量子位
Lisa Su Speaks in Shanghai: AI is Redefining Every Layer of Computing
7.5Score
TL;DR · AI Summary
At the AMD AI Developer Conference in Shanghai, CEO Lisa Su stated that AI competition is shifting from model capabilities to systems engineering and full-stack optimization. Developers need a deployable, optimizable, and continuously evolving engineering system. AMD, centered on its ROCm open-source platform, provides full-stack computing power from cloud to edge, while continuously strengthening its developer ecosystem in China.
Key Takeaways
- AI industry competition is shifting from model capabilities to systems engineeri
- AMD provides full-stack computing from cloud to edge, centered on the ROCm open-
- China is leading the open ecosystem, with AMD investing in Greater China for ove
Outline
Jump quickly between sections.
The AMD AI Developer Conference debuted in Shanghai, with Lisa Su noting the competition focus is shifting from model capabilities to systems engineering and full-stack optimization.
AI deployment faces a cost crisis, requiring deployable, optimizable, and continuously evolving engineering systems to address cost, complexity, and deployment challenges.
Agent workflows generate cumulative computing consumption through multi-round planning, tool calls, and verification, with cost structures shifting from linear to exponential as scale grows.
The Agent era requires AI to accomplish tasks, with systems running multiple models, modalities, and distributed computing simultaneously, fundamentally increasing engineering difficulty.
Edge and device deployment scenarios are fragmented, requiring toolchain and optimization strategy rebuilds for each hardware platform, accumulating hidden costs.
Lisa Su stated China is leading the open ecosystem, with AMD providing cloud-to-edge full-stack computing centered on the ROCm open-source platform.
Conference topics cover inference optimization, training engineering, edge deployment, and underlying infrastructure, with AI development moving from model usage to system construction.
The AMD AI Developer Program - China officially launched, supporting Chinese AI developers through technical resources, community engagement, and developer activities.
Mindmap
See how the topics connect at a glance.
查看大纲文本（无障碍 / 无 JS 友好）
- AMD AI开发者大会
- 行业趋势判断
- AI竞争焦点转移
- 工程化挑战
- 成本危机应对
- AMD战略布局
- 全栈算力方案
- ROCm开源平台
- 开放生态定位
- 中国市场投入
- 开发者生态建设
- 本地化适配
- AI开发者计划
Highlights
Key sentences worth saving and sharing.
The focus of competition is shifting from model capabilities to systems engineering and full-stack optimization capabilities.
While single-request costs decrease, total costs transform from a single line item on bills to harder-to-see accumulation at the system level.
The Agent era requires AI to accomplish tasks, with a single system often running multiple models, modalities, distributed computing, and tool calls simultaneously.
Innovation accelerates only when the industry comes together around open infrastructure and shared standards.
Chinese developers have never been merely consumers of AI applications, but builders of infrastructure.
#AMD#AI Engineering#ROCm#Lisa Su#Open Ecosystem
Open original articlemarkdown
2026-05-20 08:48:27 Source: [QbitAI](https://www.qbitai.com/)
AMD Continues to Invest in China's Developer Ecosystem
##### Krissie Reporting from Shanghai
QbitAI | QbitAI Official Account
"The AI revolution is redefining every layer of computing."
This was the latest assessment from AMD's AI Developer Conference, which made its debut in Shanghai today. AMD Board Chair and CEO, Dr. Lisa Su, shared this perspective on the AI industry during the event.
At the invitation of AMD, QbitAI attended this conference to observe and learn from the event firsthand.
After participating in the day's activities, it became clear that Dr. Su's perspective, as well as the conference's topics and lineup, reflect the accelerating transformation of the AI industry.
The focus of competition is shifting from model capabilities to system engineering and full-stack optimization capabilities.
The challenges faced by developers in each stage of inference, training, and fine-tuning are becoming more specific and more engineering-oriented.
What developers truly need is an engineering system that can be implemented, optimized, and continuously evolved.
This insight is particularly relevant in China.
DeepSeek, Qwen... Over the past two years, the most active AI projects globally have never been without Chinese involvement.
Chinese developers are not just consumers of AI applications but are also builders of infrastructure.
AMD's response today aligns with this trend.
## AI Developers Need a New Engineering Framework
The issue of implementation costs for AI is becoming an unavoidable core topic in the industry.
