The Trillion-Dollar Con: Why AI Companies Are Betting You’ll Get Addicted Before the Math Catches Up
By Seven Labs | June 2026
Why AI Companies Are Betting?
OpenAI is reportedly valued at over $300 billion. Anthropic crossed $60 billion. Microsoft has sunk more than $13 billion into OpenAI alone. Analysts throw around projections like “AI will add $15.7 trillion to the global economy by 2030.”
And yet, OpenAI reportedly lost over $5 billion in 2024 on roughly $3.7 billion in revenue. Anthropic is burning capital at a pace that keeps investors writing cheques just to keep the lights on. The compute costs to run these models are staggering — and they’re not coming down fast enough.
So here’s the question nobody in the hype cycle wants to answer cleanly:
If these companies are barely managing compute costs today, where exactly does the trillion-dollar ROI come from?
The honest answer is not comforting.
The Numbers Don’t Add Up — Yet
Running a frontier LLM at scale is brutally expensive. Every ChatGPT query costs fractions of a cent in compute, but at hundreds of millions of daily users, fractions become tens of millions of dollars per month. Training a single frontier model costs hundreds of millions in GPU hours. The next generation will cost more.
The classic tech startup playbook is: lose money acquiring users, achieve lock-in, then raise prices once alternatives disappear. Amazon ran this play on retail for a decade. Uber did it on taxis. Streaming services did it on cable.
AI companies are running the same play — just on a much larger scale, with much higher infrastructure costs, and against a backdrop of openly hostile open-source alternatives (Meta’s Llama models, Mistral, DeepSeek) that make lock-in genuinely hard.
The trillion-dollar ROI projections assume one or more of the following:
- AI replaces enough human labor that the productivity gains justify the cost
- AI platforms achieve deep enough workflow lock-in that switching costs become prohibitive
- Compute costs fall dramatically through new hardware and efficiency gains
- AI unlocks entirely new economic activity that doesn’t exist today
Some of these are plausible. Some are more speculative than the projections let on.
The Addiction Playbook
Here’s where the strategy becomes easier to read if you’ve watched consumer tech for the last two decades.
The goal is not to sell you a tool. The goal is to make you structurally dependent before the free trial ends.
Phase 1 — Habituation. Make the product so useful, so fast, that it becomes part of your daily workflow. GitHub Copilot in every IDE. ChatGPT in every browser tab. Claude as your thinking partner. The friction of not using it grows every week.
Phase 2 — Integration. Move beyond chat. Get into your calendar, your email, your codebase, your customer data. The deeper the integration, the higher the switching cost. This is why every major AI company is racing to build agents, memory, and connectors to enterprise software.
Phase 3 — Lock-in. Once your team’s workflows, institutional memory, and muscle memory are built around a specific platform, migrating is a multi-month project. This is when pricing power returns.
Phase 4 — Monetization at scale. Raise prices. Introduce tiered enterprise plans. Charge per seat, per token, per workflow. The ROI projections start to make sense — but only at this stage, and only if you’re still the platform people are locked into.
This is not a conspiracy. It is a business model. It is rational, and every major technology transition has followed a version of it. The question is whether AI companies will survive long enough to reach Phase 4 before compute costs, open-source competition, or regulatory pressure disrupts the path.
What the Skeptics Are Getting Right
There is a credible bear case, and serious people are making it.
The core argument: AI produces impressive outputs but doesn’t yet reliably produce verifiable business value at the scale the valuations require. Demos are spectacular. Production deployments are harder. Hallucinations in enterprise contexts aren’t just embarrassing — they’re expensive. The ROI on AI investments, when measured rigorously, is uneven and often disappointing outside of specific narrow use cases.
Gary Marcus, Timnit Gebru, and others in the “AI skeptic” camp have been arguing for years that the gap between benchmark performance and real-world reliability is being obscured by motivated reasoning and investor enthusiasm. They’re not wrong that the gap exists. Where the debate continues is whether it’s a fundamental ceiling or an engineering problem that continued investment will solve.
