The real danger is not that AI will fail, but that it will succeed too well. Monopolies, dependencies and explosive costs: LLMs are reshaping a risky battlefield for businesses.
Imagine depending on a strategic supplier who burns 14 billion dollars per year to sell you a subscription for 20 euros. This is the current paradox of AI for businesses. The real risk is not thatartificial intelligence fails technologically, but the business model of its publishers succeeds so well that it becomes an absolute monopoly – or it collapses, leaving you orphaned of the critical processes you integrated. OpenAI, Anthropic and Mistral embody three different trajectories, but all rely on rare resources and raise questions of sovereignty that the Comex can no longer ignore.
The main lines of expenditure and revenue
The revenues of the main providers come mainly from two sources: subscriptions (fixed monthly payment for extended access) and programming interfaces (APIs) (payment per use, billed by the number of “tokens”, units of text processed).
At OpenAI, subscriptions (ChatGPT Plus, Pro and enterprise versions) largely dominate, often representing 55 to 85% of revenue according to estimates. This general public strategy ensures massive adoption and strong visibility. APIs complete the picture (15 to 30%), particularly for professional integrations. Annualized revenues reached around $20-25 billion at the end of 2025, with rapid growth but significant losses (projections around $14 billion in 2026).
Conversely, Anthropic derives most of its revenue (70 to 75%) from APIs, particularly for code generation and agent uses. Direct subscriptions remain in the minority (10-15%). This “business” orientation gives revenues more concentrated on large, potentially more stable customers.
Mistral adopts a hybrid model: API for use, subscriptions (Le Chat Pro) and licenses for private deployments. With around $400 million in annualized revenue at the start of 2026 and a target of over $1 billion, the French company is banking on agility and European sovereignty.
The common expenses are colossal: inference costs (running models), training new models, infrastructure and salaries. Structural losses remain high, because the race for performance requires massive investments before margins stabilize.
Critical resources: chips, infrastructure and talent
All players rely heavily on high-performance computing chips, dominated by NVIDIA. OpenAI and Anthropic have committed hundreds of billions of dollars to data centers. Mistral, for its part, is investing in Europe (data centers in France and Sweden, acquisition of Koyeb) to limit this dependence and strengthen sovereignty.
The war for talent is just as intense: engineers specializing in artificial intelligence are rare and highly sought after, which is causing salary costs to explode. Energy constitutes a third bottleneck, with data center power consumption raising environmental and availability issues.
This dependence on chips, however, has an unexpected downside: it has pushed isolated actors, notably Chinese, to drastically optimize software efficiency, calling into question the American “brute force” model.
The evolution of the strategies of the three players
OpenAI has evolved from an open research logic to aggressive commercialization and a focus on practical adoption (autonomous agents, outcome-based pricing).
Anthropic focuses on security and alignment from the start, successfully conquering key accounts via robust APIs and code tools.
Mistral is developing a European path: combination of open source and proprietary models, sovereign cloud and industrial partnerships. The company optimizes costs and regulatory compliance rather than purely racing for scale.
Each strategy engages client companies differently: progressive lock-in among American players, or greater flexibility (but less maturity) at Mistral.
Distribution models and closed ecosystems (Microsoft and xAI)
Microsoft is not a direct model competitor, but a powerful distribution channel and infrastructure provider.
- Copilot (integrated with Microsoft 365) is the daily assistant that helps employees in Word, Excel, Teams, etc.
- Copilot Studio is a development platform that allows you to create personalized and autonomous AI agents, based on several models (OpenAI, Anthropic, Grok, etc.).
The main trap of these closed ecosystems is ‘Shadow IT’ (unauthorized use of AI by company employees and partners). If Copilot (M365) is relatively mastered by the IT department, Copilot Studio allows any business department to create an AI agent connected to company data in a few hours, spectacularly increasing the attack surface and the invisible dependence on Microsoft infrastructure.
xAI (Grok) offers a more disruptive approach, with integration into the X and Tesla ecosystem, “maximum truth” positioning and aggressive pricing. Its availability via Copilot Studio strengthens its access to businesses.
