In sum – what we know:

• An aggressive roadmap – Meta plans to release four new generations of MTIA chips within two years: MTIA 300, 400, 450, and 500.
• The shift to inference – Later generations prioritize generative AI inference costs over training, unlike mainstream chips adapted from training hardware.
• Quick release cadence – Meta has developed the capacity for a six-month release cycle, significantly faster than the industry standard of one to two years.

Meta has laid out a roadmap for its upcoming MTIA chips, with a hefty four new chips set to roll out over the next two years. While it’s been clear that Meta is working hard on in-house silicon, this represents a pretty significant gear-change, and Meta has clearly come a long way from the experimental efforts it announced in 2023.

Here’s a look at the new chips and what makes them unique. The real question going forward is whether this aggressive push actually delivers the efficiency and cost improvements the company is counting on — and what it does to the relationship with suppliers like Nvidia.

The new chips

The new chips span workflows, and arc from specialized to general-purpose. The first, called the MTIA 300, is already in production, and is purpose-built for ranking and recommendations training — the bread and butter of Meta’s ad and content systems. It’s the most narrowly scoped chip in the lineup and represents where Meta’s custom silicon efforts stand today.

Things get considerably more ambitious from there. The MTIA 400 is the first chip designed to handle all workloads, with generative AI inference as its primary target. The MTIA 450 follows the same all-workload philosophy but pushes further on gen AI inference optimization, with ranking, recommendations, and training as secondary priorities. The MTIA 500 closes out the roadmap with the same optimization focus as the 450, presumably bringing additional performance or efficiency improvements.

The 400, 450, and 500 series are slated to roll out by the end of 2027, with all four generations expected to be running in production by the end of that window. That is an extraordinary volume of silicon to bring online in such a compressed timeframe, and it’ll be worth watching closely whether Meta can sustain quality and reliability while moving that fast.

Technical architecture

Modularity is a defining principle to Meta’s efforts. New MTIA generations are engineered to slot into Meta’s existing rack infrastructure without forcing full system redesigns, which obviously makes sense given how quickly new chips are being announced. On top of that kind of modularity, it’s also worth noting that the chips are built on a chiplet architecture — the MTIA 300, for instance, comprises one compute chiplet, two network chiplets, and several HBM stacks. Each compute chiplet houses a grid of processing elements, or PEs, with some redundancy baked in to improve yield. Each PE packs two RISC-V vector cores, a dot product engine for matrix multiplication, a special function unit for activations, a reduction engine for inter-PE communication, and a DMA engine for moving data in and out of local scratch memory.

The MTIA 400 doubles compute density by combining two compute chiplets and introduces support for enhanced MX8 and MX4 low-precision formats important for efficient gen AI inference. A rack of 72 MTIA 400 devices, connected via a switched backplane, forms a single scale-up domain. The MTIA 450 then doubles HBM bandwidth over the 400 — jumping from 9.2 TB/s to 18.4 TB/s — while boosting MX4 FLOPS by 75% and adding hardware acceleration to alleviate Softmax and FlashAttention bottlenecks. The MTIA 500 pushes the modular philosophy even further, adopting a 2×2 configuration of smaller compute chiplets surrounded by HBM stacks, two network chiplets, and an SoC chiplet providing PCIe connectivity to the host CPU and scale-out NICs. It bumps HBM bandwidth another 50% to 27.6 TB/s and offers up to 512 GB of HBM capacity.

From MTIA 300 to MTIA 500, Meta says HBM bandwidth increases by 4.5x and compute FLOPS jumps 25x when comparing MX4 performance across the lineup. For context, the MTIA 400 already hits 12 PFLOPs in MX4, and the MTIA 500 reaches 30 PFLOPs — alongside 5 PFLOPs in BF16. Those are numbers that put it in the neighborhood of leading commercial accelerators, though independent validation is still lacking.

On the software side, the chips are built on widely adopted standards, including PyTorch, vLLM, Triton, and Open Compute Project specifications. That’s a notable choice because it signals Meta isn’t trying to build a fully proprietary, walled-off ecosystem. Since PyTorch originated at Meta, the MTIA stack takes a PyTorch-native approach — developers can use torch.compile and torch.export to capture and optimize model graphs without any MTIA-specific rewrites, and models can run on both GPUs and MTIA simultaneously. Under the hood, MTIA-specific compilers built on Torch FX IR, TorchInductor, Triton, MLIR, and LLVM translate those graphs into optimized device code. Meta has also integrated vLLM support through a plugin architecture, replacing key operators like FlashAttention and fused LayerNorm with MTIA-specific kernels and inheriting features like prefill-decode disaggregation and continuous batching. The runtime, notably, uses a Rust-based user-space driver rather than a traditional in-kernel Linux driver, with firmware written in bare-metal Rust for low latency and built-in memory safety.

