Trends›The rack is the unit of compute
The rack is the unit of compute
At the AI frontier the integrated rack (an NVLink-domain of GPU, CPU, DPU, interconnect and liquid cooling), not the individual chip, is the unit of design, sale and competition, because interconnect bandwidth, power density and cooling now bind harder than per-chip FLOPS, so margin and…
83 ↓1 confidence, computed 23 Oct 2026
- Technical
- High weight
What rests on thismarkets it underwrites or threatens
Underwrites
Evidencethe schedule the figure draws
The evidence, newest first
- Jun 2026
for · primary AMD's MI400 Helios is a rack-scale system connecting 72 MI400-series GPUs in a tightly coupled scale-up domain using UALink, in qualification at hyperscalers as of mid-2026 with broader availability expected in H2 2026. AMD MI400 Helios Rack-Scale System and UALink Interconnect - Jun 2026
for · observable Vera Rubin VR200 rack-scale systems quoted ~$5-7M (NVL144 Ultra up to $8.8M); NVIDIA 'moving closer to shipping entire full-scale systems'. Vera Rubin NVL72 racks up to $8.8M apiece (Tom's Hardware) - May 2026
for · primary NVIDIA Q1 FY27 data-center revenue $75.2B (92% of total) is led by sales of the Grace Blackwell NVL72 rack-scale system. NVIDIA Q1 FY27 — data center led by NVL72 (CNBC) - Apr 2026
for · observable Meta's Open Rack Wide double-wide rack-scale standard gained AMD/Wiwinn/Rittal designs — the industry standardizing at the rack, not the chip. Delivering an Open Data Center Ecosystem for AI (OCP) - Mar 2026
against · claim Much rack-scale hardware remains proprietary/locked to NVIDIA's NVLink domain — the rack is the unit, but ownership of it is concentrated, not commoditized. Open Compute still locked down? (Panel Builder) - Feb 2026
for · observable Liquid-cooling components total ~$49,860 per GB300 NVL72 rack, rising ~17% to ~$55,710 for NVL144 — rack-scale designs inseparable from direct liquid cooling. Cooling for a single Blackwell Ultra NVL72 rack ~$50k (Tom's Hardware) - Feb 2026
against · observable Cerebras completed its IPO in early 2026 after restructuring, positioning its wafer-scale chip for inference workloads, while OpenAI contracted Cerebras to provide 750 MW of inference compute through 2028 and designed custom 'Titan' chips with Broadcom. Cerebras Wafer-Scale Inference and Hyperscaler Custom Silicon Trend - Jan 2026
for · primary NVIDIA Rubin platform, a rack-scale AI architecture built around six tightly co-designed chips (Vera CPU, Rubin GPU, ConnectX-9 SuperNIC, BlueField-4 DPU, NVLink 6, Spectrum-6 Ethernet) entered production with partner systems expected in H2 2026. Rubin Platform Rack-Scale AI Architecture Unveiled at CES 2026 - Oct 2025
for · observable AMD's Instinct MI455X 'Helios' rack-scale system (up to 72 accelerators, H2 2026) is built on Meta's ORW standard — NVIDIA's main competitor also competes at the rack level. AMD+Meta Helios AI rack on Open Rack Wide (OCP Summit 2025)
What would move itthe measures watched
- watchRack-scale system pricing / per-rack ASP (GB300 NVL72, Vera Rubin, AMD Helios). Rising per-rack prices and vendors quoting/selling at the rack level confirm the rack is the unit of sale
- watchRack-scale vs 8-GPU-node (HGX) revenue mix at NVIDIA. If NVL72 rack-scale becomes the dominant share of DC revenue (vs HGX baseboards), the unit has shifted to the rack
- watchLiquid-cooling attach rate / cost per rack. Rising liquid-cooling attach is a proxy for rack-scale displacing air-cooled node design
- watchOCP Open Rack Wide adoption. Multiple vendors + hyperscalers converging on a wide rack-scale form factor confirms designing at the rack level