Meta bets on openness

Meta is not backing into open AI. It is trying to turn openness into distribution, and distribution into control. The company is preparing to release open versions of its next models because it wants developers building on Meta’s tools, Meta’s formats, and eventually Meta’s infrastructure, not just using a chatbot with a Meta logo on it (axios.com). That strategy is not new. Meta has spent the past two years pushing Llama as the model family that anyone can download, fine-tune, and run, and in March 2025 it said Llama had passed 1 billion downloads (about.fb.com). That number matters because open models win differently from closed ones. OpenAI and Anthropic can charge for access. Meta can seed an ecosystem. If a startup builds on Llama, if a cloud provider hosts Llama, if a government agency standardizes on Llama, Meta gains influence even when it is not collecting API fees. The company has said this outright in more polished language, arguing that open-source AI spreads capability widely and helps set the terms of the market (about.fb.com). The catch is that Meta’s models are “open” in the company’s preferred sense, not in the stricter one used by the Open Source Initiative, which has argued that Llama does not qualify because Meta does not fully release the training data and other ingredients needed to reproduce the system (axios.com). That distinction sounds academic until you look at what Meta actually ships. When Meta released Llama 4 in April 2025, it put out Scout and Maverick for developers while keeping its largest model, Behemoth, in training (techcrunch.com; cnbc.com). That is the pattern now coming into focus again. Meta can release enough of the stack to attract developers while holding back pieces that are too expensive, too strategic, or simply not ready. Openness, here, is not surrender. It is selective disclosure. Selective disclosure only works if the economics hold. Training frontier models is brutally expensive, and serving them at scale can be worse. So Meta’s open-model push is being paired with a hard turn into hardware. In March 2026 the company said it was accelerating its custom silicon roadmap, with four new generations of Meta Training and Inference Accelerator chips planned over two years to support ranking, recommendations, and generative AI workloads (about.fb.com). CNBC reported that the newer MTIA chips are aimed at inference, including image and video generation, and are not meant to replace the giant GPU clusters used to train the biggest language models (cnbc.com). That is the real shape of the bet. Meta is not trying to escape Nvidia overnight. It is trying to keep the most expensive parts of AI from becoming a tax paid forever to someone else. Even as it builds in-house chips, Meta has signed huge infrastructure deals with both Nvidia and AMD. In February 2026, Meta announced a long-term partnership with Nvidia for AI-optimized data centers, then days later announced a multi-year agreement with AMD to power up to 6 gigawatts of AI infrastructure (about.fb.com; about.fb.com). The company is diversifying suppliers while building its own silicon because no single lane is safe enough for the scale it wants. Put those pieces together and the strategy stops looking contradictory. Meta wants its models everywhere and its costs nowhere near retail. Open releases help spread Llama as a default layer for developers. Custom chips help make that spread affordable. The more widely Meta’s models are adopted, the more valuable it becomes to tune hardware, software, and data centers around them. One of Meta’s own recent infrastructure posts described custom silicon as being “at the center” of its AI strategy, which is a dry way of saying the company wants the stack from model weights down to the rack (about.fb.com).