When Nvidia CEO Jensen Huang announced in March 2019 that the company would spend $6.9 billion acquiring Israeli networking chip maker Mellanox Technologies, Wall Street was baffled. Nvidia was a graphics chip company riding out a bruising post-cryptocurrency hangover, with GPU demand in freefall and analysts already penciling in forecast cuts. Paying a king's ransom for a maker of InfiniBand networking equipment seemed, at best, an expensive distraction.
Few people saw what Huang was actually buying.
Seven years later, that acquisition has been recast as one of the shrewdest deals in semiconductor history — and a key reason Nvidia now sits at the center of the global AI infrastructure boom.
What Did Jensen Huang See in Mellanox That Wall Street Completely Missed?
To understand why the deal looked so strange in 2019, it helps to recall how the industry thought about AI at the time. Machine learning was growing, but the dominant assumption was straightforward: more AI power meant more GPUs. Networking — the cables, switches, and interface cards connecting servers — was considered plumbing, not strategy.
Mellanox made high-speed plumbing. Its InfiniBand interconnects and Ethernet adapters were used in supercomputing centers, but the technology had little obvious connection to Nvidia's gaming and graphics business. Analysts worried the acquisition would stretch the company thin at precisely the wrong moment.
Huang disagreed — publicly and emphatically. At the time of the announcement, he argued that future data centers would not function like collections of individual servers but like single, unified computing systems, with computation flowing across the entire network fabric. The bottleneck of the AI era, he predicted, would not be raw processing power. It would be the speed and efficiency with which processors could talk to each other.
That prediction proved correct in ways even his supporters may not have anticipated.
Why Does a Slow Network Make Even the World's Most Powerful GPUs Useless?
Training today's frontier AI models requires thousands — sometimes tens of thousands — of GPUs working in parallel. Those processors must continuously synchronize, passing gradients and parameters back and forth in massive collective communication operations. When that process runs smoothly, a cluster might achieve 90 percent utilization. When the network lags, that figure can collapse to 20 or 30 percent. Industry estimates suggest that in large H100 clusters, as much as 15 to 30 percent of total training time is spent simply waiting for data to move across the network.
Mellanox's InfiniBand was engineered precisely for this environment: ultra-low latency, lossless data transmission, and extreme bandwidth. Paired with Nvidia's own NVLink chip-to-chip interconnect and its Spectrum-X Ethernet platform, the combined portfolio gave Nvidia something no other GPU maker could offer — a complete, vertically integrated system spanning compute, communication, and the software stack tying them together via CUDA and NCCL.
The result is what Huang calls the "AI factory": not a server rack, but an end-to-end computing environment operating as a single coherent machine. Without high-speed networking, even the most powerful GPU is, in Huang's framing, an island. Mellanox supplied the bridges.
The Numbers Tell the Story
When the acquisition closed in 2020, Mellanox was generating roughly $1.3 billion in annual revenue. By Nvidia's fiscal year 2026, the company's Networking division had surpassed $31 billion — a more than twentyfold increase in six years. A single quarter's networking revenue peaked above $11 billion, representing year-on-year growth of 263 percent. Nvidia CFO Colette Kress described networking as "core to our data center scale infrastructure" on a recent earnings call, citing what she called "standout performance" in the segment.
The division has, by some measures, made Nvidia the largest networking company in the world — a title the company didn't hold, and few would have predicted it seeking, when Huang signed the Mellanox deal.
At the time, Nvidia was still primarily perceived as a graphics chip company serving the gaming market. It had just weathered the collapse in cryptocurrency-driven GPU demand, which had led to multiple downward revisions to its financial forecasts. Analysts questioned the strategic rationale: what was the point of paying such a premium for a maker of InfiniBand networking chips? Few observers grasped the long-term implications of the move.
By 2026, the answer is unambiguous. The acquisition stands as one of the most consequential in Nvidia's history. Mellanox's networking technology — once a sideshow — has become a central engine of what Nvidia describes as the "AI factory." In fiscal year 2026 alone, the Networking division exceeded $31 billion in revenue, reaching $11 billion in a single quarter, representing 263 percent year-on-year growth in Q4. Nvidia has, in effect, become the world's largest networking company by this metric.
A Shift Reshaping the Industry
The broader significance of Nvidia's move is now becoming visible in competitors' strategies. At Google Cloud Next in April 2026, the company unveiled its eighth-generation TPU with a notable structural change: for the first time, training and inference functions were split across two dedicated chips — the TPU8t, codenamed Sunfish, designed by Broadcom for training workloads, and the TPU8i, codenamed Zebrafish, designed by MediaTek for inference. The announcement underscored a wider industry shift: as AI models grow from billions to trillions of parameters, the question of how compute is connected and customized matters as much as the compute itself.
Nvidia recognized that shift seven years before it became consensus. The Mellanox acquisition was, in retrospect, less a bet on networking as a product category than a bet on a particular theory of how AI infrastructure would evolve — and a move to control the layer that theory identified as decisive.
Huang has since stressed that "compute equals revenues," a formulation that collapses the distinction between processing and communication that once made the Mellanox deal seem puzzling. Nvidia is now investing in silicon photonics and optical interconnects, positioning for a generation of AI deployments that span multiple data centers at scales the industry is only beginning to imagine.
The skeptics of 2019 have gone quiet. The $6.9 billion that once looked like an overprice turned out to be the foundation of a platform — and a reminder that in technology, the most consequential bets are often the ones nobody else understands at the time.
Original Article in Chinese （Related： TSMC's 2nm Secrets Were Stolen From the Inside. A Court Just Handed Down Its Verdict. ｜ Latest ）
