Introduction

In September 2025, NVIDIA quietly made one of the most important acquisitions of the AI era, spending over $900 million in cash and stock to acquire Rochan Sankar‘s Enfabrica, not for its products, but for its people, its patents, and its position at the intersection of compute and memory.

To most, it looked like another startup being folded into NVIDIA’s massive AI machine.

To those paying attention, like me, it was something else entirely. Entirely huge.

If there were ever a soundtrack for moments like this, it might echo Don McLean’s “The Day the Music Died.”Only this time, the lyrics would celebrate rebirth, not loss.

This was “The Day Memory Thrives” when the AI industry realized that performance would no longer be defined by compute, but by memory.

The End of the GPU Gold Rush

For the past three years, the AI industry has been obsessed with GPUs. Entire data centers have been built around them. Capital has been poured into them. Infrastructure teams have chased availability like miners in a gold rush.

But as I wrote earlier this year in “AI Market Inflection Point: Hype, Reality, and What Comes Next,” the GPU craze is starting to mature. The industry has reached a point of saturation. Data centers are filled with underutilized accelerators waiting for optimized workloads.

The problem is not that GPUs are too slow. The problem is that the rest of the infrastructure cannot keep up.

In that March 2025 piece, I wrote:

“AI isn’t just about GPUs. The true bottleneck is how quickly data can be accessed, moved, and processed. Enterprises are realizing that latency, not compute, often dictates AI performance.”

That statement proved truer than I expected. The shift from “more compute” to “better coordination” is happening faster than anyone predicted. NVIDIA’s decision to acquire Enfabrica shows that the company understands this and intends to own not just the GPU, but the memory path that feeds it.

What Enfabrica Built

Enfabrica was founded to solve a problem that every AI architect eventually faces: GPU starvation. Even in massive clusters, GPUs spend far too much time waiting for data.

Their solution, called Elastic Memory Fabric System (EMFASYS), introduces a new way to connect GPUs to memory. Instead of relying solely on expensive and limited High Bandwidth Memory (HBM), EMFASYS allows GPUs to pull from shared DDR5 memory pools across a fabric built on CXL (Compute Express Link) and Ethernet.

Think of it like this. HBM is the Bugatti sports car. It is incredibly fast but incredibly small. DDR5 is the freight truck. It moves more slowly but can carry much more. Enfabrica built the highway that lets them travel together efficiently.

This fabric-based approach changes the economics of AI clusters. Instead of scaling linearly with expensive HBM, organizations can now scale capacity using lower-cost DDR5 nodes. Each GPU still enjoys high-speed access to local data but can reach into large external pools without a severe latency penalty.

That is a breakthrough. And NVIDIA knows it.

A Missed Opportunity Remembered

To understand why this matters, we need to go back three years.

In 2022, I wrote an article titled “Life Without Optane.” Intel had just shut down its Optane business, walking away from one of the most disruptive memory technologies ever created: 3D XPoint.

At the time, I argued that Intel’s mistake was not the technology itself, but how it brought it to market. Optane was introduced into an already crowded SSD segment, which turned it into a performance niche rather than a platform shift. The real value was never in replacing NAND or flash. The real opportunity was in redefining persistent memory, a space that still had no clear market leader.

I wrote then:

“CXL could have potentially covered [Optane’s] complexities up and made it more broadly accessible to a mass market, becoming what I would call a ‘market maker.’”

That statement aged well.

Intel had the hardware innovation with Optane Persistent Memory (PMEM) and a central role in shaping the Compute Express Link (CXL) standard. What it never did was bring the two together. Instead of pairing persistent memory with a shared fabric interface, Intel divided its focus between SSDs and niche NVDIMM modules. That decision confused customers and cost Intel its chance to lead the next wave of memory innovation.

Micron Technology, on the other hand, exited 3D XPoint development to focus on technologies that supported CXL. At the time, that move seemed premature. In hindsight, it was both visionary and costly. Visionary for anticipating the rise of CXL, but costly for walking away from the persistent memory breakthrough that could have positioned them ahead of this moment. While it was done for cost reasons, and rightfully so, given today’s market landscape, I am curious if leadership questions the wisdom of a full retreat.

Intel may have started the memory revolution, but NVIDIA is the one finishing it. While Intel stepped away, NVIDIA watched, learned, and waited. I first wrote about this inflection back in 2020, when I said Optane represented more than a product. It was a bridge to a new computing era. Five years later, that bridge is finally being built.

From Optane to Enfabrica

The irony is hard to ignore. What Intel abandoned in 2022, the idea of fabric-connected and shareable memory, is exactly what Enfabrica built and what NVIDIA just acquired.

