Study: Nvidia Mellanox Acquisition
Introduction
Case studies of wildly successful acquisitions are fun to think about. The definition of a great capital allocator doesn’t just stop at funds that buy and hold something that returned 100x.
Particularly in tech investments where the nature of business is fast-moving and disruptive. There are a few main reasons why high-tech companies make acquisitions:
To deter competitors.
When a technology is clearly going to become a standard.
To quickly gain access to technology that would otherwise be time consuming to develop.
The NVDA-Mellanox acquisition displays all 3 factors and makes for a great case study.
Short History of NVDA
NVDA was founded by Jensen Huang, Chris Malachowsky, and Curtis Priem in 1993. The story took place in Denny’s restaurant where they agreed to start a high performance computing company.
There was no lack of talent amongst these 3 men. Jensen, having most recently worked at LSI Logic. In fact, when it became clear that he was going to quit LSI Logic, Jensen was so well regarded that the founder of LSI personally recommended Jensen to Don Valentine of Sequoia Capital.
Jensen apparently gave a poor presentation to Don Valentine, but due to his good reputation, Sequoia Capital decided to fund the NVDA startup.
Jensen is a brilliant thinker when it comes to business strategies, one of his favourite is to “focus on zero billion dollars markets.” Because one of the things you can definitely guarantee is where there are no customers, there are also no competitors!
NVDA would see opportunities and enter into markets ahead of competitors. For example, in the 1990s when the gaming industry was not respected by the investing community, NVDA went into developing graphic cards for games.
The gaming industry began to birth a use-case for graphics chips, demanding huge performance and technology breakthroughs to get better gameplay and visualisation. NVDA catered to this market, and in the process, built itself a competitive advantage.
Then, in 2006, a breakthrough came when Jensen had a meeting with a Stanford professor Andrew Ng, who explained how he was using GPUs to process large data sets for machine learning. He discovered that what previously took him weeks using CPUs, could be done in days on GPUs.
Jensen realized that GPUs had to be opened up to be programmable by external developers. This gave rise to CUDA (Compute Unified Device Architecture) in 2007, and NVDA fifth chip GeForce 256 was the first to be programmable.
As the saying goes, “The rest was history.” NVDA would enter the machine learning space when there were virtually no customers, and now enjoys the near-monopoly position in AI.
Solution to Moore’s Law
To advance Moore’s Law, everybody is familiar with the increasing difficulty to squeeze more transitors into a finite chip. Jensen had an insight that although it is harder to get performance at chip-level, there was still plenty of compute power to extract from a datacenter level.
It stands to reason that compute scaling must come from efficient connectors between chips.
And this brings us to the Mellanox acquisition.
History of Mellanox
Mellanox was founded in Israel in 1999 with a focus on high-performance networking. Their initial technology focus was called “Infiniband”. The company IPO in 2007, with a market cap of $500m. Around this time, they started to work on another technology called “Ethernet”, which more people are familiar with, and a greater number of competitors played.
Eventually, Mellanox products became a key component to datacenters, high-performance computers (HPC), and more recently, AI acceleration.
NVDA was a big customer of Mellanox’s Infiniband, and so after years of collaboration, NVDA decided to make an acquisition bid.
Why Mellanox?
There are mainly 3 reasons:
Vertical integration of hardware systems.
Competition deterence.
Ethernet’s lower performance versus Infiniband.
Vertical Integration
When operating at the datacenter level, making efficiency gains is no longer at single chip level, but it is across hundreds of chips.
At this scale, it was beneficial for NVDA to take in valuable R&D resources in order to realize this way of compute.
Vertical integration was deemed to deliver true synergies.
Deter Competitors
Mellanox was the dominant supplier of Infiniband and probably invested the most R&D dollars in the industry. If one of NVDA competitors were to buy Mellanox, they would have to pay a toll to access Infiniband, or have it bundled with other technologies that NVDA would be forced to purchase. This is obviously not a good outcome.
Marvell, Xilinx and Microsoft have reportedly shown interest in Mellanox.
Intel went on a bidding war with NVDA and gave up at a final bid of $6b, losing to NVDA’s $6.9b offer.
Ethernet Performance
Ethernet had a much broader ecosystem than Infiniband with many competitors and suppliers. No single company owned the technology. Therefore, there was no supplier risk.
However, in 2019, new variants of Infiniband were simply much better than Ethernet cables in terms of latency.
This latency is unbearable as compute workloads scaled exponentially. So the acquisition made a lot of sense if NVDA could secure this tech for itself.
Result of Acquisition
After completing the $6.9b deal in April 2020, just 2 years later ChatGPT was launched, and demand for NVDA chips increased quickly.
NVDA would start reporting their “Compute & Networking” segment where Mellanox was in. Prior to 2019, there was no such segment, and as time progressed, other business lines were included such as DGX cloud, autonomous cars solutions, robotics.
This segment in fiscal year 2021 reported revenues of $6.8b with $2.5b operating profits. Very impressive quantum and margins!
It wouldn’t take long for this investment to payback itself.
Alternatives
With the benefit of hindsight, what if NVDA spent this $6.9b repurchasing shares instead?
Could NVDA done well without Mellanox even if owned by a competitor?
Maybe it would have faced higher prices for Infiniband and had less integration with the technology. But NVDA overall advantage in AI was so great that we think they would have thrived regardless.
On the other hand, it’s clear that acquiring Mellanox allowed NVDA to move faster with performance increases and AI compute scaling.
At worst, Mellanox can be thought of as a hedge against losing exclusive access to Infiniband tech. This reason alone could justify the decision.
In the tech world, M&A is different from buying a railroad. Technologies have a shorter lifespan. NVDA knows this and has developed Spectrum-X networking solution in Ethernet, specifically designed for AI workloads.
Conclusion
It is interesting to think about capital allocation in high tech. As described, this acquisition looked more defensive than offensive. Keeping a technology open rather than letting a competitor determine the standards on how this tech will be used.
It also opens up the discussion of how much capital NVDA needs to reinvest back into the business to keep margins high.
NVDA produces a lot of cash now, and probably will continue as long as AI spending continues.
When looking at post-mortem of capital allocation, the industry makes a difference. The tech world is unlike Costco/Walmart where capital spent on opening new stores will yield almost guaranteed known returns.
With high-tech firms, moving fast is a must and they must face a shorter lifespan of investment.
You want to position yourself near the tree. Even if you don’t catch the apple before it hits the ground, so long as you’re the first one to pick it up.
Jensen Huang