AMD Announces the Helios AI Platform and Forecasts 60% of AI Compute for Inference by 2026

At the Advancing AI 2026 conference, AMD unveiled the Helios rack-scale AI platform, the sixth-generation EPYC Venice server CPU, the Instinct MI430X and MI350P, marking AMD’s entry into the AI factory competition with a complete system comprising GPUs, CPUs, networking, and software. Lisa Su anticipates that approximately 60% of global AI computing power in 2026 will be used for inference, requiring agents to leverage vast amounts of GPUs for inference and substantial CPUs to orchestrate each step. She emphasized AMD’s differentiated approach through Chiplet technology and product portfolio architecture, collaborating with leading companies such as OpenAI, Meta, and Anthropic to develop Helios, with the first products scheduled for shipment by the end of the third quarter.

Author and source: Tencent Technology

Lisa Su summarized AMD’s strategy in her keynote into three points: computing leadership, an open platform, and “making AI ubiquitous.”

Advancing AI is AMD’s annual event unveiling its AI computing ecosystem. On July 24, 2026, Beijing time, AMD unveiled a comprehensive lineup including the Helios rack-scale AI platform, the sixth-generation EPYC “Venice” server CPU, the Instinct MI430X and MI350P accelerators, ROCm.AI, and the “Gorgon Halo” and Kria AI Robotics developer platforms designed for local agents and robots. At the heart of the event was Helios, powered by the MI455X, marking AMD’s entry into the AI factory competition with a complete system integrating GPUs, CPUs, networking, and software.

Lisa Su summarized AMD’s strategy in her keynote into three points: computational leadership, an open platform, and “making AI ubiquitous.” Behind these three strategies is AMD’s clear assessment of shifts in the AI industry: training will continue to expand, but the focus of compute consumption has accelerated toward inference; agents have expanded a single model invocation into continuous inference, tool calls, code execution, and data access, transforming AI infrastructure from a GPU issue into a systemic engineering challenge encompassing CPUs, networking, software, power, and data centers.

Lisa Su expects that by 2026, approximately 60% of global AI computing power will be used for inference, emphasizing that agents require both "massive GPUs to perform inference" and "massive CPUs to orchestrate each step."

On July 24, at the AMD Advancing AI 2026 event, Dr. Lisa Su, Chairman and CEO of AMD, Vamsi Boppana, Senior Vice President of AMD’s AI Business Unit, and Dan McNamara, Senior Vice President and General Manager of AMD’s Computing and Enterprise AI Business Unit, engaged in a nearly half-hour discussion with global media, including Tencent Technology. Beyond the product announcements made during the keynote, this conversation addressed several of AMD’s most critical current issues: whether AI demand will continue to grow at a rapid pace, whether Helios will be delivered on schedule, how AMD plans to catch up with NVIDIA, and whether an open ecosystem is a viable business strategy.

Lisa Su also provided further insight into AMD’s assessment of demand: “The more useful AI becomes, the more people will want to use it.” In her view, corporate investment in AI has shifted from initial cost reduction toward more direct value creation, such as product development and business growth. Therefore, even as companies begin to tighten their token budgets, the long-term upward trend in compute demand will remain unchanged.

Faced with the direct question of whether AMD is still following NVIDIA, Lisa Su did not avoid it. She stated that AMD has chosen a different path: leveraging Chiplet and product portfolio architecture to address diverse workloads, and connecting customers and partners through an open ecosystem. She emphasized that AMD is not satisfied with natural market growth, but “expects to grow faster than the market average in every segment it participates in”; the transition from MI450 to MI500 is “not a minor upgrade, but another major leap.”

This competitive approach is also reflected in Helios’s development methodology. Lisa Su stated that Helios was not developed independently by AMD and then handed over to customers; instead, it was co-designed with leading companies such as OpenAI, Meta, and Anthropic. Customers are involved earlier in the product definition process—from chips and software to data centers and supply chains. She referred to this collaboration as a key “secret weapon” enabling Helios’s on-time mass production.

Lisa Su: That's a great question.

