ByteDance and Oracle Adopt Arm AGI CPU; Arm Forecasts $20 Billion in Revenue by 2028

On June 3, Xinxidongxi reported that Arm CEO Rene Haas delivered a keynote speech at Computex 2026 yesterday, announcing that ByteDance and Oracle have adopted Arm’s in-house data center CPU chip, Arm AGI.

Last month, Arm doubled its demand forecast for Arm AGI CPUs, expecting $2 billion (approximately RMB 13.5 billion) in revenue during the fiscal years 2027 and 2028, and anticipates the product will generate annual revenue of $15 billion (approximately RMB 101.6 billion) within about five years.

Rene Haas said in an interview with foreign media yesterday that it is “nearly impossible” for the U.S. to block AI CPU exports to China, because AI CPUs have broad applications, making it difficult to identify which CPUs are specifically designed for AI, and hard to set precise performance thresholds and memory bandwidth limits as with AI chips.

On Monday, NVIDIA unveiled the RTX Spark superchip and Vera data center CPU based on Arm architecture; that evening, Arm's stock price rose steadily, closing up 16% on Tuesday. Year to date, Arm's stock has gained 263%.

NVIDIA’s founder and CEO Jensen Huang dropped by Rene Haas’s Tuesday talk and joked upon taking the stage: “Look at his stock price—every time I release a product, his stock goes up, while mine doesn’t move at all.”

Rene Haas responded shrewdly: "You were a shareholder, and then you sold your shares."

Jensen Huang immediately jumped in: "Yes, yes, oh, I need cash."

The two seemed to be old friends, chatting enthusiastically for 15 minutes, frequently breaking into comedic bits, trading punchlines that had the entire audience roaring with laughter—they themselves often laughed so hard they showed all their teeth.

This has been the most lively tech industry conversation I've seen recently.

For example, after gushing about Arm CPUs for quite some time, Huang Renxun concluded: “The keyword is 'Arm is perfect.'”

Rene Haas chimed in: “Another keyword is 'thank you.'”

Jensen Huang immediately spoke Chinese: "Oh, nowhere, nowhere, don't be so polite."

Then Rene Haas complained, "This contest isn't fair anymore." (implying that Huang's use of Chinese was unfair)

Then Huang Renxun kindly added, "You're welcome."

Huang also joked that “one of Arm’s greatest advantages is not having to worry about supply chain issues”—the supply chain for IP is electrons, and you can have as many electrons as you want.

“So I love its business model,” Huang began, reflecting on the past. “You know, I tried—I once attempted to become Arm. I worked with Rene before, and we tried to collaborate again, but it wasn’t meant to be. I still feel sad about it.”

Rene Haas said, "If the two companies merge, we will become the largest company in the world."

“I like this,” Huang said with a laugh. “That’s a great idea.”

It seems both parties regret that NVIDIA failed to acquire Arm.

Finally, during the gift-giving segment, Rene Haas triggered a wave of nostalgia by presenting Jensen Huang with a Microsoft Surface RT laptop powered by the NVIDIA Tegra 3 chip, even signing it with Huang’s signature.

The NVIDIA Tegra 3 is NVIDIA's globally首款 Arm-based mobile quad-core processor, introduced several years ago.

Jensen Huang pointed to a photo on the large screen and boasted, “What happened when we were young? I have to say, I think I look younger. Do you agree? I feel like I’ve taken very good care of myself.”

Rene Haas laughed until it was blurry.

Then Huang Renxun snatched the gift and raised his voice: “This is for me? If I sign it and give it back to you, it’ll become a treasure.”

Rene Haas said: "No, you sign it and give it back to me—there’s a contract and an invoice here; we can’t do that. We know that game."

Returning to serious industry topics, during this talk, Rene Haas asked Huang Renxun several key questions:

1. Why develop RTX Spark?

2. How to balance and trade off between local agents and cloud-based agents?

3. Can agents truly operate independently, without relying on the underlying operating system?

4. How does Jensen Huang view the constraints on growth over the next few years?

Huang also painted a broad vision for market growth: Currently, the computer industry is limited by the number of people using computers; with intelligent agents capable of autonomously using computers, we will no longer have just one billion people using computers, but tens of billions—and possibly even more intelligent agents, robots, and autonomous vehicles than people.

So the question is, just how large can the scale of computer products be?

“I feel that by now, the outcome is already sealed—this industry, currently worth trillions of dollars, could grow tenfold, and we’re on the way there,” said Jensen Huang.

Rene Haas also shared the latest advancements and future plans of Arm in the fields of agent PCs and data center CPUs.

He also mentioned chatting this week with C.C. Wei, Chairman and CEO of TSMC, and Victor Hsu, Senior Vice President and COO, who said they had never seen the semiconductor industry cycle remain so prosperous for four consecutive years.

Huang Renxun addressed several key questions raised by Rene Haas, and these insights are highly relevant to the future development of AI PCs and chip design strategies.

1. Why was the RTX Spark product developed?

PCs and operating systems have existed for 40 years; manual programming will be replaced by agent applications that use tools within PCs. How then should we redesign architectures, transform operating systems, and reinvent computers?

NVIDIA recognizes that agent systems require high-performance CPUs, which is why it has adopted Arm.

The RTX Spark superchip features a 20-core CPU with excellent single-thread performance, requiring substantial memory to store numerous parameters.

NVIDIA has created a new data format called NVFP4 to compress large language models, maximize model size, and integrate highly intelligent AI directly into system memory.

NVIDIA also aims to integrate CUDA and CUDA Tile for accelerated computing, combining tensor core processing into a single processor.

2. How to balance and decide between locally deployed agents and cloud-based agents?

These Arm PCs will become autonomous agents that operate continuously.

Today, if you leave your laptop at home or in a hotel, you won’t be able to use it.

In the future, you’ll simply need to pick up your phone to remotely communicate with your PC and instruct agents to get things done.

Jensen Huang said: "The essence of a personal computing device is that you can do anything with it without spending time."

Use cloud APIs when necessary; otherwise, perform tasks locally on your computer.

3. When running an agent, is the operating system important? If we consider the agent itself as an operating system, can it truly operate independently with minimal reliance on the underlying operating system?

The importance of the operating system remains unchanged, and may even be more critical than before.

This is also the point of contention often raised when people talk about the emergence of AI—“software is dead”—but Huang believes nothing could be more absurd.

People may only understand about 10 to 20 percent of the features of many tools.

But now, you can tell the agent what you want.

The agent clearly knows how to use these tools because it has read the Skills files, which essentially serve as the user manuals for each tool. Now, it can leverage the associated MCP or CLI to unlock all these tools and fulfill your needs.

These tools will be more valuable than ever; they run on operating systems, so we need Windows, and we will need these APIs and tools for a long time to come.

4. What are the constraints on growth in the coming years?

“We’ve seen constraints in almost every area,” said Huang. NVIDIA planned ahead and managed its supply chain effectively, achieving nearly 100% year-over-year growth this year, with very rapid growth expected next year—our supply chain is capable of supporting NVIDIA’s expansion.

But demand is actually higher.

Huang Renxun said that new computing application models indeed require a new architecture, and a major breakthrough now is that agents can produce practical AI—this is why everyone’s growth has been so astonishing.

When AI becomes practical, the tokens it generates can produce profits. When token generation becomes profitable, everyone wants to create tokens worth trillions of times more.

AI is no longer just a chatbot that answers questions—it can think, use tools, read, continue thinking, plan, and experiment, leading to a significant increase in the number of tokens required. The profitability of tokens is driving demand for computing power, creating a compounding effect.

In the PC sector, companies such as Google, Apple, NVIDIA, and Qualcomm have developed PC chips based on the Arm architecture, and Arm has been collaborating with Apple, Google, Microsoft, and others for decades.

Rene Haas said Arm is honored to collaborate with NVIDIA on the development of the RTX Spark superchip based on the Arm architecture. This custom Grace CPU features 20 cores, each built on the Arm architecture.

