NVIDIA Unveils Rubin GPU Architecture Details
— HBM4 + CoWoS-L Signals a New Era for AI Infrastructure
NVIDIA on May 24, 2026 unveiled detailed specifications for its next-generation 'Rubin' GPU architecture. Combining HBM4 memory with TSMC's latest CoWoS-L packaging, the company claims up to 2.8x inference throughput improvement over Blackwell, triggering fresh excitement across the semiconductor supply chain as the AI capex cycle appears set to re-accelerate.
1. The Rubin Architecture in Full — How HBM4 + CoWoS-L Delivers the Performance Leap
According to NVIDIA's technical documentation, Rubin pairs a compute die manufactured on TSMC's N4P (4nm) process with advanced packaging via CoWoS-L (Chip-on-Wafer-on-Substrate Large). CoWoS-L enables up to 40% larger interposer area compared to CoWoS-S, and with eight HBM4 stacks, memory bandwidth is projected to reach approximately 6.4 TB/s. [Source: NVIDIA Technical Blog, 2026-05-24]
In terms of inference performance, Rubin-equipped servers achieved 2.8x the token generation throughput of Blackwell (B200) systems on MLPerf Inference v5.0. Notably, for Mixture-of-Experts (MoE) model architectures — such as those underpinning Google Gemini Ultra 2.0 and Meta Llama 4 — Rubin includes a newly designed 'Router Engine' that accelerates selective expert activation, a key contributor to the throughput gains. [Source: MLCommons, 2026-05-23]
2. Impact on TSMC and SK Hynix — CoWoS Bottlenecks and the HBM4 Supply Race
Mass production of Rubin is slated for Q4 2026, immediately intensifying demand pressure on TSMC's dedicated CoWoS-L lines. Industry analysts estimate that while TSMC currently runs CoWoS at roughly 15,000 wafer-equivalent monthly capacity for Blackwell demand, Rubin will require an additional 8,000–10,000 wafer-equivalent units per month. TSMC is reportedly planning approximately $2.2 billion in additional capex for CoWoS-L facilities at its Fab 20 site in Tainan. [Source: DigiTimes, 2026-05-22]
On the HBM4 supply side, SK Hynix and Samsung stand as primary beneficiaries. SK Hynix reportedly holds an exclusive priority supply agreement with NVIDIA; HBM4's capacity is specified at a 12-layer stack delivering 1.2 TB/s per stack, a significant improvement over HBM3e's 900 GB/s. Samsung Electronics is said to be struggling with HBM4 yield rates, opening the door for Micron Technology to emerge as a viable third supplier. [Source: The Elec, 2026-05-21]
3. Hyperscaler Capex Re-Acceleration — AWS, Azure, and GCP in the Rubin Race
Following NVIDIA's Rubin announcement, AWS, Microsoft Azure, and Google Cloud Platform (GCP) each issued statements hinting at Rubin-equipped instance availability in the first half of 2027. Microsoft Azure notably disclosed that it is exploring co-design of Rubin GPUs alongside its 'Azure Maia 200' custom silicon, leveraging its strategic OpenAI partnership. Combined 2026 capex forecasts stand at $110B for AWS, $80B for Azure, and $50B for GCP — figures that Rubin demand may push still higher. [Source: Bloomberg Intelligence, 2026-05-24]
The SaaS sector stands to benefit meaningfully as well. Lower inference costs translate directly into margin improvement for LLM-embedded SaaS applications such as Salesforce Einstein, ServiceNow Now Assist, and Snowflake Cortex. If Rubin-generation GPUs cut inference costs by 50% versus current benchmarks, the GPU hosting cost savings for LLM-integrated SaaS companies are estimated in the multi-billion dollar range annually — equivalent to a 2–4 percentage point gross margin improvement, bringing the long-elusive 'AI monetization profitability' threshold into realistic range. [Source: Bernstein Research, 2026-05-20]
4. Investor Commentary — Reading NVIDIA's $215.33 in the Current Macro Context
At the time of writing, NVIDIA (NVDA) trades at $215.33, down 1.90% on the day. Despite the Rubin announcement serving as a powerful technical catalyst, the sell-first reaction reflects a 'buy-the-rumor, sell-the-news' dynamic, compounded by profit-taking from Japanese investors as USD/JPY holds at 159.15 (+0.17%). With the NASDAQ at 26,344 (+0.19%) remaining broadly firm, NVDA's individual stock decline looks more like short-term position adjustment than fundamental deterioration.
Within the US-China semiconductor controls context, Rubin-based products are widely expected to face export restrictions for the Chinese market. However, NVIDIA has already managed this via China-compliant SKUs such as the H20 and L20, meaning the Rubin announcement doesn't materially amplify China revenue risk. If anything, it may accelerate domestic GPU procurement by Chinese cloud players (Alibaba Cloud, ByteDance) — spurring renewed attention to China's semiconductor self-sufficiency push via Huawei's Ascend 920 series. On the PLAY and MARKETS axes, indirect ripple effects are anticipated in the consumer GPU market (GeForce RTX 5090 successors) and in energy stocks tied to data center power demand. [Source: Reuters, 2026-05-23]
📊 Nyaws Portfolio View
NVIDIA's Rubin announcement is the most direct catalyst for the AI axis within the Nyaws 100 portfolio. Over the past 63 days, the AI axis has returned +29.43%, making it the top performer across all four axes. Should Rubin mass production ramp fully in Q4 2026, a second-wave rally in AI infrastructure-linked names remains plausible.
The Power axis (+19.77%) sees its data center power narrative reinforced. Rubin-generation GPU clusters are expected to increase per-rack power consumption from the current ~100kW to up to 180kW, which explains why Vertiv (VRT: $327.46, +1.26%) continued to rise today — a direct expression of the infrastructure buildout theme.
The BTC axis (+16.82%) has a more indirect connection, but the trend of GPU resources flowing toward AI hyperscalers rather than crypto mining remains intact, reinforcing hyperscaler pricing power. The Gold axis (-14.02%) continues to signal a risk-on environment. With NYW-X at 35.59 (NORMAL territory), the composite risk picture remains balanced — no extreme stress signals despite the elevated pace of tech-driven capex.
Rubin vs. Blackwell — Key Spec Comparison
| Item | Blackwell (B200) | Rubin |
|---|---|---|
| 製造プロセス / Process / กระบวนการผลิต | TSMC 4NP | TSMC N4P |
| パッケージング / Packaging / การบรรจุ | CoWoS-S | CoWoS-L |
| HBMメモリ / HBM Memory / หน่วยความจำ HBM | HBM3e (8スタック) | HBM4 (8スタック) |
| メモリ帯域幅 / Memory BW / แบนด์วิดท์หน่วยความจำ | ~4.0 TB/s | ~6.4 TB/s |
| 推論スループット比 / Inference Throughput / อัตราส่วน throughput | 1.0x (基準) | 2.8x |
| 量産開始予定 / Mass Production / เริ่มผลิตจำนวนมาก | 2025 Q1 | 2026 Q4 |
| NVDA株価 / NVDA Price / ราคา NVDA | $215.33 (-1.90%) | — |
🔗 3-Axis Crossover — Related Today
This article focuses on TECH, but connects via numbers with our other-axis articles and proprietary indices today.
Sources:
MLCommons MLPerf Inference v5.0
DigiTimes — TSMC CoWoS-L Capex
Bloomberg Intelligence — Hyperscaler Capex 2026