Meta is moving swiftly to deploy its proprietary AI chips, aiming to reduce its dependence on NVIDIA and lower data center operational expenses.

The company confirmed that its third-generation chip, the MTIA 450 (codename Arke), will begin entering data centers in the first half of next year. Meanwhile, the fourth-generation MTIA 500 (codename Astrid) is expected to complete its design phase within roughly a month, with a broader rollout slated for data centers by the end of 2027.

This chip roadmap marks a decisive step forward in Meta's efforts to curb AI infrastructure spending.

According to Yee Jiun Song, Meta's Vice President of Engineering, the custom silicon delivers greater efficiency than current NVIDIA offerings when running AI models, precisely because the company is handling a substantial portion of the engineering work in-house.

Over the next 12 months, Meta's commitment to deploying its own chips already exceeds 1 gigawatt of electricity equivalent. Song indicated that expansion is expected to accelerate further thereafter, assuming no major downturn in AI market demand.

As of Tuesday's intraday trading, Meta rose 0.18%, while NVIDIA gained 0.74%.

Partnership Approach and Technology Roadmap

Meta is collaborating with Broadcom for chip design and TSMC for manufacturing, reinforcing its broader strategy to reduce reliance on NVIDIA's industry dominance.

Meta first unveiled its in-house chip initiative in 2023. The current third-generation product received its initial batch of 12 sample units from TSMC on September 1, with measured performance deviating only 2% to 3% from simulation results. On the first day of testing, the engineering team successfully ran Meta's own models using the chip.

Song noted that test results reveal no obvious design flaws, though several more months of debugging and optimization are required, while TSMC's manufacturing yields continue to ramp up.

He emphasized that the chip series relies heavily on high-bandwidth memory and targets general inference scenarios, rather than ultra-fast response-time inference markets. "This is our workhorse chip for general inference," he said.

Training Chip Program Scrapped in Favor of Inference Cost Savings

Meta had previously planned a chip codenamed Olympus designed to support both AI model training and inference phases, with a target release around 2028 to 2029.

However, as Song disclosed, that project has been cancelled. The company is now concentrating exclusively on inference chips, with cost considerations serving as a primary driver.

"When you begin scaling up to multi-gigawatt capacity, cost becomes absolutely critical," he said. He pointed out that if a chip handling both training and inference were roughly 30% more expensive, it would be "completely unacceptable" given the scale of deployment.

Meta Superintelligence Labs is contributing to fine-tuning the chips by providing demand forecasts for future AI models, helping boost inference-stage operational efficiency.

Clear Roadmap Ahead with Speed and Throughput as Priorities

Following the completion of Astrid's development, Meta says subsequent chip research will shift focus toward enhancing computational speed and throughput, meaning the volume of AI tasks processed per unit of time. The introduction of optical fiber technology is also expected to further improve chip performance.

"We have a very solid roadmap," Song said. "In the coming years, we expect to keep producing chips that can compete with supplier offerings."

This stance reflects Meta's ambition to progressively gain core technical control over its AI infrastructure through proprietary silicon, while establishing long-term competitive advantages in both cost and energy efficiency.