- Resource allocation is best determined through an undemocratic private market making poor investment decisions.
- That data centres are productive assets for a length of time where there is value transferred to the public via bankruptcy.
- That bankruptcy does not, itself, destroy significant social value.
- That there is not a cheaper public and democratic option for "necessary" AI compute.
Unfortunately, because of the mechanism of TCC/OCC and Real Competition, it is more likely that none of the above are true.
Poor economic analysis of competition, an incorrect understanding of real costs, over-estimating productivity improvements from using AI, and retail investors being lied to will result in a net transfer from the public to the extremely wealthy. And, data centre assets that were (partially) built without regard for demand will not be socially useful because they will be dead capital.
Commercially viable data centres and firms running AI models will be operated and their services sold to fill the market demand where there are benefits. This includes the government, military, banking, and research institutes. So, the dead capital that are "left over" after a crisis will not have a market to fill, even at some expected reduced cost.
There is no real reduced cost from a bankrupt data centre either. Just as one cannot nationalize production owned in a different country, this infrastructure failing will not be captured by the "public good" organizations who have no skill in running them and no money to operate them. Those assets will simply be lost as other options will over-ride them.
The net result is a net transfer from the public, not to the public. The reasons are two-fold:
- Money should have been spent on something else, a net loss of potential value creation and a net loss of value that was created.
- The borrowing from the future to reduce a costs to build today includes the jobs, environment, and impacts used to subsidize the build-out. Those jobs cannot be re-captured through use of dead capital.
This second reason above is somewhat specific to large capital projects that are supposed to lead to increases to productivity but do not.
The drive to increase profits at the expense of the future does not create value, it simply cannibalizes future costs for profits today. This process is the real benefit of financialized models. Even the crisis from over-production does not necessarily produce even a net benefit to society on capital or network capacity build-out.
Alternative
If we are going to talk about public goods of AI technologies, it is necessary to build them out from public research institutes that still have some connection to democratic norms.
Leveraging assets built for an intended purpose will create both the value creation and the expertise to run it for other things. This model is tried and has created results without wasteful investment and poor, crisis-resulting capitalist economics.
Purpose-built systems also have the added benefit of being possible even if capital fails to drive people out of work and be correctly costed. That is, if we want data centres for the public good, it is the public that will have to build them.