From Ryuta Hamamoto at TIMEWELL
This is Ryuta Hamamoto from TIMEWELL Corporation.
In 2019, NVIDIA paid $7 billion for Mellanox — an Israeli networking company and the pioneer of InfiniBand technology. The strategic logic was clear in retrospect: GPU performance alone wasn't sufficient for AI-scale computing. What mattered was the ability to connect hundreds, then thousands, of GPUs into a coherent system. Network performance was the binding constraint.
In 2026, that $7 billion looks prescient. NVIDIA's networking business generates $8.2 billion in annual revenue — up 162% year over year.
NVIDIA Mellanox: 2026 Snapshot
| Item | Detail |
|---|---|
| Acquisition price | $7 billion (2019) |
| 2026 networking revenue | $8.2 billion (+162% YoY) |
| NVLink Gen 5 bandwidth | 1.8 TB/s bidirectional per GPU |
| NVLink Gen 5 scale | 72 GPUs fully interconnected |
| Spectrum-X | $10 billion annual revenue run rate |
| Blackwell NVL72 | SemiAnalysis AI inference benchmark leader |
| Bandwidth roadmap | 800 Gbps → 1.6 Tbps (2026), 3.2 Tbps (2030) |
| CEO assessment | "Lead in scale-up networking will continue for years" |
Looking for AI training and consulting?
Learn about WARP training programs and consulting services in our materials.
The Acquisition Logic: Why Networking Was the Missing Piece
The bottleneck that silicon improvements couldn't fix
Before the Mellanox acquisition, NVIDIA's GPU performance improvements were hitting a structural limit: even faster individual GPUs couldn't deliver proportional gains at scale because the communication between compute nodes was the bottleneck. Moving to larger models — the direction AI was clearly heading — required GPUs to work together seamlessly, not operate in parallel isolation.
Mellanox co-founder Michael Kagan became NVIDIA's CTO following the acquisition, and articulated the new design philosophy: "Design the entire datacenter as a single system."
What Mellanox brought:
- InfiniBand technology leadership
- RDMA (Remote Direct Memory Access) expertise
- Deep engineering capability in low-latency, high-bandwidth interconnects
- A technology foundation for connecting GPU clusters across racks and datacenters
NVIDIA's Networking Technology Stack
NVLink Gen 5: GPU-to-GPU Inside the Rack
NVLink is NVIDIA's proprietary high-speed interconnect for GPU-to-GPU communication within a rack.
Gen 5 specifications:
- 14x the bandwidth of PCIe Gen 5
- 1.8 TB/s bidirectional bandwidth per GPU
- Up to 18 NVLink connections at 100 GB/s each
- 72 GPUs fully interconnected within a single rack
This is what makes the NVL72 possible: 72 GPUs operating as a single virtual processor, sharing compute and memory resources without bottleneck.
InfiniBand: Datacenter-Scale AI Fabric
InfiniBand connects racks into datacenter-scale AI systems — turning individual rack-scale NVL72 units into coordinated clusters.
Technical characteristics:
- "Designed from scratch for synchronous high-performance computing"
- RDMA: bypasses CPU for direct memory access between nodes, eliminating CPU jitter
- Adaptive routing: dynamically selects optimal data paths
- Congestion control: prevents throughput degradation under heavy load
Competitive position:
Jensen Huang has stated directly: "There is no comparable low-latency, high-bandwidth connectivity solution available from competitors — not from AMD, not from the cloud providers' custom chips." InfiniBand is the standard for AI high-performance computing, and no equivalent exists.
Spectrum-X: Ethernet at AI Scale
NVIDIA entered the Ethernet market in 2024 with Spectrum-X, targeting organizations that prefer Ethernet infrastructure over InfiniBand.
Growth trajectory:
- $10 billion annual revenue run rate achieved
- Spectrum-XGS announced: connects multiple gigascale AI factories
Spectrum-X addresses the market that won't adopt InfiniBand — either due to existing Ethernet infrastructure investments or preference for open standards. NVIDIA now covers both segments.
