It has been over three years since ChatGPT emerged.

If AI is destined to be a feast, then which dish has arrived at the feast now?

The answer may disappoint many impatient diners: just served cold dishes.

In the past three years, the entire industry has been busy laying tablecloths, setting tableware, and determining seating arrangements. Computing power manufacturers are building kitchens, chip companies are forging cookware, large model manufacturers are preparing ingredients, and entrepreneurs from various industries are looking at the door: which table should I sit at, what dishes should I eat, and how should I use chopsticks?

Everyone could smell the aroma wafting from the kitchen, but the truly steaming "main course" had not yet been served.

A true feast never ends hastily in the cold dish stage. The more delicate and well prepared the appetizers are, the more it indicates that the later hard dishes are worth waiting for.

So, the question is not 'Can AI really work?' but 'How should I use AI?'?

Which layer of the five layered cake did you eat?

In March 2026, Huang Renxun did a rare thing.

Five days before the opening of GTC 2026, he published a long article titled 'AI Is a 5-Layer Cake'. The article proposes that "AI is a five layer cake": from the bottom layer of energy, to chips and hardware, to infrastructure, then to models, and the top layer is applications.

This metaphor reduces AI to an infrastructure project that follows strict physical laws and requires heavy investment in assets.

Intelligence does not happen out of thin air. Behind every token, there is current, heat, cooling, and conversion efficiency.

But if we look at it from a broader perspective, AI is far more than just a "five layer cake", but a "Manchu Han feast"——

Besides energy, there are chips.

Besides chips, there is a network.

Outside of the network, there is storage.

Besides storage, there is security.

In addition to security, there is also operations and maintenance.

In addition to operation and maintenance, there are specific needs for algorithms, data, scenarios, talents, policies, ecology, and various industries.

AI is the new foundation for all industries. Each link is enough to become a separate "dish", and each "dish" may give birth to a company with a market value of billions or even trillions.

More importantly, this feast has just begun.

Since ChatGPT ignited global enthusiasm, China's daily token usage has increased by 1400 times, and global AI infrastructure investment has exceeded trillions of dollars.

As Yu Yingtao, Chairman of Tsinghua Unigroup and President and CEO of New H3C Group, said at the NAVIGATE 2026 Leaders Summit, "Currently, the entire AI industry is permeated with a peculiar collective emotion - excitement and anxiety coexist, equally strong. ”

On the one hand, big models are moving from the laboratory to various industries, and a large number of enterprises are fully committed to AI; On the other hand, the investment in computing power is huge but the ROI is vague, and the utilization rate of GPU in some data centers is less than 60%.

Yu Yingtao used a metaphor to describe the current industry mentality: "It's like standing on a track that has never been seen before, with the starting gun firing and the engine roaring, but the track ahead is shrouded in thick fog, and no one can see where the finish line is

Yu Yingtao delivers a speech at the NAVIGATE 2026 Leaders Summit

Under this collective sentiment, every company and individual is asking the same question: How should I move forward in this AI feast?

New H3C's Seven Hard Dishes

When AI changes from a "model story" to a "construction story", and from "algorithm competition" to "engineering competition", the value of infrastructure is rediscovered.

What we need is not only smarter models, but also high-speed networks that can connect thousands of GPUs, data centers that can support training with trillions of parameters, and underlying architectures that can stably deliver computing power to every terminal.

This is precisely the track that New H3C has been deeply cultivating for 23 years. At the NAVIGATE 2026 Leaders Summit, New H3C released seven core products at once, covering the entire chain of "computing network storage cloud security maintenance". These seven 'hard dishes' are not isolated single product releases, but system level solutions for the AI era.

First step: UniPoD S80000 full series supernodes

This is the flagship product released by New H3C this time, a high-performance computing platform for trillion level models. Covering a full range of products from 32 cards to 1024 cards, with the highest flexibility to expand up to 16384 cards, fully meeting the needs of various business scenarios from small and medium-sized training to large-scale clusters.

New H3C defines it as a system level answer that integrates computing, networking, storage, cloud computing, security, and maintenance. It is not simply a hardware stack, but an ultimate engineering presentation that deeply integrates the six major capabilities. Through deep optimization of software and hardware collaboration, training performance has been improved by 70%, and inference performance has been increased by three times.

Second step: High density fully liquid cooled complete machine S90000

1U 6CPU， The highest density in the industry. A single cabinet can provide over 30000 cores of CPU computing power output. The whole machine is designed with full liquid cooling and no fan, and the PUE can be as low as 1.04.

In today's world where 'the end of computing power is electricity', energy efficiency ratio is competitiveness. As global data center power demand continues to rise and energy consumption constraint policies are introduced in various regions, a device that can push PUE close to the theoretical limit means that customers can obtain more computing power output with less electricity cost.

Third track: Full range of single-chip 102.4T intelligent computing switches

The network is the most vulnerable bottleneck in AI infrastructure. The intelligent computing switch released by New H3C achieves the industry's highest single-chip bandwidth, with an overall bandwidth rate increase of 100%, bandwidth utilization rate increase of 68%, and latency reduction of 15%.

You should know that in a 10000 card cluster, the efficiency of inter card communication directly determines the overall speed of training. A congested network can turn the most expensive GPU into a 'high-end decoration'. New H3C has simultaneously released intelligent DCI switches and deterministic computing routers, forming a complete network solution from within the cluster to across data centers.

