NVIDIA Blackwell Ultra Ramps Up
— HBM4 Tightness and TSMC CoWoS Expansion Signal a New AI Infrastructure Cycle
Multiple supply chain sources indicate that NVIDIA's next-generation GPU, Blackwell Ultra (B300), is on track for full-scale mass production in the second half of Q2 2026. Simultaneous HBM4 supply constraints from SK Hynix and TSMC's CoWoS-L capacity race are converging to drive the next phase of AI data center investment.
1. What Is Blackwell Ultra — Technical Evolution from B100
Blackwell Ultra (B300) is positioned as an enhanced iteration of NVIDIA's current Blackwell architecture (B100/B200). Industry analyst estimates suggest that FP8 inference performance will be approximately 1.4x that of the B200, while the adoption of HBM4 is expected to deliver up to 1.8x memory bandwidth compared to the previous-generation HBM3e. The GPU is manufactured on TSMC's 3nm (N3P) process and achieves higher inter-die connectivity density through CoWoS-L (Chip-on-Wafer-on-Substrate Local) technology. [Source: DigiTimes / TechInsights estimates]
The most critical change is in the memory configuration. The B300 is expected to feature 12 stacks of HBM4, bringing total HBM capacity to approximately 288 GB — roughly 3.6x that of the H100 (80 GB). This substantially alleviates KV cache constraints during large language model (LLM) inference. As generative AI applications increasingly shift toward long-context processing (100K–1M tokens), this memory capacity expansion represents a genuine technical threshold breakthrough.
2. HBM4 Supply-Demand Reality — The Three-Way Battle: SK Hynix, Samsung, and Micron
At present, only SK Hynix has begun mass shipping HBM4, with Samsung and Micron targeting full-scale shipments in Q3 and Q4 2026 respectively. SK Hynix's HBM4 production capacity is estimated at approximately 6 million units (8Hi equivalent) for full-year 2026, which is insufficient to meet NVIDIA B300 demand on its own. As a result, multiple reports indicate that hyperscalers including Microsoft, Google, and Amazon have been accelerating long-term agreements (LTAs) to secure supplier allocations ahead of schedule. [Source: The Elec / Bloomberg]
What differentiates HBM4 from current HBM3e is manufacturing complexity. HBM4 narrows the TSV pitch by approximately 30% compared to previous generations and expands the I/O bandwidth to 1,024 bits. This requires more time to stabilize yield, and effective HBM4 supply in the first half of 2026 is likely running 15–20% below initial industry expectations. This is one reason why NVIDIA's ASP (average selling price) remains elevated.
3. TSMC's CoWoS-L Expansion — The Advanced Packaging Race in H2 2026
TSMC is reportedly planning to expand its CoWoS advanced packaging capacity to approximately 35,000 wafers per month (12-inch equivalent) by end of 2026 — roughly 2.3x the level at end-2024 — according to multiple reports. NVIDIA is estimated to account for 60–65% of total allocation, with the remainder split among AMD (MI400X series), Google (TPU v7), and Amazon (Trainium3). [Source: TrendForce / DigiTimes, May 2026]
CoWoS-L uses a localized interposer design compared to conventional CoWoS-S, improving cost efficiency while maintaining tight coupling between the GPU and multiple HBM stacks. Industry estimates suggest that CoWoS packaging costs account for approximately 8–10% of the total cost of a B300-equipped NVL72 rack, placing it alongside HBM4 (estimated at 22–23%) as a key cost driver. Capacity tightness across the supply chain provides a structural basis for NVIDIA's gross margins to remain elevated over the next two quarters.