In early 2026, David Patterson, Turing Award laureate and Google Distinguished Engineer, issued a warning. He stated that the large-scale implementation of AI is facing a cost crisis.
This crisis has a somewhat contradictory appearance: token costs continue to decline, but enterprise AI budgets are increasing instead of decreasing.
The underlying reason for this is a fundamental shift in the working paradigm of AI.
The rapid rise of frameworks like OpenClaw and Hermes within months indicates that projects are transitioning from single-query interactions to agent workflows.
Under the new workflow, completing a task requires multiple planning steps, repeated tool calls, and constant validation, each of which incurs computational costs.
While individual costs are reduced, overall costs shift from a single line item on a bill to a more complex cumulative cost at the system level.
Such systemic issues require systemic solutions, marking a new phase in AI competition.
In this phase, the true test is whether a system can run stably, efficiently, and sustainably at scale.
This challenge can be broken down into three layers.
First, at the cost layer: the larger the scale, the more pronounced the cumulative effect of token consumption.
The difference between running ten and one thousand concurrent agents is not just token usage, but also redesign in system scheduling, fault tolerance, and resource allocation, leading to a cost structure that shifts from linear to exponential.
Under the agent workflow, each task requires a complete chain of multi-round planning, tool calls, and validation, with computational costs incurred at each step.
As scale increases, this cumulative effect becomes more pronounced, shifting the cost structure from linear to exponential.
Second, at the complexity layer: the leap in engineering difficulty stems from fundamental changes in AI application forms.
In the previous chat format, one model corresponds to one capability with clear boundaries. However, the agent era requires AI to "get things done," often running multiple models, multimodal capabilities, distributed computing, and tool calls within a single system. Delays or failures in any环节 will impact the entire chain.
Engineers face the challenge of maintaining a system that is continuously evolving and may need to scale at any moment.
Finally, at the deployment layer: fragmented scenarios are becoming a new engineering burden.
Cloud-based inference cannot meet all needs, as some scenarios require data to remain local, others are highly sensitive to latency, and some have unstable networks.
These demands push developers toward edge devices and on-device deployment. However, each hardware platform change often requires redoing toolchains, optimization strategies, and debugging environments, with hidden costs accumulating due to fragmentation.
The叠加 of these three layers of pressure points to a single conclusion: developers need an implementable, optimizable, and continuously evolvable engineering framework.
## Dr. Lisa Su: China is Leading Open Ecosystem Development
At this AI Developer Conference, AMD Board Chair and CEO Dr. Lisa Su shared AMD's perspective on these needs:
In the agent era, individuals may have 5, 10, or even 100 agents, fundamentally altering the structure of computational costs. Relying solely on GPUs is no longer sufficient; a full-stack computing solution combining GPUs and CPUs is needed to truly meet requirements.
AMD's strategy is to provide end-to-end computing power from the cloud to edge devices, centered around the ROCm open-source software platform, ensuring developers have the right tools for every deployment scenario.
This perspective has deeper roots.
In terms of global strategy, AMD has a consistent approach in response to the trend of AI becoming more engineering-oriented.
At CES 2026 earlier this year, Dr. Lisa Su already pointed to the direction: open ecosystems will be the next infrastructure for AI.
Only when the industry unites around open infrastructure and shared standards can innovation accelerate.
Infrastructure implies that it belongs to no single company but is a foundational layer that the entire industry depends on, builds, and benefits from collectively.
By choosing to use the term "open ecosystem," AMD is making a statement: the future of AI should not be locked into a single closed system.
This stance is also reshaping AMD's positioning.
From selling chips to becoming a platform, AMD aims to become a platform that developers can trust long-term,跨越 hardware iterations.
Software openness and hardware iteration form a mutually reinforcing relationship, creating a ecosystem where a large developer community is willing to stay engaged long-term.
Specifically in China, AMD has been deeply involved in the Greater China region for over 30 years, with its Shanghai R&D center being one of the largest globally.
In Dr. Su's view, China is not only an important market for AMD but also a crucial component of AMD's global roadmap, covering chips, AI software, and platform engineering.
In terms of open-source ecosystems, Dr. Su directly pointed out that China is leading in this area. This openness is the core force driving the entire AI ecosystem to evolve as quickly as possible and aligns closely with AMD's strategic direction.
Hosting this conference in Shanghai is a demonstration of this strategy落地 and continuity in the Chinese market.
