NVIDIA Unveils Blackwell Ultra (GB300)
— A New Chapter in AI Inference Infrastructure
NVIDIA has officially detailed its next-generation Blackwell Ultra (GB300) GPU, boasting up to 4× inference throughput over the H100 and 288 GB of HBM3e memory, signaling the dawn of a new AI infrastructure investment cycle.
1. GB300 Specifications — What Changed and by How Much
Unveiled on May 20, 2026, the Blackwell Ultra (GB300) delivers up to 40% higher FP8 inference performance over the previous Blackwell B200, with HBM3e memory expandable to 288 GB via NVLink. Single-GPU TDP reaches 1,200W, making direct liquid cooling virtually mandatory. [Source: NVIDIA Official Press Release, 2026-05-20]
Particularly noteworthy is the adoption of NVLink Switch Gen5, enabling up to 576 GB300 units to be combined in a single cluster. This makes single-cluster inference of ultra-large models (10-trillion-parameter scale) feasible for the first time — something impossible with the H100 generation. Architectural changes include 5th-generation Tensor Cores and native FP4 format support. [Source: NVIDIA GTC 2026 Tech Brief]
2. TSMC and HBM Supply Chain — CoWoS Bottleneck Risk Returns
The GB300 is manufactured on TSMC's 3nm process (N3E) and uses CoWoS-S packaging. The CoWoS capacity shortfall that plagued the B200 generation risks re-emerging in H2 2026 as GB300 ramps up. Industry analysts estimate monthly CoWoS capacity at approximately 120,000 wafers in Q3–Q4 2026, which could fall 20–30% short of current demand backlog. [Source: TechInsights Supply Chain Report, 2026-05]
On the HBM supply side, SK Hynix maintains its position as NVIDIA's primary supplier. The company has ramped mass production of 8-High HBM3e (24 GB/stack) in Q2 2026 to support GB300's 288 GB configuration. Samsung continues to struggle with yield improvements, further deepening reliance on SK Hynix. Micron is accelerating capex for entry into the market from late 2026 onward. [Source: Bloomberg Tech, 2026-05-19]
3. Hyperscaler Response — Microsoft, Google, and Amazon Race to Procure
Microsoft (Azure), Google (GCP), and Amazon (AWS) have all reportedly secured priority procurement contracts for GB300, with GPU-related spending set to account for an even greater share of their 2026 capex plans. Microsoft plans over $80 billion in total datacenter capex for full-year 2026 (+40% YoY), with approximately 60% earmarked for GPUs and AI accelerators. [Source: Microsoft Investor Day 2026]
Google, meanwhile, continues to blend GB300 with its own TPU v6 (Trillium), positioning the NVIDIA chip specifically for low-latency LLM inference workloads. AWS announced it will begin GB300 UltraCluster deployments in Q3 2026. These moves signal a symbolic shift: the era of training-dominated AI infrastructure is giving way to an era defined by inference cost optimization. [Source: AWS re:Invent Preview, Reuters 2026-05-18]
4. Cross-Axis View — Ripple Effects on Play and Markets
On the Play axis, NVIDIA's GB300 announcement may also influence the GeForce RTX 6000 series consumer roadmap. If a die-shrunk Blackwell Ultra variant reaches gaming GPUs, it could reshape the next-gen gaming PC market in H2 2026 to early 2027. Major game publishers like Nintendo — heavily reliant on NVIDIA Tegra-based SoCs — stand to benefit indirectly from the generational AI processor leap.
On the Markets axis, NVDA shares traded at $220.61 (-0.77% day-on-day), a modest decline that likely reflects some degree of 'priced-in' expectations ahead of the GB300 announcement. WTI crude's steep drop to $104.03 (-4.26%) offers potential relief on datacenter power cost outlooks. With USD/JPY at 159.02 and yen weakness persisting, Japanese semiconductor names like Shin-Etsu Chemical and Tokyo Electron benefit from elevated dollar-denominated revenues. Tomorrow's FOMC minutes (April) could shift discount rates for growth stocks and SaaS valuations — a key variable to monitor.
📊 Nyaws Portfolio View
Today's GB300 announcement is a direct event for the AI infrastructure risk axis within NYW-X v1.0. With NYW-X currently at 34.13 (NORMAL range), the index reflects a balance between supply-side risk (CoWoS tightness) and robust demand-side momentum.
Looking at Nyaws 100 performance, the AI axis leads with a 63-day return of +21.90%, driven by heavy weighting in NVIDIA and TSMC-related names. As GB300 enters full-scale mass production, this trend has strong potential to persist.
The Power axis at +20.56% is also notable. GB300's 1,200W TDP further amplifies demand for datacenter liquid cooling and power infrastructure. Names like Vertiv (VRT) remain in focus despite today's pullback to $322.63 (-5.03%) — the structural growth theme remains intact.
The Gold axis (-10.21%) shows classic underperformance in a risk-on environment, with a clear inverse correlation to the AI and Power axes. The BTC axis (+15.86%) remains resilient amid macro noise; BTC/USD at $76,910 partly reflects heightened investor interest in digital assets driven by AI infrastructure enthusiasm. Should NYW-X rise due to supply-side risks like CoWoS tightness, a reassessment of the portfolio's AI axis weighting may become warranted.
Today's Data
| Item | Value |
|---|---|
| NVDA 株価 | $220.61 (-0.77%) |
| GB300 推論性能（対H100比） | +4倍 / ×4 vs H100 |
| GB300 HBM3e容量 | 288 GB |
| GB300 TDP | 1,200 W |
| CoWoS月産見込み (Q3-Q4 2026) | ~120,000 枚 |
| SK Hynix HBM3e 8-High容量 | 24 GB/スタック |
| Microsoft データセンターCapex 2026 | $80B超 (+40% YoY) |
| WTI原油 | $104.03 (-4.26%) |
| USD/JPY | 159.02 (+0.10%) |
| VRT 株価 | $322.63 (-5.03%) |
🔗 3-Axis Crossover — Related Today
This article focuses on TECH, but connects via numbers with our other-axis articles and proprietary indices today.
Sources: