NVIDIA Rubin vs AMD Helios: The Duel for Rack-Scale AI Computing Dominance

Summary: NVIDIA unveiled Rubin GPU's complete architecture at SIGGRAPH 2026 (3nm dual-die, 336B transistors, 288GB HBM4), directly confronting AMD's simultaneously launched Helios rack system (2nm MI455X, 432GB HBM4, 31TB rack memory). AMD challenges NVIDIA's dominance with a 50% memory advantage and 30% cost efficiency claims, while NVIDIA defends its crown with 3.6 EFLOPS rack compute, mature NVLink 6 ecosystem, and first-mover production advantage. The divergence between closed integration and open platform approaches is fundamentally rewriting the rules of AI infrastructure competition.

---

1. Event Recap

Between July 20-29, 2026, the AI chip sector experienced its densest competitive product launch cycle in recent years.

NVIDIA revealed the complete technical specifications of the Rubin GPU at SIGGRAPH 2026. This next-generation AI accelerator, manufactured on TSMC's 3nm process, uses a dual-die package design with each die measuring approximately 800mm-squared, near the reticle limit. The two dies are interconnected via the NV-HBI interface, totaling 336 billion transistors. Inference performance reaches 25 PFLOPS (NVFP4 precision), with sparsity acceleration equivalent to 50 PFLOPS; training performance stands at 35 PFLOPS. It features 896 Tensor Cores (third-generation Transformer Engine architecture) and 288GB of HBM4 memory with 22 TB/s total bandwidth. The NVLink 6 switch provides 28.8 TB/s bandwidth and 3,600 GB/s GPU interconnect, double the previous generation.

The NVL72 rack configuration comprises 72 Rubin GPUs, 36 Vera CPUs (88-core Arm architecture), and 36 NVLink switches, delivering 3.6 EFLOPS total compute (NVFP4), 20.7TB total HBM4 capacity, and 1,580 TB/s aggregate bandwidth.

Concurrently, AMD launched the MI455X accelerator and Helios rack system at its Advancing AI 2026 event on July 23. The MI455X uses a TSMC 2nm and 3nm hybrid process with 12 compute and I/O chiplets, exceeding 320 billion transistors on the CDNA 5 architecture, delivering 20 PFLOPS FP8 performance with 432GB HBM4 and 23.3 TB/s bandwidth, single-card memory capacity surpassing Rubin's 288GB.

The Helios rack integrates 72 MI455X GPUs and 18 EPYC Venice CPUs, achieving 31TB total HBM4 capacity, 50% higher than NVL72's 20.7TB. AMD CEO Lisa Su claimed Helios delivers 10-15x better performance per power budget and 30% more tokens per dollar versus Vera Rubin NVL72 (tested on Kimi K2 model).

On the customer deployment front, Anthropic announced plans to procure up to 2GW of Helios racks (1GW delivered in H1 2027), with AMD simultaneously investing $5 billion in Anthropic. OpenAI confirmed it has been running Helios racks in its data centers for three months, with GPT-class workloads operational on MI455X.

TrendForce reported on July 28 that NVIDIA may need to reduce Rubin annual production from 2 million to 1.5 million units due to insufficient HBM4 supply, while also halving Vera CPU SOCAMM memory from 192GB to 96GB to control costs. Bernstein estimates a single Vera Rubin NVL72 rack could cost up to $9.1 million, with HBM4 pricing projected to reach $53/GB by 2027.

---

2. Technical Deep Dive

Architecture Comparison: Two Fundamentally Different Design Philosophies

NVIDIA Rubin follows an "extreme integration" approach: 3nm dual-die GPU packaging, Vera CPU directly connected via NVLink-C2C, NVLink 6 switches providing full intra-rack interconnect. The entire NVL72 is designed as a tightly coupled "rack-scale accelerator" with 72 GPUs operating in a unified performance domain.

AMD Helios takes an "open disaggregated" approach: MI455X uses 12 chiplets (2nm compute + 3nm I/O), connects intra-rack via UALink over Ethernet, and uses Ultra Ethernet for external networking. Rack specifications are fully public, allowing hyperscalers to build custom variants.

Four-Vendor Competitive Matrix

> Note: Google TPU v6 and Huawei Ascend 920 data are partially based on previously disclosed information and industry estimates; complete specifications have not been officially published.

