Redistributed

Every time you call a commercial AI API, you pay rent. Not a fee for a product you now own. Not a license that grants you permanent rights. Rent — paid to the platform owner for temporary access to capability you will never possess.

This is the inference tax: the ongoing extraction of value from everyone who uses AI through commercial APIs. Unlike buying software, you never finish paying. Unlike open source, you never gain independence. The meter runs forever.

Understanding this mechanism matters because it reveals how value flows in the AI economy — and who captures the surplus when AI creates something useful.

The Shift: From Ownership to Access

Software used to be something you bought. You paid once for Microsoft Office, installed it on your computer, and used it until you chose to upgrade. The relationship ended at purchase.

The SaaS revolution changed this. Salesforce pioneered the model in the late 1990s: instead of buying software, you subscribe to it. You pay monthly or annually for access. Stop paying, lose access. The vendor captures recurring revenue instead of one-time sales.

AI APIs extend this logic further. You don't even subscribe to a capability — you pay per use. Every token processed, every image generated, every request served incurs a charge. Your costs scale with your success. The more value you create, the more rent you pay.

Nick Srnicek documented how platforms create new forms of economic power through control of infrastructure and data (Srnicek, 2017). AI APIs represent perhaps the purest form of this dynamic. The platform owns the model, the compute, the infrastructure. You own nothing except the outputs — and even those come with terms of service.

The distinction matters economically. When you buy software, the vendor must earn your next purchase by shipping improvements. When you rent AI by the token, the vendor extracts value from every use regardless of whether they've improved anything. The incentive structure shifts from earning future sales to maintaining current dependency.

What Inference Costs: The Economics of the Token

The inference tax has a precise structure. Current prices as of March 2026:

Model	Input (per 1M tokens)	Output (per 1M tokens)
GPT-4o	$2.50	$10.00
Claude Opus 4.5	$5.00	$25.00
Gemini 2.5 Pro	$1.25	$10.00
Llama 4 Maverick (via API)	$0.31	$0.85

These numbers look small. A million tokens is roughly 750,000 words — more than most people write in a year. The fractions seem negligible.

But fractions compound. A coding assistant handling 100,000 developer requests daily might generate 50 million tokens. At GPT-4o rates, that's $500 per day just for output — $180,000 per year. A customer service bot processing 10,000 conversations per day could easily reach $50,000-100,000 annually. The "pennies per request" become real money at scale.

The asymmetry between input and output pricing reveals something about the business model. Output tokens cost 3-5x more than input tokens, reflecting higher computational demands. But this also means applications that generate content — chatbots, writing assistants, coding tools — pay a premium that search interfaces and classifiers do not. The pricing shapes what gets built.

Compare this to traditional software economics. Microsoft Office cost perhaps $300 one time. A heavy user creating millions of words over years paid the same as someone who opened it once. The marginal cost of use was zero.

With inference pricing, marginal cost is everything. Success is taxed. Growth incurs bills. The relationship between value created and rent extracted is direct and unending.

The Lock-In Mechanisms

Why don't developers simply switch to cheaper providers or run their own models? Because the inference tax comes bundled with switching costs designed to make exit expensive.

System prompts and fine-tuning. Every serious AI application develops carefully crafted system prompts that elicit good behavior from a specific model. These prompts often represent months of iteration. They don't transfer cleanly — what works for GPT-4 may fail on Claude, and vice versa. Fine-tuned models are even stickier; the training investment is specific to one provider's infrastructure.

Integration costs. APIs have different structures, rate limits, error handling patterns, and quirks. A codebase built around OpenAI's API requires real engineering work to migrate to Anthropic or Google. The deeper the integration, the higher the cost of change.

Evaluation and testing. Before switching, you need to know the new model performs adequately for your use case. This means building evaluation pipelines, running comparisons, and validating outputs — work that doesn't directly serve customers but is required to even consider alternatives.

Behavioral consistency. If your application promises certain behaviors to users, changing the underlying model risks breaking those promises. The new model might refuse requests the old one handled, or handle them differently. Regression testing against user expectations is expensive.

Ecosystem dependencies. Providers offer adjacent services — embeddings, speech-to-text, image analysis — that create additional lock-in. If you're using OpenAI for text generation and embeddings, switching text generation means either running two providers or migrating embeddings too.

Langdon Winner asked whether artifacts have politics (Winner, 1980). API design certainly does. The technical choices that make switching expensive are not accidents. They're features from the provider's perspective — each one increasing the moat around recurring revenue.

