Back to the AI frontline? Meta jumps 15% in one week

Author: Bao Yilong; Source: Wall Street Insights

This week, Meta recorded its strongest weekly performance since February 2024 thanks to a series of AI strategic developments, and market confidence in its AI capabilities and commercialization path is significantly rebounding.

On Friday this week, Meta’s stock rose 6% in a single day; its total gain for the week was 14.8%, marking at least the best weekly performance since February 2024, and lifting its year-to-date return to about 1.4%.

The last time Meta posted a similarly strong week was in February 2024, when investors responded positively to the early results of the company’s “Year of Efficiency” cost-cutting plan—intended to counter the negative image created by its earlier, heavy bets on the metaverse through financial discipline.

This stock breakout suggests Meta may be gradually shedding the market label of an “AI laggard,” which would also create room for it to further ramp up its AI strategy.

Wall Street Insights mentioned that on July 9, Meta launched its flagship model Muse Spark 1.1, which has already surpassed Google’s Gemini model across multiple test projects such as Agent capabilities, programming, and multimodality.

Meanwhile, according to Reuters, Meta is advancing its in-house chip mass production plan and greatly expanding its compute infrastructure. Based on this, Deutsche Bank analyst Benjamin Black raised his estimate for potential incremental revenue from Meta’s third-party cloud services from $17 billion to $24 billion.

Research firm SemiAnalysis released a report estimating that Meta’s Meta Superintelligence (MSL) could overtake Google in rankings of leading-edge AI capabilities within the next six months, shifting the AI competition landscape from being dominated by Google and OpenAI to a three-way standoff among Meta, OpenAI, and Anthropic.

Muse Spark 1.1’s low-price strategy aims at a price war

Meta’s Muse Spark 1.1 model launched this week is its first commercially available model with agent programming capabilities at near-state-of-the-art level, and it comes with a paid API interface.

On Thursday, CEO Mark Zuckerberg posted on the X platform, emphasizing that the model’s pricing is “very affordable,” sparking widespread speculation that Meta may deliberately initiate an AI inference price war to pressure competitors.

The Meta Model API offers a $20 free credit per account, and charges on a pay-as-you-go basis: the input price is $1.25 per million tokens, and the output price is $4.25 per million tokens.

Richard Windsor, founder of Radio Free Mobile, noted in a research report on Friday that the launch of Muse Spark 1.1 confirms recent reports that Meta plans to launch a new business selling compute capacity.

Windsor wrote:

There is growing evidence that, given the attractive return rates currently available, Meta will roll out a new business that sells compute capacity to third parties.

He further pointed out that Muse Spark is close to top-tier models in AI programming capabilities, “but the price is only 25% of the latter,” making it highly appealing to the mass market.

In-house chip development and compute expansion, significantly boosting cloud revenue upside

Wall Street Insights mentioned that Meta has planned to start mass production of its in-house AI chip codenamed “Iris” this September. The chip is co-designed with Broadcom, manufactured by TSMC, completes testing in just six weeks, and has already signed multi-year supply agreements with Samsung, SanDisk, and Sumitomo Electric.

In terms of compute scale, Meta plans to deploy 7 gigawatts of compute this year, and to double that figure to 14 gigawatts in 2027.

Supporting these goals are five gigawatt-class “titan” ultra-large data center clusters that Meta is building in parallel, as well as its in-house “AI-Backbone” network architecture. The latter allows Meta to asynchronously scale complex training tasks across geographic distances of thousands of kilometers.

In a research report on Thursday, Deutsche Bank analyst Benjamin Black said the above compute expansion plan implies potential incremental revenue for Meta’s third-party cloud services of about $24 billion—significantly higher than the previous forecast of $17 billion.

He also noted that Meta’s in-house chip could open up a practical path for the company to reduce costs and improve efficiency.

SemiAnalysis: Meta AI could surpass Google within half a year

Wall Street Insights mentioned that research firm SemiAnalysis believes that after a year of aggressive capital investment and an architectural restructuring, MSL is expected to overtake Google in the ranking of leading-edge AI capabilities within the next six months.

The report states that the current “two-horse” pattern led by Google and OpenAI will be rewritten into a three-way standoff among Meta, OpenAI, and Anthropic.

SemiAnalysis’s core assessment hinges on the speed of compute expansion: Meta’s growth trajectory in AI compute scale will cause it to surpass the combined compute of OpenAI and Anthropic by year-end.

Citing an internal memo, Reuters reported that Meta’s cap on capital expenditures for AI infrastructure this year could be as high as $145 billion.

On talent, Meta has reassigned 3,000 engineers to an internal reinforcement learning environment factory to build proprietary data that is difficult for commercial data providers to replicate. Meta also invested $14.3 billion in Scale AI, using this to bring in large numbers of top research talent from institutions including OpenAI and Anthropic.

SemiAnalysis believes that evaluating MSL based solely on current benchmark test results is “seeing the trees but not the forest.” The truly key factor is the momentum for future development, not the highs and lows of the starting point right now.

The report points out that if Zuckerberg maintains the current pace of capital investment, Google could be permanently pushed out of the first tier of global AI ultra-large-scale players.