Written by: Techub News Compilation

Introduction

In March 2026 Beijing time, NVIDIA's annual technology event GTC kicked off at the SAP Center in San Jose. Before the keynote speech by founder and CEO Jensen Huang, a three-hour "opening show" gathered dozens of top developers, entrepreneurs, scientists, and investors from around the world. This pre-event discussion was hosted by Conviction founding partner Sarah Guo and Atreides Management founding partner Gavin Baker, interspersed with live interviews by reporter Tiffany Janzen with several guests. The conversation focused on how accelerated computing is transitioning from training to inference and delving into the physical world, becoming the infrastructure reshaping various industries. This article compiles the core viewpoints from this high-density collision of ideas, presenting readers with a panoramic view of the frontlines of the AI revolution.

Summary

• Accelerated computing has transcended AI, becoming a universal computing paradigm from graphic rendering and data processing to scientific simulation, with its market expanding for the first time into design and engineering.
• The core challenges of enterprise AI applications are shifting from “computational constraints” to “capability constraints” and “data value unlocking,” with reliable and interpretable inference results becoming crucial.
• The “deconstruction” of inference workloads (such as separating pre-fill and decoding) will greatly improve the utilization and residual value of existing GPUs, extend infrastructure lifecycles, and lower barriers to innovation.
• Physical AI (robotics, autonomous driving) is the next breakout point, with its development path gradually moving from structured enterprise scenarios to unstructured consumer scenarios, where safety and simulation validation are critical.
• Specialized, multi-agent collaborative AI models (such as digital twin experts in healthcare) will become mainstream in complex fields, driving a transformation in knowledge work paradigms.

Accelerated Computing: From Dedicated Engines to Universal Infrastructure

Mark Edelstone, Managing Director at Morgan Stanley and an analyst during NVIDIA's IPO, reviewed the company's path to success. He pointed out that NVIDIA established its vision of building a “platform” early on, integrating architecture, algorithms, and software compatibility, and established a sustained leadership capability in the fierce competition of graphic chips through the “three teams addressing two product cycles” R&D rhythm. This platform mindset smoothly transitioned to the field of accelerated computing.

Desh Nirmal, Senior Vice President of IBM Software, interpreted AI platforming from the perspective of data evolution. He believes that data is evolving from the “end point” of the past (stored in databases waiting for queries) to today’s “tools” (accessible through various interfaces) and is about to become tomorrow’s “code” or “skills.” In the future, enterprises will be able to dynamically generate programs to process specific data tasks through natural language instructions, achieving unprecedented flexibility and scale. Ninety percent of enterprise data is unstructured; how to relate it to transactional data and extract value is the core of generative AI's implementation.

Akash “Aki” Jain, President and CTO of Palantir’s U.S. Government Business, shared the company's experience in transitioning from data governance to AI applications. Palantir's starting point is to help human teams collaborate in processing data like software engineering teams. As AI technologies such as computer vision became practical due to NVIDIA GPUs and CUDA, Palantir began using these capabilities for mission-critical tasks, such as ensuring the safety of military personnel using unstructured data like video and audio. Today, as generative AI matures, the focus is shifting to how to safely and controllably integrate "AI teammates" into human workflows to achieve scalable outcomes. Aki made an interesting point: the limiting factors for organizations are shifting from “neurons” (human labor) to “electrons” (computing power).

Anirudh Devgan, President and CEO of Cadence, elaborated on the impact of AI in the foundational field of chip design. He likened it to a three-layer cake: the top layer is agent AI, the middle layer is the “ground truth” of physics and chemistry, and the bottom layer is computation and data. Successful application requires vertically slicing the entire cake. In the field of chip design, agent AI can already automatically generate register transfer language code and testing platforms, while the powerful underlying computing power (like GPUs) allows simulation speeds and scales to reach levels previously unimaginable. Although the efficiency of design tools has improved a hundredfold, the exponential growth of chip complexity demands that tool performance needs to be improved tenfold again, and AI is key to achieving that goal.

AI Infrastructure: Building a Global Intelligent "Five-Layer Cake"

Michael Dell, founder, chairman, and CEO of Dell Technologies, pointed out that computing will be ubiquitous, from edge devices to hyperscale data centers. The AI factories and data platforms launched by Dell focus on solving the data orchestration problem, delivering data to GPUs at extremely high speeds. He emphasized that with the emergence of multi-agent systems and complex queries, tiered memory systems like KV caching have become essential, and the growth rate of memory and stored data may even exceed that of computing power itself. Unlocking the “cold data” within enterprises that has never been used is the direction currently explored by thousands of customers.