The main trap: these closed ecosystems offer massive distribution and ease of integration, but increase dependence on a single supplier and its strategic choices.
The Chinese case: efficiency through constraint, a business model through the ecosystem
The “DeepSeek” shock and the illusion of brute force:
Contrary to the Western illusion that quality depends solely on the number of GPUs (NVIDIA), Chinese models (DeepSeek, Alibaba’s Qwen) have demonstrated that a massive investment in software engineering (optimization of MoE architectures, fine-grained data management) makes it possible to match the performance of GPT-4 or Claude with a fraction of the hardware resources.
Governance implication: The narrative of the “chip shortage” justifying high prices among Westerners is now contested. The scarce resource is no longer the chip, it is optimization talent.
The ecosystem business model (the Chinese “Loss Leader”):
Chinese model APIs are often discounted (or even free). Their business model is not based on selling compute, but on locking in massive B2B and B2C ecosystems (DingTalk for Alibaba, WeChat for Tencent, Baidu Search). AI is an acquisition and retention cost, not a direct profit center.
Governance implication: A European company that integrates a “free” or low-cost Chinese API does not pay with money, but with the leak of its usage patterns to foreign actors.
The strategy of “Open Weights” versus Open Source:
China has massively adopted the “open weights” model (the code is not necessarily visible, but the model parameters are). This allows Western companies to download these models to run them “on-premise” (locally).
The trap of sovereignty: For a French mid-sized company seeking to escape the American cloud, a local Chinese model seems attractive. But the dependence is shifting: instead of depending on American infrastructure, the company depends on algorithmic updates and cognitive alignment (values, censorship) dictated by Beijing.
Comparative table “War of models: US vs China vs Europe”:
- US (OpenAI/Anthropic): Direct monetization (subscriptions and API), autonomous agent premium, raw dependence on hardware.
- China (Qwen/DeepSeek): Indirect monetization (ecosystems), bonus for software efficiency, geopolitical “open weights”.
- Europe (Mistral): Hybrid monetization, premium for compliance and sovereignty, search for regulatory niches.
Uncertainties and major future trends: pitfalls for managers
Services and prices will inevitably change. We expect an increase for advanced uses (autonomous agents, measurable results) and stronger differentiation. Competition and optimizations should reduce the cost of standard uses.
The major uncertainties relate to:
- Real profitability (high cash burn despite growth).
- Resources (chip shortages, energy costs, talents).
- Geopolitics (sanctions, dependencies).
Outlook for Enterprise Users: Yes, expect changes. The practical recommendations are: diversify suppliers, rigorously measure return on investment, invest in internal skills and implement strong governance (CSRD compliance, dependency audits, sovereignty strategy).
This dependence on the business models of AI publishers is not just a technical subject, it is a major financial and extra-financial risk. Yet, as we recently analyzed, the majority of CAC 40 companies still fail to mention their AI suppliers in their universal registration documents (DEU). Tomorrow, not declaring its dependence on a third-party subscription model (often non-European) or the integration of Chinese open weights locally will be considered by auditors and investors as a breach of duty of care. AI governance also involves shareholder transparency.
Conclusion and recommendations for the Comex
Whatever the choice, American ecosystem, Chinese efficiency or European path, there is a sacrifice: cost, data, values or autonomy. Governance is no longer an option, but the condition for transforming AI into a real sustainable competitive advantage.
Actionable recommendations:
- Map: Immediately assess the overall dependency (models, data, infrastructure), including uses via Copilot Studio or Chinese APIs ignored by the IT department.
- Arbitrate: Distinguish between commodity uses (where cost takes precedence) and critical processes (where sovereignty and stability dictate the choice, for example via Mistral on-premise licenses).
- Reporter: Integrate these dependencies into the next annual report (DEU/CSRD) not as a constraint, but as a signal of maturity in your governance for investors.
The companies that will keep control of their AI strategy will be those that have been able to treat these publishers not as technological partners, but as strategic infrastructure suppliers subject to the same level of requirements as energy or raw materials.
