Following criticism over its limited gains during the recent surge in AI demand, AMD is moving to more directly compete with Nvidia in the AI accelerator market.
According to a detailed breakdown by SemiAnalysis, the initiative is being led by CEO Lisa Su and spans multiple areas, including software improvements, custom-built data center hardware, and aggressive pricing strategies. SemiAnalysis analyst Dylan Patel described the shift as "AMD 2.0" and referred to the company's new posture as a transition into "Wartime Mode."
CEO Lisa Su initiates "Wartime Mode"
The changes follow a December 2024 report from SemiAnalysis that described AMD's AI software stack as "plagued with bugs that make training AI models nearly impossible without significant debugging." While the hardware was considered competitive on paper, the report stated that ROCm remained "very difficult to work with."
Lisa Su reportedly contacted both SemiAnalysis and AMD engineering leadership within hours of the report's publication. According to the firm, she acknowledged the issues and initiated a broader internal realignment. SemiAnalysis characterized the result as a company-wide shift into “Wartime Mode.”
One of the main priorities is strengthening AMD's ROCm software stack. The company is investing in improvements to usability, stability, and performance. SemiAnalysis reports that AMD is focusing on better Python integration and support for widely used AI frameworks such as PyTorch. Additional efforts include upgrading developer tools for data center use, with work underway on Docker compatibility, monitoring features, and support for multi-GPU configurations.
In January 2025, AMD launched a dedicated developer relations program, led by Anush Elangovan, to address longstanding gaps in ecosystem support. According to SemiAnalysis, Elangovan has been directly engaging with developers through online platforms and in-person events.
AMD is also preparing a free developer cloud to provide broader access to its GPUs. The platform is intended to simplify adoption and mirrors elements of Google's TPU Research Cloud, which helped drive early community engagement around TPUs.
MI450X targets Nvidia's next-generation rack systems
AMD's hardware roadmap centers on the MI450X series, expected in late 2026. The chips are designed to compete with Nvidia's upcoming VR200 NVL144 system. AMD plans to build its own rack-scale systems using 64 or 128 GPUs connected via Infinity Fabric over Ethernet.
The design approach shares similarities with Nvidia's GB200 NVL72 system, but the MI450X targets Nvidia's next generation directly. To support the rollout, AMD acquired ZT Systems in March 2025. The system builder is expected to provide key capabilities in rack-scale architecture and manufacturing.
Due to ongoing gaps in ROCm software, AMD is relying in part on pricing to attract customers. The MI300X is currently positioned as a lower-cost alternative, but its appeal is limited when compared against Nvidia's newer offerings. The MI355X series, meanwhile, is positioned closer to Nvidia's air-cooled HGX platforms than to its high-end rack-scale systems.
Software and hiring remain persistent challenges
SemiAnalysis identifies talent retention and recruitment as major roadblocks to AMD's progress. The firm notes that "AMD has so far paid too little to retain or attract top talent," with compensation structures more aligned to legacy semiconductor firms than to leading AI hardware companies.
The report recommends AMD increase stock-based compensation to improve long-term incentives. "We strongly believe that if AMD does not significantly increase their AI Software Engineer pay, AMD will continue losing to Nvidia," SemiAnalysis wrote.
Internal infrastructure is also a limiting factor. AMD has access to "less than one twentieth of the number of GPUs of Nvidia," according to the report. Unlike Nvidia, which operates long-term, dedicated GPU clusters, AMD's development relies heavily on short-term leased capacity. SemiAnalysis says this "short-sighted focus on quarterly earnings compromises its capacity for long-term competitiveness."
The report concludes: "AMD needs to maintain (if not intensify) their sense of urgency to even have a chance at being on par with Nvidia."