Michael Intrator, co-founder, chairman, and CEO of cloud computing provider CoreWeave, expressed excitement about the “deconstruction” of inference workloads. He believes that breaking down inference into different stages (such as pre-fill and decoding) allows companies like CoreWeave to optimize deeply for each stage. This specialization was not economically feasible in previous waves of computing but has become viable now due to scale. More importantly, deconstruction allows parts of the workload to be routed to older generation GPUs for execution, thereby extracting more value from existing infrastructure and extending asset lifetimes, which will have positive impacts on financing and credit models.

Lynn, co-founder and CEO of Fireworks AI, envisions 2026 as the “Year of Efficiency” and “Year of AI Innovation Explosion.” She pointed out that currently, having product-market fit does not equate to having a sustainable business, and many companies may “expand to bankruptcy” due to high inference costs. Fireworks helps clients utilize their private data to enhance model intelligence, speed, and cost-effectiveness through a “three-dimensional optimization” approach, aiming to reduce total ownership costs by 5 to 10 times. She envisions a future where training and inference are integrated, enabling models to continuously learn during service processes and achieve greater efficiency.

Joe Creed, CEO of Caterpillar, represents the base of the AI infrastructure pyramid——energy and physical construction. He candidly stated that he is witnessing the largest wave of infrastructure development in his career. In the “five-layer cake” proposed by Jensen Huang (chips, systems, software, AI factories, physical world), the first three layers (energy, chips, infrastructure) are exactly the areas empowered by Caterpillar. The explosion of AI is expected to consume far more electricity than anticipated, and Caterpillar is increasing the capacity of power generation and engineering machinery globally to support this intelligent revolution.

Physical AI and Agents: When Intelligence Merges with the Real World

Daniel Nadler, co-founder and CEO of Open Evidence, discussed the necessity of specialized agents in the medical field. The human body itself is highly differentiated, and medical knowledge has become highly specialized. Therefore, future medical AI will not be a single general intelligence, but rather a collaborative network composed of numerous digital twin experts (such as cardiologists, neurologists). Doctors will be able to dynamically call on these “digital experts” to assist in diagnoses. He pointed out that the explosive growth of biomedical knowledge (doubling every five years) has already far exceeded the learning ability of human doctors, and AI digital twins will become essential tools for doctors to navigate the ocean of knowledge. He revealed that Open Evidence, a company with fewer than a hundred employees, will indirectly support the diagnosis and treatment of 300 million patients in the U.S. this year, showcasing a new paradigm of “small teams, big impact” in the AI era.

Danping, Vice President of R&D at Siemens, shared the application of AI in the industrial physical world. Through agent AI and model context protocols, industry experts’ domain knowledge can be embedded within workflows, empowering employees without technical backgrounds. Siemens' collaboration with NVIDIA covers four major areas: simulation, supply chain, manufacturing, and electronic design automation. For example, using GPUs to accelerate physical simulations, building high-precision process models with AI, and even using generative AI to control workflows, enabling software to possess the ability to “think and act.” Digital twin tools allow global teams to conduct photorealistic collaborative design and decision-making in virtual environments.

Raquel Urtasun, founder and CEO of Waabi, provided lessons learned from autonomous driving for physical AI. She noted that early autonomous driving used a “robotic expert” approach, attempting to manually code everything, which was not scalable. Waabi adheres to the principles of “AI-first” and “simulation-first,” able to exhaustively test corner cases in virtual environments by constructing world models and mixed-reality testing, which is the only way to ensure the safety of critical systems. She believes this new simulation-based approach, which allows for end-to-end verification and inference, will be widely adopted in other physical AI fields.

Deepak, founder of a robotics company, discussed the path to the implementation of general robots. He believes robots will follow a gradual development path from enterprise to consumer scenarios. Initially deployed on a large scale in structured, high willingness-to-pay, and high fault tolerance enterprise environments (like factories); then gradually expanding to semi-structured scenarios such as hospitals and grocery stores using accumulated data; and finally entering completely unstructured environments like homes. This process relies on a “data flywheel”—using data from the previous stage to drive the enhancement of general capabilities in the next stage. He previewed that there would be important corporate collaboration announcements to achieve a “one-time large-scale deployment” of robots in enterprise scenarios.

Alex Kendall, co-founder and CEO of autonomous driving company Wayve, showcased the progress of “end-to-end AI” driving autonomous driving. Wayve built world models to conduct “zero-shot” driving tests (driving in cities where training data has never been collected) in over 500 cities, proving that its system possesses genuine generalization capabilities. He emphasized that every future vehicle will be autonomous, and the automotive industry will be the first stage for embodied AI to be showcased on a large scale. Wayve has partnered with Uber, Stellantis, and others, planning to launch robotaxi pilot operations in over 10 cities globally.

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。