AMD Helios бросает вызов Nvidia: стоечная AI-система выйдет в конце 2026 года

AMD is preparing to ship Helios — a rack-scale AI system based on Instinct MI400 accelerators that directly competes with Nvidia's server platforms. According to TechCrunch on July 23, 2026, the first shipments to customers will begin before the end of 2026.

What Helios is

Helios is not a single chip, but an entire server rack designed by AMD as a unified accelerator for training and running large AI models. At its core are AMD Instinct MI400-series accelerators, united into a single unit by high-speed interconnects.

Each MI400 accelerator carries memory of the new HBM4 standard, which is critical for loading models with hundreds of billions of parameters without constant access to external nodes. Helios continues the lineup started by the MI300 and MI350 series accelerators: AMD is consistently increasing memory capacity and bandwidth to catch up with Nvidia in data centers.

According to AMD, the system's key specifications are:

• Shipments to customers — before the end of 2026
• Foundation — AMD Instinct MI400 accelerators
• 72 GPUs in a single rack
• 31 TB of HBM4 memory and about 1.4 PB/s of bandwidth
• Open connectivity standards instead of proprietary NVLink

The rack-scale format means the customer receives a ready-made compute node, not a set of cards for self-assembly. It is precisely this approach that helped Nvidia entrench itself in data centers: its GB200 NVL72 racks are sold as a single system, and AMD is responding for the first time with a product of the same class.

How Helios differs from Nvidia

Helios's main distinction is its bet on open standards. For connectivity between accelerators, AMD uses UALink and UltraEthernet, whereas Nvidia builds its racks on the closed NVLink. By AMD's design, this allows customers to avoid being locked into a single supplier and to combine hardware from different vendors.

"Helios is designed to operate as one giant compute engine," —

Lisa Su, CEO of AMD, at the system's announcement.

Openness is a strategic response to Nvidia's closed ecosystem, where chips, networking, and CUDA software are all tied to one another. AMD counters it with a combination of its own hardware, open interconnects, and the ROCm platform.

Why this matters for the market

Nvidia controls more than 90% of the AI accelerator market, and Helios is AMD's most serious attempt yet to capture a share of that demand. The largest compute buyers — Microsoft, Meta, Oracle — are looking for a second supplier to reduce dependency and prices, and AMD's rack-scale system gives them a comparable alternative for the first time. Even a few percentage points of this market would mean billions of dollars in AI accelerator revenue for AMD.

Demand for training large models continues to grow faster than Nvidia can supply hardware. Nvidia is already preparing a response — the next-generation Vera Rubin racks — so Helios is arriving in the middle of a direct platform race, not into an empty niche. For AMD, this is a window: competitive performance and total cost of ownership could secure the company a place in hyperscaler procurement.

What it means

The race for AI hardware is ceasing to be Nvidia's monopoly. The rack-scale system is becoming the new unit of sale, and with Helios's launch, the largest data centers gain a real second supplier for the first time — meaning pressure on prices and delivery timelines.

Frequently Asked Questions

When will AMD Helios launch?

The first shipments to customers will begin before the end of 2026 — as reported by TechCrunch on July 23, 2026. The system is built on AMD Instinct MI400 accelerators.

How does AMD Helios differ from Nvidia's systems?

Helios uses the open connectivity standards UALink and UltraEthernet instead of Nvidia's proprietary NVLink. It is a rack-scale system with 72 GPUs, which AMD positions as a unified compute node for training large AI models.

*Meta has been designated an extremist organization and is banned in Russia.