First and foremost, predicting markets is one of the most difficult tasks, as predictions are unlikely to be entirely accurate.

What we can be certain of is that we’ve spent considerable time engaging with customers to understand the fundamental shifts occurring at the workload level. What we’re seeing now is extremely strong demand for computing power. This holds true in the GPU and accelerator space, but the pace of change in the CPU sector is even more aggressive.

As for why we believe this trend will continue, you’ve also heard today from OpenAI, Meta, and other major clients. We must plan years in advance to prepare for chips, power, data centers, and the entire supply chain. Therefore, we are now collaborating more closely with our clients than ever before.

AI is a classic example: the more useful it becomes, the more people want to use it. Of course, everyone will continue to optimize it.

Dan previously discussed what we’ve observed when working with enterprise clients. Nonetheless, the overall trend toward AI adoption continues to rise. We’re seeing AI play an increasingly significant role in product development and other key areas where businesses create value.

The value brought by AI has far exceeded cost reduction. While cost reduction may have been the initial motivation for companies to deploy AI, it is now increasingly used to create genuine business value.

Therefore, we have strong confidence in the demand curve.

At the same time, people are increasingly recognizing that this is no longer just an issue of a single chip—it requires a complete family of chips, a comprehensive product portfolio, and a robust hardware ecosystem. This has always been AMD’s distinctive insight.

Lisa Su: First, I want to clearly acknowledge this. As you’ve already seen from the hardware displayed on stage, Helios is truly remarkable.

It is an incredible engineering achievement involving powerful computing, power supply, system integration, density, and many other factors.

But remember, we have been preparing for this moment for years.

This includes capabilities developed internally, such as ROCm and platform software; capabilities built by our networking team; as well as our acquisitions of Pensando and ZT Systems and the integration of these capabilities.

One “secret weapon” I hope everyone feels today is that we are fully aligned with our partners. In other words, AMD doesn’t develop products independently and then hand them off to partners—we co-develop them together.

OpenAI is co-developing with us, Meta is co-developing with us. You also heard from Anthropic’s co-founder, Tom Brown. Because of this, we are very confident in our product ramp-up.

"Full-scale production" means we are ready to ship. We will begin shipping at the end of the third quarter and continue scaling up through the fourth quarter and the first half of next year.

The first products will be deployed in specific data centers, and we have already completed a clear plan. We are also closely collaborating with our ODM partners. One key difference with Helios compared to previous products is that every stage of the supply chain has been fully aligned—this is a major reason for our confidence.

I feel very positive about Helios's progress.

Lisa Su: We’ve indeed launched many products today. But seriously, we believe we’re paving the way for the true direction that future AI computing requires. GPUs are certainly crucial, and we have strong confidence in our GPU roadmap.

As I hinted today, moving from MI450 to MI500 is not a minor upgrade, but another major leap.

We believe AMD will lead in scale-up computing expansion. This is a very important assessment, but it’s the direction we currently see.

We chose a different path. Over the past decade, and especially the last five years, we have been investing in chiplet architecture and product portfolio architecture.

Using EPYC as an example, everyone is now starting to find CPUs interesting, but we’ve been preparing for this moment for many years.

Mark Papermaster and his team have been thinking about how to build an ecosystem where different workloads can use the most suitable chips. This concept is also embedded throughout our overall roadmap.

Therefore, we are very confident in our current position. This is a massive market. AMD believes that the total market opportunity we are targeting will reach $2 trillion by 2030, which is an enormous figure.

But we don’t just aim to share in market growth—we expect AMD to grow faster than the market average in every segment we participate in.

AMD will compete in its own way—through leading technology, an open ecosystem, and selecting the right computing for different workloads. This requires both breadth and depth.

Vamsi Boppana: Lisa has already covered it thoroughly.

New niche areas constantly emerge in the market. We must stay focused and execute rigorously on the most important tasks at hand, or we risk being distracted by countless directions.

But at some point, we will choose a specific niche and clearly state that we believe we will achieve undisputed leadership there. We are very confident about this.

