A former Amazon product leader recently made the most compelling Marxist argument I've heard all year. He probably wouldn't put it that way.
Nate B. Jones, an AI analyst, made a sharp case that the fundamental unit of computing has shifted from instructions to tokens. His Substack piece digs into what this means for developer careers, enterprise budgets, and organizational structure. Go watch it. But the video skips past a question that kept nagging me: if intelligence is now a commodity you purchase by the token, who's the landlord?
For Sixty Years, Software Ran on Human Cognition
The means of production in software used to be relatively democratized. A talented person with a laptop and the right skills could compete with teams ten times their size. The barrier was knowledge, not capital. Open-source tools, commodity hardware, and cloud compute made the playing field surprisingly level. You owned your productive capacity. It lived in your head.
That's changing. The means of production are migrating from human cognition to purchased inference. Tokens are the raw material. API access is the factory floor. And the numbers make the shift concrete.
In the US, where most of this spending is concentrated, the average enterprise LLM spend hit $7 million in 2025, up from $4.5 million just two years prior, and is projected to blow past $11 million in 2026, according to A16Z's Enterprise AI survey. StrongDM's CTO Justin McCarthy disclosed that his three-person team targets $1,000 per day in token spend and writes no code by hand. The A16Z survey found the share of US organizations planning to spend over $100,000 per month on AI has doubled from 20% to 45%.
Marx would recognize the dynamic instantly. The laptop-and-skill model gave developers genuine ownership of their productive capacity. The token model makes them purchasers of someone else's.
Three Layers of Control
The political economy of tokens isn't monolithic. There are three distinct layers where power concentrates, and understanding which one matters most changes the whole picture.
The infrastructure layer is the most capital-intensive chokepoint. AWS, Google Cloud, GPU clusters. Anthropic spent $2.66 billion on AWS through September 2025, against an estimated $2.55 billion in cumulative revenue over the same period. More than 100% of top-line revenue going to a single infrastructure provider. Perplexity spent 164% of its entire 2024 revenue across AWS, Anthropic, and OpenAI combined. For those keeping score at home, that's spending $1.64 on ingredients for every dollar of revenue. Not a business model. A bonfire with a logo. These numbers, reported by journalist Ed Zitron, reveal that even the companies building the models are tenants on someone else's capital infrastructure.
The model layer is where pricing power lives. Whoever trains and serves frontier models sets the pricing regime. Here's the case study that makes this visceral: Cursor, Anthropic's single largest customer and a billion-dollar-revenue AI coding editor, had its AWS costs spike from $6 million to over $12 million between May and June 2025 after Anthropic raised priority tier and caching prices. Cursor was forced to gut its unlimited $20/month plan and introduce a $200/month tier. Users revolted. Cursor eventually started building its own model.
A landlord raised rent on a tenant who had nowhere else to go. The tenant was the landlord's single largest customer and still had zero leverage. Textbook political economy.
The context and domain layer is where things get interesting for individuals. The person who knows the dental practice workflow, the construction scheduling constraints, the insurance compliance landscape. This is the layer where human expertise translates intelligence into value. Marx wouldn't have anticipated this cleanly because in his framework, labor-power had "no other repository than human flesh and blood" — knowledge wasn't easily separated from the laborer. Now it can be. Cheaper tokens make more niche problems economically viable, which makes domain expertise more valuable, not less.
The Open-Source Wrinkle
Here's where the concentration thesis runs into a genuine complication. DeepSeek demonstrated frontier-competitive performance from dramatically lower training budgets. Llama is free. Mistral, out of Paris, punches well above its weight. Hugging Face made model deployment accessible to small teams. Doesn't open source dissolve the whole landlord problem?
Not as much as it looks.
Open weights are not open means of production. When Meta releases Llama, they release the artifact of an enormously expensive production process. Training a frontier model still costs tens or hundreds of millions of dollars in compute. The weights are open. The capacity to produce the next generation of weights is not.
Think of it this way. Imagine Andrew Carnegie published the exact chemical formula for his best steel and let anyone use it royalty-free. Generous. But it doesn't give you a steel mill. The ability to produce the next breakthrough alloy still requires the blast furnace, the ore supply chain, the metallurgical research lab. Open-source models give you access to this generation's intelligence but not a seat at the table for producing the next generation. In a field moving this fast, last generation's model is a depreciating asset.
Then there's the operational reality. Running your own inference at scale requires GPUs (NVIDIA supply chain concentration), MLOps expertise, reliability engineering, latency management, safety filtering. It's a serious make-versus-buy decision, and the history of technology suggests most organizations eventually buy. We all ran our own email servers once. Nobody misses it. If Cursor, a deeply technical company, couldn't easily self-host its way out of Anthropic dependency, most enterprises certainly won't.
And the punchline: Meta's open-source strategy is itself a power play. Meta gives away Llama because Meta's competitive interest is served by commoditizing the model layer so value migrates to layers Meta controls: its social graph, advertising infrastructure, and data assets. This is the classic "commoditize your complement" strategy. IBM open-sourced the PC architecture and lost the value to Microsoft and Intel. Google open-sourced Android and captured the value in search and advertising. Meta is running the same play at the AI layer.
Open source doesn't dissolve concentration. It shifts where concentration happens.
The Race That Actually Matters
The honest answer is we don't know which future we're heading toward. What we're looking at is a layered oligopoly: compute is highly concentrated, model training is concentrated but leaky (open weights keep escaping into the commons), inference serving is competitive but trending toward consolidation, and the application layer is genuinely open.
If open-source models stay within striking distance of the frontier, as DeepSeek suggested they might, the concentration thesis weakens significantly. Buying tokens from OpenAI remains a choice rather than a necessity. Competition keeps the landlords honest.
If the frontier pulls away and requires billion-dollar training runs that only three or four organizations can fund, openness at the model layer becomes a polite fiction. Free to use yesterday's intelligence while real power consolidates around tomorrow's. Satya Nadella invoked Jevons' paradox by name after the DeepSeek moment: as AI gets more efficient and accessible, total consumption explodes. Cheaper tokens don't mean lower bills. They mean more tokens consumed.
When GPT-4 launched in March 2023, it cost $36 per million tokens at blended rates. Today, GPT-4o mini matches that performance at roughly $0.24 per million. A 150x drop. Enterprise AI spending doubled. The paradox holds.
The Power Question
Jones's video covers the career strategy question well: what skills matter when tokens replace keystrokes. But underneath the career question sits a power question that the industry mostly ignores.
Who sets the price of inference? Who captures the margin when intelligence becomes a utility? The entities that control token supply are structurally analogous to the steel barons and railroad owners of the Gilded Age. Same structural position, better wardrobe. Rockefeller never gave a keynote in a gray crewneck about democratizing petroleum.
The Gilded Age eventually got antitrust. The EU, in character, already passed an AI Act. Washington is still debating whether regulation is necessary. The question isn't whether the token age gets governance. It's whether governance arrives before the concentration becomes irreversible.