The Platform Cut: Historical Comparisons

The inference tax isn't new. It's the latest instance of a recurring pattern: infrastructure owners extracting rent from everyone who depends on the infrastructure.

Shipping and ports. Before containerization standardized freight, port operators and longshoremen captured significant value from global trade through control of essential infrastructure. The landlord position — controlling the point through which value must flow — has always been lucrative.

Telecommunications. AT&T's monopoly on telephone infrastructure allowed it to extract rents from every call for decades. The pattern persists in modern telecom: carriers charge both customers and content providers for access to their networks.

App stores. Apple and Google take 15-30% of every transaction in their mobile ecosystems. The "app store tax" drew antitrust scrutiny precisely because platform control enables rent extraction at scale. The cases against Apple (Epic v. Apple) and Google center on whether this control is anticompetitive.

Payment processors. Visa and Mastercard charge 2-3% of every transaction flowing through their networks. Credit card interchange fees are effectively a tax on all commerce, enabled by network effects and infrastructure control.

AI APIs fit this pattern. The providers control the infrastructure — the trained models, the serving infrastructure, the API surface. Everyone building on top pays the toll. The question is whether the toll is fair compensation for capability provided or excessive rent extraction enabled by market position.

Current evidence suggests both. Prices have fallen dramatically — GPT-4o costs 92% less than GPT-4 did at launch — indicating competition and efficiency gains. But the fundamental structure remains: you pay, forever, with no path to ownership.

Who Pays the Inference Tax

The tax falls differently on different actors:

Developers and startups. Small builders pay list price, facing rate limits and capacity constraints that larger customers avoid. A startup competing against an enterprise using the same API pays more per token while having less access. The pricing structure compounds existing advantages.

Enterprises. Large customers negotiate volume discounts of 20-40% and get priority access, dedicated capacity, and service level agreements. They pay less but at greater scale — the absolute dollars flowing to providers are enormous.

Consumers. End users rarely see the inference tax directly, but they pay through subscription prices, feature limitations, and which products get built at all. Applications that can't afford API costs at consumer price points simply don't exist. The economics push AI products toward enterprise markets where customer lifetime value can absorb infrastructure costs.

Developers in lower-income countries. Global pricing means developers worldwide pay the same rates, but those rates represent vastly different fractions of local budgets. A pricing structure set for San Francisco excludes builders in Lagos or Lima not by rule but by economics.

The redistribution runs upward. Value created by applications built on AI APIs flows partially to the applications' creators and partially to the API providers. The platform's share — the inference tax — compounds across all uses. Each successful application means more rent paid. The provider wins when anyone wins, which is the point.

Alternative Models: Escape Routes from the Tax

The inference tax is not inevitable. Alternative approaches exist, each with tradeoffs:

Open-weight models. Llama, Mistral, and other open-weight releases can be self-hosted. You pay for compute directly — renting GPUs from cloud providers or purchasing hardware — but escape the per-token tax. At high volumes, self-hosting becomes economically dominant. But it requires infrastructure expertise, capital for hardware, and acceptance of whatever limitations the open model carries relative to frontier APIs.

As Widder et al. document, "open" in AI doesn't mean what it meant in traditional open-source software (Widder et al., 2023). You get the model weights but not the training data, not the full recipe to reproduce the model, not the human feedback that shaped behavior. It's better than pure proprietary control but not liberation.

Local inference. Running smaller models on personal hardware — laptops, phones, edge devices — eliminates dependency on any provider. This works for some applications: simple text generation, summarization, classification. But frontier capabilities require hardware most users don't have and models most users can't access. Local inference is escape for some, not all.

Cooperative and public alternatives. In principle, AI infrastructure could be operated as a commons or public utility. The Open Source Initiative and various academic efforts pursue truly open models. Some propose public AI services funded by taxes rather than usage fees — Mariana Mazzucato's logic of the "entrepreneurial state" applied to AI (Mazzucato, 2021). Public investment created much of the foundation for current AI capabilities; the returns could flow back to the public.

These alternatives remain marginal. Most AI applications in production today pay the inference tax to one of a few providers. The alternatives constrain pricing at the margins but haven't displaced the dominant model.

The Trajectory: Will the Tax Rise or Fall?

Prices have collapsed since GPT-4's launch in 2023. Will this continue?

Pressures toward lower prices:

Moore's Law continues for inference hardware. Specialized chips from Nvidia, Google (TPUs), and startups improve efficiency. Running models gets cheaper.
Competition intensifies. Anthropic, Google, Meta, Mistral, and Chinese labs like DeepSeek all vie for developers. Price wars attract market share.
Open-weight alternatives anchor pricing. Providers cannot charge dramatically more than self-hosting costs without losing volume customers.
Efficiency improvements. Better model architectures, quantization, distillation, and other techniques reduce compute requirements.

