Main content

Pure Storage seeks to calm AI performance and capacity anxieties with Evergreen//One for AI

Derek du Preez Profile picture for user ddpreez June 19, 2024
With buyers figuring out how they want to use new AI technologies, Pure Storage is hoping it can alleviate storage stress with Evergreen//One for AI

Cloud computing technology and online data storage for global data sharing © NanoStockk -
( © NanoStockk -

Enterprises around the world are assessing the proliferation of Artificial Intelligence technologies (as well as navigating their hype), but are oftentimes unsure about which use cases to pursue. There is pressure from boards to not miss the boat when it comes to AI and business leads are scrambling to figure out where the value lies - whilst technology teams are left with the job of ‘just making it work’. 

However, as has been documented extensively, training AI models is an expensive endeavour. Whilst not all companies will seek to build their own models, many will train their data using inference or RAG technologies, which still comes at a high price. The cost of GPUs - as is evidenced by NVIDIA’s share price - is noteworthy, for instance. However, these GPUs also require high performance infrastructure, which when coupled with capacity ‘unknowns’ due to uncertain use cases, is leaving buyers with an unpredictable challenge on their hands. 

Imagine having to write a RFP to build out infrastructure for a variety of unknown use cases, over an unknown period of time, with uncertain usage, and unpredictable results. A potential recipe for a costly disaster. 

Knowing this, Pure Storage is seeking to gain traction with AI enterprise buyers by alleviating much of the worry of the ‘unknowns’ with its new storage-as-a-service offering, Evergreen//One for AI. 

But first, by way of background, Evergreen//One (the product umbrella, which now includes Evergreen//One for AI) has been on the market for a number of years and essentially aims to make storage ownership as compelling as a cloud operating model - with Pure going as far as paying users for their racks and power use. 

The idea behind the original Evergreen//One is that buyers pay for what they use, get seamless upgrades, have a wide variety of SLAs, as well as get the ability to move workloads easily between hybrid cloud/multi-cloud environments. Pure believes that whilst compute and networking -as-a-service offerings have become mainstream, storage had been left behind. 

Coupling this with the trend towards data intensive workloads running on cloud native environments, with Kubernetes adoption proving popular, Pure Storage’s acquisition of Portworx has allowed the vendor to be able to provide buyers with a persistent, scalable storage option. This has also resulted in a pull towards a cloud-like operating model for storage.

Essentially, data is not just sitting on some storage kit somewhere, application developers need it to be usable. As Stantiago Navonne, Director of Product Management for Evergreen//One told diginomica at the vendor’s user event in Las Vegas this week: 

Storage for the longest time has been a cost, a service, that the application owner doesn't particularly care about. Despite the applications becoming cloud first, container first, and more modern, the way that storage was provisioned for that application has not kept up with that. 

But the value of that type of deployment is becoming more and more common. Containerizing applications is becoming more and more common. So there’s a pull. 

With Evergreen//One, what we have tried to achieve is the benefits of that operating model, while still maintaining the control, and other advantages that you get - like performance - by running infrastructure on prem. 

Navonne said that Evergreen//One isn’t about thinking of -as-a-service on premise as a new type of purchasing model, but rather as a new operating model that allows buyers to subscribe to something completely different. For example, Evergreen//One includes a number of SLAs, such as cyber recovery and resilience. Navonne said: 

I'll give you one concrete example of a customer very recently that purchased Evergreen//One. It was a state government in the United States. They have been purchasing with Pure Storage for years, but the lightbulb moment for them to switch to Evergreen//One was the broader and richer cyber recovery and resilience SLAs. 

It caught the attention of somebody at that organization whose challenge was that in order to guarantee a recovery time from any cyber event, for every piece of appliance they bought, they would buy another copy of that appliance to have on standby - so that they would be ready to recover if they needed to. 

You can imagine how expensive that is. And what they realized there is that with the cyber recovery guarantee, by subscribing to that, they didn't need to do that anymore.

And then from there, they turned around and thought, ‘how do I think about everything else about this infrastructure in terms of what makes sense for our consumption model?’ For example, they realized that purchasing capacity for three years from now, with a buffer for an unplanned spike, could just be a subscription (to Evergreen//One). 

A new AI use case

So, this week Pure Storage has now announced Evergreen//One for AI - what it calls the ‘first purpose-built AI Storage-as-Service’ - to extend the operating model principles of Evergreen//One to new enterprise AI workloads. And this is being driven by the uncertainty facing enterprise buyers: 

You can imagine there's a lot of noise and a lot of requests around deploying infrastructure for AI. What we've seen is really a level of unpredictability and unknown that I haven't seen in storage before. 

