The dominant narrative around Nvidia has shifted. Investors who once questioned whether custom chips from Amazon and Google would erode Nvidia's GPU monopoly are now asking a different question: can anyone else run an entire AI data center as efficiently as Nvidia can?
The answer, so far, is no. And Nvidia knows it.
After growing its market cap tenfold between early 2023 and mid-2025, Nvidia shares have traded on a more modest trajectory over the past year as concerns about GPU competition mounted. The company's latest earnings report, delivered Wednesday, has pushed investors toward a new framework for evaluating Nvidia's moat, as detailed in TechCrunch's reporting. The focus is no longer just on silicon performance. It is on how well the company orchestrates the entire data center ecosystem surrounding the GPU.
The Vera Rubin system
Nvidia is rolling out its Vera Rubin architecture, a system that goes far beyond a single graphics processor. The Rubin GPU pairs with the Vera CPU, the Groq 3 LPX inference accelerator, and dedicated racks for storage and networking. These are not bolt-on components. They are purpose-built subsystems designed to minimize the performance losses that occur outside the GPU itself.
If the GPU is the engine, the rest of the system is the car.
Jason Hardy, Nvidia's vice president of storage technology, has emphasized that memory capacity within a single server has hard limits. The critical challenge is getting data to the GPU at exactly the right moment, not simply moving more of it. "We saw upwards of 3x improvement in these operations, where the Vera CPU is allowing for acceleration," Hardy told TechCrunch. "So now we can use our flash to its fullest potential, because we can get all that performance out of it without bottlenecking." This continues the trend of Nvidia expanding beyond GPUs, a pattern visible in its earlier Groq 3 LPX inference accelerator announcement.
The Vera CPU focuses on orchestrating data flow rather than executing the heavy mathematical workloads that define traditional CPU performance. It manages memory staging, storage access patterns, and network routing for thousands of concurrent AI inference requests. That is a fundamentally different problem than what a standard x86 processor was designed to handle.
Storage becomes active, not passive
Nvidia's approach to storage represents a more fundamental shift. The company's technical blog, authored by Hardy himself at the Future of Memory and Storage conference in early August, described how GPUs can now initiate storage requests directly, generating thousands of concurrent operations. Traditional storage systems cannot keep pace with that demand.
Benchmarks presented at the Future of Memory and Storage conference showed the Vera CPU paired with the BlueField-4 STX storage processor delivering up to 3.21 times higher throughput than an x86 CPU in a two-stage compression and encryption pipeline, according to Nvidia's own technical blog. The implication is stark: storage stops being a passive repository and becomes an active participant in the data path.
This changes decades-old tradeoffs between memory and storage. Forty years ago, the decision of where data lived was measured in minutes of access time. On today's GPUs paired with Nvidia's storage solutions, the same calculation plays out in microseconds.
Nvidia has also open-sourced its cuFile APIs, which let GPUs read from and write to storage directly without CPU mediation. Google, Intel, Nvidia, and Meta are listed as inaugural maintainers. The move signals a broader industry alignment around GPU-driven storage architectures.
The competition moves to systems
OpenAI takes a different approach to the same problem. When it introduced its Jalapeño custom chip earlier this month, the company emphasized minimizing data movement entirely rather than managing it more efficiently. "We designed Jalapeño to minimize data movement and communication delays," OpenAI said in a blog post. "Its large domain allows the entire workload to remain within one connected system."
Both approaches share the same underlying logic: efficiency gains in AI infrastructure are increasingly coming from smarter data management, not raw compute cycles. This opens a new layer of competition that benefits neither pure-play GPU manufacturers nor cloud providers building custom silicon in isolation. Success requires integrated system design.
Nvidia's lead in this new layer appears substantial at this stage. The company's Storage-Next initiative brings together more than 40 storage and flash vendors, including DDN, Kioxia, and Micron, to define GPU-driven storage standards. Its SCADA framework enables GPUs to pull directly from storage into their own memory, while its Spectrum-X Ethernet networking and NVLink protocols create the connectivity layer that binds everything together.
What this means for the market
The shift from GPU-only competition to system-level competition has several implications for the broader market.
First, it raises the barrier to entry. Building a competitive AI chip is difficult. Building a competitive AI data center requires mastering GPU design, CPU orchestration, storage architectures, networking protocols, and the software stacks that bind them together. Few companies have that breadth.
Second, it changes the competitive dynamics with hyperscalers. Amazon and Google continue investing in custom silicon, which remains a rational strategy. But their in-house chips must still integrate into a larger ecosystem, and Nvidia is positioning itself as the system integrator rather than just the component supplier.
Third, it creates new partnership opportunities. The open-sourcing of cuFile and participation in Storage-Next suggests Nvidia is building an ecosystem where third-party storage and networking vendors benefit from its architecture, creating lock-in through mutual dependence rather than proprietary closure.
The earnings reaction suggests investors have absorbed this narrative. Nvidia's shares have been consolidating for roughly a year. The new focus on system orchestration gives the company a story that extends well beyond the next GPU release cycle. Whether the systems layer proves as durable as the GPU layer remains an open question, but for now Nvidia appears to be moving first and building the rules of engagement.