Nvidia's Pain Points: From Circuit Board Physics to a Rival's Inferencing Ambitions

Nvidia's Kyber rack delayed to 2028 as memory bandwidth, software efficiency gain importance; SK Hynix lists on Nasdaq, Micron margins soar.

The AI infrastructure market is pivoting. For months, the playbook was simple: pile into the companies making the fastest chips. That era is giving way to a more complex calculus in which memory bandwidth, manufacturing precision, and software efficiency all carry equal weight. Nvidia, still the undisputed leader in server GPUs, now finds itself squeezed by forces that no amount of CUDA lock-in can fully offset.

The most tangible symptom is a delay in the Kyber NVL144 rack, the centerpiece of the company's upcoming Rubin platform. Originally slated for 2026, it has been pushed back to 2028. The holdup isn't a coding bug — it is a physics problem. Engineers are wrestling with 78-layer circuit boards carrying 25-micron traces, precision that pushes conventional PCB fabrication to its absolute limit. The smaller Vera Rubin NVL72, a sibling system, remains on track for autumn 2026, but the Kyber slip signals that the days of effortless generational leaps in compute density are numbered.

Meanwhile, capital is rotating toward the memory suppliers that keep those GPUs fed. SK Hynix listed on the Nasdaq on Friday, cementing its role as a linchpin in the high-bandwidth memory chain. Micron, another key player, recently reported a gross margin of 87 percent in its data-center business — a stark indicator that memory is currently the most lucrative layer of the AI stack. Spot prices for Nvidia's H100 compute have eased as supply catches up, while High Bandwidth Memory prices could double by 2027. For Nvidia, that means a structural cost headwind: the very components that power its GPUs are eating into its own margin.

A separate threat is forming on the software side. DeepSeek, a Chinese AI firm, is developing its own inferencing chips, tapping open-source models and ultra-efficient training methods to deliver massive compute at a fraction of the usual cost. The broader industry trend is unmistakable. Amazon, Google, and Microsoft are all designing custom processors to reduce reliance on external vendors. Nvidia's CUDA ecosystem remains a formidable moat, but the industry's push toward cheaper inferencing — the phase where AI actually executes tasks for users — threatens to erode the premium pricing that high-end GPUs currently command.

Nvidia is not standing still. The company announced an $80 billion share buyback program and raised its quarterly dividend, a dual signal of confidence aimed at a market that has grown more sensitive to valuation. In the first quarter of fiscal 2027, revenue hit $81.62 billion, comfortably above expectations, with the data-center segment alone contributing $75 billion. The stock closed on Friday at €176.50, down 0.51 percent on the session. That leaves it 25.8 percent above the 52-week low of €140.30 set exactly one year ago, and 12.84 percent below the mid?May high of €202.50. Year to date, the shares are up 9.56 percent. The market capitalization stands at €4.17 trillion.

Technically, the stock is treading water. It sits 2.52 percent below its 50-day moving average of €181.06, and the relative strength index at 49.9 points to a neutral reading — neither overbought nor oversold. Analysts remain broadly bullish, with a consensus price target of €263.69, implying upside of roughly 49 percent from current levels. Reaching that target, however, will depend on how quickly Nvidia can resolve the manufacturing hurdles in its Rubin-era road map and how effectively it can manage the creeping cost pressure from memory and the competitive threat from inferencing rivals.

The underpinning demand for AI infrastructure is still immense. Hyperscaler capital expenditure continues to climb, and Nvidia's architecture runs nearly all of the leading models. But the easy gains of the past two years — where each chip generation arrived faster and each forecast proved too conservative — are giving way to a period where physics, procurement, and upstart competitors all demand a seat at the table.