“I believe this has the highest number of CPU cores available in any laptop on the market today,” said Rene Haas. When paired with the Blackwell GPU, this chip delivers 1 PFLOPS of FP4 AI performance, 128 GB of unified memory, and full native support for Windows on Arm.

Arm’s role is to work closely with NVIDIA and MediaTek using Arm’s compute subsystem strategy.

The compute subsystem integrates all the necessary components—CPU, GPU, system IP, and memory controller—to build a complete end-to-end solution.

Arm collaborated with MediaTek to complete this work, enabling MediaTek to provide a complete solution.

Rene Haas also presented Arm’s CSS roadmap for agent PCs, with the next generation optimized for custom CPU cores designed specifically for PCs.

Rene Haas said that over 25 billion Arm chips were manufactured in Taiwan, and Arm’s first self-designed CPU, the Arm AGI, released in March this year, was produced by TSMC in Taiwan.

The Arm AGI CPU is designed for AI agent infrastructure, built on TSMC’s 3nm process technology with a dual-chiplet architecture. It integrates 136 Arm Neoverse V3 high-performance cores, each with 2MB of L2 cache, supporting a clock speed of up to 3.7GHz and delivering 6GB/s memory bandwidth per core with memory latency under 100ns. It features a 96-lane PCIe Gen 6 interface and supports CXL 3.0 protocol, with a TDP of up to 300W.

Arm AGI CPU partners include OpenAI, Meta, Cerebras, SAP, SK Telecom, Rebellions, and others. Rene Haas announced that market demand for this chip has grown even stronger since its launch, with Oracle and ByteDance also joining the ecosystem, validating that the Arm AGI CPU can solve real-world problems.

Of course, not all companies want to buy an Arm AGI CPU. For companies interested in developing their own chips, Arm offers a variety of IP and compute subsystems (CSS) dedicated to providing any solution our customers desire.

In the data center, the Axion CPU, connected to Google’s latest AI chips, the TPU 8T and TPU 8I, is a chip based on Arm Neoverse technology that reduces power consumption by up to 60% compared to x86 CPUs, without compromising performance.

Amazon’s in-house CPU, Graviton, also uses the Arm architecture. Amazon CEO Andy Jassy revealed: “Two major customers asked if they could purchase all of our Graviton instances for 2026.”

NVIDIA also recently released its new Vera CPU based on Arm this week.

Arm plans to make its custom CPU development a long-term initiative and has unveiled a three-year roadmap.

The second-generation Arm AGI CPU is currently in development, featuring more cores, higher energy efficiency, and improved performance compared to the previous generation.

The third-generation Arm AGI CPU is also coming soon.

These are all based on the compute subsystems that Arm plans to deliver alongside the chips.

This week, speeches by chip industry leaders such as Jensen Huang, Chen Liwu, and Rene Haas reflected common trends in the CPU industry—agents are transforming computational logic, opening a brand-new window of market opportunity for CPUs.

In recent years, the focus of computing power competition has primarily been on GPUs, which are essential for AI training. However, since the surge in agent applications this year, demand for agent inference has grown significantly, requiring substantial state management, tool invocation, and workflow orchestration—tasks that are CPU-intensive.

Intel and AMD continue to solidify the dominance of x86 processors in the data center market. Meanwhile, emerging players such as Amazon, Google, and NVIDIA are largely betting on the Arm CPU architecture. Even Arm has made a "departure from tradition," officially entering the data center CPU market this year.

An interesting phenomenon is that the semiconductor industry is forming a new trend toward vertical integration.

Chip giants with diverse product lines, such as NVIDIA, Intel, and AMD, are increasingly emphasizing their full-stack capabilities, and their respective advantages ultimately converge on higher energy efficiency, comprehensive solutions, and greater cost savings.

Leading companies across various industries are also "crossing boundaries": cloud giants are developing their own chips, chip companies are moving upstream to offer complete system solutions, and semiconductor IP companies are advancing upstream to design chips.

As tokens become new competing currencies and demand for computational power explodes, maximizing effective computation per watt will be the central focus of the next generation of chip competition.

This article is from the WeChat public account "Xin Dongxi," authored by ZeR0 and edited by Mo Ying.