Is 90% of AI Subsidized?

Is 90% of AI actually subsidized; and what happens to your stack when the subsidies end?

The $20/month Claude subscription that gives a $15M distribution company access to frontier AI is not priced at cost.

Anthropic’s compute costs, researcher salaries, safety work, and infrastructure dwarf the revenue it generates at that price point. The gap is funded by Amazon, Google, venture capital, and the strategic logic that whoever establishes market position now will recoup it later.

Whether “later” means your prices double next year or the competitive dynamics keep prices low forever is genuinely uncertain.

What is not uncertain is how to build an AI stack that is resilient to price changes rather than exposed to them.

The subsidy question: what the evidence actually says

What is known:

The economics of training and serving frontier AI models are not publicly disclosed in detail. What is available in public filings, analyst estimates, and industry reporting:

• OpenAI’s reported losses have run in the billions of dollars annually, significantly exceeding revenue
• Anthropic has raised multi-billion dollar rounds from Amazon and Google specifically to fund compute costs for training and serving Claude
• Google DeepMind and Google Brain’s AI expenditure is embedded in Alphabet’s broader AI investment, which runs at tens of billions per year
• Sam Altman has publicly referenced the extraordinary compute costs of running frontier models

The gap between what users pay and what it costs to serve them is being funded by investors and strategic partners who believe AI market position is worth the subsidy. This is not a secret; it is the explicit business strategy.

What is uncertain:

Whether and when the pricing will normalize depends on factors that are genuinely unpredictable:

• How fast inference costs drop as hardware (custom AI chips, more efficient serving infrastructure) matures
• Whether open-source models reach frontier-competitive quality and remove the pricing power of closed model providers
• How the competitive dynamic between OpenAI, Anthropic, Google, and Meta evolves
• Whether model providers achieve the usage scale that makes current pricing sustainable through volume

The honest calibration:

The AI subsidy argument is not fringe contrarianism; it is mainstream industry analysis. The question of when and how much prices normalize is genuinely uncertain. The question of whether current pricing reflects sustainable economics is not particularly contested.

The three scenarios: what price normalization could look like

Scenario 1: Gradual normalization through efficiency gains (most likely)

As inference hardware matures (NVIDIA’s next-generation chips, custom AI ASICs from Google and others), the cost of serving frontier models drops significantly. Price increases are gradual and partially offset by efficiency gains.

API costs increase 50–100% over three to five years but remain accessible to mid-market companies. Consumer subscription prices increase from $20 to $40–$60/month over the same period.

Implication for mid-market companies: the ROI of AI automation remains positive. Some workflows that were economically marginal at current prices become uneconomical. The overall stack cost increases meaningfully but does not break the business case.

Scenario 2: Competitive commoditisation (second most likely)

Open-source models reach frontier-competitive quality for most business workflows (already partially happening with Llama 3.1). Commercial providers are forced to compete with free or near-free alternatives. Prices stay low or decrease as the market commoditises the model layer.

Implication for mid-market companies: AI costs remain affordable. The competitive advantage shifts even more strongly toward companies with well-built context layers and workflow systems; the model is free, the foundation that makes it useful is the differentiator.

Scenario 3: Sharp price normalization (least likely but highest impact)

The competitive dynamic shifts; one or two providers establish dominant market position, open-source alternatives stall at below-frontier quality, and the remaining providers increase prices sharply to move toward sustainable unit economics. API costs triple or quadruple over 12–18 months. Consumer subscriptions increase to $80–$150/month.

Implication for mid-market companies: the economics of some automation workflows change significantly. Companies with model-agnostic architectures can migrate to the most cost-efficient provider. Companies hard-coded to a specific provider face higher migration costs.

The three resilience principles: how to build a stack that survives price changes

Resilience principle 1: Model-agnostic workflow architecture

The most important price resilience decision: design workflows to be model-agnostic. The workflow logic; the context pack, the prompt structure, the output format requirements, the human checkpoint; should be documented in a format that can be run on any model, not locked to a specific model’s features or behavior.

What this looks like in practice: prompts that rely on explicit, structured instructions (not on specific model idiosyncrasies), context packs stored in plain text (not in a proprietary AI tool’s format), and workflow documentation that could be rebuilt on a different platform in hours rather than months.

What to avoid: workflows that rely on specific features only available from one provider, workflows where the prompt was tuned so specifically to one model’s behavior that it breaks on another.

Resilience principle 2: Usage efficiency through right-sizing

Not every task requires frontier model capability. Matching the task to the appropriate model tier reduces cost exposure significantly.

The efficiency framework:

Building the workflow architecture with right-sized model selection creates natural price resilience: if frontier model prices increase, shift more tasks to mid-tier; if mid-tier increases, shift more tasks to lightweight.

Resilience principle 3: Foundation portability (the most important)

The value in a well-built AI system is not in the model subscription. It is in the context pack, the workflow documentation, the team fluency, and the adoption data. These are company assets; they exist in documents and practices the company owns, not in any AI provider’s system.

A company with a well-built, portable context layer can migrate its AI stack to any provider; or to local models if the economics justify it; in days rather than months.

The portability test: if you canceled every AI subscription today and signed up with a different provider tomorrow, how long would it take to be fully operational?

