Mark Zuckerberg wants Meta to own more of the technology that powers its artificial intelligence systems. The social media giant plans to start producing its custom Iris chip in September. That move forms one piece of a broader effort to cut dependence on suppliers like Nvidia.
But the strategy comes with a staggering price tag. Meta now expects to spend as much as $145 billion on AI infrastructure this year. The figure nearly doubles previous guidance. Shares fell 3.5 percent in premarket trading after the update. Investors remain unsure how the company will earn enough to justify the outlay.
The Iris chip belongs to Meta’s MTIA project. Engineers launched the first versions in 2023. Their goal was clear from the start. Reduce reliance on outside chipmakers. Internal tests of Iris wrapped up after six weeks. No major problems surfaced. Production will ramp through a partnership with Broadcom for design and Taiwan Semiconductor for manufacturing. Yahoo Finance first detailed the timeline and partners.
Meta has locked in other pieces of the supply chain too. Long-term deals cover memory from Samsung, flash storage from Sandisk, and fiber optics from Sumitomo Electric. The company aims to double computing capacity from seven gigawatts this year to 14 gigawatts next. Those numbers reflect ambition. They also highlight risk.
Zuckerberg has talked openly about building toward artificial general intelligence. The spending supports that vision. Yet Meta lacks the cloud revenue streams that help offset similar costs at Microsoft or Google. This difference leaves analysts uneasy. Recent reports suggest Meta may develop its own cloud offerings to monetize excess capacity. The stock recovered 9.47 percent over five days after those stories circulated. Still, year-to-date performance trails the S&P 500.
Meanwhile Meta’s approach to model releases shows signs of strain. The company built its reputation in AI partly through open-source Llama models. Llama 4 Scout and Maverick arrived last year as natively multimodal open-weight releases with massive context windows. They drew more than a billion downloads. Developers embraced them. Meta’s own blog celebrated the milestone.
But the launch of Muse Spark in April marked a shift. This model came from the new Meta Superintelligence Labs led by Alexandr Wang, recruited from Scale AI. Muse Spark stayed fully proprietary. No open weights. No free downloads. The company spent $14.3 billion on the underlying rebuild. It tore down the old stack and started fresh. Artificial Intelligence News reported the details and the resulting tension with the developer community that fueled Llama’s success.
Meta has signaled it may open-source versions of future models. Sources told Axios the company wants to keep certain pieces proprietary at first to manage safety risks. That hybrid path reflects caution. It also risks alienating the open-source advocates who once praised Meta’s stance.
Recent product moves have stirred fresh controversy. On July 7 Meta introduced Muse Image, its first image generation model from the Superintelligence Labs. The tool lets users create visuals tied to their own content and world. It rolled out across Meta AI, Instagram Stories, and WhatsApp in select markets. About.fb.com carried the announcement.
One feature allowed referencing public Instagram accounts in prompts. Backlash hit within days. Critics called it prone to misuse. Meta pulled the capability on July 10. The company cited feedback that the option missed the mark. It also added a hidden provenance signal to images so people can verify their origin. A detection tool now sits at meta.ai/identification. BBC covered the swift reversal.
These episodes reveal the tightrope. Meta pushes hard on AI capabilities while trying to contain downsides. The Iris chip and supply chain investments promise more control. They also lock in enormous fixed costs. If open-source momentum slows, the company may lean harder on proprietary offerings. That pivot could boost margins. It might also invite regulatory heat over market power.
Industry watchers point to the larger picture. Open models from Meta have already shaped competition. They force closed labs to justify their premiums. Yet frontier performance still favors the biggest spenders. Muse Spark reportedly competes at the top tier. Its closed nature protects the secret sauce. Meta’s AI blog described the new family as built for high visual fidelity and reasoning.
Executives face questions about returns. Capital expenditure at this scale demands results. Advertising remains Meta’s cash engine. AI enhancements to ad targeting and content moderation deliver value there. But the infrastructure bet extends far beyond. Zuckerberg has framed it as necessary infrastructure for the next decade. Skeptics see a potential cash sink.
And the confusion persists. Meta sells access to spare capacity even as it races to build more. That dual track puzzles some investors. One theory holds that excess capacity will feed a future cloud business. Another suggests it simply reflects overbuilding. Either way, the numbers are huge.
Broader conversations on X highlight the trade-offs. Users debate whether Meta’s open-source contributions still matter now that flagship models turn closed. Some praise the efficiency gains from Llama derivatives. Others argue the real money stays with proprietary frontier systems. One recent thread noted how cheaper capable models could expand total AI usage and ultimately lift demand for chips and servers.
Meta’s stock reaction shows the market isn’t convinced yet. The rebound after cloud rumors offers a hint of optimism. But sustained gains will require proof that all this spending translates into durable competitive advantage. Iris entering production next month will mark a tangible step. Whether it delivers the independence Zuckerberg seeks remains the open question.
The coming quarters will test the thesis. If Meta can integrate its custom silicon, control costs, and balance open and closed releases, it may emerge stronger. Missteps on safety, backlash over features, or failure to monetize the build-out could force a rethink. For now the company presses ahead. The scale of its ambition leaves little room for half measures.
WebProNews is an iEntry Publication