In that same 2022 post, I referenced Intel Fellow Dr. Debendra Das Sharma, who described connecting Optane over CXL so it could be accessed across servers, hot-pluggable, and failover-ready. He described a future where persistent memory could live on the bus and be accessed by any processor in the rack.

That vision never came to fruition at Intel. But it is alive inside NVIDIA today.

Enfabrica’s EMFASYS architecture makes that vision practical. It allows GPUs to treat remote memory as an extension of local capacity, providing near-HBM performance for workloads that previously had to live on slower storage tiers.

This is not just a clever optimization, it is the foundation of a new memory hierarchy for AI. Data can now move freely, workloads can be rebalanced dynamically, and capacity can scale without breaking cost models.

Below: Intel’s original “memory-to-storage” pyramid was a blueprint for bridging DRAM and NAND that hinted at what was to come. Enfabrica and NVIDIA are finally completing what this graphic only began to imagine.

Validation of the CXL Thesis

When I wrote in March that “CXL is on the horizon,” it wasn’t speculation. It was an observation based on trajectory because CXL had already proven its potential in concept. It just needed a catalyst.

That catalyst just arrived. Thank you, Jensen Huang

With Enfabrica now part of NVIDIA, the market no longer sees CXL as optional. It has become the backbone of the next generation of AI infrastructure.

CXL matters because it finally breaks the static bond between compute and memory. It introduces memory composability, which allows memory to be pooled, shared, and allocated dynamically across nodes.

That capability changes everything.

It improves GPU utilization, eliminates the need for over-provisioning, and enables hybrid memory architectures that combine HBM, DRAM, and DDR5 pools. It also gives CPUs, GPUs, and accelerators a common language that was never possible before.

NVIDIA did not just endorse the CXL standard. It validated it. It validated the vision I outlined years ago and made it real.

The Gen5 Line in the Sand

What’s the catch, right?

Here is the new reality. Organizations that want to participate in the next phase of AI infrastructure must operate on Gen5 PCIe or higher.

The bandwidth, latency, and coherency requirements of CXL fabrics cannot be met on older Gen4 systems, rather, Gen5 opens the performance lanes that make true memory pooling a possibility. It provides the low-latency communication needed between GPUs, CPUs, and memory nodes to support modern AI workloads of today and the advanced workloads of tomorrow. Without it, shared memory fabrics like Enfabrica’s cannot perform at scale.

However, the hardware readiness goes well beyond PCIe. CXL also requires CPU architectures that include native CXL protocol support, including Type 1, Type 2, and Type 3 device compatibility. That support only became available in the latest generation of x86 processors from Intel and AMD . Without those CPUs, even Gen5 systems cannot take full advantage of the memory pooling, coherency, or fabric-aware resource sharing.

This is one of the reasons I have been such a strong advocate for Gen5 PCIe. Until now, it was difficult to justify the move because CXL had not yet reached mainstream adoption. Without broad CPU, software, and hardware support for memory pooling, the business case simply did not add up. In my opinion, that has changed!

Let me be clear, this transition is not about vendor differentiation. It is about industry readiness and the customer. Gen5 and CXL are now baseline requirements for any organization that intends to remain competitive in the AI data center. Companies that embrace this shift early will be positioned to lead the next era of infrastructure. They will be able to support composable memory architectures, adapt faster to workload diversity, and integrate more easily into fabric-centric ecosystems.

The buying criteria for “AI-ready” infrastructure are changing. Performance will no longer be measured by component speed but by architectural balance and readiness for memory-centric design. The focus is shifting from hardware specifications to architectural enablement, the foundation of a true AI Factory now made possible by the integration of memory, compute, and fabric.

Memory Redefined is the Next Era of Compute

If the 2010s were defined by the GPU, the late 2020s will belong to memory. We were so close five years ago, but the reality is we are here now. The Enfabrica acquisition signals that traditional scaling is over. GPUs cannot deliver more performance without rethinking how they access data.

I have said for years that if NVIDIA did not own the memory pooling market, it would eventually be ownedby it. In other words, NVIDIA risked being marginalized by companies like Enfabrica once memory pooling reached mainstream adoption. This acquisition changes that completely.

And it is beyond brilliant.

NVIDIA’s move is not about adding another technology to its lineup. It is about removing the final bottleneck that limits AI performance so their GPUs will be used to their fullest extent. Once a company controls compute, interconnect, and memory, it controls the entire performance envelope. That is the real strategic intent behind this acquisition: NVIDIA is making AI realities more achievable with Enfabrica.

The bigger picture should be crystal clear, performance leadership will no longer be defined by raw compute or clock speed. It will be defined by architectural balance and the harmony between compute, memory, and data movement.

In this new model:

• Compute becomes the execution layer.
• Memory becomes the scalability layer.
• The fabric becomes the connective layer that makes it all work.

