The LLM Ecosystem — mapping the forces at play

In brief

The LLM ecosystem is not a unified market: it is a set of interdependent layers — frontier labs, infrastructure providers, open-source actors, Chinese labs — whose equilibria shifted profoundly between 2024 and 2026. The DeepSeek disruption, the rise of open source, and US-China geopolitical competition have rendered obsolete the maps drawn in 2022-2023. What was once called the absolute dominance of closed Western labs is now a position contested on multiple fronts simultaneously.

In short: the LLM ecosystem is a three-story building. On the ground floor, the elevator manufacturers (Nvidia, hyperscalers) keep the whole house running. On the first floor, the architect-laboratories (OpenAI, Anthropic, Google DeepMind, Meta) draw the models. On the second floor, the distributors (Hugging Face, OpenRouter) connect to the tenants. If any floor collapses, the other two cannot function.

Three layers, mutual dependence

Understanding the LLM ecosystem requires distinguishing three levels that stack up without merging.

The model layer groups the labs that train and publish models. This is the visible layer — the one of press releases, benchmarks, and funding announcements. It concentrates most media attention.

The infrastructure layer is less visible but more decisive. It includes chip manufacturers (Nvidia above all, with more than 70% of cloud AI compute hours in 2024) and the three hyperscalers — AWS, Azure, Google Cloud — which hold over 60% of the global cloud market. These players do not build models, but no frontier model gets built without them. The hyperscalers were collectively spending more than $380 billion on AI infrastructure in 2025.

The distribution layer includes platforms like Hugging Face (13 million users, 2 million public models) and inference aggregators like OpenRouter. They are the ones that make models accessible and measurable — the Open LLM Leaderboard, developed with EleutherAI, has become the de facto benchmark for comparing open-source models.

These three layers are linked by asymmetric financial and technical flows. Microsoft invests heavily in OpenAI while providing the Azure infrastructure on which GPT runs. Google builds its own TPUs to reduce its Nvidia dependency while distributing Gemini via Google Cloud. Anthropic raised $30 billion from investors that include Google and Amazon — its two main cloud providers. These entanglements blur the boundary between competitors and partners.

In short: the tenants (the architect-laboratories) must rent their floor from the landlords (the hyperscalers) to exist. Microsoft rents to OpenAI, Google and Amazon rent to Anthropic. The commercial competitor is also the lessor — every time OpenAI wins a customer, Microsoft takes a cut; every time Anthropic wins one, Google or Amazon takes a cut. This is not a war, it is coopetition.

Western frontier labs: a contested hegemony

Between 2022 and 2024, three players clearly dominated: OpenAI, Google, and Anthropic. In 2025-2026, this hierarchy held in broad strokes, but with significant internal redistribution.

OpenAI remains the leader in brand recognition and consumer reach. ChatGPT holds 68% of the consumer market share. With GPT-5 (August 2025), the company reached $13 billion in annual revenue and 800 million weekly users. But its enterprise market share fell from 50% in 2023 to 27% in 2025.

This decline primarily benefited Anthropic, which captured 40% of enterprise LLM spending in 2025 (versus 12% in 2023). The trajectory is clear: Claude’s growth in the coding segment — estimated at 54% market share — drove the whole. Anthropic reached $14 billion in annualized revenue run-rate in early 2026, with a valuation of $380 billion after a $30 billion Series G raise.

Google DeepMind occupies a structurally different position. Its 450 million Gemini users and 2 billion AI Overview users in Search give it a distribution that no other lab can replicate. Its progression to 21% enterprise market share reflects this integrated distribution, more than any isolable technical superiority.

Beyond these three actors, Meta AI plays a distinct part. Llama 4 is published in open weights, by default, in 200 languages. The strategy is explicit: Meta is not trying to monetize the model itself, but to prevent a closed proprietary ecosystem from becoming the global standard. This is a geopolitical stance as much as a commercial one.

Mistral AI is the only European lab of significant scale. With 276 employees, $100 million in revenue in 2025, and a valuation of €14 billion, the Paris-based company proves that a niche positioning — European sovereignty, open-weights, enterprise API — is viable. But the question of its sustainability against the massive capital of American players remains open.

The DeepSeek shock

In January 2025, a Hangzhou startup of around 200 employees published DeepSeek R1 under the MIT license. On mathematical reasoning benchmarks (AIME, MATH-500) and code (SWE-Bench), R1 matched or exceeded OpenAI o1. Inference cost: $0.55 per million input tokens, versus $15 for GPT-4o — a 97% reduction.