• “A few days of testing”: the foundation is portable
• “Months of rebuilding”: there is a portability risk

The cost of building portably is essentially zero. The cost of not doing so is real.

The practical decision: what to do (and not do) about AI subsidy risk

What to do:

Build model-agnostically from the start. Every workflow designed with portability in mind is cheaper to migrate than one that is not. This costs nothing additional to implement; it is a design discipline.

Document your workflows and context pack as company assets. The context pack, workflow documentation, and adoption standards belong to the company; store them in the company’s own systems (Google Drive, Notion, internal documentation), not only in an AI provider’s proprietary format.

Right-size model selection for existing workflows. If a high-volume classification workflow is running on Claude Opus because that is the default, moving it to Claude Haiku reduces current costs and creates price buffer if Opus increases.

Develop familiarity with the leading open-source alternatives. Running occasional test batches through Llama 3.1 70B for your high-volume, low-complexity workflows gives you a realiztic sense of whether local models are viable alternatives; so you are not starting that evaluation from zero if prices change.

What not to do:

Do not cancel AI subscriptions or stop building. The subsidy argument is not a reason to pause AI adoption; it is a reason to build more portably. The ROI of AI automation at current prices is clear; building toward price resilience now is cheaper than rebuilding after prices change.

Do not build elaborate local model infrastructure to avoid price risk. As covered in the local AI hardware article, the economics of local models for most mid-market workflows do not currently justify the maintenance burden. Building local infrastructure as insurance against price increases that may not materialise is likely a worse investment than building a portable cloud-based stack.

Do not lock into long-term AI tool contracts that assume current pricing. Multi-year SaaS contracts with AI-powered tools at current pricing create exposure if the underlying model costs change significantly. Prefer month-to-month or short-term agreements for AI tools where the pricing is most exposed.

Common questions on AI pricing and subsidy risk

”Should I lock in AI pricing contracts now before prices rise?”

No; the contract would lock you in to a specific provider, reducing the portability that is the most valuable resilience asset. The portability thesis says: stay flexible. The right response to potential price increases is model-agnostic architecture and right-sized model selection; not locking into a long-term contract with the provider whose prices might increase.

”Which AI providers are most likely to raise prices?”

The providers most likely to raise prices are those that: establish dominant market position in specific workflow categories, have the clearest path to sustainable unit economics, and have the least competition from open-source alternatives. The providers most likely to hold or reduce prices are those facing strong competition from capable open-source alternatives or from other commercial providers.

”How exposed is a company that uses AI heavily to pricing changes?”

Exposure is proportional to: how much of the AI value comes from frontier models (versus mid-tier or lightweight models), how model-agnostic the workflow architecture is, and how portable the context layer is. A company that follows the three resilience principles has low exposure; a company that has built dozens of workflows hard-coded to one provider at one model tier has higher exposure.

”Are open-source models a viable backup if prices rise?”

For specific workflow categories: yes. High-volume, low-complexity tasks (classification, structured data extraction, simple summarisation) are already viable on well-configured Llama 3.1 70B for some companies. Judgment-intensive tasks (nuanced proposals, complex analysis, strategic recommendations) are not yet viable on local open-source models for most mid-market use cases. Build the local model familiarity now; do not depend on it before you have tested it.

”What is the actual cost of running frontier AI at scale?”

Rough estimates for frontier model serving at the inference scale required for broad deployment: cost-per-token is orders of magnitude above current API pricing at the volumes that would make it economically self-sustaining without investment subsidies. The publicly available compute cost estimates from industry analysts place the fully-loaded cost of frontier model inference at 5–20x current API pricing for most providers; before accounting for training costs.

”How do I explain AI pricing risk to my board or investors?”

Frame it as a build quality question, not a risk mitigation question. “We are building our AI infrastructure model-agnostically and with portable foundations; which means we can migrate providers if pricing changes significantly, and our competitive advantage is in the context layer and workflow design that is a company asset, not in any specific model subscription.” This framing positions the resilience principles as quality decisions rather than defensive reactions.

Want an AI system built with portability and resilience from the start?

Yes, current AI pricing is almost certainly below sustainable long-run cost for frontier models. Whether and how fast prices normalize is genuinely uncertain.

The right response is not to stop building; it is to build with the three resilience principles in mind: model-agnostic architecture, right-sized model selection, and portable foundations that are company assets rather than tool dependencies.

The company that builds this way has the same AI capabilities at current prices and significantly lower migration costs if the pricing landscape changes. The company that builds without these principles has the same capabilities now and a more expensive transition if prices change.

Path one: apply the portability test to your current stack. If you canceled every AI subscription today and signed up with a different provider tomorrow, how long would it take to be fully operational? If the answer is “months,” identify which element (non-portable context, provider-specific workflow features, lack of documentation) is creating the dependency and address it.

Path two: bring in a partner. If you want the AI system built model-agnostically from the start; with a portable context pack, documented workflow library, and right-sized model selection built into the architecture; that is the work Phos AI Labs does by default. We’ve helped 400+ businesses run their entire organization on AI. The fastest way to know if it is the right fit is a conversation. Thirty minutes, no deck. Start here.