4. Ripple Effects on SaaS and Cloud — Falling Inference Costs Reshaping the AI App Economy
As Blackwell Ultra transitions into mass production, AI cloud inference costs are estimated to fall a further 30–40% compared to the current Blackwell generation. OpenAI and Anthropic had already achieved a roughly 90% reduction in inference costs over the prior 18 months, and B300's full deployment will push this trend to the next stage. For AWS, Azure, and GCP, this represents an opportunity to deliver higher-performance services at the same cost, while substantially lowering the barrier for SaaS vendors — including Salesforce, ServiceNow, and Snowflake — to implement agentic AI features across their platforms. [Source: ARK Invest estimates / Anthropic public materials]
[Cross-Axis: Markets & Play] From a macro perspective, USD/JPY is trading at 158.93 (-0.07%) and the S&P 500 stands at 7,446 (+0.17%), reflecting a firm tone. While elevated rate concerns persist, the structural decline in AI inference costs positively revises the forward free cash flow outlook for growth stocks, providing a logical foundation for capital inflows into SaaS and AI infrastructure equities. NVDA is down at $219.51 (-1.77%) today, which can be interpreted as a short-term reaction to front-loaded production costs in the Blackwell Ultra ramp. On the Play axis, game engines such as Unreal Engine and Unity are also set to benefit from falling inference costs, potentially accelerating the commercialization of interactive NPCs and procedural content generation.
5. US-China Regulatory Risk and the Rise of DSA (Distributed System Architecture)
ja
en
th
📊 Nyaws Portfolio View
Today's NYW-X (cross-risk index) is holding at 34.28 (NORMAL zone), with no excessive risk-off or risk-on signals. The Blackwell Ultra ramp-up news is acting as a tightening force on sentiment, but with NVDA stock itself down 1.77%, there appears to be a short-term divergence between near-term noise and medium-term fundamentals.
Within the Nyaws 100 portfolio, the AI axis leads with a 63-day return of +25.52%. HBM4 and CoWoS tightness continue to underpin the fundamentals of AI infrastructure-related names (NVDA, SK Hynix, TSMC), providing a structural basis for the AI axis to maintain its leadership position. The Power axis (+18.68%) also remains firm on the back of expanding data center power demand, a trend that Blackwell Ultra's production ramp is set to reinforce.
On the other hand, the Gold axis (-11.02%) remains under pressure from persistently elevated real interest rates, and the stability of NYW-X at 34.28 reflects a market environment in which risk sentiment is concentrated in AI-related assets. The strength of the BTC axis (+14.32%) signals continued risk appetite for crypto assets, and the three-axis structure of AI, Power, and BTC as the primary portfolio return drivers is likely to persist for the near term.
Vertiv (VRT) posting a notable gain of +2.45% to $323.40 today is also worth highlighting. The Blackwell Ultra production ramp directly implies expanded demand for liquid cooling and power management infrastructure in data centers, placing power and cooling infrastructure stocks like VRT in a position to indirectly capture semiconductor supply chain tailwinds. This is also why the Power axis of the Nyaws 100 remains firm in tandem with the AI axis.
Today's Key Data (May 22, 2026)
| Item | Value |
|---|---|
| NVDA | $219.51 (-1.77%) |
| VRT (Vertiv) | $323.40 (+2.45%) |
| NASDAQ | 26,293 (+0.09%) |
| S&P 500 | 7,446 (+0.17%) |
| USD/JPY | 158.93 (-0.07%) |
| WTI原油 | $98.00 (-0.26%) |
| GOLD | $4,544/oz (+0.28%) |
| BTC/USD | $77,716 (+0.33%) |
| HBM4供給（2026年通年推計） | ~600万個（8Hi換算） |
| TSMCのCoWoS月産目標（2026年末） | ~35,000枚（12インチ換算） |
| B300 HBMキャパシティ | ~288GB（12スタック×HBM4） |
| B300 vs H100 メモリ容量比 | 約3.6倍 |
🔗 3-Axis Crossover — Related Today
This article focuses on TECH, but connects via numbers with our other-axis articles and proprietary indices today.
Sources:
DigiTimes (TSMC CoWoS expansion)
TrendForce (CoWoS capacity estimates)
TechInsights (B300 performance estimates)