The True Cost of 'Free' AI
When a product is free, the product is you. That adage was coined for social media. It applies, with even higher stakes, to AI.
ChatGPT, Google Gemini, and their peers offer extraordinary capability at no upfront cost. The zero-dollar price tag makes them feel like a neutral utility, like running water or electricity. They are not. Every conversation, every document you paste in, every question you ask contributes to a commercial operation built on knowing you better than you know yourself. The cost is real. It just doesn't appear on your credit card statement.
What you give away when you type
Read the fine print. OpenAI's terms of service state that content you submit may be used to "develop and improve our Services." Google's terms are similarly permissive. Unless you are an enterprise customer who has negotiated a data processing agreement — and most individual users and small businesses have not — the text you send to these services is training data.
This is not a privacy scare story. Large language models become more capable and more personalized with more data. So, the inputs you provide to it are valuable. Unfortunately, that value flows to the platform, not to you.
The consequences are most acute when you input things that should be private. Medical questions. Legal drafts. Financial projections. Client communications. Business strategies. The designed use of your data is only part of the picture. There is also accidental exposure — which turns out to be just as hard to control. In January 2026, a bug in Microsoft 365 Copilot caused it to read and summarize emails marked with confidentiality labels — the very labels IT departments use to block automated tool access. The bug bypassed data loss prevention policies entirely, processing restricted emails without triggering any alerts. Microsoft didn't disclose the issue until weeks after it began.
Between what AI vendors are permitted to do with your data and what their software accidentally does with it, the attack surface is larger than most organizations realize.
The attention economy is putting on a new hat
"Free" AI products are not primarily AI products. They are data products with an AI interface.
According to Alphabet's full-year 2025 SEC filing, Google's annual revenue exceeded $400 billion — roughly 17% of which came from Google Cloud, the division that includes Gemini. The remainder came overwhelmingly from advertising: from knowing what users want before they know it themselves, and selling access to that knowledge. When you use Gemini to draft an email or search the web, you are interacting with a system whose commercial purpose is building a detailed model of your interests, habits, and intentions.
OpenAI's structure is different but the incentive is similar. The company has raised approximately $168 billion across multiple rounds — including a $110 billion raise in February 2026 alone. Investors at that scale expect returns. The pathway to those returns, for a company that gives its core product away for free, runs through the data generated by that product.
None of this is secret. These companies have filed detailed disclosures with regulators explaining their business models. The problem is that users rarely read those disclosures before pasting their client list into a chatbot.
The switching cost that isn't visible yet
There is a longer-term cost that is harder to price: dependency.
AI tools improve with use. The more you rely on a free AI assistant, the more your workflows, your institutional knowledge, and your muscle memory adapt around it. When that product changes — and all free products change, because their operators need to monetize more aggressively over time — the switching cost is painful.
OpenAI deprecated its Codex API with roughly two days' notice, and its Assistants API is being retired on August 26, 2026. Google deprecated Gemini 3 Pro Preview with eleven days' notice — falling short of its own stated two-week minimum. Teams that had built production workflows on these tools faced unexpected rewrites.
Cory Doctorow coined the term "enshittification" to describe how platform businesses attract users with generosity, then systematically degrade the product to extract more value once users are locked in. In February 2026, the Norwegian Consumer Council published Breaking Free: Pathways to a Fair Technological Future, dedicating an entire chapter to how generative AI is accelerating this pattern. Their conclusion: "The question is not whether generative AI systems will be enshittified, but rather how and when."
The evidence is already visible. OpenAI has announced it will introduce advertising into ChatGPT and allow sponsored product recommendations — outputs that the report notes "may be particularly difficult to detect" for users who treat the system as a neutral advisor. GPT-4 migrated behind $200/month paywalls. API pricing rises as the product improves. Rate limits were applied to paying subscribers.
What privacy-first AI actually costs
The honest answer is: it depends on scale, and the comparison is more nuanced than it used to be.
Cloud API prices have fallen sharply. As of early 2026, GPT-4o costs $0.0025 per 1,000 input tokens and $0.010 per 1,000 output tokens; Claude Sonnet runs $0.003/$0.015; Gemini 2.5 Pro runs $0.00125/$0.010. These are not the 2023 prices your CFO may remember. The raw per-token cost of cloud inference is genuinely cheap.
Local inference has its own economics. NVIDIA's inference benchmarking analysis and independent unit-economics work from Introl show that a busy H100 running a 7B model at ~70% GPU utilization costs roughly $0.013 per 1,000 tokens — comparable to mid-tier cloud API pricing. At higher utilization or with more efficient models, local inference can undercut cloud pricing meaningfully. At low utilization (a server sitting mostly idle), local becomes more expensive per token.
The crossover happens with volume and time. A November 2025 ArXiv paper, "A Cost-Benefit Analysis of On-Premise Large Language Model Deployment", models breakeven scenarios in detail: organizations with consistent, high-volume inference workloads typically recover hardware costs within one to two years, after which the marginal cost per query drops to electricity.
The privacy benefit doesn't wait for breakeven. From the moment the hardware is running, your data stays inside your network — no terms of service that permit training use, no third-party servers, no accidental exposure through someone else's bug. The cost crossover is a financial argument; the data sovereignty is immediate.
The calculus for individuals versus businesses
For individual users without sensitive data, the free AI tools are a reasonable trade. You get a capable assistant. You pay with anonymized data. The risk-benefit is probably positive.
The calculus changes materially for anyone who regularly processes:
- Client or patient information
- Proprietary business data or trade secrets
- Legal, financial, or medical documents
- Competitive intelligence
- Personal communications
For these users — which includes most knowledge workers and virtually all businesses — the true cost of free AI is the gradual erosion of informational sovereignty. Your competitive edge, your clients' trust, your negotiating position: these are what you pay, one query at a time.
The good news is that the open-source AI ecosystem in 2026 is extraordinary. The models are genuinely excellent — DeepSeek-V3's technical report shows it outperforming GPT-4o on MMLU (88.5 vs 87.2) and matching Claude-3.5-Sonnet across multiple benchmarks, while being fully open-weight. The tooling to run them locally has matured significantly: Ollama, with over 163,000 GitHub stars, lets anyone run capable models with a single terminal command — no ML engineering background required. Owning your AI stack is no longer the province of large enterprises; a European marketing agency deployed a self-hosted Mistral instance on a budget Hetzner server in 22 business days, motivated by keeping client data off foreign servers. It is increasingly practical for any organization that handles information it wants to keep private.
Free has a price. Knowing it is the first step toward not paying it.
Shobdo builds privacy-first AI solutions — from speech and audio software to GPU infrastructure you own. No data harvesting. No attention manipulation. Just tools that work for you.