IBM Should Buy Mellanox Before HP Or Cisco Does

July 23, 2012 Timothy Prickett Morgan

Maybe IBM should have bought Mellanox Technologies as well as Blade Network Technologies back in the fall of 2010. Or perhaps just after Mellanox ate rival-partner Voltaire that November and positioned itself to be a player in both Ethernet and InfiniBand switches and adapter cards.

Intel is certainly hot to trot to own chip designers who make network switch and router chips as well as silicon and full adapters for servers that allow them to link to switches, and IBM would do well to get more serious about networking than just buying up Ethernet switch maker Blade Network, itself a spinout of now defunct switch maker Nortel Networks. That acquisition has been a good one for Big Blue and is the foundation of its Systems Networking business, a response to Cisco Systems getting into the server racket and Dell, Hewlett-Packard, Huawei Technologies, and Oracle (to a lesser extent) offering both servers and networking, and more importantly, converged systems that merge the two together. Buying BNT was an estimated $400 million well spent. But IBM should have done more, and it may have to if it wants to blunt a very serious attack by Intel on the system interconnect front that has implications for the attainment of exascale systems–those capable of processing 1,000 petaflops and holding 1,000 petabytes of data–by the end of the decade and for general-purpose systems in the near-term and long-term.

Intel bought Ethernet ASIC chip maker Fulcrum Microsystems back in July 2011, getting its hands on an Ethernet switch maker that has designed clockless, low-power, high-bandwidth chips for switches and routers. In January of this year, Intel paid $125 million to buy the InfiniBand switch and adapter business from QLogic, giving Mellanox, which pretty much owns the InifniBand market, a very big potential competitor down the line. Then in April, Intel paid $140 million to buy the interconnect business from supercomputer maker Cray, which retained rights to the current “Gemini” XE and future “Aries” interconnects while basically giving this market over to Intel. These three deals give Intel a large number of serious network propellerheads and established customers from which to make money selling raw chips.

It also gives Intel a chance to do in networking what it has come to do in processing: absolutely dominate (in terms of volumes and revenues) and therefore suck most of the profits out of the base for itself, leaving the actual vendors who use its silicon whatever money they can make on services and software (if they have it).

Mellanox is riding the next InfiniBand wave right now, with Fourteen Data Rate, or FDR, switches and adapters, which run at 56 Gb/sec, riding a wave in the supercomputing, hyperscale data center, database cluster, and clustered storage spaces concurrent with Intel’s launch of its Xeon E5 processors, which support PCI Express 3.0 peripheral slots and therefore have enough bandwidth for FDR InfiniBand to go full-tilt-boogie. And the company is also making headway in 10 Gigabit and 40 Gigabit Ethernet switches and adapters.

In the quarter ended in June, Mellanox had $133.5 million in sales, which were up 110 percent compared to the year ago period, with net income of $32.1 million, a whopping 15 times higher than in the year ago quarter. Almost all of that incremental growth came from FDR InfiniBand products, and slower InfiniBand products (QDR runs at 40 Gb/sec and DDR runs at 20 Gb/sec) pretty much held steady as customers either extended existing clusters or simply went with cheaper options. Mellanox also saw its Ethernet business grow by a factor of 2.5, but it only brought in $9.3 million.

The smallness of the Mellanox Ethernet switch and adapter business does not really matter. What does matter is that Mellanox has invented a switch ASIC called SwitchX and an adapter ASIC called ConnectX-3 that are able to speak either InfiniBand or Ethernet, and this is valuable technology that is going to make money and generate profits in our increasingly networked and bandwidth-hungry world. IBM might have to isolate Mellanox circuitry inside of its Microelectronics Division, treating Systems and Technology Group as just another client, to make a Mellanox deal work. But Big Blue did that with the Cell Power-based processor it used for some time in supercomputers as well as selling to game console and electronics makers.

Incidentally, IBM itself drove 19 percent of sales at Mellanox in the second quarter, with HP driving 30 percent of sales. Buying Mellanox might drive HP and Dell into Intel’s loving arms, therefore crimping potential revenues for a Mellanox that is not isolated somehow inside Big Blue. (Intel is probably praying someone buys Mellanox just for this very reason.)

Thanks to the killer Q2 numbers that Mellanox turned in, buying the Israeli chip, switch, and adapter maker would not come cheap. As The Four Hundred goes to press on Friday, Mellanox shares are up like a rocket by 35 percent, giving the company a market capitalization of $3.73 billion. Mellanox would not be a cheap acquisition at all, and even with only a slight premium above this–let’s call it $4.5 billion–that is nine times revenues.

But the question is this: Does IBM want to be a networking innovator, as it was decades ago, or does it want to be at the mercy of Intel or someone else?