Participants:
- Raghu Raghuram, former CEO of VMware, now a16z Partner for Enterprise and Infrastructure
- Ben Horowitz, co-founder of a16z
- Martin Casado, Partner at a16z, former co-founder of Nicira
- David George, Partner at a16z
1. Overview
Raghu Raghuram’s career spans three major eras: the Internet, virtualization, and artificial intelligence. From the browser wars of Netscape to VMware defining virtualization computing, and now to the infrastructure rebirth driven by AI, he has always been at the center of technological paradigm shifts.
In this discussion, Raghu reviewed the evolutionary logic of enterprise computing over the past thirty years and proposed a core judgment:
“AI is triggering the largest infrastructure rebirth since virtualization.”
He pointed out that the development of artificial intelligence is breaking the traditional abstract boundaries of cloud computing, promoting the integration of computing, networking, energy, and physical automation. The future infrastructure will no longer be merely “software-defined”, but rather “agent-defined”.
The main themes of the discussion revolved around the following four topics:
1. From virtualization to AI computing: the cyclical reconstruction of infrastructure;
2. VMware’s growth logic and the strategic insights from acquiring Nicira;
3. The new paradigm of AI-driven infrastructure: energy consumption, robotics, and autonomous systems;
4. Opportunities for startups in the new “AI + physical world” landscape.
2. Thematic Discussion
(1) From Virtualization to AI Computing: The Cyclical Reconstruction of Infrastructure
Raghu believes that the evolution of information technology can be divided into four cycles, each centered around the “elevation of resource abstraction layers”:
|
Technology Cycle |
Core Abstraction Layer |
Representative Companies/Platforms |
Economic Driver |
|
1990s Internet Era |
Network Communication Abstraction (HTTP, Browser) |
Netscape, Microsoft |
Information Dissemination |
|
2000s Virtualization Era |
Computing Abstraction (Virtual Machines, Cloud) |
VMware, AWS |
Cost Efficiency |
|
2010s Cloud-Native Era |
Application Abstraction (Containers, API) |
Google, Kubernetes |
Scalability and Automation |
|
2020s AI and Robotics Era |
Intelligent Abstraction (Agent, Autonomy) |
OpenAI, NVIDIA, Anthropic |
Learning and Reasoning |
He pointed out that each technological reconstruction is accompanied by an industrial reorganization: Netscape challenged Microsoft’s desktop monopoly; VMware redefined the data center; and now, AI companies are redefining energy density, computing architecture, and development models.
Ben Horowitz commented that the immense demand for computing resources driven by AI is pushing data centers to become “new power plants”, as infrastructure shifts from virtualization to “energy quantification”. Raghu emphasized that this is not just a hardware issue, but a resource orchestration problem—how to create a dynamic autonomous system of computing power, storage, networking, and energy.
(2) VMware’s Growth Logic and Strategic Insights from the Acquisition of Nicira
Raghu’s experience at VMware reveals a fundamental rule of enterprise-level infrastructure: “The deeper the abstraction, the stronger the stickiness.”
He recalled three key stages in VMware’s early development:
1. Survival Stage ($0 → $40M): Gained initial market through virtualization desktop applications;
2. Expansion Stage ($40M → $1B): Established a standardized ecosystem through enterprise data center deployments;
3. Explosive Stage ($1B → $13B): Achieved cross-platform advantages with hybrid cloud and software-defined networking.
During this process, VMware’s biggest turning point was the acquisition of Nicira. The deal, completed for $1.3 billion, brought tens of times the long-term return (Nicira’s network virtualization technology became core to vSphere and NSX). Raghu revealed that at the time, Cisco had made an acquisition offer, but VMware chose a strategic path of “cultural fit first”—the founding team of Nicira, Martin Casado and Nick McKeown, represented the future direction of “understanding networking from a software perspective.”
“We are not buying technology; we are buying the future operating system mindset.”
The integration of Nicira propelled the rise of Software-Defined Networking (SDN), transforming VMware from a “virtual machine company” into a “virtual data center company.” This experience offers direct insights for current AI infrastructure entrepreneurs: in a rapidly evolving paradigm, defining interfaces is more important than owning resources.
(3) AI-Triggered Infrastructure Rebirth: The Fusion of Computing, Energy, and Physicality
Raghu refers to the current AI era as the “second great infrastructure reshuffle (Infrastructure Rebirth).” He pointed out that AI computing is shifting from “cloud abstraction” to “physical constraints”:
- Increased coupling of computing power and energy consumption, making energy density a key variable;
- Network bottlenecks of GPU clusters prompting redesign of communication layers;
- AI training pipelines are driving systemic innovations in cooling, wiring, and storage structures.
Raghu summarized this trend in one sentence:
“We are entering an era of overlapping energy, algorithms, and automation.”
