Based on the opinions of multiple analysts, a few days ago an inconspicuous news may become a watershed in the AI hardware landscape. Meta plans to mass produce a self-developed chip called "Iris" starting from September. Internal memos show that the testing of the chip only took six weeks and no major problems were found, which is almost a miracle in the chip industry. Meta's goal is straightforward - to reduce dependence on Nvidia GPUs. So the logic behind this is worth remembering, when a company's core cost item is monopolized by a single supplier, self-developed is always a matter of time, not a matter of whether or not to do it. Meta is not the first, nor will it be the last. And Meta has also done something it has never done before - charging for its first AI model release. Its reasoning model Muse Spark 1.1 scored 51 in the Artificial Intelligence Analytics Index, an improvement of 8 points within three months, leading Gemini 3.5 Flash by 50 points. The pricing is $1.25 per million tokens input and $4.25 per million tokens output, which is about a quarter of the price of Claude Opus 4.8 and GPT 5.5. Overall, these two measures may represent a strategic shift for Meta, rather than a show of skill. Meta has been offering the Llama series for free for many years, but now it is charging for one of the most cost-effective orchestration models in the market, while vertically integrating and significantly reducing inference costs through self-developed chips. Its goal is not to top the evaluation list, but to take the initiative in pricing.