Opinion: The golden age of storage hardware startups is over and storage software rules the world.
B&F is including hyper-converged infrastructure appliance vendors amongst the storage startups and we recognise that storage hardware startups always had a software component as well.
There are storage array startup technology eras, such as the all-flash array era, the deduplicating appliance area, the NVMe array era and the HCI era. Generally speaking, deduplicating arrays solved a backup disk capacity problem, all-flash arrays fixed the slow disk access problem, NVMe SSDs fixed a slow SSD issue, NVMe over Fabrics sorted out the slow storage networking problem and HCI was a response to SAN complexity difficulties.
Here is a rough, ready list of HCI and storage array startups recorded in our files since 1999:
- 2019 – 2021: None
- 2018: Nebulon
- 2017: None
- 2016: VAST Data
- 2015: Fungible
- 2014: E8, Excelero, Pavilion Data
- 2013: Apeiron, Stratoscale, Vexata
- 2012: Datrium, Qumulo, SoftIron, Symbolic IO
- 2011: Cloudian, Coho Data, Mangstor, StorONE
- 2010: Exablox, Infinidat, SolidFire, Tegile
- 2009: Gridstore, Maxta, Nutanix, Pure Storage, Scality, SimpliVity
- 2008: CTERA, Kaminario, Nasuni, Nimble, Tintri, Whiptail
- 2007: Scale Computing
- 2006: None
- 2005: Violin Memory
- 2004: None
- 2003: Pivot3
- 2002: Compellent, Exagrid, X-IO
- 2001: Data Domain, EqualLogic, Pillar Data
- 2000: None
- 1999: 3PAR
It’s immediately obvious that the seven years between 2007 and 2014 were the golden years. Since then, only VAST Data has been started up and even it has recently moved away from hardware; it knows which way the wind is blowing.
What big storage issues are left that can be fixed with better array system hardware?
The single largest problem is data movement, making it quicker or avoiding it altogether. There are clever software tricks to make it quicker, such as Nvidia’s GPUDirect and its CPU/storage IO software stack bypass technology. There is also a hardware dimension to this with compute on storage drives from NGD and Pliops – but these are examples of drive-level technology not system or array-level.
Another hardware angle to the data movement problem is to keep the majority of data in memory and reduce storage IO that way. It appears to need the CXL bus and and storage-class memory. These are technologies that will be adopted by the mainstream server vendors in two to three years when they mature, and probably won’t provide scope for HW/SW startups.
How things have moved on
In general the big incumbents – Dell, HPE, and NetApp – have become faster at adopting new technology; that is why, in my view, no NVMe-over Fabrics startup has made it to an IPO event. They have also adopted Optane quite quickly and that has closed off opportunities for most storage startups using Optane, except ones that twinned it with clever software, such as VAST Data.
Another factor lowering interest in hardware-focused storage startups is the public cloud, which can throw enormous capacity and compute scale at storage-related problems.
There’s plenty of scope for storage innovation though. New resistive RAM technologies are being developed and penta-level cell flash may have promise. Software abstraction layers are needed in the hybrid cloud world to unite the on-premises and hybrid environments. Kubernetes storage is a hotbed of development. The data management area has enormous scope as does the whole analytics and AI field.
But for now it appears that a hardware-based approach to storage system-level problems has come to an end. So, alas, a golden age of storage HW/SW startups is coming to a close.