By the way, NVIDIA has been a member of the CXL consortium, which was founded by Intel, from the beginning, so it makes sense that leadership would see the horizon more clearly based on the conversations with customers they have every day. In my opinion, NVIDIA recognized the real customer dilemma, and it was never about GPU performance.  It was about memory accessibility for the AI workloads, again, something I have been saying for the last two years.

Whoever owns that fabric owns the future of AI.

A Lesson in Market Timing

If you have spent a decent amount of time in this industry, you know how this goes, there is often a clear pattern in how innovation unfolds.

• In 2022, Intel had the right technology but failed to align it with market direction. (lacked vision)
• In 2023 and 2024, hyperscalers rushed to buy every GPU they could find, believing that raw compute could buy intelligence. (lacked vision and restraint)
• In 2025, the market finally corrected. (I can see clearly now the rain is gone “vision”)

The first wave invents. The second wave misuses. The third wave operationalizes. And aren’t we all happy we are in the operationalizing phase! NVIDIA’s Enfabrica acquisition represents that third wave. It takes fragmented ideas about disaggregated memory and persistent computing and turns them into a cohesive business model.

Once again, NVIDIA did not follow the market. It redefined it. Again.

And, I have to say it, “I told you so.”

Beyond NVIDIA: My predictions

What NVIDIA just did with this acquisition is create a ripple effect that will uplevel the entire AI market. It is a rising tide that will lift or drown every boat in the ecosystem, with lasting effects across every layer of AI infrastructure.

• Cloud Providers will begin designing rack-level systems with shared memory pools. Expect to see “memory-as-a-service” models emerge by 2026.
• OEMs will accelerate Gen5 and CXL support or risk losing relevance. Expect to see entire product refreshes built around “AI memory fabric” positioning.
• Storage Vendors will need to rethink their story. Performance alone will not sell. Data mobility and orchestration will. We call it Data Singularity.
• Software Developers will begin writing CXL-aware frameworks for scheduling, caching, and AI pipeline management.

For everyone building or selling infrastructure, this is not a small shift. It is a generational change.

From Prediction to Proof: A Bridge Over Troubled Waters

When I look back over the past three years, the story is remarkably clear to me, and always has been. In 2022, I said the industry lost something important when Optane was discontinued. The packaging was wrong. Intel should have focused on the persistent memory market not compete with NAND. Someone knew better, but was overruled. Optane represented a bridge between volatile and persistent memory that could have redefined computing.

In March 2025, I said that CXL would become the next major inflection point because AI’s real constraint was not compute but data movement and memory bandwidth.

Now, in late 2025, NVIDIA has validated both ideas.

They built the bridge Intel started, but never finished.

They operationalized the standard that others treated as theoretical.

They turned memory into the most strategic layer of the AI stack.

With great power comes great responsibility.  Well done, NVIDIA.

What Comes Next

If the hardware story was about interconnects, the next story will be about orchestration. It is always about orchestration. CXL fabrics will make possible a new class of memory-aware software that can dynamically allocate resources across CPUs, GPUs, and accelerators. This will change how pipelines are built, how applications scale, and how customers realize results and outcomes.

It reminds me of Nebula, the company Chris Kemp founded after leaving NASA. They built one of the first on-prem cloud appliances, trying to make orchestration simple before the world was ready for it. Their vision was right, but the ecosystem and hardware were not there yet.

CXL changes that and now the hardware has finally caught up with the orchestration vision.

Long term, this will intersect with the movement toward Sovereign AI. The ability to pool and control memory locally will matter for national data strategies and privacy frameworks. Memory locality will become as important as data residency. Customers will no longer have to choose between sovereignty and time to outcome. The decision to move data to a public cloud will finally become a business choice, not a performance compromise.

The next era will be defined by memory autonomy, the freedom to decide where, how, and when data lives in the memory fabric.

Closing Thoughts

Have we officially entered the post-GPU era of AI infrastructure? In my opinion, yes, we have. Performance gains are no longer coming from adding more compute, but from redefining memory.

I remember a conversation years ago with one of the admins running a supercomputer at the Missile Defense Agency in Huntsville, Alabama. He said, “If we can’t feed those cores, we’re burning cash.” In other words, the value of that system wasn’t in how fast it could go on paper, but in how consistently it could stay busy. The same holds true for AI infrastructure today. The goal isn’t just to build faster GPUs, it’s to keep them working efficiently, without starving them of data.

Memory is no longer a supporting actor. It has taken the lead.

In 2020, Optane could have been the bridge. By 2025, CXL became the bridge. And now, with Enfabrica, NVIDIA finally built it.

Oh, and by the way, memory didn’t replace compute.  It redefined it.