Navonne explained that, for traditional workloads, buyers are typically able to come up with some sort of realistic scenario for their storage requirements - looking at the average capacity and performance needs over time and allowing for a bit of headroom. This is no longer true of AI: 

We basically see a number of customers that say, ‘I know I need some GPUs, I have no idea what I'm going to be doing with them, because there are seven different workflows that are being developed in that industry’. 

This quickly results in them not knowing how much storage they’re going to need. For example, if an enterprise is training on text, which is relatively light, versus inferring on images, which is much more data heavy, the needs and requirements are going to be vastly different. Also, with the speed at which AI is advancing, who knows what will be required in two, three or five years time? Navonne said: 

When you ask them the question of: how much capacity do you need? They say, ‘I don't know’. Or even worse, they make a number up.

As such, Pure found itself in a position that it was trying to solve a problem for customers where there were too many unknown variables in the mix. This was coupled with the fact that GPUs, typically a requirement for training for AI, are expensive and have unique characteristics as it relates to infrastructure needs. For example, a GPU vendor often has a recommendation of storage performance - but that’s a difficult buying decision when coupled with unknown capacity. As Navonne explained: 

That's the unpredictability and flexibility aspect we’re trying to address. If you look at the performance requirements of AI, of these GPU clusters, they are extremely expensive. If you think about the most valuable asset in this deployment, it’s the GPU. And you want to keep it 100% utilized. 

To keep it 100% utilized, you have to minimize the amount of time you wait on storage. That's where an extremely high performance requirement comes from the vendors’ perspective.

So extremely high performance requirements, on extremely small capacity requirements. Storage is typically not good at that. If you think about how scale out systems work, you increase performance by growing the number of nodes, growing the amount of capacity with that. 

What Evergreen//One for AI does is tell the customer ‘we're going to separate these two metrics for you’. Not only because one of them is known and the other one is unpredictable, but also because one of them is extremely high and the other one is extremely low.

Charges based on performance

Essentially what Pure Storage is allowing buyers to do with Evergreen//One for AI is get access to scalable storage, where you only pay for what you use, but is actually priced on performance: 

Let's not even try to have a conversation about how much data you're going to store. Instead, we're going to commit to you that every marginal terabyte you store on this, it is going to be extremely cost effective. We're going to agree on the rate for those terabytes and you know that at no point will this grow to cost you more than what you could have otherwise done. 

So we don't want you to have to pay for the capacity, unless you use it - but we have to deploy to hit your performance target. So we're only going to charge you for the performance aspect on a monthly basis. And then, even though we may be deploying 10 petabytes of storage to hit your performance numbers, you're not going to pay for those 10 petabytes unless you actually do use them. That's our risk and our next cost.

Equally compelling is that Evergreen//One for AI is like any other Evergreen//One contract, in that buyers are only signing up for a minimum of one year. Navonne said that if you, as a buyer, decide AI isn’t what you need anymore, this will save you from a scenario where you have bought an amount of infrastructure that you then have to depreciate over five years. 

The core of what Navonne is describing is that organizations, more likely than not, don’t know what they’re buying for when it comes to AI. Buying an oversized solution, which may need upgrading in a year or two anyway, feels like a risky purchasing decision. Pure Storage is hoping to take away these concerns on the storage front with Evergreen//One for AI: 

The goal is for you to not have to think about storage. What I'm trying to put together with Evergreen//One for AI is a solution for the customer, as they're deploying this infrastructure - and they're probably scrambling to do so - where they do not have to think about the storage. 

Just show them that this model will deliver whatever you want it to deliver, in a cost effective manner, so that you don't have to think about it today. Just let it work and change it as you move forward.

My take

As is evidenced by Pure Storage’s recent earnings, the demand for the vendor’s as-a-service offering is strong and is driving revenue growth for the company. What rings true about the new Evergreen//One for AI pitch from Pure is that enterprises are trying to plan for a number of unknown scenarios that are potentially expensive and require heavy investment - attempting to do this up front does seem to be a fool's game. Another positive is that many organizations will want to do their AI training in a private cloud operating model scenario, given how much company data could be involved (alongside privacy concerns). It will be interesting to see which deployment models win out and it won’t be a universal approach. And it’s worth noting that Evergreen//One for AI has just been announced and we don’t have the customer proof points available yet. That being said, I think for enterprises feeling uneasy about their AI storage needs, this will read as a very compelling option.

A grey colored placeholder image