He categorized AI infrastructure into three levels:
|
Level |
Core Objective |
Representative Trends |
Investment Logic |
|
AI Fabric |
Dynamic scheduling of computing and networking |
High-density GPU clusters, Infiniband, optical interconnects |
Cloud computing replacement and computing power integration |
|
AI-Aware Systems |
Self-tuning energy consumption and task allocation |
Energy-Aware Compute, intelligent cooling |
Hardware + software joint optimization |
|
Physical AI |
Data center robotics and autonomous operations |
Automated wiring, robotic arm management |
Robotics + infrastructure integration |
Especially at the third level, Raghu emphasized that “Physical AI” will become a strategic focus for the next decade. This concept refers to a new type of system: robots performing maintenance, wiring, and inspection in data centers, improving energy efficiency and safety through a closed loop of perception-decision-action.
Martin Casado further pointed out that this means AI is no longer just “virtual intelligence” but has entered the stage of autonomous execution, extending the boundaries of enterprises into the physical world.
(4) AI and Energy: The Evolution from Virtual Machines to Energy Machines
Raghu noted that the expansion of AI has transformed “energy consumption” from a cost item into a strategic variable. The goal of the virtualization era was to “do more computing with fewer servers”; now, the goal of the AI era is to “do smarter computing with more energy.”
He proposed a key trend: “Energy-Aware Compute.” This system dynamically allocates computing tasks by real-time monitoring of power and workload, thereby reducing waste and extending hardware lifespan.
|
Comparison Dimension |
Virtualization Era |
AI Era |
|
Computing Objective |
Maximizing Efficiency |
Optimizing Intelligence |
|
Resource Bottleneck |
CPU and Memory |
GPU and Energy |
|
Optimization Strategy |
Load Balancing |
Dynamic Energy Allocation |
|
Operation and Maintenance Method |
Manual and Scripted |
Autonomous Control by Agents |
He believes that this new concept of “energy as logic” will reshape the economic structure of data centers: data centers will no longer be just computing factories, but energy autonomous networks.
Ben Horowitz summarized: “Virtualization liberated hardware; AI is liberating energy.”
(5) Robotics and Autonomous Infrastructure: The “Re-Materialization” of Infrastructure
Raghu pointed out that under the wave of AI, the tech industry is experiencing a “re-materialization.” For the past decade, Silicon Valley has adhered to the belief that “software is eating the world,” but now AI training and reasoning need to return to physical boundaries:
- Data centers need to redesign spatial layouts and energy flows;
- Automated robots are beginning to take over physical maintenance;
- Agents are forming collaborative networks at the infrastructure level.
He proposed the future trend of “vertical robots”: AI-driven robots will focus on specific scenarios, such as server wiring, cooling control, and wind turbine inspection. These areas combine AI vision, force control, and autonomous algorithms, with significant potential for efficiency improvement.
“Every automated screw tightening action is backed by a stream of energy-saving bits.”
Raghu emphasized that the combination of robots and AI is not about replacing human labor but making operational systems safer and more sustainable. The future “AI factory” will consist of physical robots, software agents, and energy optimization systems working together.
(6) Opportunities for Startups: Solving Enterprise-Level Problems from Day 1
Raghu’s core advice is:
“New generation infrastructure startups must think about enterprise-level problems from day one.”
This means AI entrepreneurs should not only pursue model innovation but also focus on the scalability, security, and energy resilience of infrastructure.
He identified three categories of entrepreneurial directions with long-term value:
1. AI-native Infrastructure: Infrastructure built around the model lifecycle (e.g., training scheduling, model deployment, load adaptation);
2. Autonomous Operations: Autonomous agents for data centers and industrial systems;
3. Energy-Tech Convergence: Intelligent scheduling layer that integrates energy management with computing management.
Martin Casado concluded: “Today’s infrastructure startups find opportunities not above the abstraction layer, but below the physical boundaries.” This means the AI revolution is not just an algorithm revolution but a paradigm shift in infrastructure.
3. Conclusion and Outlook
Raghu Raghuram’s career trajectory represents the evolution path of technological infrastructure from software abstraction to intelligent autonomy. From the Netscape browser opening the internet economy to VMware virtualization leading enterprise computing, and now to AI and robotics driving physical intelligence, the definition of infrastructure is being completely rewritten.
First, AI has become the “demand engine” for infrastructure innovation. Its computing density and energy consumption scale force the entire ecosystem to restructure, from chips to power networks, from cooling equipment to scheduling algorithms.
Second, the fusion of energy and computing is becoming a core trend. Future data centers will replace traditional “task flow management” with “energy flow management,” and energy awareness will become the underlying logic of infrastructure design.
Third, the combination of robotics and AI opens the era of physical intelligence. Infrastructure is no longer a static asset but a “dynamic organism” with perception, decision-making, and execution capabilities.
Fourth, the shift in investment logic. Capital will shift from “virtual expansion” to “physical intelligence,” returning from the boundaries of cloud computing to the complexities of the real world.
Raghu concluded:
“Virtualization allowed us to rebuild reality in the digital world; AI and robotics will allow us to rebuild digital logic in the real world.”