MiniMax has announced its most ambitious model yet: an LLM with 2.7 trillion parameters, the largest ever built by a Chinese company, which it intends to release as open source. The news, reported by The Information, lands as the tug-of-war between American and Chinese labs over the AI frontier intensifies, with open source becoming a battleground both commercial and geopolitical.

The staggering figure – 2.7 trillion parameters – isn’t just a trophy. It rewrites the engineering and economic rules for anyone wanting to run the model themselves. At FP16 precision, loading the weights into VRAM alone would demand over 5 terabytes of memory, a requirement that today forces sharding across dozens of cutting-edge GPUs like A100s or H100s, linked by ultra-high-bandwidth interconnects. Even with aggressive quantization to INT8 or INT4, the footprint remains such that multi-GPU nodes are inevitable, with capital costs for acceptable inference latency easily exceeding six figures in dollars. Fine-tuning multiplies the computational burden further, confining the operation to a handful of specialized data centers.

This scenario directly impacts on-premise deployment decisions. An LLM of this scale isn’t suited to lightweight adoption in air-gapped or edge environments: those choosing the self-hosting route face an exploding TCO, spanning hardware procurement, energy consumption, and cooling. At the same time, the open license shifts incentives: instead of paying API fees to a cloud provider, an organization with sufficient resources can internalize the model, but only if it owns the infrastructure to tame it. Data sovereignty, a hot topic in Europe and beyond, returns to center stage: hosting a Chinese model on local servers may appear a regulatory paradox, while entrusting it to a Western hyperscaler raises questions about GDPR compliance and technological dependencies.

For US labs, MiniMax’s move is another squeeze on margins. If the model sees wide adoption, it would shrink the market for proprietary API calls and raise the bar for anyone competing with closed parameters. But there’s a flip side: open source doesn’t equal free. The hardware barrier naturally selects users, favoring those who already own GPU clusters and relegating smaller players to cloud service consumers. A dynamic arises where code availability alone isn’t enough to democratize access; what’s needed is computing power that few can afford.

Structurally, the announcement signals that the parameter race is far from over, despite recent emphasis on efficiency and smaller models. MiniMax is choosing extreme scale to seize global relevance, at a time when multiple Chinese players are pushing open source as a strategic lever. Who wins, in this landscape, isn’t necessarily whoever writes the largest model, but whoever controls the infrastructure layer – the chips, the interconnects, the cooling systems – on which those models will run. This is a game now played in foundries and machine rooms, more than on GitHub repositories.