Large Language Models

Chinese open-weight models seize global developer mindshare

The competitive center of gravity in LLMs is shifting eastward faster than most Western investors anticipated. On OpenRouter's routing platform, the top six most popular models are all open-weight models from Chinese firms — Tencent, Xiaomi, DeepSeek, MiniMax, and Z.ai — a striking data point that undermines the narrative of OpenAI and Anthropic dominance at the developer layer. Moonshot AI's Kimi K2 (1 trillion total parameters, MoE architecture) and Alibaba's Qwen family (200,000+ derivative models on Hugging Face) exemplify the scale and quality these labs are now reaching. Meanwhile, Zhipu AI's public market debut has already broken 1 trillion RMB in market cap within six months, demonstrating that Chinese LLM commercialization is translating into capital markets impact, not just GitHub stars.

A structural shift is underway: enterprises are choosing to own and run AI models rather than pay recurring API fees, driven by cost pressure as AI token spend approaches headcount-scale operating costs, and by control and strategic-risk concerns. This dynamic directly benefits open-weight model providers and infrastructure companies like SGLang (backing 400,000 GPUs at xAI, Microsoft Azure, and Nvidia deployments) and Together AI, while pressuring closed-API labs on retention. The Cohere–Aleph Alpha merger, valued at ~$20B, is best read through this lens — combining enterprise model ownership capabilities across two geographies to serve buyers who want sovereign, on-premise options.

Why it matters · Infrastructure and fine-tuning platforms that enable model ownership — rather than API middlemen — are structurally positioned to capture the next wave of enterprise AI budget.

Both Demis Hassabis (DeepMind) and Dario Amodei (Anthropic) have issued high-profile calls for binding oversight — Hassabis proposing an independent standards body modeled on FINRA to review frontier models before release, ideally stood up before year-end, and Amodei calling for a federal agency with power to block unsafe models. Voluntary pre-release model sharing up to 30 days before launch is emerging as a proposed industry norm. This convergence from the two most credible safety-focused CEOs signals that self-regulation is becoming a competitive and reputational strategy, not just altruism — particularly as the Commerce Department demonstrated willingness to block specific models (Anthropic's Fable and Mythos) over national security concerns.

Why it matters · Labs that build safety-review processes into their release cadence now will hold a structural advantage when formal regulation arrives, while those that don't face binary regulatory risk on major product launches.

Anthropic's Claude Code has been identified as one of the fastest-growing AI products ever, having transitioned from an L2 to an L4 coding agent, which analysis from 晚点聊 LateTalk suggests expanded the addressable market from ~$10B to potentially $10 trillion. Anysphere (Cursor) — reportedly being acquired by SpaceX in a landmark $60B all-stock deal — and Claude Code together signal that agentic coding is graduating from developer novelty to enterprise infrastructure. OpenAI's $400M investment in a company at a $3.8B valuation (signal [9]) further underscores frontier labs' urgency to lock in the agentic workflow layer.

Why it matters · The coding-agent category is bifurcating into foundation-model-native tools (Claude Code, Cursor) that can compound network effects through codebase context, and commodity completions — investors should weight the former heavily.

Of the $141B deployed across LLM-adjacent deals in the last 90 days, $54.8B sits in 'unknown' stage rounds — a classification that overwhelmingly reflects sovereign, strategic, and pre-IPO vehicles that defy standard VC categorization. The $1.5B growth round at a $71B valuation (signal [8]) and the $1B compute deal to accelerate open-source AI (signal [31]) are emblematic: these are infrastructure-scale commitments by hyperscalers and sovereign funds, not traditional venture. Nvidia's 31 deals and Amazon's 13 deals as the most active investors confirm that semiconductor and cloud platforms are using equity as a customer-acquisition and ecosystem-lock-in tool at a scale that dwarfs traditional VC activity.

Why it matters · LPs and fund managers benchmarking against 'LLM deal count' metrics are systematically underweighting the strategic capital distorting valuations and overweighting the signal value of traditional VC rounds in this category.