The trillion-dollar projections also tend to measure gross economic activity — not net. If AI automates $1 trillion worth of work, but that displaces $800 billion in human wages, the net economic gain is $200 billion. A large number, but considerably less than the headline.
What the Bulls Are Getting Right
To be fair, the skeptics have also been consistently underestimating capability jumps. GPT-2 was dismissed as a party trick. GPT-4 is running medical diagnostics, legal document review, and software architecture design at a level that would have seemed implausible five years ago.
The compute cost problem is not static. Inference efficiency is improving. Custom silicon (Google’s TPUs, Amazon’s Trainium, Groq’s LPU) is making inference meaningfully cheaper per token every year. The curve that matters is not today’s cost — it’s where costs are heading as the hardware ecosystem matures around AI workloads.
And the addiction hypothesis — whatever you think of the ethics of it — is already working. Developers genuinely cannot imagine going back to coding without autocomplete. Knowledge workers who use AI for drafting, research, and synthesis are measurably faster. The dependency is real and growing.
The Honest Assessment
Here is what we believe at Seven Labs, after three years of building production AI systems for real clients:
The trillion-dollar number is probably not wrong in the long run. It’s just wrong about the timeline.
The companies currently burning capital are making a bet that the lock-in will stick long enough for the economics to flip. That bet could be right. It could also collapse if open-source models catch up fast enough, if regulation forces data portability, or if enterprises realize they can run smaller specialized models on their own infrastructure at a fraction of the cost.
What concerns us more than the financials is the behavioral layer. The addiction-then-monetize playbook has a structural incentive to prioritize engagement over genuinely useful outputs. A tool that makes you feel productive is not the same as a tool that makes you actually productive. The metrics that matter to an AI company’s valuation — DAU, session length, messages sent — are not the same metrics that matter to your business.
The trillion-dollar ROI is real. Some company will capture it. But it will go to whoever builds the most indispensable workflows — not whoever has the best benchmark scores.
For businesses building on AI today, the strategic question is not “which AI company will win?” It’s “how do I extract the real productivity gains available right now, without building dependencies that will cost me more than those gains in 18 months?”
That is exactly the kind of question we exist to answer.
What This Means If You’re Building on AI
A few practical conclusions:
Avoid single-vendor AI dependencies for core workflows. Build abstraction layers. Use orchestration frameworks (LangChain, LlamaIndex) that let you swap underlying models. The model that’s best today will not be best in 12 months — and prices will fluctuate.
Measure actual output quality, not just speed. AI makes things faster. That’s real. But faster wrong answers are not better. Build evaluation pipelines that measure accuracy and business outcomes, not just response latency.
Own your data and your pipelines. The companies that build proprietary training data and fine-tuned models on their own infrastructure will have significantly more leverage than those who are pure API consumers when pricing pressure comes.
The economic value is real in specific places. RAG-powered knowledge retrieval, document processing, code generation assistance, customer support routing — these have measurable, auditable ROI today. The trillion-dollar aggregate projections are not evenly distributed across all use cases.
The question is not whether AI is worth it. It is which AI, implemented how, measured against what outcomes.
Anyone selling you on the trillion-dollar number without answering those questions is selling you the addiction, not the outcome.
Seven Labs builds production-grade AI systems, automation infrastructure, and secure platforms for businesses that want real outcomes — not demos. If you’re trying to figure out where AI actually makes sense in your operations, let’s talk.
📅 Book a call: calendly.com/sevenlabsolutions/30min
🌐 Website: sevenlabs.site
💻 GitHub: github.com/SevenLabSolutions
🔗 LinkedIn: linkedin.com/company/115781914
Tags: AI strategy, AI economics, OpenAI, Anthropic, enterprise AI, automation, Seven Labs
Top comments (0)