Evergreen//One for AI: Modern Storage Economics for the AI Era

The industry’s most talked-about AI partnership just validated what forward-thinking organizations already know: the era of massive upfront infrastructure investments is over.

Last month, NVIDIA and OpenAI announced a groundbreaking $100 billion partnership structured around leasing rather than purchasing GPU hardware. Instead of OpenAI writing enormous checks upfront for chips that could become obsolete within years, they’re spreading costs over time while transferring depreciation risk back to NVIDIA. The result? An estimated 10-15% reduction in total hardware costs and the agility to scale with AI’s unpredictable demands.

This isn’t just smart financial engineering—it’s a fundamental shift in how AI-driven organizations approach infrastructure procurement. The traditional “buy everything upfront and hope it lasts five years” model simply doesn’t align with AI workloads that can evolve dramatically in a matter of weeks, not years.

While the headlines focus on GPU leasing, the same economic pressures and solutions apply across the entire AI infrastructure stack—especially storage. Just as OpenAI recognized that owning depreciating GPU assets made little financial sense, organizations worldwide are discovering that traditional storage procurement models create similar inefficiencies and risks.

The solution is less about leasing hardware and more about embracing infrastructure models that guarantee you always have access to the latest technology without the burden of ownership, obsolescence, or unpredictable scaling costs. In storage, this evolution is already happening.

Applying Consumption-based Economics to the Storage Layer

We all know how supply and demand works: forecast what you need, obtain enough to cover your needs, and revisit each cycle as needs fluctuate. But when it comes to the highly experimental and iterative nature of AI, this rigid procurement model is creating a fundamental mismatch—especially for infrastructure teams ramping up resources in what is, quite frankly, pretty uncertain territory.

And, it changes every day.

Yesterday’s architectures have IT teams handcuffed to traditional supply and demand models that lead to either costly overprovisioning or performance bottlenecks that waste valuable GPU resources. Traditional approaches simply aren’t equipped to offer the flexibility they need—that’s reserved for public cloud solutions, but even those have been double-edged swords.

The Financial Reality of Infrastructure for AI at Scale

The financial burden of implementing and maintaining AI infrastructure has become a significant concern for organizations of all sizes. Consider that leading AI companies can spend upwards of $700,000 per day just to maintain their infrastructure and run flagship products. In 2024, some companies’ total spending on inference and training could reach $7 billion, driven by increasing computational demands.

The challenge: Traditional storage procurement approaches simply don’t align with AI’s unpredictable growth patterns.

Traditional Procurement Models vs. the AI Era

The status quo has been defined by traditional storage technology vendors with technical constraints and obsolescence built-in to their products. This has trapped organizations into a mid-to-longterm forecasting approach to storage infrastructure that falls short for several reasons:

AI workloads are inherently experimental and iterative.
Data requirements, performance needs, and capacity demands evolve unpredictably.
The three-to-five-year tech refresh cycle can’t keep pace with rapidly advancing AI capabilities.
Sizing storage for AI is incredibly difficult to forecast accurately.

Deploy too little storage and create bottlenecks for expensive GPU resources. Overprovision and waste capital that could be invested in more valuable AI resources.

This isn’t a new problem in IT, but AI can definitely make it worse—unless a new model (and mindset) is embraced.

Consumption-based Models in the AI Era

The traditional expenditure model can’t always deliver the agility needed for AI workloads. When data requirements change in weeks, not years, organizations need consumption-based models that scale on their terms—not their vendors’.

Instead of making long-term bets on capacity for AI, as-a-service consumption models change the whole operational dynamic. It’s a new paradigm where business leaders can focus less on trying to predict requirements and purchasing, owning, managing, and operating infrastructure. Rather than struggling to service the business’s AI needs using what they’ve already purchased, they can focus on accessing outcomes.

Ultimately, data becomes the focus, not capacity—a step in the right direction when you consider every successful AI strategy starts with a successful data strategy.

Evergreen//One: A New Economic Model for AI

True cloud-first flexibility has landed in the data center—but this next-gen storage solution isn’t just another buying model.

Evergreen//One™ transforms how organizations acquire and manage storage for AI workloads. It was specifically designed to address the unpredictability of infrastructure requirements for workloads like AI and is certified to work with NVIDIA OVX servers, DGX BasePOD, and DGX SuperPOD clusters, ensuring compatibility with industry-leading AI infrastructure. You can read the ESG Economic Validation to learn more.

Performance and Efficiency Guaranteed by SLA

Evergreen//One guarantees performance based on the maximum bandwidth requirements of your GPU clusters. This means:

You get all the performance needed to keep valuable AI resources fully utilized, delivered through a service level agreement (SLA).
There is no performance ceiling—even for multiterabyte-per-second requirements.
AI infrastructure is notorious for its energy consumption. Evergreen//One for AI addresses this concern with an energy efficiency SLA that guarantees service delivery at specified watt-per-unit measurements. Pure Storage pays for rack space and power for its technology when deployed your data center.

Predictable Economics for Unpredictable Workloads

What truly sets Evergreen//One for AI apart is its economic model that aligns with AI’s unpredictable nature:

Predictable monthly spending with SLA-backed guarantees
Pay only for what you use with simple, transparent billing
No need to delete valuable artifacts, checkpoints, or logs
Freedom to scale without the financial penalties of traditional models

Eliminate Technical Debt with Continuous Innovation

For OpenAI, leasing the hardware frees them from the risk of the chips becoming outdated sooner than expected. Customers with Evergreen//One for AI get the same benefits, with continuous innovation delivered non-disruptively. As AI workloads evolve, Pure Storage customers get:

Latest technology advancements delivered seamlessly
Performance enhancements without downtime
Feature updates without costly forklift upgrades

Platform-wide Data Management with Pure1

Included with the Evergreen//One for AI subscription is Pure1®, an AI-driven data-services platform that provides:

Predictive intelligence through Pure1 Meta®
Accurate forecasting of application and infrastructure needs. The platform even includes AI-powered reserve expansion in Pure1 to model your usage and recommend expansion of reserve commitments before reaching on-demand usage levels, helping optimize costs.
Continuous monitoring and proactive resolution
Full-stack analytics and global visibility

Focus on AI Innovation, Not Storage Management

The traditional approach of guessing future storage needs at the beginning of a three-to-five-year refresh cycle simply doesn’t work for AI. With Evergreen//One for AI, you get the storage performance and capacity you need today, but with the flexibility you need when AI capabilities and requirements can evolve weekly rather than yearly.

Discover the economic and operational agility organizations need to stay competitive without breaking the bank with Evergreen//One for AI.