Nvidia's AI moat is shifting from GPUs to data center orchestration, TechCrunch argues

For most of the AI boom, Nvidia's story was simple: it was the sole supplier of top-end GPUs, and it profited enormously as the industry scaled up. A new TechCrunch analysis argues that this framing is now out of date. Following the company's earnings report on Wednesday, investors have begun reassessing where Nvidia's advantage actually sits — in the machinery that surrounds the GPU rather than in the GPU alone.

The story investors told until now

TechCrunch's Russell Brandom lays out the familiar version first: Nvidia's near-monopoly on training hardware made it wildly profitable in the early years of the boom, but hyperscalers such as Amazon and Google have since started building their own chips, ending Nvidia's run as the only real option. The numbers track that anxiety. After growing its market cap tenfold between the start of 2023 and mid-2025, Nvidia's shares have followed a much flatter path over the past year, weighed down by concerns about GPU competition.

The newer story, which Brandom says took hold after Wednesday's earnings, is that AI compute is heading toward gigawatt-scale deployments, and keeping a data center of that size running at peak efficiency has become genuinely difficult. Nvidia has spent years building much of the leading hardware for that job, which leaves it strong in the systems around its GPUs even as the GPUs themselves attract more rivals.

What is actually in the rack

The clearest evidence is what Nvidia is now shipping. Its Vera Rubin architecture, currently rolling out, pairs the Rubin GPU with a set of companion parts: the Vera CPU, the Groq 3 LPX inference accelerator and comparable rack-level systems for storage and networking. Brandom describes these as heavily specialized units that do not process tokens at all — their job is to keep everything outside the GPU working at maximum efficiency, with the GPU as the engine and these components as everything else the vehicle needs to run.

The Vera CPU targets data orchestration in particular. "Vera is important because there's only so much memory that you can put in a single server or any sort of compute platform," Jason Hardy, Nvidia's VP of storage technology, told TechCrunch. Memory capacity has grown alongside compute — one reason firms like Micron have profited in the second wave of the infrastructure buildout — but delivering the right data to a GPU at the right moment is far from straightforward. As operators push tokens-per-watt ever lower, that traffic management has become a first-order engineering problem. Hardy said Nvidia measured "upwards of 3x improvement in these operations" when the Vera CPU handles the acceleration, which lets flash storage perform at full potential without bottlenecking.

OpenAI is attacking the same bottleneck

The problem is visible beyond Nvidia. OpenAI's Jalapeño chip, detailed in a blog post earlier this month, takes the opposite route: rather than moving data more efficiently, it tries to move as little as possible. "We designed Jalapeño to minimize data movement and communication delays," OpenAI wrote, adding that its large domain allows "the entire workload to remain within one connected system, minimizing data movement and helping the complete request stay fast and efficient from beginning to end."

The method differs, but the logic matches. Efficiency gains are now coming from smarter data handling rather than raw processor cycles, and that opens a fresh layer of infrastructure for chipmakers and hyperscalers to fight over.

Why it matters

TechCrunch's analysis reframes the central competitive question in AI hardware. If the contest is no longer just about who builds the fastest accelerator, but about who can make an entire gigawatt-scale system run efficiently, then custom silicon from Amazon and Google is a narrower threat to Nvidia than it appears. The shift is not an automatic win — Nvidia will face rivals at this layer too — but Brandom judges the company to be well ahead in the early going. For anyone buying AI infrastructure, it also suggests that real pricing power may sit with whichever vendor masters orchestration, not merely whoever ships the most compute.