Meta is committing up to $145 billion（約23兆円） to AI infrastructure this year and building custom chips to overtake Google in frontier AI capability within six months, according to research firm SemiAnalysis. The company is redirecting 3,000 engineers to create proprietary training data through internal reinforcement learning, giving it an advantage over competitors who depend on public data sources. Meta plans to deploy 7 gigawatts of computing power in 2026, doubling to 14 gigawatts in 2027, supported by its custom Iris chip entering production in September.
Summaries like this, in your inbox every morning.
Sign up free →What happened
Meta is deploying massive compute infrastructure—planning to spend up to $145 billion（約23兆円） on AI infrastructure this year, with 7 gigawatts of computing power in 2026 and 14 gigawatts in 2027. The company released developer access to its upgraded Muse Spark 1.1 model and will begin production of its custom AI chip, code-named Iris, in September.
Why it matters
According to SemiAnalysis, Meta is positioned to leapfrog Google in frontier AI within the next six months by building proprietary data pipelines through internal reinforcement learning environments and reallocating 3,000 engineers—advantages that give it an edge competitors relying on public data cannot match. This shift could reshape the hierarchy of AI capability leaders.
What to watch
Meta plans to deploy 7 gigawatts of computing power in 2026 and double that to 14 gigawatts in 2027. The custom Iris chip, designed alongside Broadcom and manufactured by TSMC, cleared bug testing in six weeks and has multi-year supply agreements with Samsung, SanDisk, and Sumitomo Electric.
Meta's aggressive capital and engineering deployment marks a strategic pivot in the AI landscape. By redirecting 3,000 engineers to build a proprietary reinforcement learning environment and tracking employee workflows for training data, the company is sidestepping the bottleneck that constrains competitors: the scarcity of high-quality public data. According to SemiAnalysis, this internal supply chain advantage, combined with unprecedented compute infrastructure—five gigawatt-scale datacenter clusters with custom AI-Backbone networking—positions Meta to scale training workloads in ways commercial data brokers and publicly-dependent rivals cannot replicate.
The infrastructure timeline underscores this commitment. Meta's plan to deploy 7 gigawatts in 2026 and 14 gigawatts in 2027 is paired with the September production launch of its custom Iris chip, designed with Broadcom and manufactured by TSMC. Multi-year supply agreements with Samsung, SanDisk, and Sumitomo Electric lock in the silicon and components needed to sustain this buildup. SemiAnalysis projects Meta will surpass both OpenAI and Anthropic in total AI compute by year-end, a shift that reflects not just spending scale but operational control—designing, manufacturing, and deploying custom silicon rather than relying on third-party hardware. The release of developer access to Muse Spark 1.1 signals that these infrastructure investments are already feeding new model capabilities aimed directly at OpenAI and Anthropic's paying customers, suggesting Meta's hardware and data advantages are beginning to materialize in product.
No discussion yet for this article
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
1 minute a day. The AI essentials.
200+ sources · Email / LINE / Slack
