💰 OpenAI ($6.6B), 🔒 Anthropic (90%) and 🏛️ Google ($250M)
Meta launches Muse Wonder and Wonder Full, its first models it calls frontier, with active memory across conversations.
Microsoft unveils Maia 300: 40% more tokens per megawatt than leading GPU systems and multi-gigawatt deployments from 2027.
IDC data shows Nvidia’s share of AI data-center chips in China falling to 34% in the first half of 2026.
Today’s three stories draw the same line: anyone who wants to play at the frontier has to control the model, the silicon and the market at once.
🧠 Meta (Wonder), the jump to the frontier
Meta Superintelligence Labs released Muse Wonder and Muse Wonder Full on October 12, the first models the company presents as frontier. Wonder costs $4 per million input tokens and $20 output. Wonder Full is only available to Muse Code subscribers. Both were trained at the Prometheus data center in Ohio. The bet is Active Memory (carries context across conversations) and Active Focus (pursues a goal across multiple steps and checks in before consequential actions).
Meta claims state-of-the-art results on agentic tasks, coding, law, finance and health. Artificial Analysis places Wonder Full near the top. Zuckerberg says Meta has reached the frontier. The strategy shift is clear: the best model is no longer open-weight. It is sold closed and integrated into the Muse agent.
⚡ Microsoft (Maia 300), own silicon
Microsoft announced Maia 300, its third-generation in-house inference accelerator. Each chip carries eight HBM4 stacks (288 GB at 18 TB/s) and a 950 W TDP. A 96-accelerator rack draws about 167 kW and, according to Microsoft, delivers up to 7x the tokens of a Maia 200 rack. The key figure: 40% more tokens per megawatt than leading GPU systems. Multi-gigawatt deployments from 2027. The company talks about more than a million chips next year.
Maia 300 is not an experiment. Microsoft says it will become the primary inference system for its products. Less reliance on Nvidia, more control over cost per token and the agent roadmap. The bottleneck is no longer only the model. It is how many tokens you can serve per watt.
📉 Nvidia (34%), the market that is closing
IDC data released around October 12–13 shows Nvidia shipped 678,000 AI accelerators to data centers in China in the first half of 2026, down 44% year-on-year. Its market share falls to 34%. Huawei rises to 28%. U.S. export restrictions and the rise of Chinese domestic silicon are redrawing the map.
Nvidia still dominates outside China, but the largest growth market for AI infrastructure is no longer theirs by default. Whoever controls the silicon controls the pace of deployment. China is proving it with numbers, not press releases.
🌐 The thread that connects them
Meta closes the best model and sells it as a personal agent product. Microsoft builds the chip that lets it stop depending on Nvidia to serve those agents. Nvidia loses ground in the market that most needs new capacity. The three stories say the same thing in different words: the frontier is not only a benchmark score. It is model + silicon + access to demand. Whoever does not control all three pieces is left counting rankings.