Aligning with common frameworks makes it straightforward to move workloads between custom and third-party hardware, which is a smart hedge given the inherent risks of any custom silicon program. Meta says these chips deliver better compute and cost efficiency than general-purpose silicon for its specific ranking, recommendation, and content delivery workloads, though independent benchmarks to back those claims are still hard to come by.

The inference strategy

Strip away the details and the MTIA push boils down to reducing dependence on external GPU suppliers, though that has long been the case with in-house silicon. Nvidia dominates the AI hardware market, and supply constraints have been a persistent headache for every major tech company trying to scale AI infrastructure. By shifting massive inference workloads onto custom chips, Meta gets to control its own infrastructure margins and drive down cost-per-workload without being at the mercy of someone else’s production schedules or pricing.

Meta, however, frames this as a “portfolio approach” rather than a full replacement of third-party hardware. The company acknowledges no single chip covers all its needs and says it’ll keep sourcing from multiple suppliers. MTIA sits at the center of the strategy, but it’s not the only chip in the company’s stack. Still, the direction is clear — Meta wants far more control over its silicon destiny.

The rapid iteration cycles the company is targeting — six months or less between generations — are built around the idea that AI techniques and workload demands can shift dramatically in a matter of months. Being able to iterate on hardware faster than traditional semiconductor timelines could be a genuine edge. But faster cycles also mean more surface area for things to break, and the industry’s longer timelines exist for good reason. After all, chip design and fabrication are extraordinarily complex.

Meta’s announcement fits into an accelerating industry-wide shift toward proprietary AI infrastructure. Google has TPUs, Amazon has Trainium and Inferentia, Microsoft has Maia. MTIA’s expansion slots right into that trend. Meta, however, does still have a multi-billion-dollar relationship with Nvidia. For now, Meta appears to be threading the needle by running both tracks simultaneously. As MTIA scales, though, the balance of that relationship will inevitably shift.

In sum – what to know:

AI demand drives early commitments – More than 90% of DC Tuas capacity was secured before launch, reflecting strong demand for AI-ready infrastructure in Singapore.

Hybrid cooling for AI workloads – The facility combines air cooling and direct-to-chip liquid cooling to support rising power densities from AI and high-performance computing deployments.

High-density AI infrastructure – DC Tuas supports rack densities of up to 200 kW and expands Nxera’s Singapore capacity to 120 MW.

Strong demand for AI-ready infrastructure helped drive early commitments at data center firm Nxera’s new DC Tuas facility in Singapore, with more than 90% of its capacity secured before the site officially opened, Nxera CEO Bill Chang told RCR Wireless News.

Nexera is the regional data center arm of Singtel Group. The new facility represents Nxera’s largest data center in Singapore and expands the company’s local capacity as demand grows for infrastructure capable of supporting artificial intelligence workloads. The site delivers 58 megawatts of AI-ready capacity and increases Nxera’s total data center footprint in the country to 120 MW.

According to Chang, the facility was designed to support a wide range of computing needs, including traditional enterprise applications as well as more demanding AI and high-performance computing deployments.

“DC Tuas is purpose-built to be able to support a broad spectrum of workloads – from conventional enterprise and cloud applications to next-generation AI and high-performance computing,” Chang said.

The eight-storey facility spans approximately 120,000 square feet and was built to accommodate the higher rack densities required by advanced AI infrastructure. As a carrier-neutral, multi-tenant facility, it allows multiple operators and enterprises to deploy computing resources within the same site.

Chang explained that traditional enterprise and cloud applications typically require less power density and can be supported using conventional cooling technologies.

“Traditional enterprise or cloud workloads typically operate at lower power densities and can be supported through air cooling,” he said.

“In contrast, advanced AI workloads require significantly higher power and heat densities, which call for liquid cooling solutions,” Chang said.

To address this shift, DC Tuas was designed with hybrid cooling capabilities that combine air cooling with direct-to-chip liquid cooling. This approach allows operators to run both conventional workloads and more intensive AI systems within the same facility.

“DC Tuas’ hybrid cooling capabilities offer customers the flexibility to deploy today’s workloads while future-proofing their infrastructure as AI requirements continue to evolve,” Chang said.

Nxera says the facility hosts Singapore’s largest deployment of direct-to-chip liquid cooling within a multi-tenant data center. The technology is designed to remove heat more efficiently from high-performance servers while reducing energy and water consumption compared with conventional cooling systems.

Chang said the company expects computing demands to continue increasing as artificial intelligence models become more complex and computationally intensive.

“As AI adoption accelerates and models become more compute-intensive, power density requirements are expected to continue rising,” he said.

In response to this trend, liquid cooling technologies will become increasingly important in supporting large AI clusters.

“Liquid cooling technologies would thus need to be able to manage higher thermal loads more effectively and efficiently,” Chang said.

The DC Tuas facility was designed to accommodate this trajectory, enabling operators to increase rack power densities as AI infrastructure evolves.