Blackwell NVL72: Rack-Scale AI Computing
Architecture:
| Component | Detail |
|---|---|
| Configuration | 18 Compute Trays + 9 Switch Trays |
| GPU count | 72 Blackwell chips, fully interconnected |
| Form factor | Single cabinet |
| Interconnect | NVLink + NVLink Switch + rack-scale network fabric |
Performance:
SemiAnalysis AI inference benchmarks ranked NVL72 first in its class — meaningfully ahead of competing AMD systems.
Roadmap: Bandwidth Through 2030
| Timeline | Technology | Bandwidth |
|---|---|---|
| 2026 | InfiniBand / Spectrum-X Gen 2 | 800 Gbps |
| 2026 | Next-gen InfiniBand | 1.6 Tbps |
| 2030 | Future generation | 3.2 Tbps |
Jensen Huang's view on competitive positioning: "We achieved a scale-up network that will take a long time for anyone to catch up to. This lead will continue for years."
The Million-GPU Problem
As AI clusters scale toward millions of GPUs, networking challenges compound:
Latency distribution:
- A 1-GPU task that takes 1 second takes 1 millisecond across 1,000 GPUs
- At this scale, high-speed communication between every GPU with tight latency variance is non-negotiable
Reliability at scale:
- Systems with hundreds of thousands to millions of components will always have some components failing
- At 99.999% uptime, a million-component system has ~10 failures per day
- Software-layer failure detection and transparent routing around failed components is required infrastructure
DPU (Data Processing Unit): The DPU offloads OS processing and infrastructure management functions from the CPU, freeing CPU resources for application computation. At datacenter scale, this efficiency gain is significant.
Then vs. Now: The Mellanox Transformation
| Item | 2019 (Acquisition) | 2026 |
|---|---|---|
| Acquisition cost | $7 billion | Networking revenue: $8.2B/year |
| NVLink | Early generations | Gen 5, 1.8 TB/s, 72-GPU clusters |
| InfiniBand | Niche HPC market | AI HPC standard |
| Spectrum-X | Didn't exist | $10B annual revenue run rate |
| Roadmap | Unclear | 800G → 1.6T → 3.2T Tbps |
| Competitive situation | Multiple options available | "Years of lead" |
| Datacenter scale | Megawatt-class | Gigawatt-class planning |
| AI workloads | Experimental | Billions of simultaneous users |
Competitive Advantages and Enterprise Considerations
What makes NVIDIA's networking position durable:
- No competitor has equivalent low-latency, high-bandwidth interconnect technology
- CUDA API scales seamlessly from 1 GPU to multi-GPU clusters — the same codebase
- Vertical integration: GPU + NVLink + InfiniBand + Spectrum-X + DPU
- Software ecosystem depth built over years of CUDA deployment
Adoption considerations:
| Factor | Detail |
|---|---|
| Cost | Premium pricing at every layer of the stack |
| Energy | Gigawatt-scale datacenter operations require significant power infrastructure |
| Vendor dependency | Deep integration with NVIDIA creates switching costs |
| Operational complexity | Large-scale systems require specialized management capabilities |
Summary
The $7 billion Mellanox acquisition in 2019 was a strategic inflection point. NVIDIA recognized that GPU performance wasn't the binding constraint for AI-scale computing — interconnect was. The company bought the solution.
Seven years later, networking is an $8.2 billion business growing at 162% annually. NVLink Gen 5 connects 72 GPUs at 1.8 TB/s per chip. InfiniBand is the AI HPC standard. Spectrum-X has captured $10 billion in Ethernet revenue.
Jensen Huang's claim of "years of lead" in scale-up networking is not marketing language. There is no competing system with equivalent capability, and the Mellanox engineering depth accumulated over decades doesn't replicate quickly.
For enterprises planning AI infrastructure, the network layer is no longer an afterthought. It is, increasingly, the capability that determines whether large-scale AI deployments work.
