NeuralStackThe latest round of big tech announcements points to a decisive shift in the AI race: from headline-grabbing benchmark scores to the hard work of building national-scale infrastructure and operational foundations. A new report on official disclosures from major players frames the "main battlefield" as the integration of power, semiconductors, data centers, business systems, robots and safety measures, rather than just bigger models. At the same time, a growing ecosystem of open and connected models is expanding the choices available to enterprises that want to put AI to work across their operations.
One of the clearest examples of this infrastructure-first strategy is NVIDIA's plan to build the "NVIDIA Vera Rubin AI factory" in Japan together with Noetra. Described as a national-scale physical AI facility, it will serve as the computing backbone for the FRONTia Project led by Japan's Ministry of Economy, Trade and Industry. The system is designed around large numbers of NVIDIA's latest CPUs and GPUs connected via the Vera Rubin NVL72 and DSX platform, and it will also incorporate technologies like Spectrum-X and BlueField DPUs to support demanding workloads.
NVIDIA positions this AI factory as a hub for so-called physical AI, spanning AI agents, digital twins and robotics in manufacturing, logistics, healthcare and telecommunications. The plan is not just to offer raw compute, but to provide pre-trained weights and software suites such as Nemotron, Cosmos, Isaac GR00T and NeMo broadly to domestic companies and developers. In effect, the project is framed as a core pillar of Japan's AI industrial policy, aiming to accelerate the deployment of multimodal foundation models in real-world sectors and lower the barrier for local firms to adopt advanced AI.
Chip giant TSMC underscored how this infrastructure push is reshaping the semiconductor business with a sizable increase to its 2026 capital expenditure plan. In its latest quarterly earnings report, the company raised its planned investment range and highlighted that expanding AI demand is the driving force. TSMC said it will devote the bulk of this spending to advanced processes, with smaller but still significant portions earmarked for specialty technologies as well as advanced packaging, testing and mask manufacturing, reflecting the complexity of AI-era chip production.
Within that outlook, TSMC pointed to the rise of agentic AI as a key reason demand is spreading beyond GPUs to include CPUs as well. To support this surge, the company detailed an additional multibillion-dollar investment in Arizona in the United States, building on its existing footprint there. Together, these moves signal that AI workloads are now central enough to justify not just incremental fab expansions but an aggressive ramp in global manufacturing capacity and geographic diversification.
Meta is pursuing a similar scale-up on the data center side, announcing plans to grow its AI facility in Richland Parish, Louisiana, to 5GW of compute capacity. The company said the total investment in the region will exceed tens of billions of dollars, including more than a billion dollars for local infrastructure such as roads, water and sewage. Meta also emphasized the regional economic impact: contracts with in-state companies already total well over a billion dollars, the project is expected to create more than a thousand jobs once operational, and the company will fund related power and water infrastructure as well as scholarships at a local community college for data center-related fields.
Alongside these hardware and infrastructure bets, NVIDIA is trying to redefine how AI factories are evaluated, arguing that "performance per watt" is the key metric rather than raw computational output alone. In a company blog, it explained that the amount of tokens a system can generate under a given power cap ultimately determines both revenue and margins for AI services. Its GB300 NVL72 platform is described as delivering substantial performance-per-watt gains over the previous Hopper generation by combining large-scale NVLink-based scale-up, advanced liquid cooling, fine-grained power control through DSX MaxLPS and software techniques like NVFP4 quantization, decoupled serving and KV cache optimization.
Open and connected models are the other major front in this week's announcements. Moonshot AI introduced Kimi K3, a mixture-of-experts model with a total parameter count in the trillions that selects a subset of experts on each call. The company says Kimi K3 incorporates proprietary mechanisms like Kimi Delta Attention and Attention Residuals, supports image understanding and offers a very long context window, making it suitable for extended coding sessions, knowledge-intensive work and complex reasoning tasks. Kimi K3 is available through Moonshot's own services such as Kimi.com, Kimi Work and Kimi Code, as well as via an API for developers.
Why this matters
Taken together, these announcements illustrate that AI is moving from experimental pilots to deep integration with national infrastructure and core business systems. Governments are treating AI factories as strategic assets, chipmakers are retooling their investment plans around AI demand and hyperscalers like Meta are tying AI data centers to local economic development and workforce training. At the same time, advances in open models and system-level efficiency mean enterprises can choose from a broader range of tools, while also being pushed to think about safety issues such as autonomous attacks, shadow AI and prompt injection as first-order management concerns rather than afterthoughts.
In the background, the report notes that enterprise AI has already advanced beyond simple answer support into autonomously handling customer service, core operations, system migration and ongoing operational management. Physical AI, encompassing vehicles, factories and robots, is expanding in parallel with these software and data center investments, hinting at a future where AI systems act directly on the physical world as much as the digital one. Looking ahead, the companies that can align performance-per-watt, robust infrastructure, open ecosystem support and strong safety practices are likely to set the pace in the next phase of the AI race, as the industry moves from isolated models to full-stack, nationally significant AI capabilities.