Nvidia Rewires Its AI Roadmap: A Cancelled Chip, a Rising CPU, and a Record Quarter

Nvidia shelves its Rubin CPX inference chip, pivots to a platform strategy with Vera CPU and $20B Groq licensing, as Q1 revenue jumps 85% to $81.6B.

When Jensen Huang touched down in Taipei last week, his mission was not ceremonial. The Nvidia chief executive was there to lock down manufacturing capacity for the company’s next-generation Vera Rubin GPU platform, meeting with TSMC and SK Hynix on the sidelines of Computex 2026. But behind the scenes, a quieter but equally consequential shift is underway: Nvidia has effectively shelved a planned inference chip, reorienting its strategy toward a broader, platform-based model that includes licensing, specialized infrastructure, and a new internally designed central processor.

The Rubin CPX chip, originally slated for the second half of this year with 128 GB of GDDR7 memory, has disappeared from Nvidia’s public roadmap. Supply-chain sources report no visible orders for the required memory or substrates, and the project was absent from the GTC 2026 roadmap altogether. Industry observers now regard it as functionally canceled. Nvidia has not framed the move as a retreat, but the pivot is unmistakable: instead of pushing another discrete GPU, the company is leaning into a $20 billion licensing deal with Groq 3 LPX and betting on low-latency, real-time inference architectures.

The financial firepower for this reorganisation comes from a quarter that left little room for doubt. Nvidia’s fiscal first-quarter revenue surged 85 percent to $81.6 billion, comfortably above its own forecast of $78 billion. Data-centre sales — the engine room — jumped 92 percent to $75.2 billion. Gross margins leapt from 43 percent to 72 percent, underscoring the pricing leverage the company enjoys in a structurally supply-constrained market. For the current quarter, management guided for $91 billion in revenue, nearly 95 percent growth and well ahead of the $87.2 billion consensus.

That cash muscle is being redeployed aggressively. Nvidia has authorised an $80 billion share buyback programme and hiked its quarterly dividend to $0.25 per share. At the same time, it is deepening ties with Marvell Technology via a $2 billion investment to develop custom XPUs and networking solutions under the NVLink Fusion umbrella. Outstanding purchase commitments stand at $119 billion — a measure of how deeply embedded Nvidia has become in the global AI supply chain.

The Vera CPU is the most tangible symbol of the new direction. Benchmarks show the 88-core Olympus ARM-based processor achieving memory bandwidth of 1.2 TB/s, outperforming Intel’s Xeon 6980P by 55 percent and AMD’s EPYC 9575F by 10 percent, with significantly better energy efficiency. Nvidia has already shipped early samples to Anthropic, OpenAI, and SpaceXAI. Oracle Cloud Infrastructure plans to deploy hundreds of thousands of Vera CPUs starting in 2026 — the first hyperscaler to commit at that scale. Huang has described the global CPU market as a $200 billion opportunity the company had previously ignored. Analysts estimate CPU-related revenue could reach $20 billion by the end of fiscal 2027.

None of this optimism, however, has insulated Nvidia’s stock from near-term pressure. The shares closed at €181.32, down 1.7 percent on the day and 5.63 percent lower on the week. They sit roughly 9 percent below the all-time high of €201.05 hit in mid-May, with the relative strength index hovering near 40 — a technical cooling after a blistering run. Huang himself injected a note of caution at Taoyuan Airport, warning that rising memory prices represent “a very important form of inflation” that could meaningfully push up consumer electronics costs. He urged memory makers to expand capacity.

SK Hynix is already responding with a new technology called iHBM, which integrates cooling elements directly into the memory package to manage heat-related performance degradation. The solution is expected to debut with the eighth-generation HBM standard, primarily for Nvidia’s next GPU platform.

At the same time, Nvidia is expanding beyond pure hardware into industrial AI workflows. On a Rescale-hosted event, the company unveiled PhysicsNeMo, a platform for agent-based digital development. Early data from partners such as McLaren Automotive suggests the approach can triple productivity, improve cost efficiency 30-fold, and shorten development cycles by 60 percent, while allowing four times as many design candidates to be evaluated in the same period. This agentic AI push fits neatly with the broader strategy: less dependence on discrete GPU refreshes, more control over end-to-end AI infrastructure.

Huang’s keynote at the Taipei Music Center on June 1 will be the next high-profile test of the narrative. The Vera Rubin NVL72 system — linking 36 Vera CPUs with 72 Rubin GPUs over sixth-generation NVLink — promises up to ten times better inference performance per watt and ten times lower cost per token. Huang has called the launch “probably the biggest product introduction in Taiwan’s history.” Whether the market buys the new, broader story — one where a canceled chip is a feature, not a bug — will determine if the stock can recapture its highs.