Lisa Su: Today, we can clearly state that AMD holds an undisputed leadership position in the CPU market.

Additionally, what is the current progress of OpenAI's deployment? According to the previous plan, the relevant deployment is scheduled to begin in the second half of 2026.

Lisa Su: First, we are thrilled, honored, and excited to collaborate with the world’s leading frontier AI companies, including OpenAI, Anthropic, Meta, and others.

I would explain it this way: Each transaction is unique.

We have a very special relationship with OpenAI. You can also sense this from today’s presentation by Sachin and Philippe.

We have been collaborating with OpenAI since the MI300. Both parties have jointly developed roadmaps and advanced software. Frankly, OpenAI has made a significant commitment to AMD, including deployments of up to 6 GW.

We previously stated that the first 1 GW would begin deployment in the second half of 2026 and continue into 2027. I can confirm that this plan is progressing exactly as expected.

Without targeting any specific client, overall demand for the MI450 series has exceeded initial expectations.

Customers have seen the actual performance of the hardware and are pleased with the results. As a result, we are increasing production capacity and supply scale over the coming quarters.

Regarding Anthropic, I briefly mentioned it on stage—AMD has been closely following Anthropic for several years, ever since Dario and Tom first founded the company.

However, such collaborations always require finding the right timing. For any company, adopting a new chip platform demands significant engineering investment. Therefore, we understand that Anthropic’s decision to dedicate its engineering capabilities, resources, funding, and time to the AMD platform is a major one.

Each transaction is unique. Anthropic's agreement primarily centers on MI450, with a maximum scale of up to 2 GW. The first 1 GW is planned to begin deployment in the first half of 2027. We are also discussing longer-term plans with Anthropic.

Therefore, the two transactions simply use different structures.

Lisa Su: Of course. AMD's relationship with Anthropic is first and foremost built on providing substantial computing power to support Anthropic's ongoing scaling efforts.

Users can expect Anthropic to utilize AMD's GPUs and CPUs through multiple channels, including cloud services and potentially its own data centers.

We are working closely with Anthropic to support these various deployment formats. This collaboration extends far beyond simply purchasing chips—we will jointly handle system deployment and launch, and collaborate on running Claude on AMD platforms.

From a software perspective, Claude and Codex are among the most powerful and popular platforms today.

We want developers to be able to develop on AMD hardware very efficiently through these platforms, which is also an important part of the collaboration.

In terms of data centers, we are also actively ensuring sufficient deployment capacity in the market. This is indeed one of AMD’s key focus areas, but it is not limited to Anthropic—it addresses the overall demand across the entire system and ecosystem.

Lisa Su: I do not view any partnership as a temporary stopgap. We do not approach partnerships with the mindset of “just a temporary measure.”

We partnered with Cerebras because Cerebras has highly innovative technology. There are many different approaches to accelerating specific workloads, and Cerebras’ technology is one of the most compelling, seamlessly complementing Helios.

AMD's open ecosystem means we will collaborate with multiple companies that possess valuable technologies, as long as those technologies align with our assessment of long-term growth.

In the long term, I believe there will be increased decoupling and separation of workloads. The pace of this trend will depend in part on the cost structures and pricing levels of different solutions.

From AMD’s own perspective, we have a robust roadmap, including the MI400, MI500, and MI600. At the same time, we also have the capability to mix and match different computing technologies.

We are very excited about our partnership with Cerebras. You can expect AMD to continue driving further workload partitioning and targeted acceleration in the future.

Vamsi Boppana: Let me quickly add one point—there’s no need to look far ahead; AMD’s current product portfolio already includes numerous small AI engines capable of handling ultra-low-latency, ultra-low-power, and real-time response tasks.

You can explore our embedded product portfolio and its applications in automotive, medical devices, and other fields. The key point is that AMD believes AI will integrate closely into various computing form factors, each with distinct characteristics. We provide tailored solutions for these different scenarios.

This article is from Tencent Technology, authored by Xiao Jing, edited by Xu Qingyang, and published with permission from 36Kr.