Version 17

What

Why it matters

Open questions

• L12 cluster validation — the first full compute cluster with scale-out networking — has not been reported for either CoreWeave or Microsoft; when does the first L12 milestone occur?

• Jensen said at COMPUTEX on June 1 that rack-level mass production had not yet begun [4]; when does it start, and does that gap create delivery risk against H2 2026 commitments?

• AgentPerf is produced by Artificial Analysis and promoted by NVIDIA [13][14] — does any independent body validate the benchmark methodology or run it on non-NVIDIA hardware?

• SemiAnalysis's Vera SOCAMM report drew fake-news accusations; SemiAnalysis defends it with physical evidence at the SK Hynix Computex booth [16] — what exactly does the report claim and what are the architectural or supply-chain implications?

Narrative

Timeline

• 2026-01-05: NVIDIA debuts Rubin chip at CES: 336 billion transistors, 50 petaflops AI performance. [32]
• 2026-01: Jensen Huang announces at CES 2026 that Vera Rubin NVL72 is in full production. [33][34]
• 2026-02: SK Hynix begins HBM4 mass production shipments to NVIDIA, holding approximately 70% of NVIDIA's HBM4 orders. [35][28]
• 2026-03-17: Nscale acquires 8GW Monarch Compute Campus in West Virginia; Microsoft signs 1.35GW LOI co-announced with NVIDIA and Caterpillar. [36][37][23][38]
• 2026-05: Samsung sells out entire 2026 HBM4 supply; rack prices reach $8.8M; shortage projected until 2028. [39][10][40][9]
• 2026-05: NVL72's 600kW per-rack power requirement documented as incompatible with existing data centers, requiring greenfield construction. [1][29][30]
• 2026-05: NVIDIA equity stake in Nscale confirmed at approximately $674M; Microsoft's Rubin GPU deployment via Nscale revised upward to 130,000 units. [41][31][24]
• 2026-05-18: First Vera CPUs hand-delivered to OpenAI, Anthropic, and other leading AI labs. [42][43][44]
• 2026-05-21: NVIDIA reports Q1 2026 earnings: $81.6B revenue, up 85% year-over-year. [17][45]
• 2026-05-21: GTC Taipei: Vera Rubin NVL72 wins Computex Best Choice Golden Award; Meta, Google Cloud, and Microsoft formalize partnerships; $2B NVLink Fusion investment in Marvell announced. [46][47][48][49][22][18]
• 2026-05-31: Dell delivers world's first fully validated Vera Rubin NVL72 rack to CoreWeave; L11 diagnostics confirm scale-up domain operational. [25][26][3][2][27][50][51]
• 2026-06-01: Jensen Huang at COMPUTEX announces Microsoft completed bring-up of first Rubin VR200 NVL72 via Foxconn; wafer-level production started, rack-level mass production not yet begun. [4]
• 2026-06-01: SemiAnalysis rates Jensen's COMPUTEX keynote 'F tier': no new AI datacenter products; Rubin FP4/FP8 gains documented at ~3.5x over GB200 while FP16 gains are ~1.6x. [20][19]
• 2026-06-04: Jensen Huang identifies agentic AI as a $200B market, framing Vera CPU as an active orchestration layer rather than a scheduler. [15]
• 2026-06-05: Bloomberg reports NVIDIA CEO certified all three HBM4 suppliers — Samsung, SK Hynix, and Micron — for Vera Rubin; Jensen describes the platform as 'in full production' during Seoul supply-chain visit. [11][5]
• 2026-06-05: NVIDIA open-sources Rubin NVSwitch Tray BoM; SemiAnalysis reports design includes AMD EPYC 3151 embedded CPUs — nine per VR200 rack. [21]
• 2026-06-08: Analyst Rohan Paul reports HBM4 market share: SK Hynix ~60-70%, Samsung ~25-30%, Micron remainder, all in full production for Vera Rubin; SemiAnalysis defends Vera SOCAMM report against fake-news accusations. [6][16]
• 2026-06-12: NVIDIA publishes AgentPerf benchmark results via Artificial Analysis: GB300 NVL72 runs up to 20x more agents per megawatt than H200 on agentic workloads. [13][14]

Perspectives

Tensions

• Jensen stated at COMPUTEX on June 1 that wafer-level production had started but rack-level mass production had not yet begun [4]; Bloomberg reported him describing Vera Rubin as 'in full production' four days later in Seoul [5][11] — the two statements use different production language and leave actual rack delivery timelines ambiguous. [4][11][5]
• NVIDIA promotes AgentPerf as the first agentic AI infrastructure benchmark and claims GB300 NVL72 delivers 20x more agents per megawatt than H200 [13]; the benchmark is produced by Artificial Analysis under NVIDIA direction with no independent validation. [13][14]
• NVIDIA markets Vera Rubin on a 10x cost-per-token reduction, but SemiAnalysis documents FP4/FP8 gains of ~3.5x over GB200 while FP16 gains are ~1.6x and HBM capacity is flat [19] — the efficiency claim is workload-specific, not uniform. [19]
• SemiAnalysis published a Vera SOCAMM note that critics called fake news; SemiAnalysis defends it by citing physical evidence at SK Hynix's Computex booth [16] — the substantive content remains incompletely described in available sources. [16]
• NVIDIA's promotional materials present the NVL72 as a NVIDIA-integrated system, but the open-sourced NVSwitch Tray BoM discloses AMD EPYC 3151 embedded CPUs — nine per VR200 rack [21] — a cross-vendor dependency absent from NVIDIA's public architecture descriptions. [21]
• NVIDIA holds approximately $674M in Nscale equity [31] while describing Nscale publicly as a commercial partner, making large deployment announcements co-publicized by both companies non-arm's-length transactions [23]. [31][23]