The Memory Race: Why 432GB vs 288GB Matters

Memory capacity has become the core competitive dimension in modern AI accelerators. Larger memory means:

• Bigger models fit entirely on single cards, reducing cross-node communication
• Longer context windows can be processed directly (1M+ tokens)
• Larger batch sizes improve throughput
• Less quantization needed, preserving model accuracy

AMD Helios holds a 50% memory advantage with 31TB rack total (vs NVIDIA's 20.7TB). When running massive MoE models like Kimi K2, this directly translates to fewer nodes and lower latency.

NVIDIA's counter-strategy is "system-level memory": Vera CPUs' SOCAMM (LPDDR5X), though reduced from 55TB to 28TB, combined with 20.7TB HBM4, still provides 48.7TB total available memory pool. NVIDIA has created a tiered memory architecture in NVL72: HBM4 for extreme bandwidth, LPDDR5X for large-capacity buffering.

NVLink 6 vs UALink: The Interconnect Architecture Debate

NVIDIA NVLink 6: 3,600 GB/s per GPU, 28.8 TB/s switch, fully liquid-cooled, proprietary protocol. Advantages: mature software stack, SHARP in-network reduction reduces 50% congestion. Disadvantage: locks into NVIDIA ecosystem.

AMD UALink over Ethernet: 260 TB/s rack-scale, based on open standards. Advantage: hyperscalers can customize, compatible with multi-vendor equipment. Disadvantage: software ecosystem maturity lags NVLink.

---

3. Financial Logic

Cost Structure Breakdown

According to Bernstein and GF Securities analysis, the Vera Rubin NVL72 rack's cost structure reveals key economic dynamics of AI infrastructure:

HBM4 pricing is projected to reach $53/GB by 2027 (Bernstein), meaning 288GB per-GPU HBM4 cost is approximately $15,200, with 72 GPUs totaling ~$1.1M. Memory cost as a share of system BOM declining from 29% to approximately 20% is NVIDIA's target threshold.

Tokens per Dollar: AMD's Core Weapon

AMD claims Helios delivers 30% more tokens per dollar than Vera Rubin NVL72 on Kimi K2. Key drivers:

• Memory advantage: 432GB vs 288GB per card reduces cross-node communication
• Open platform: UALink enables custom optimization
• Chip cost: 2nm chiplet design may be more economical than 3nm dual-die

However, AMD uses MXFP4 format while NVIDIA uses NVFP4: they do not measure identical arithmetic operations, so direct comparison requires caution. Additionally, AMD's actual FP4 throughput is approximately half of its peak specifications.

Revenue Projections

---

4. Strategic Depth

NVIDIA's Triple Moat

• Production first-mover: Vera Rubin has been in production since Q1, with partner deliveries beginning H1 2026. AMD Helios first deployment not until Q4.
• CUDA ecosystem: Despite ROCm's rapid progress, many specialized libraries still default to CUDA. AMD demonstrated ROCm progress in PyTorch and inference servers at Advancing AI, but the ecosystem gap still requires 12-18 months to close.
• Full-stack integration: Rubin platform's six chips (GPU+CPU+NVLink+SuperNIC+DPU+Ethernet switch) are co-designed; AMD must rely on Broadcom switch chips and third-party networking.

AMD's Differentiated Breakthrough

• Memory leadership: 432GB HBM4 per card, 31TB per rack: a hard advantage for massive models and long-context workloads.
• Open platform: Public fabric specifications allow hyperscaler customization: a flexibility NVIDIA does not offer.
• Strategic customer lock-in: Anthropic 2GW / $5B investment + OpenAI 6GW deployment directly counterbalances NVIDIA's core customer base.
• AMD + Cerebras combination: Large GPU throughput + Cerebras low-latency inference = 5x fastest token throughput.

