S3DAT NEWS
· 20 stories RSS s3dat.com

AMD reports 97% Data Center growth as Helios racks come online

AMD reported 97% annual growth in Data Center revenue on November 3, alongside Helios workloads at five major customers. The update distinguishes early customer deployments from finished systems available to buyers.

Spec sheet

Data Center revenue
$8.5 billion in Q3; up 97% year over year
Named users
Anthropic, Meta, Microsoft, OpenAI, Oracle Cloud Infrastructure
Manufacturer shipments
Expected in Q4 2026, AMD says
Rack design
72 MI455X GPUs and 18 Venice CPUs

Chips, racks and cloud instances

The revenue figure includes EPYC processors as well as Instinct accelerators; it does not measure Helios sales alone. In prepared earnings remarks, AMD said volume shipments of MI455X chips began during the third quarter. HPE, Dell, Lenovo and Supermicro are expected to begin shipping Helios systems in the fourth quarter. Venice processor shipments also started in the third quarter, with cloud instances planned for early 2027. These are separate delivery milestones: a chip shipment does not establish when a developer can rent a machine.

An update to an existing rollout

AMD’s July 23 launch described Helios as an integrated architecture: 72 MI455X accelerators, 18 sixth-generation EPYC Venice processors, networking and ROCm software. It already reported Meta testing workloads, while OpenAI expected to bring systems online in the fourth quarter. November’s customer list is therefore an update to a rollout already under way. The practical difference from buying individual accelerators is the surrounding system. Processing, communication and software have to operate together, and the company’s partners turn that design into deployable equipment.

What the manufacturer supplies

HPE’s September 30 announcement illustrates that division of work. Vultr placed a $1.2 billion order for Helios systems at US data centres. HPE’s design includes six Juniper QFX5252 scale-up Ethernet switch trays per rack, alongside direct liquid cooling and deployment services. The order describes a provider building capacity; it supplies no rental price for an application team. Buying a rack and renting compute also put different work on the buyer. A system operator takes responsibility for installation and maintenance, while a cloud customer buys access under the provider’s service terms.

Components can hold back capacity

AMD’s November 4 quarterly filing makes the boundary explicit: the company supplies products and licenses the design, but does not manufacture or sell completed Helios racks. It warns that shortages, long lead times or quality problems in third-party components could delay completed systems and AMD’s associated revenue. AMD also says it has supported procurement of some outside components, taking on additional inventory and financial risk. A rollout therefore depends on more than accelerator supply. One missing component can hold back a complete system even when the main chips are ready.

The earnings remarks describe the cost of preparing for expansion. Inventory rose by roughly $3 billion during the quarter to $11.5 billion, which AMD attributed to Data Center demand. Management expects Data Center revenue to grow about 50% sequentially in the fourth quarter. For developers, the useful evidence will be an available configuration with a quoted price and delivery date, followed by results from their own workload.

Claim check

  • Company claim AMD says five companies are running Helios workloads. [1]
  • Company claim Manufacturer system shipments in Q4 and 50% sequential Data Center growth are AMD forecasts. [4]
  • Confirmed AMD’s filing says it does not manufacture or sell completed Helios racks. [5]
  • Unclear The earnings materials do not establish generally rentable capacity or comparative application costs. [1]

Sources