According to multiple analysts, a seemingly insignificant news item from a few days ago may have become a watershed moment in the AI hardware landscape. Meta plans to mass-produce its self-developed chip, "Iris," starting in September. Internal memos show that the chip underwent testing in just six weeks without any major issues, a speed that is almost miraculous in the chip industry. Meta's goal is straightforward—to reduce its reliance on Nvidia GPUs. Therefore, the underlying logic is worth remembering: when a company's core cost items are monopolized by a single supplier, self-development is always a matter of time, not a question of whether to do it. Meta is not the first, nor will it be the last. Moreover, Meta has done something else it has never done before—charging for its released AI model for the first time. Its inference model, Muse
Spark 1.1, scored 51 in the AI analytics index, an 8-point improvement in three months, outperforming Gemini 3.5
Flash's 50 points. The pricing is $1.25 per million token inputs and $4.25 per million outputs, approximately a quarter of the price of Claude Opus 4.8 and GPT
5.5. In summary, these two moves likely represent a strategic shift for Meta, rather than a mere display of technical prowess. For years, Meta has offered its Llama series for free; now, it's charging for one of the most cost-effective orchestration models on the market, while simultaneously vertically integrating its self-developed chips to significantly reduce inference costs. Its goal isn't to top the benchmark charts, but rather to gain control over pricing.