Fourth track: New generation AI native storage X20000 series

In AI training scenarios, data IO is often an invisible efficiency killer. 40% of GPU idle data time is a staggering waste. New H3C's new generation AI native storage is based on a distributed storage engine, with a single node bandwidth of 200GB/s and IOPS of 3 million.

Of particular note is the KV Cache (Key Value Cache) inference acceleration technology. Although KV Cache can cache large model historical context into GPU memory, avoiding duplicate calculations, long texts and high concurrency can lead to a significant increase in data volume, far exceeding the GPU's memory limit.

New H3C has launched XCache acceleration software based on the concept of "converting to storage", which has built a memory pool from shared memory to CXL (computing fast interconnect), reducing the first token latency (TTFT) by 90% and increasing user concurrency by 10 times. This significantly reduces the data waiting time for GPU computing.

Fifth track: New generation AI intelligent cloud

If the first four dishes are compared to "hardware ingredients" such as CPU, GPU, network, and storage, then AI intelligent cloud is the "pot" that cooks them together. New H3C's new generation AI intelligent cloud achieves integrated scheduling of general and intelligent computing, supports unified management of over 100 GPUs/NPUs, and has a floating-point operation utilization rate of up to 82% in terms of performance. It also has management capabilities for over 100000 nodes.

This means that customers do not have to be bound to a single chip or hardware, and can flexibly choose the optimal solution in a diverse and heterogeneous environment. For government and enterprise customers facing uncertainty in chip supply, the value of this "not being kidnapped" ability is no less than a single point of performance improvement.

Sixth track: Industry's highest performance AI firewall M9000-X

As AI gradually participates in business decision-making and automated execution, the demand for security protection has been raised to unprecedented heights. The AI firewall M9000-X launched by New H3C has a throughput of 16T and introduces AI threat detection engine and semantic intent recognition engine.

Simultaneously released are the intelligent agent security protection system and the Agentic SOC autonomous security operation platform. New H3C pointed out that AI empowering security has become the main theme of technological development, and the security system is evolving from traditional boundary protection to application protection represented by intelligent agents.

Seventh track: Lingxi operation and maintenance intelligent agent

Although this dish ranks last, it may be the one with the most long-term value. Lingxi operation and maintenance intelligent agent embodies New H3C's 23 years of ICT operation and maintenance experience, 120 million network devices' operation and maintenance accumulation, as well as a million level case knowledge base and over 50 million real operation and maintenance events. It truly achieved the transition from a "business tool" to a "smart partner".

New H3C will transform decades of operational experience and sufficient high-quality knowledge data into the core capabilities of intelligent agents. The efficiency of operation and maintenance troubleshooting has been improved by 90%, the deployment time at the beginning has been saved by 60%, and repeated alarms have been compressed by 90%.

This is a typical case of accumulating wealth over time.

With 23 years of experience in safeguarding critical infrastructure, AI technology is now encoded into intelligent agents, allowing every customer to enjoy expert level operation and maintenance services. This is also the ultimate embodiment of New H3C's "AI in ALL" technology strategy: not using AI to replace people, but using AI to empower the entire lifecycle of infrastructure.

Let AI truly run

When the seven dishes are served, many people may ask: Why do we have to make so many? Isn't it good to specialize in one dish?

According to Xinhua News Agency, the real battlefield of AI is not in the laboratory, not in the press conference, and not in the ranking of computing power. The real battlefield of AI is in the industry. Behind every industry empowered by AI, three things are needed: real scenarios, massive amounts of data, and engineering capabilities.

China is best at these three things.

In China, there are over 6000 AI companies with the world's largest industrial categories and richest application scenarios. But there is a huge gap between transforming AI from a vision in PPT into a production line in a factory, a medical assistant in a hospital, and a digital civil servant in the government. The name of this gap is' Engineering Landing '. This requires not only the most advanced technology in a certain aspect, but also the ability to integrate the six major capabilities of computing power, network, storage, cloud, security, and operations into a system.

New H3C NAVIGATE 2026 Navigator Peak Exhibition Area Site

This is also the key to why New H3C can firmly establish its leading position in digitalization and AI solutions. Because no matter how AI algorithms change or models iterate, the demand for infrastructure is always there. Moreover, with the popularization of AI, this demand will only become larger and more rigid.

The AI industry is a complete chain from energy to applications, and it cannot do without a single layer. And when we open this framework, we will find that there are countless subdivisions within each layer, each of which requires professional players. This AI feast is not a solo show by a few giants, but a major reshuffle of industries with a new division of labor and repositioning.

New H3C's position in this feast is extremely clear: it is the builder and enabler who ensures that computing power can stably produce tokens and that AI can truly land in every industry scenario.

As Xinhua News Agency said, "Cultivate pragmatically and infuse wisdom into the times." This is not a slogan, but a definite coordinate that an ICT leading enterprise has found for itself in the AI era.

The AI feast has just begun, and the highlights are still ahead.

One day, people will forget the hustle and bustle of competing for seats, forget the dazzling menus, and even forget which dish was served in which year. But people will remember one thing: what did they eat?

The only commitment that cannot be absent from the entire feast is to keep the computing power from cooling down, keep the tokens from being congested, and ensure that every piece of data can reach its destination.

What New H3C needs to do is not to grab the most prominent chair, but to ensure that every dish can be served steaming hot.

Let every attendee have their own feast.