Competitive Landscape Evolution

The AI accelerator market is shifting from "NVIDIA dominance" toward a "duopoly + custom ASIC" tri-polar structure:

---

5. Challenges and Concerns

NVIDIA Risks

• HBM4 supply bottleneck: Production may drop from 2M to 1.5M units; SK hynix and Micron HBM4 qualification delays are key constraints. NVIDIA receiving only approximately 60% of LPDRAM demand.
• Memory cost escalation: SOCAMM halving is a reluctant move, but weakens Vera CPU performance: 28TB vs 55TB means constrained host-side memory.
• FP64 scientific computing gap: Rubin's FP64 at only 33 TFLOPS, lower than four-year-old Hopper. HPC customers may shift to AMD MI430X.
• Pricing pressure: $9.1M per rack faces customer resistance amid hyperscaler FCF turning negative (Google Q2 FCF first negative ever).

AMD Risks

• Production timeline: Helios first deployment Q4 2026 vs NVIDIA already in production. Q4 delay would impact Anthropic and OpenAI deployment plans.
• HBM4 supply: Same challenge as NVIDIA: most HBM4 capacity already committed to hyperscalers. AMD may face "design ready, no supply" dilemma.
• ROCm software maturity: Despite PyTorch improvements, specialized libraries, profiling tools, and custom kernels still assume CUDA as default. Full migration requires 12-18 months.
• Financial viability: $5B Anthropic investment + 2GW Helios deployment = massive upfront commitment. AMD datacenter GPU revenue must reach $20B+ to justify strategy sustainability.

Shared Industry Risks

• Energy constraints: TSMC consumes 9% of Taiwan's total electricity; natural gas depends on Strait of Hormuz transit. Geopolitical risks affect all chip suppliers.
• AI Capex sustainability: Six major hyperscalers' 2026 Capex totals $785B (+62% YoY), free cash flow compressed from 28% to 11%. If AI revenue growth underperforms, rack procurement pace may slow.
• Export control escalation: Both Rubin and Helios are subject to US export controls; Chinese market cannot access latest products.

---

6. Conclusion

Multi-Layer Significance

Technical level: NVIDIA Rubin and AMD Helios represent two fundamentally different AI infrastructure evolution paths: extreme integration versus open disaggregation. The advantages and disadvantages of both approaches will be validated through actual deployments from Q4 2026 to H1 2027. The triangular game of memory capacity (AMD advantage) versus software ecosystem (NVIDIA advantage) versus production pace (NVIDIA advantage) will determine market structure.

Enterprise value: For enterprise IT decision-makers, a genuine "dual-source strategy" choice emerges for the first time. For the past two years, NVIDIA was the only high-performance option; now AMD Helios offers a viable alternative for massive model and cost-efficiency scenarios. However, migration costs (ROCm adaptation, operations process reconstruction) remain the primary obstacle.

Investment perspective: NVIDIA currently trades at approximately 20x forward PE, near multi-year valuation lows. If Rubin production proceeds as planned (60,000+ NVL72 racks per year), valuation has room for recovery. If AMD's datacenter business successfully scales Helios, revenue could jump from $12B to $20B+, but execution risk remains elevated.

Key Predictions

Within 12 months (2026 Q4 - 2027 Q3):

• NVIDIA Rubin actual shipments approximately 1-1.5 million GPUs (below original 2M plan), constrained by HBM4 supply ✅High confidence
• AMD Helios completes Anthropic's first 1GW deployment, but full production delayed 1-2 quarters ⚠️High confidence
• Independent benchmarks will show Helios leading 15-30% on memory-intensive workloads, NVIDIA leading 10-20% on general training scenarios ⚠️Vendor claim

Within 24 months (2027 Q4 - 2028 Q3):

• Market structure shifts from NVIDIA 85% share to NVIDIA 70-75% + AMD 15-20% + Custom ASIC 10-15% ⚠️High confidence
• Next-gen NVIDIA Rubin Ultra and AMD MI500 will compete head-to-head in 2027, with performance gaps further narrowing ✅High confidence
• "Tokens per dollar" becomes standard procurement metric, replacing traditional FLOPS comparison ✅High confidence

Within 36 months (2028 Q4 - 2029 Q3):

• AI accelerator market enters "tri-polar + custom" structure; NVIDIA no longer holds absolute pricing power ⚠️High confidence
• Rack-level competition expands from GPU to complete systems (cooling, power, networking); Total Cost of Intelligence (TCI) replaces TCO as the standard metric ⚠️Vendor claim
• TSMC 1.4nm production drives next-generation iteration of Rubin Ultra and MI500 ⚠️High confidence