Jensen Huang did not walk onto the GTC 2026 stage holding a chip. He wheeled out a system.

That visual shift — from a single GPU to a full rack-scale supercomputer — captures everything about where NVIDIA is taking the AI hardware industry. At CES 2026, NVIDIA officially launched the Vera Rubin platform; at GTC in March, Huang revealed it would generate at least $1 trillion in cumulative revenue from 2025 through 2027, doubling the previous forecast announced just a year prior (TradingKey).

That number either sounds impossible or inevitable, depending on how closely you're watching the hyperscalers' capital expenditure plans.

What Vera Rubin Actually Is

Vera Rubin is not a single chip. It is a seven-chip, five-rack-scale computing platform designed to replace Blackwell as NVIDIA's flagship AI infrastructure system.

The core of the platform is the Rubin GPU, built on TSMC's 3nm process with 336 billion transistors — 1.6x the density of Blackwell. The headline performance numbers, per NVIDIA's CES announcement and confirmed by ServeTheHome's technical analysis:

  • 50 PFLOPS of FP4 inference compute per GPU — 5x faster than Blackwell
  • 35 PFLOPS of FP4 training compute — 3.5x faster than Blackwell
  • 288GB of HBM4 memory per GPU, with 22 TB/s bandwidth — 2.8x the bandwidth of Blackwell's HBM3e
  • 8x better performance-per-watt than Blackwell for inference workloads

Paired with the Vera CPU — 88 custom ARM cores, 1.5TB LPDDR5X memory, 1.2 TB/s bandwidth — a single Vera Rubin node houses one Vera CPU and two Rubin GPUs connected via NVLink-C2C at 1.8 TB/s.

The NVL72 Rack: Scale-Up Made Physical

The Vera Rubin NVL72 rack integrates 72 Rubin GPUs and 36 Vera CPUs into a unified architecture via NVLink 6, which doubles the available bandwidth to 3.6 TB/s per GPU versus the previous generation's NVLink 5.

At the rack level, the numbers are staggering:

Spec Vera Rubin NVL72 Blackwell NVL72
Total HBM Memory 20.7 TB 13.8 TB
FP4 Inference 3.6 EFLOPS 1.4 EFLOPS
FP8 Training 2.5 EFLOPS 0.72 EFLOPS
NVLink Total BW 260 TB/s 130 TB/s
Rack Power 120–130 kW 120 kW

Sources: Introl technical analysis, ServeTheHome

Eight NVL72 racks interconnected form a DGX SuperPOD Vera Rubin, delivering 28.8 EFLOPS of FP4 compute with a combined 600TB of memory.

The Groq Wildcard

At GTC 2026, Huang unveiled the Groq 3 LPU — NVIDIA's first product built since its $20 billion acquisition of Groq in late 2025 — as a complementary inference processor to the Vera Rubin platform. The Groq 3 features 500MB of on-chip SRAM, enabling inference without external HBM memory, with end-to-end latency "an order of magnitude lower than traditional GPUs," according to the GTC announcement (TradingKey).

NVIDIA's Dynamo intelligent scheduling system assigns work across this hybrid architecture: Rubin handles the compute-intensive "prefill" stage, while Groq 3 handles latency-sensitive "token decoding." The result, NVIDIA claims, is 35x throughput improvement for trillion-parameter models at equivalent power — a number aimed squarely at the economics of inference at scale.

"This is our moment. This is a reinvention. This is a renaissance of the enterprise IT." — Jensen Huang, GTC 2026 keynote (NVIDIA Blog)

The $690 Billion Context

Vera Rubin is not being built in isolation. Futurum Group's analysis projects that Microsoft, Amazon, Alphabet, Meta, and Oracle will collectively spend $660–690 billion on capital expenditure in 2026, nearly doubling 2025 levels — with the vast majority directed at AI compute, data centers, and networking.

Microsoft has deployed hundreds of thousands of NVIDIA Grace Blackwell GPUs in liquid-cooled Azure data centers over the past year and was the first hyperscaler to power up Vera Rubin NVL72 systems. AWS announced it will offer NVIDIA RTX PRO 4500 Blackwell Server Edition instances and is collaborating on a major Rubin platform expansion. Both cloud providers are committing to Rubin volume in the second half of 2026.

Meanwhile, an AI memory chip shortage is compounding the hardware race. IDC's March 2026 analysis, reported by Yahoo Finance, describes a high-bandwidth memory (HBM) supply crunch "like no other," with production capacity unable to keep pace with AI server demand. NVIDIA's HBM4 supply for Rubin is constrained, with estimates suggesting total 2026 output at 200,000–300,000 Rubin GPUs despite high-confidence demand far exceeding that.

What's Next: Feynman

At GTC 2026, Huang also teased NVIDIA's next platform after Rubin: Feynman, which will pair a new CPU called Rosa with a new LPU called LP40, built with NVIDIA's MVFP4 computing structure for further speedups. Rubin Ultra — delivering 15 EFLOPS in an NVL576 configuration with 576 GPUs per rack — is expected in H2 2027.

NVIDIA also confirmed plans for Vera Rubin Space-1, a Rubin-based computing platform designed for orbital data centers — the first serious commercial attempt to move AI infrastructure off-planet.

Takeaway: Vera Rubin is not an incremental GPU upgrade — it's a platform bet that NVIDIA wins the next decade of AI compute by designing the entire stack, from silicon to rack to data center reference design. With $1 trillion in cumulative demand forecast and hyperscalers already committing, the constraint isn't demand. It's whether TSMC can build enough chips, fast enough.

---