Inside Advancing AI – AMD Pushes AI Rivalry With Helios Rack And MI400 Chips

AMD Pushes Harder Against Nvidia in AI Chip Race

At the “Advancing AI” event in San Jose on June 12, 2025, AMD CEO Lisa Su unveiled the Instinct MI350 series—featuring MI350X and MI355X—boasting up to 4× performance gains compared to the MI300X. Su emphasized AMD’s mission to redefine value by claiming 40% higher performance-per-dollar in AI inference tasks compared to Nvidia’s Blackwell GPUs (B200 series). This strategic positioning directly targets the cost-sensitive, high-volume inference segment where AMD sees substantial market opportunity.

The new chips are based on AMD’s CDNA 4 architecture and manufactured using TSMC’s 3nm process. They are expected to begin shipping in Q4 2024, with widespread data center integration forecasted by early 2025.

Introducing Helios: AMD’s Open Rack-Scale AI Server

Alongside the new chip launch, AMD revealed Helios—a 2026-bound, open-standard, rack-scale AI server platform designed to house up to 72 GPUs per rack. Helios will integrate AMD’s upcoming MI400 series GPUs (built on the future CDNA 5 architecture and expected in 2026), Zen 6 CPUs, and Pensando’s Vulcano data processing units for high-speed networking.

Unlike Nvidia’s proprietary NVLink-based NVL72 and Vera Rubin systems, Helios leverages open industry standards for networking and server management. This open architecture is expected to appeal to hyperscalers and enterprises looking to avoid vendor lock-in.

Su emphasized that AMD’s vision is about “enabling an open ecosystem for AI innovation,” calling Helios a foundational leap for hyperscale inference and training at scale.

Strategic Alliances and Infrastructure Commitments

OpenAI and Open Collaboration

One of the biggest surprises at the event was the appearance of OpenAI CEO Sam Altman, who confirmed a deepening collaboration with AMD.

OpenAI, which has historically depended heavily on Nvidia chips, will adopt AMD’s MI350 and future MI400 series for select inference and training tasks.

“When you first told me about the specs… it just sounded totally crazy,” Altman said onstage, referring to the MI400 chip roadmap. He added, “We’re incredibly excited about what we’re building together—it’s an amazing thing.”

This marks one of the most high-profile validations yet of AMD’s AI ambitions, particularly as OpenAI prepares to scale its infrastructure in response to the explosive growth of ChatGPT and its API services.

Hyperscalers and Startups Join the Fold

Beyond OpenAI, AMD confirmed strong traction among hyperscalers during Advancing AI event. Companies including Meta, Microsoft, Oracle, xAI (Elon Musk’s AI startup), Marvell, and Astera Labs have committed to adopting MI350-based infrastructure within the next 12 months.

In a notable move, Crusoe, a cloud provider focused on sustainable AI infrastructure, announced a $400 million order for roughly 13,000 MI355X chips to power its new U.S.-based inference data center, launching in Q4 2025.

Crusoe’s CEO Chase Lochmiller stated that AMD’s cost-efficiency and open standards made the decision straightforward.

Competitive Strategy and Market Reactions

Cost Gains vs. Stock Pressure

Despite the flood of announcements, AMD stock dipped slightly by 1.7% following the event. Some analysts attributed the decline to investors expecting even bigger surprises or concrete financial forecasts.

However, tech industry observers remain optimistic about AMD’s ability to capitalize on growing demand for AI chips.

“AMD continues to position itself as the viable second-source alternative to Nvidia,” said Daniel Morgan, senior portfolio manager at Synovus Trust. “They still face an uphill battle in software, but the performance metrics are now credible enough to win deals.”

Expanding Stack via Acquisitions and Hiring

In a calculated effort to scale vertically, AMD has made a series of strategic acquisitions:

• ZT Systems (March 2025) – for rack-scale systems integration.

• Untether AI (late 2024) – for energy-efficient AI acceleration.

• Lamini’s engineering team – to enhance software tooling for fine-tuned LLM deployment.

AMD also expanded its AI ecosystem with 25+ investments in startups including Nod.ai (open-source software optimization) and Silo AI (Europe’s largest private AI lab).

These investments strengthen AMD’s position in both hardware and software innovation—critical for competing with Nvidia’s CUDA-powered stack.

Software Ecosystem: ROCm vs. CUDA

Su acknowledged that while ROCm (AMD’s software development stack) has matured significantly, it still trails Nvidia’s CUDA in terms of developer support and tools.

To address this, AMD pledged deeper collaboration with open-source communities and said ROCm now supports over 90% of popular AI frameworks out-of-the-box.

Despite ongoing U.S. export restrictions on advanced chips to China, Su stated that AI infrastructure growth “remains robust” globally and AMD expects “double-digit growth” in its data center segment.

Market Outlook and Industry Impact

AI Accelerator Market Set to Explode

AMD predicts that the AI accelerator market will surpass $500 billion by 2028, growing at an annualized rate of 60%.

The company’s aggressive roadmap—with the MI350 series this year and MI400 by 2026—is tightly aligned with this massive infrastructure buildout.

Data center operators are seeking alternatives as Nvidia supply constraints persist. AMD’s push for open standards and better price-performance may help tip the scales in its favor.

Open Standards as a Competitive Edge

One of the core differentiators AMD continues to emphasize is interoperability. Whereas Nvidia favors a vertically integrated, proprietary model, AMD wants to be the foundation for an open, flexible AI infrastructure ecosystem.

Lisa Su closed the keynote by stating: “The future of AI won’t be built by one company alone. Our job is to enable choice, scale, and innovation across the board.”

Looking Ahead

• The MI350 series begins volume shipments in Q4 2024, with large-scale deployments expected by Q1 2025.

• The MI400 series and Helios server racks are scheduled to launch in early 2026.

• AMD plans continued investment in software, developer tools, and open standards to strengthen adoption.

• Key partners like OpenAI and Crusoe will likely serve as flagship use cases for AMD’s evolving ecosystem.

• While export controls and supply chain pressures remain challenges, AMD is now viewed as the most credible Nvidia alternative in the AI chip market.

Quick Comparison Table

With powerful new chips, open server architectures, and deepening alliances with top AI players, AMD has clearly signaled that it is no longer playing catch-up—it’s aiming to lead the next wave of AI infrastructure transformation.

The coming year will reveal whether AMD can turn its technical wins into lasting market leadership.