I didn’t post a response to the article written a little more than a year ago claiming that OCP had “blown up the server market“. Yes, that was really in the title. I’ll ignore most of the obvious issues with this, but lets review a year later, shall we?
Open hardware designs are great in concept. Share your design with the world, and lower your customers costs … er … whoops.
They are great, if you are a large scale consumer, and you can get what you need in that form factor. They are not so great if a portion of your value offering is in the hardware side of the equation.
I mention this as we were at OCP in San Jose last week, meeting customers. Not presenting, not with a booth. Just meeting people and talking through their issues.
What emerged from these discussions is that there is strong interest in value. We discussed our offerings with customers after they defined their need. We asked them point blank if OCP designs were required, or even needed. The answer, in general, was an unambiguous “no”.
The people populating the OCP market providing servers and storage are not, in general, the ones designing the systems. They are assemblers of gear, fairly low end, high volume folks, whom can build vast quantities of cheap gear. The designs are being done by Facebook, and they solve, not so curiously, Facebook’s engineering challenges.
Which, as we discovered in asking the question of the utility of the designs for our customers, not so useful outside of Facebook’s domain.
Indeed, on the face of this, it appears to me to be the case that OCP is an attempt to lower the cost of Facebook’s engineering (and not just internal IT costs to service its own designs). The more people that use this system, the lower the cost will be for Facebook to acquire them. Economies of scale and all that. They have created what they needed, and now they need more people to buy them.
The problem is that not every business is a hyperscale web data center. And many of the problems faced by our customers in particular, data motion, high performance computing and networking with storage and applications, aren’t well served by existing designs.
We asked our customers if they wanted us to work on OCP based designs. We pointed out that fundamentally, we’d need help on the engineering costs, as once the spec was published, our “investment” was effectively un-recoupable. That is, it is actually dangerous for a small shop to invest in building value into OCP systems, as they’d be unlikely to ever recoup that investment.
This is completely at odds with the open source software market, where releasing open source software (which we do) is a great way to get hired for doing something. You have the expertise and knowledge, and you can make it work well. There is an economy of scarcity, which is how normal capitalist economies function.
So what you get instead of innovators, are a) design for hire groups, b) manufacture for hire groups. The folks in the first group need to at minimum recoup their investment in the design, and turn at least a small profit for this to make sense to them, before they turn their design over to the rest of the world + dog to manufacture without their involvement, and with no revenue being paid to them for their expertise.
This dis-aggregates the supplier value chain; it breaks business models. This isn’t such a bad thing, as some business models are broken (most VARs don’t add any V). But for the folks with real value to offer, it strongly disincentivizes them from being in this market.
Completely opposite of blowing up the market, this drains the market of value, and pushes the smart people onto other things. Its the invisible hand of the market, acting as it should under pressure.
I don’t see this as disruptive now, one year after that article was written. OCP designs represent a miniscule fraction of servers and storage sold. Most of them supplied by Quanta. This report details the challenges ahead. Indeed, this is what Moor says as to their risks
No independent architecture R&D, minimal IP: Quanta is playing the role Dell
played for Intel in the 1990s: great at selling and fulfilling products largely based
on Intel R&D. Their current ?better together? solution story is solely based on Intel
infrastructure and does not yet comprehend emerging, non-Intel innovations from
other technology providers. Also, Quanta does not have a significant amount of
unique in-house IP which could limit their stickiness and value versus the larger
Basically they are a manufacturer for hire based upon designs from Intel and Facebook. They sell to hyperscale web folks like Facebook. But growing beyond this is going to be very hard for them.
We did spend time with other vendors, meeting some of our counterparts. We asked them their opinion on OCP. Many of them share the same view, that this is a cost minimization strategy on the part of Facebook, and less about real value. Its less about disruption and more about control of costs, which makes perfect sense for Facebook and other hyperscale folks to do.
In fact, pretty much all cloud vendors of appropriate scale need to be designing and building their own servers/storage when they cross various magic thresholds. Best way to do this is acqui-hire a firm that does specialty systems now, knows how to build and scale computing. These are HPC firms, people who have lived and breathed all of these issues over the last 20+ years.
All of this gets back to that there are no silver bullets, and the article implied, to a degree, that OCP is a silver bullet. Many people are going to spend a great deal of money to learn this lesson. One that many others had learned in previous generations.
This also noted, I like some of the innovations I did see, all driven by Facebook, that their contract manufacturers are pushing out now as OCP designs. Some of them perfectly mimic stuff we had done in 2006 or so, and couldn’t get any investors interested in then (lowering the cost of hardware by making it simpler/better). The investors I spoke with in CA had a rather negative view of the whole OCP concept. Marc Andreessen was up on stage with Andy Bechtolsheim talking it up. But the reality remains that there aren’t business models available to deliver open hardware value that allow you to protect your design. Anyone, anywhere can replicate it. They have the spec. How will this lead to business success?
One way is as a small design house, kicking out specific point designs and relevant software. The path would be an acqui-hire, not an IPO. Sort of like the fab-less semiconductor houses of the preceding two decades. Fabs became expensive, and few players remain in that business. The IP in this case isn’t even open source. Now make all the IP open, and this gets extraordinarily hard to create a niche.
Which means that this will drive innovators to other areas where they can carve a niche out.
So, at the moment I have not drunk the koolaid. I do not see a path or business model for a small company with significant value to play in that (quite small) market. Indeed, the only players were really the “box stackers”.
And worse for the OCP organizers, they had to apologize every now and then to the hardware folks in the audience when a presenter made a comment about designing their own CPUs or disks or … Yeah, this is not the way to sell the market concept to the people whom will attempt to offer value, that they will be required to give away for free, and then compete with others to service.
A market is a bidirectional exchange of about equal value. Its not unidirectional. OCP is pretty much unidirectional. They need to make sure that participants in the market providing goods can at least recoup investment and some profits. Its not organized this way at the moment. Its early, and I am not writing it off.
But disruptive in a “creative destruction” sort of mode? Nope, not even close.