NewsTechnical

Helios Is AMD’s First AI System To Rival Nvidia Vera Rubin — We Got An Exclusive, First Look

CNBC

AMD has unveiled Helios, its first rack-scale AI system designed to compete with Nvidia's dominant position in AI infrastructure. With major commitments from Meta, Microsoft, OpenAI, and Oracle, Helios aims to capture market share by offering superior efficiency, customizable configurations, and open-source software alternatives to Nvidia's proprietary CUDA ecosystem.

Summary

AMD announced Helios, a 7,000-pound, $5 million rack-scale system containing 72 GPUs and 18 CPUs per rack, representing the company's first major challenge to Nvidia's near-monopoly in AI infrastructure. The system integrates four core AMD technologies: GPUs, CPUs, networking, and software into a cohesive platform. According to AMD leadership, Helios offers superior inference performance, greater memory bandwidth, and more customization options compared to Nvidia's competing systems like Grace Blackwell and Vera Rubin.

AMD has secured significant early commitments from major tech companies: Meta committed to deploying 1 gigawatt of AMD GPUs initially on Helios racks, Microsoft signed up for systems in its data centers, and both OpenAI and Oracle made major deployment commitments for 2025. These deals provide AMD with crucial market validation and momentum as it enters a market where Nvidia currently controls over 95% of data center GPU share.

A key differentiator for Helios is its open architecture and use of open standards, contrasted with Nvidia's vertically integrated approach. AMD emphasizes that Helios uses partnerships across the ecosystem rather than proprietary lock-in. The company also highlights its newer ROCm software stack as a necessary open-source alternative to Nvidia's ubiquitous CUDA software, though executives acknowledge CUDA currently has a larger and more mature ecosystem.

AMD's path to this moment involved significant investments and acquisitions. The company acquired XYlinks for $50 billion and ZT Systems for $4.9 billion to gain expertise in AI networking and server building. AMD also made several software acquisitions to develop its ROCm stack. The company deployed 8,000 employees across multiple new lab facilities in Austin, Texas, working on Helios development, testing, and validation.

The transcript reveals that AMD faces multiple supply chain constraints. The MI455 GPUs are manufactured at TSMC's most advanced 2nm node, where Nvidia has already reserved majority capacity for advanced packaging. AMD committed $10 billion to alternative Taiwanese packaging companies like ASSE to secure capacity. Memory supply is another challenge, with each GPU requiring 432 GB of HBM memory, though AMD states it has secured relationships with all three major memory suppliers. Power consumption represents another hurdle, with each rack requiring approximately 225,000 to 245,000 watts and consuming 750 gallons per minute of chilled water in closed-loop systems.

Historically, AMD dominated data center CPUs in 2003 with nearly 25% market share, but lost this position through delays and missteps. Under CEO Lisa Su's leadership since 2014, AMD has rebuilt its data center business to become its fastest-growing segment and now represents the majority of revenue. AMD achieved a record 46% revenue share in x86 server CPUs in Q1 2025, though Intel remains the overall leader. The company estimates it will generate tens of billions of dollars in data center AI revenue starting next year, primarily from Helios.

AMD positions Helios not as a zero-sum replacement for Nvidia systems but as addressing massive market growth in AI compute. The company estimates each Helios system costs $5-5.5 million compared to Nvidia's Vera Rubin at $3.5-4 million, with AMD arguing superior total cost of ownership and lower cost-per-token metrics justify the premium. Future interconnect technology, specifically fiber optics replacing copper cables, could further differentiate systems within 2-3 years. Trade tensions with China present additional constraints, as 21.5% of AMD's 2025 revenue came from China, and regulatory restrictions may limit Helios availability in that market.

Key Insights

  • AMD's 8,000 employees working on Helios represent a massive organizational commitment, with the company holding quarterly town halls where all Helios team members meet together.
  • AMD claims data center AI revenue will reach tens of billions of dollars starting next year, with the majority coming from Helios specifically.
  • Nvidia has reserved the majority of TSMC's leading co-assemble packaging capacity, forcing AMD to commit $10 billion to alternative Taiwanese packaging companies like ASSE.
  • AMD achieved a record 46% revenue share in the x86 server CPU market in Q1 2025, demonstrating significant gains in the data center CPU space despite Intel remaining the overall leader.
  • AMD positions Helios as succeeding through delivering leadership performance per competing products and better total cost of ownership rather than asking customers to replace Nvidia systems entirely.

Topics

AMD Helios rack-scale AI system specifications and designCompetition with Nvidia in AI infrastructure and CUDA vs ROCm software ecosystemsMajor customer commitments from Meta, Microsoft, OpenAI, and OracleSupply chain constraints including TSMC wafer capacity, advanced packaging, and HBM memoryAMD's historical comeback in data center under Lisa Su's leadershipHelios architecture advantages including open standards, networking technology, and cost-per-token efficiency

Transcript

[0:00] Chip giant AMD is in a race with Nvidia to make these graphics processing units to power AI. Now with its first rack scale system for AI, Helios finally here. AMD is aiming to give the world's most valuable company its first real competition in years. >> We're in our newest mega lab. So this is a facility that we use to validate and bring up Helios, get it ready for production. Also known for its industry-leading central processing units, AMD is bringing it all together in [music] this 7,000lb, $5 million [0:32] behemoth. >> This is Helios. It's our baby. It's 72 GPUs and 18 CPUs per rack. [music] >> We came to Central Texas for the world's…

Full transcript available for MurmurCast members

Sign Up to Access

More from CNBC

Get AI summaries like this delivered to your inbox daily

Get AI summaries delivered to your inbox

MurmurCast summarizes your YouTube channels, podcasts, and newsletters into one daily email digest.