Advanced Computing in the Age of AI | Wednesday, December 6, 2023

Contract Manufacturer Jabil Sells Straight Into Hyperscale Datacenters 

The hyperscale systems market is getting a new supplier, and this one is a big one with plenty of expertise in manufacturing complex servers and storage and integrating gear at the rack level. The startup is called StackVelocity, and it is a business unit of contract manufacturer Jabil Circuit.

You may not know Jabil by name, but many of your IT equipment suppliers do. The company, which is based on St Petersburg, Florida, is among the largest contract manufacturers in the world, a group that includes Foxconn, Flextronics, Quanta Computer, Celestica, Sanmina-SCI, Benchmark Electronics, and many others.

Jeff DiCorpo, who is general manager of the new StackVelocity unit, says that the enterprise and infrastructure manufacturing segments of Jabil account for between $6 billion to $7 billion in sales each year, with the company bringing in over $18 billion making additional products, including mobile handsets, industrial gear, and medical equipment as well as IT gear. System and storage makers usually don't talk about who makes their gear, but SGI recently announced that it was shifting manufacturing operations to Jabil, and Cisco Systems uses Jabil to make its Unified Computing System blade servers, just to name two that are relevant to the EnterpriseTech world.

Up until now, Jabil has made servers, storage, switches, and other gear on behalf of vendors who slapped their brands on them and sold them as their own. (Well, technically speaking, Jabil actually does the brand slapping, too.) Jabil does everything from board-level assembly to rack-level integration on behalf of IT supplier customers, including a modest amount of configure-to-order and parts supply chain purchasing and management. For several of its customers, Jabil does joint development. Thus far, says DiCorpo, this work has all been done on behalf of IT system suppliers.

With StackVelocity, Jabil is setting up a separate business unit that is aimed specifically at hyperscale datacenter operators and, importantly, at selling gear directly to them rather than through a middleman.

"We have seen a lot of change in the marketplace, and as we evaluated that, we considered multiple ways that we could serve that market, and we decided that one way that we will serve that market is as a large-scale system integrator," explains DiCorpo to EnterpriseTech. "This is principally focused around rack integration for hyperscale datacenters, and we will do our own as well as Open Compute solutions."

This move echoes those of many OCP Solution Providers – that is a brand that is bestowed by the Open Compute Project to vendors who have put their modified OCP designs and manufacturing processes through a certification process to ensure compliance. Quanta QCT, the hyperscale system division of Taiwanese manufacturer Quanta Computer, makes its own storage, servers, switches, and racks as well as supplying OCP gear, and thus far has been the dominant supplier to the top several dozen hyperscale datacenter operators. But there are other suppliers, including Hyve Solutions, AMAX, Penguin Computing, Racklive, CTC, and now StackVelocity.

As for the custom gear that StackVelocity will make, the scenario is much the way it works at Quanta or the Data Center Solutions division of Dell, which also has a big slice of the custom server and storage market for hyperscale customers. (Facebook, in fact, started out using custom servers from Dell for the first couple of generations of its systems, which ran in co-location datacenters, before striking out on its own to design systems, storage, and datacenters and founding the Open Compute Project three years ago this month.) DiCorpo says that StackVelocity will work with customers to define their requirements, and work with them at whatever level they need, from joint design through manufacturing.

The key thing for StackVelocity is that Jabil has a worldwide network of manufacturing locations and, importantly for some customers, it does manufacturing in several locations in the United States and Canada as well as throughout Europe and Asia. The issue is not nationalism – hyperscale customers have operations and applications that span the globe and think of themselves as multinational  – but rather the availability of relatively local manufacturing to meet the needs of nearby datacenters. The company with the broadest manufacturing base and deepest supply chain is going to win when it comes to hyperscale system making.

"We are leveraging a fairly extensive ecosystem of partners, so we are not tied to any particular technology or hardware choice. We will do whatever the customer needs. We will help them with designing and validating their rack solutions." The other value add that StackVelocity brings is a warranty on the gear, which is something all of the tier one system makers provide, as well as integration of the software. The other part of scale that StackVelocity can provide is financial. The company is perfectly willing for hyperscale customers to negotiate their own parts supply contracts, but for the next several hundred customers – many of them large enterprises that don't have the scale of Google, Amazon, Facebook, Yahoo, and others – it might be best to leverage the buying power of Jabil.

The thing to remember is that hyperscale is a hardware and software design philosophy and it is not just strictly about size and scale. That said, for the numbers to work out, hyperscale system manufacturers have to have a certain amount of volume for a deal to make sense. At Dell DCS, as EnterpriseTech has previously reported, a custom server engagement requires at least 5,000 nodes if the customer wants a design-to-order engagement and at least 1,000 nodes if Dell is building to an existing design. Many customers roll out new systems in lots of 10,000, all pre-racked and ready to have power and network plugged in. StackVelocity's definition is a little broader, and it is thinking that it can do engagements from anywhere from one to 1,000 racks, which works out to between 40 and 40,000 nodes. Obviously, a custom server engagement will require a larger commitment on the node count, just like at Dell DCS.

"We are looking for customers who are thinking in units of racks," explains DiCorpo. "Clearly, where we are going to be able to provide a lot of value is where the customer has a repeatable, homogeneous, large footprint where we can apply manufacturing processes. As a contract manufacturer, one of the things we can bring is scale, and there is going to be a natural fit as customers have more scale. The other aspect of this is that we are really aimed at customers who are looking to streamline their IT infrastructure and looking to take advantage of open source software and commodity hardware, but could use help in design, test, and validating and maybe even services and support for these products in their datacenters."

The company is also looking to cash in on the explosive growth in the hyperscale datacenter market, which DiCorpo says is a multi-billion dollar opportunity growing at 20 to 30 percent compounded annually. By some estimates, these hyperscale datacenters account for around 20 percent of server shipments today and could double that slice of the server pie in the next three to five years. (Quanta says it will only take three years, but IDC says it will be more like five.)

Right out of the chute, StackVelocity has cooked up an OpenStack private cloud reference design with OpenStack supplier Mirantis. This initial design was created specifically for pilot programs at enterprises who want to see what OpenStack can do, and it supports 450 virtual machine guests and is expandable to 1,100 VMs as the rack is fully populated with servers and storage.  The setup comes with one 10 Gb/sec Ethernet switch for data, eleven compute nodes using a pair of eight-core X86 processors with 24 GB of memory and five storage nodes based on a pair of six-core X86 processors that run the Ceph distributed object store and the Cinder volume manager layer that can be deployed on top of it. The stack is configured with Red Hat Enterprise Linux and the KVM hypervisor as well as the "Havana" release of OpenStack.

StackVelocity has also created a Hadoop reference design, which is based on work it did in conjunction with Texas Tech University's Cloud and Autonomic Computing Center, which is supported with funding from the National Science Foundation.

DiCorpo says that StackVelocity is in the early stages of other customer engagements now.

It is unclear what repercussions the establishment of StackVelocity will have on Jabil's manufacturing relationships with other server, storage, and switching peddlers. But in this day and age, and with these hyperscale customers in particular, price and speed are more important than any other factor and Jabil cannot ignore this market. No system and storage vendor can.