Pressures toward stable or higher prices:

Capability premium. As models gain new abilities — better reasoning, longer context, multimodal understanding — providers can charge more for genuinely new value.
Market concentration. If consolidation continues, fewer providers means less price competition. Foundation model markets tend toward concentration due to the economics of training (Vipra & Korinek, 2023).
Demand growth. If AI becomes truly essential infrastructure, inelastic demand enables higher pricing.
Cost increases elsewhere. Energy prices, chip supply constraints, and labor costs could raise operating expenses.

The most likely path: continued price declines for commodity capabilities alongside premium pricing for frontier features. The tax falls but doesn't disappear. The structure — perpetual rent for access — persists even as the rate changes.

The Redistribution Question

The inference tax embeds a choice about how AI value is distributed. The current structure captures a portion of all value created using commercial AI models and transfers it to a handful of providers.

Is this fair? Arguments on both sides:

In favor of the current model:

Providers invested billions to create these capabilities. The tax funds ongoing research and improvement. Building and maintaining frontier AI infrastructure is genuinely expensive, and providers are not wrong that the costs must be recovered somewhere.
API access democratizes capabilities most users couldn't build themselves. A developer today can access AI reasoning that would have required a PhD team and millions in compute five years ago — for pennies per request. That is a real expansion of access.
Competition is forcing prices down. The 92% price decline since GPT-4's launch is not cosmetic — it reflects genuine engineering effort and competitive pressure. The market is working, at least on this dimension.

Against the current model:

Public investment funded foundational AI research — DARPA grants, university labs, publicly trained researchers. Private capture of publicly enabled returns is a familiar pattern in technology (Mazzucato, 2021).
The structure concentrates market power. A few providers controlling essential infrastructure positions them to extract increasing rents as dependency deepens.
Perpetual dependency is not democracy. Renting capability you can never own maintains hierarchies of access.

As Ursula Franklin observed, some technologies are "holistic" — augmenting human capability — while others are "prescriptive" — reducing humans to following machine outputs (Franklin, 1989). The inference tax is neutral on this question, but the business model it enables is not. Building dependency serves platform interests regardless of whether the underlying AI serves user interests.

What Can Be Done

For those paying the inference tax:

Optimize ruthlessly. Shorter prompts, cached responses, batch processing, smaller models where possible — every efficiency reduces rent paid.

Evaluate alternatives regularly. Open-weight models improve faster than most realize. Self-hosting calculus changes. Rebuild the math periodically.

Diversify dependencies. Running on multiple providers reduces lock-in even if it doesn't reduce cost. The option to switch has value beyond actually switching.

Support public alternatives. Academic AI labs, open-source projects, and public interest AI efforts provide counterweights to commercial concentration. They deserve support even from those not yet using them.

For policymakers:

Antitrust scrutiny. Foundation model markets show early signs of concentration. Action before entrenchment is easier than action after.

Interoperability requirements. Mandating that fine-tuned models be portable, or that data formats be standardized, would reduce switching costs and discipline pricing.

Public AI infrastructure. As with other essential infrastructure — roads, electricity, water — public provision creates an alternative that constrains private extraction.

Research funding for open alternatives. Public money that created the preconditions for AI can fund truly open alternatives rather than subsidizing private capture.

The inference tax is not a law of nature. It's a business model choice, made possible by technological structure and regulatory absence. Different choices are possible.

Every API call is a vote for the current system. Understanding what you're voting for is the first step toward demanding alternatives.

Sources

Srnicek, Nick. Platform Capitalism. Polity, 2017. Publisher
Winner, Langdon. "Do Artifacts Have Politics?" Daedalus 109 (1980). JSTOR
Franklin, Ursula. The Real World of Technology. House of Anansi, 1989. Publisher
Mazzucato, Mariana. Mission Economy. Penguin, 2021. Publisher
Vipra, Jai & Korinek, Anton. "Market Concentration Implications of Foundation Models." Brookings, 2023. Brookings
Widder, David Gray et al. "Open (For Business): Big Tech, Concentrated Power, and the Political Economy of Open AI." SSRN, 2023. SSRN
OpenAI. "API Pricing." 2026. OpenAI
Anthropic. "Pricing." 2026. Anthropic

General

The Inference Tax: How API Pricing Creates Perpetual Rent