In China, this strategy translates to these specific actions:
- Continuing to invest in local developer community building, ensuring Chinese developers can effectively and efficiently use these tools in their daily engineering practices.
- Collaborating with local open-source ecosystems to avoid繁重的 compatibility work for developers.
- Ultimately reducing the threshold for AI development and deployment, enabling more teams to turn their ideas into systems running in production environments.
## AI Enters Systematic Engineering Practice
The competition in AI engineering is now a foundational infrastructure challenge that the entire developer community must address together.
The topics covered in AMD's AI Developer Conference, with its hands-on workshops and technical presentations, represent a cross-section of current AI engineering practices.
Inference-focused topics centered on the new challenges brought by the agent era.
In the single-query era, inference costs could be measured by single-call prices. However, under the agent workflow, completing a task requires multi-round planning, tool calls, and validation, fundamentally changing the cost structure.
How to reduce token costs under the new paradigm, maintain throughput efficiency in high-concurrency scenarios, and enable automated optimization of inference are the "hard bones" the industry is currently tackling, representing the core themes of the inference-focused sessions.
Notably, topics related to training reflected the engineering pressure as AI applications deepen.
RLHF, once a research paper, has become a standard component of teams' workflows. The challenge now is how to efficiently run end-to-end alignment training on a single card.
The MoE architecture is now being widely adopted in production, with stability and efficiency during超大规模 training becoming daily engineering challenges.
Edge-focused topics showed the most noticeable changes.
Fully offline AI desktop robots, personal agents driven by local large models, and complete development workflows running on local hardware are now achievable on AMD edge hardware.
Edge AI is no longer a degraded alternative to cloud-based solutions. It has its own engineering logic in scenarios like privacy protection, low latency, and offline availability, requiring a complete support system from model quantization to local inference acceleration.
Another set of topics explored more foundational paths.
Kernel development for AI, compiler optimization, GPU kernel AI agents, and PyTorch distributed training framework adaptation for AMD GPUs—all focus on the infrastructure layer that determines how far the ecosystem can go.
AMD is reinforcing long-term connections with the developer community through hands-on workshops, open-source toolchains, and real-world engineering scenarios, driving AI development from model usage to system construction.
From these topics, it is clear that AMD aims to create a flywheel where developers move from understanding to implementation and then to continuous evolution.
Notably, the AMD AI Developer Program-China officially launched today.
This program is AMD's membership ecosystem project for AI developers, offering technical resources, development courses, community engagement, and developer events to help developers more efficiently work on AI applications and large language model development.
By joining the program, developers can connect with a broader Chinese developer community, participate in technical exchanges and workshops supported by AMD AI experts and ecosystem partners.
Post-conference, AMD will continue to provide technical content updates, community interactions, and subsequent developer events through this program.
## Ongoing Investment in China's Developer Ecosystem
Building a developer ecosystem is a long-term commitment that requires continuous improvements to toolchains, community operations, localized adaptations, and the trust built through real-world engineering practices.
Once this trust is established, the switching cost is extremely high, as the team's engineering inertia, accumulated optimizations, and validated workflows are deeply embedded.
It seems AMD has already gained this trust.
Today in Shanghai, almost all key builders of China's AI engineering ecosystem were gathered under one roof.
This density is the result of years of accumulation and a cross-section of AMD's long-term efforts in the Chinese developer community.
From timely adaptation to Chinese mainstream open-source models like DeepSeek and Qwen to ongoing community building, AMD's efforts follow a clear logic:
Ensure Chinese developers can not only use these tools but also use them effectively in their daily engineering practices.
This logic reflects AMD's judgment on the Chinese AI market: China is not just a consumer market for AI applications but also an important builder of AI infrastructure.
Contributions from the Chinese open-source community in areas like training frameworks, inference engines, and model quantization are being widely adopted by the global developer community.
AMD's deep investment in China's developer ecosystem at this time is both recognition of this reality and a bet on the future. By engaging with the open-source community, building toolchains, and connecting with local developers, AMD is making long-term investments in China's AI developer ecosystem.
In the AI era, the deepest moat is when developers choose to build their applications on your platform and don't want to leave.
What AMD is doing is making this a reality in China, one step at a time.
*This article is copyrighted. Unauthorized reproduction or use in any form is prohibited.*