This is not merely a price shock. It is a challenge to a structural assumption: the idea that frontier capabilities are proportional to capital invested in training. DeepSeek V3 was trained for $5.6 million according to its technical report — while GPT-4o represented orders of magnitude more.

The immediate impact on financial markets was swift: infrastructure provider valuations fell the day of the announcement. For the open-source ecosystem, the effect was acceleration: other Chinese labs opened their weights under competitive pressure. Baidu, historically closed, published ERNIE 4.5 in early 2025. Alibaba accelerated open-source releases from the Qwen series.

The Chinese ecosystem: a second structural pole

The Chinese ecosystem is not monolithic. Several lineages can be distinguished.

DeepSeek (Hangzhou) stands out for its technical contribution: R1 is the first open-source model to validate pure-RL for reasoning, without prior supervised fine-tuning. This is a contribution to fundamental research, not merely an efficiency demonstration.

Alibaba (Qwen series, 0.5 to 72 billion parameters) and Baidu (ERNIE) represent established industrial players, forced to open their models under competitive pressure. ByteDance and Moonshot AI target different use cases — Doubao for the Chinese consumer market, Kimi K2 for multimodal research.

The most concrete measure of their rise: Chinese models represented less than 2% of weekly tokens on OpenRouter at end of 2024. They reached 30% in some weeks during H2 2025, with an average of 13%. On Hugging Face, they reportedly account for 41% of downloads over 12 months — a figure from a single source, to be treated with caution.

The compression of catch-up time

One synthetic indicator summarizes the evolution of competitive dynamics: the delay between a proprietary frontier model’s release and its equivalent open-source reproduction. This period, which was measured in years during the GPT-3/4 cycle, is estimated at 16 months in 2025. This figure comes from an expert survey (single source, not verified by other studies), but it is consistent with observed facts.

The structural consequence is direct: competitive advantages based on pure technical superiority are temporary. Durable advantage shifts toward distribution (ChatGPT, Google’s integration into Search), enterprise trust (Anthropic in the coding segment), or investment capacity (OpenAI, with a projected valuation of one trillion dollars at its IPO in late 2026).

Infrastructure: the real bottleneck

All competition between labs plays out on a substrate that most of them do not control. Nvidia holds over 70% of cloud AI compute. The three hyperscalers (AWS, Azure, Google Cloud) represent approximately 62% of the global cloud market. The GPU-as-a-Service market is estimated at $5.7 billion in 2025 and could reach $26 billion in 2031.

This de facto monopoly has several consequences. First, it gives hyperscalers leverage over labs: Amazon is both an Anthropic investor and its infrastructure provider via AWS. Second, it shapes industrial policy: the United States imposed export controls on H100 chips destined for China to maintain this advantage. Third, it drives players with the means to develop alternatives: Google’s TPUs, AWS’s Trainium chips, and Meta’s custom chips reduce Nvidia dependency over the long term, but not in the short term.

Chinese labs, deprived of access to the latest Nvidia chip generations, responded by optimizing their computational efficiency. DeepSeek R1 is the most documented result. American export controls produced the opposite of their original intent: they stimulated algorithmic innovation among their targets.

Three unresolved tensions

Concentration vs. decentralization. 88% of enterprise API spending remains concentrated on three players (Anthropic, OpenAI, Google). But on-premises deployments exceeded 50% of the market by mid-2025. These two realities coexist: concentration is real in API usage, while decentralization advances in deployments on proprietary infrastructure.

Open vs. closed: sustainability of the boundary. The assumption of an unbridgeable gap between proprietary frontier models and open source did not survive DeepSeek R1. The question remains open: is 16-month open-source reproduction a durable trend, or is DeepSeek a special case favored by specific conditions (export controls, constrained optimization)?

Geopolitics and fragmentation. US-China competition creates two parallel ecosystems that interpenetrate through open source. The binary “tech race” framing does not capture the reality of a scientific community that continues to publish and cite transnationally. Carnegie Endowment notes that frontier model governance cannot be reduced to the open/closed opposition, nor to geopolitical rivalry.

Choosing a player by need

Key takeaways

• The LLM ecosystem has three distinct layers — models, infrastructure, distribution — linked by asymmetric dependencies that funding announcements do not make visible.
• The redistribution of enterprise shares (Anthropic 40%, OpenAI 27% in 2025) happened faster than most 2023 forecasts.
• DeepSeek R1 demonstrated that a 200-person lab can replicate OpenAI’s frontier capabilities at 97% lower inference cost: pure technical advantage is temporary.
• Nvidia holds over 70% of cloud AI compute; American chip export controls stimulated Chinese algorithmic innovation rather than constraining it.
• The estimated 16-month open-source catch-up delay challenges the durability of competitive moats based solely on model superiority.