When the first time artificial intelligence occurred for the first time in 2010, general purpose central processing units (CPU) and graphical processing sections (GPU) were sufficient to manage early nervous networks, photo generators and language models. However, until 2025, the rising of the agent AI, which has changed the equation of real-time thinking, planning and operating models.
With one click, these AI-power assistants can manage HR tickets to manage your business locations and customer surveys and manage supply chains.
“We are going to a world where hundreds of specialized, special models are known as human teams,” says Vamsi Boppana, Advanced Micro Devices (AMD) AI group SVP. “These models communicate with each other with each other, the processing of traditional data processing delay. This machine-machine interactions unlock a completely new intelligence.”
When the company integrates AI agents into live workflows, they understand that the real autonomy requires a fundamental new calculation fund.
“The transition to an unprecedented demand for a static inefficient transition to work, accounts, memory and an unprecedented infrastructure with the return requirement, the unprecedented infrastructure puts unprecedented pressure,” Boppana added. “Ultra-low delay data processing, aware of memory, a dynamic orchestra and energy efficiency is no longer optional – they are important.”
To support these requirements, the industry is progressed to a special silicone specially designed for autonomous agents. Technological leaders such as Meta, Openai, Google, Amazon and Antropic, now make silicone, infrastructure and orchestra layers to strengthen what the world’s first autonomous digital workforce can be.
“We are an Openai, Meta and Microsoft, with partners, both Microsoft’s Correspondence and Training Systems optimized for a special AI workload,” Mark Papermaster, AMD’s Chief Technology Officer Fast company. “It provides us early information to develop cooperation, models and development for their delay needs.
Invest in supercomputing systems, cooling technologies and are a AI-optimized high-density server racks to manage resources for thousands of suitable AI agents.
“When you ask you to work with you to create a study report using a few dozens of documents or asking you to work with weekly studies on a podcast, it uses AI hypercomputer [Google’s supercomputing system] Bu istəkləri dəstəkləmək üçün, “Mark Lohmeyer, vitse-prezident və Google Cloud-da hesablama və AI / Maşın öyrənmə infrastrukturunun baş meneceri və AI21, SSI, Nuro, Salesforce, Hubx, Essential AI və aktual AI və aktual AI inşaatçıları ilə birlikdə dərin tərəfdaşlıqda hazırlanmışdır.”
Extensive computation shift to the targeted silicone
Agentic systems do not work in isolation. Constantly interact with enterprise database, personal devices and even vehicles. The effect-model is a continuous requirement to apply the knowledge learned to create results.
“Agent AI, Tolga Kurtoglu in Lenovo, requires more apparatus specialty to support CTOs.
The result technological companies to prevent guards, are partners with chipmakers to build silicone designed for low delay. Openai develops special plugs and develops hardware-software engineers, and Meta optimizes memory hierarchies and parliament in MTIA accelerators and Grand Teton infrastructure.
“The last EU’s end or alphabet in our alphabet requires a large number of subsystems to provide 634 trillion trillion tokens in 2025,” says a large number of subsystem for users. “The decade of special AI silicone is a purposeful tensor processing sections (TPU), which is aimedically established for large-scale and agent AI systems.”
TPUS is set to be more efficient and faster than CPU and GPU for specific AI tasks. In April in the next 2025 in April in April, the company presented the seventh generation TPU, which was scale from 9,216 plugs with an interchip connection capacity for Advanced AI business burden. Models like Gemini 2.5 and Alphafold runs in TPU.
“Ironwood TPUS is also significantly efficient, resulting in the value of the placement of complex AI models.
Transformer-optimized GPU, accelerators such as AMD’s instinct Mi series, are an engineer for real-time adaptation, along with sparkling processing sections (NPU) and Chip (SOCS) systems. AMD recently introduced the Instinct Mi350 series of Instinct Mi350 series designed to accelerate workloads along the Agentic AI, generative AI and high-performance calculation.
“Agent AI AI requires more than single accelerators. The concert requires full system solutions with a CPU, GPU and a highly banded network,” said Amd’s Papermaster. “UCP, like Helios, we remove delay points and improve the flow.
According to the AMD, the world’s best 10 AI model builders, including meta, Openai, Microsoft and Xai have a workload of production in instinct accelerators.
“Clients are trying to solve traditional problems using either EU or invent a completely new AI-native application. It is a chiplet integration and memory architecture,” he says Boppana. “META’s 405b-parameter model Llama 3.1 is prepared in our series and Microsoft AZURE in AMD in AMD, as it prepares a large confusion in AMD, and is more on the road.”
The MI350 series, including Instinct Mi350x and Mi355X GPU, provides four-generation growth in a leap in a single leap, AI calculation and eventually.
“We are working on large genes-gen-gen-genes,” Boppana says. “Mi400, started in early 2026, we have seen the goals based on large-scale AI exercises and results, in some applications. Such a rapid progress agent is what the agent requires AI period.”
Electricity efficiency now drives to the edge of the design, data center
Despite their performance promise, the generative and agent AI systems come with high energy costs. A stanford report consumes about 1.287 megawatts of GPT-3 training, the equivalent of a small nuclear power plant that lasts an hour.
AI education and results create important heating and carbon emissions by accounting for up to 40% of the energy consumption of a data center. As a result, power efficiency is now a top design priority.
“We see stronger demand for more modules, decent and energy-efficient placements for agent-based applications. When saving costs and power, they must put AI agents,” says Lohmeyer.
Infrastructure providers such as Lenovo are now presented by AI Edge chips and data center racks that are suitable for distributed cognition. These allow devices to make quick decisions on device agents when synchronized with cloud-based models.
“The heat is the fatal enemy of the sensitive circulation and causes closures, more slowly performance and data loss. Supergents are to implement full potential hinges to develop and support the supercomputer power needed to support many agent environments. “
The future of the enterprise is AI autonomous, but the difficulties remain
Despite the growth rate, the main problems continue. Kurtoglu said that many CIOS and CTOs are still fighting to justify the value of the EU initiatives.
“Lenovo’s AI Headeiness Index 2025 is the most compacting field enterprises, one of the most or less low confidence in the EI’s territorial institutions in this area. Belief, security and control;
To solve this, Lenovo, Private, Enterprise and Public AI systems recommend a hybrid AI to support each other to be trusted and scale.
“Hybrid AI allows you to reliable and complex agenda AI due to your sensitive information, a reliable device or a reliable environment. Every question or decision meets the” round trips “,” Kurtoglu said. “Even if the cloud connection is intermediate, at least one part of the agent’s responsibilities is resistant.”
The Lohmeyer adds that a basic problem for Google Cloud helps customers, especially when the agent systems provide new use of agent systems.
“Agentic systems are difficult to predict the use of autonomous traffic,” Lohmeyer explains. “Therefore, we are working on tools such as a dynamic processing planner to help optimize and manage costs with customers. At the same time, we develop our platforms and tools to make sure agent systems are larger and correct management.”
Boppana notes that although in various stages of organizations, the interests of the AIT’s AIT are growing rapidly. “Some are aggressively leaning, others still think about how to combine the AI to work flow. But on the board, the momentum is real,” he says. “AMD has launched more than 100 built-in AI projects, including a successful placement in search of chip inspection, chief, generation and knowledge.”
Agent AI is expanding to the outside of AI Server farms, such as infrastructure, supported agents should be as smart, distributed and autonomous. In this future, AI will not be written in the code only – this will be glued to Silicone.