“DC Tuas has been built with this trajectory in mind, enabling it to accommodate increasing power densities over time and support customers as their AI infrastructure and needs grow,” Chang said.

Chang added that facilities designed for efficiency and high-density computing are particularly important in Singapore, where space and resources are limited.

The facility also forms part of the broader digital infrastructure platform operated by Nxera’s parent company, Singtel. The site is integrated with a cable landing station that provides direct access to both domestic and international networks, enabling lower latency connectivity and improved network performance for customers operating large-scale computing workloads.

In sum – what to know:

Nuclear power to back AI growth – France plans to use its large nuclear energy capacity to support the expansion of AI data centers requiring vast amounts of electricity.

Surplus low-carbon electricity available – The country exported about 90 TWh of decarbonized electricity last year, giving it room to expand computing infrastructure.

AI investment strategy – France hopes its stable nuclear power supply will attract companies building large AI data centers and high-performance computing facilities.

French president Emmanuel Macron said the country intends to rely on its nuclear energy capacity to support the rapid expansion of artificial intelligence infrastructure, particularly energy-intensive data centers.

Speaking at the World Nuclear Energy Summit in Paris, Macron said the country’s extensive nuclear power system gives it a strategic advantage as demand grows for the computing resources required to run advanced AI models.

According to Macron, France’s existing energy mix could help sustain the growing electricity needs of large-scale computing infrastructure. Nuclear power plants already provide the majority of the country’s electricity and produce a substantial volume of low-carbon power that can be directed toward digital infrastructure.

The French president also noted that the country exported roughly 90 terawatt hours of decarbonized electricity last year, most of it generated by nuclear facilities. This surplus capacity, he said, creates an opportunity for France to host additional data centers without putting pressure on domestic electricity supply.

The French government sees nuclear power as a stable and low-emission energy source capable of supporting the significant electricity requirements associated with AI systems. Training and operating large AI models requires huge computing power, which in turn drives high levels of energy consumption in the data centers that run them.

As demand for AI infrastructure accelerates globally, governments and technology companies are increasingly searching for locations that can provide reliable, large-scale electricity supply. Macron said France’s nuclear fleet places the country in a strong position in the global competition to attract investment in artificial intelligence infrastructure.

By combining its existing supply of decarbonized electricity with new investments in computing capacity, France hopes to attract technology companies seeking locations for large AI-focused data center campuses. The strategy forms part of a broader effort by the French government to position the country as a major hub for artificial intelligence development.

RCR Tech recently spoke with two experts who explained the challenges of using nuclear power in AI data centers.

In sum – what to know:

Major AI infra investment – Bridge Data Centers plans to invest between S$3B and S$5B in Singapore to expand AI-ready data center infra and support regional demand for cloud and AI computing.

Energy strategies becoming critical – The company is exploring alternative power solutions including floating hydrogen generation and research into nuclear and other low-carbon energy sources for data centers.

Regional capacity expansion – While Singapore remains a connectivity hub, additional computing capacity is being developed across Malaysia, Thailand, and India to support growing AI workloads.

APAC data center firm Bridge Data Centers (BDC) plans to invest between S$3-S$5 billion ($2.34-$3.90 billion in Singapore to expand AI-ready data center infrastructure, as demand for computing capacity accelerates across the region.

The Singapore-headquartered company, backed by Bain Capital, said the investment will support the development of next-generation digital infrastructure and help reinforce Singapore’s role as a regional hub for AI and cloud services.

Company representatives said the investment is part of a broader effort to expand computing capacity and support regional demand for AI infrastructure.

The firm currently operates hyperscale data center campuses across Southeast Asia and India and aims to expand its regional capacity to about 2 gigawatts by 2030.

Energy supply and sustainability are becoming central considerations in data center expansion as AI workloads increase power consumption. Bridge Data Centers said it is exploring alternative energy solutions and partnerships aimed at strengthening long-term power strategies for AI infrastructure.

One initiative involves a collaboration with Concord New Energy to develop what the partners describe as Singapore’s first floating hydrogen power generation system designed for data center applications.

Bridge is also collaborating with several technology vendors to develop cooling and power systems designed for high-density AI computing environments.

While Singapore serves as the company’s headquarters and a key interconnection hub, the company said it expects future AI capacity to be distributed across multiple markets in Asia-Pacific.

Bridge Data Centers is assessing nuclear energy as a potential power source for future data center infrastructure, as operators search for low-carbon energy options capable of supporting rapidly growing AI workloads.

The company recently announced it has signed a letter of intent with Singapore’s A*STAR Institute of High Performance Computing and engineering consultancy HY with the aim of evaluating the feasibility of nuclear energy for next-generation, AI-ready data centers.

Under the terms of the collaboration, the research institute will conduct physics-based simulations to analyze risks related to hydrogen storage and its use as a sustainable energy source for data center operations. The results will then inform a broader risk assessment focused on possible nuclear power plant designs and their system architectures.