AI and the future of capitalism

Until recently, most of the AI talk was about the race between the big tech players - OpenAI, Anthropic, Grok, Gemini. Now, with the recent launch of Kimi K3 by Moonshot the talk is about open vs closed models. According to Mozilla’s The State of Open Source AI report, the top models by token volume are now predominantly open-weight: their parameters published, downloadable, and runnable by anyone with the hardware. The benchmark gap between open and closed models, which stood at roughly 17 points on standard knowledge tests in late 2023, is now effectively zero. And models like Kimi K3 are achieving around 2.5x the scaling efficiency of their predecessors: more capability per unit of compute, improving fast.

This is interesting as a technology story. But I think it is more interesting as an economics story, and perhaps most interesting as a question about what kind of economy we are heading toward. In Marx’s framework, surplus value is the gap between what it costs to produce something and what that something is worth. A factory owner buys labor, raw materials, and machinery; the worker produces goods whose value exceeds those input costs including the salaries of the workers; the surplus is what the owner captures as profit. It’s the logic of every business: the value delivered to the customer exceeds what it cost to deliver it, and that gap is what makes a business a business.

AI, right now, is an extraordinary surplus value machine. A company pays for API access at a few cents per query and uses it to replace a task that used to cost orders of magnitude more. The gap between cost and value produced is enormous, and everyone along the chain is capturing some of it: AI companies through API pricing, businesses through productivity gains, their customers through lower prices or better products.

What happens if that gap closes? Models are rapidly becoming more efficient. Quantization techniques now reduce model sizes significantly with minimal quality loss, making serious models runnable on consumer hardware. Local inference is already a step-function cost structure: buy the hardware once, and your only ongoing cost is electricity. The trajectory of open models suggests that, not too many years (or months?) out, a model capable of doing most cognitive work could plausibly run on hardware that many people already own, at a marginal cost approaching only the electricity to run it.

If that happens, and if the models are open-weight, the surplus value equation changes fundamentally. The task that used to cost a hundred dollars in human labor now costs the amortized price of a GPU and a few watts of electricity. The gap between cost and value does not just shrink: for a wide class of cognitive tasks, it approaches zero. Not because the work is worthless, but because the cost of doing it becomes trivially small for anyone.

This is where the open versus closed distinction stops being a technical detail and becomes something more consequential. In the closed-source scenario, AI companies function as gatekeepers of that surplus. Because you can only access the model through their API, they sit between the cost of running the intelligence and the value it creates, collecting a toll on every query. The surplus value does not disappear: it concentrates. A small number of companies become, in effect, the owners of the means of cognitive production, and everyone else pays for access. This is capitalism, but a particularly concentrated form of it. Not many capitalists competing, but a handful of platforms that own the thing that produces value across essentially every sector of the economy simultaneously.

In the open-source scenario, something structurally different happens. If the model is free to download and cheap enough to run locally, there is no toll to charge. A cookie shop owner can produce an app at the cost of hardware and electricity. The surplus value that used to flow to an AI company now stays with whoever is doing the work, or, taken to the limit, dissolves into the general economy as simply the electricity cost of thinking.

We have been here before, maybe. Linux did not kill capitalism: it commoditized the operating system layer and shifted value extraction upward, to services, cloud infrastructure, and software ecosystems. Android made smartphone software nearly free and value migrated to app stores, advertising, and hardware. Open source has a long history of disrupting specific rent-extraction points without disrupting the broader logic of capital accumulation. But there is a reasonable argument that the intelligence layer is different from the OS or the browser. Those were infrastructure. Intelligence is the thing that produces economic value in almost everything, almost everywhere, at least in the “information economy”. Commoditizing infrastructure shifts where value is captured. Commoditizing intelligence itself may shift something more fundamental: how surplus is generated at all across the economy.

Perhaps the most important question is not whether AI will be good or bad for workers, or whether open or closed source will win as a technology choice. It is a more structural question: if open-weight AI continues on its current trajectory, are we heading toward a world where intelligence, and associated value-creators like process knowledge, as an input to production becomes effectively free? And if so, what form of economic organization emerges on the other side of that?

Maybe a new rent-extraction layer appears above or below the model: compute, data, trust, enterprise integration. Or maybe the commoditization of intelligence is genuinely different in kind, not just degree, from what came before. That seems worth thinking carefully about. We are watching a shift in who gets to capture the value that intelligence creates.