The Information Machine

NVIDIA Launches Vera CPU and Vera Rubin NVL72 at COMPUTEX / GTC Taipei

closed · v19 · 2026-06-19 · 301 items · history

What's new in v19

The new items this pass (30397, 30398, 31276, 31324, 31325) are social media amplifications of the OpenAI CFO / Fall 2026 Vera Rubin training run story already covered via item 30179 in the prior synthesis. Item 31324 is the SemiAnalysis Threads post that was the original source for that story; items 30398 and 31325 are downstream restatements. No new themes, voices, or factual claims have emerged this pass.

What

NVIDIA's Vera Rubin NVL72 has cleared rack-level (L11) validation at CoreWeave and Microsoft [3][4], and the CEO confirmed all three HBM4 suppliers — Samsung, SK Hynix, and Micron — are certified [5]. Two complications persist: SemiAnalysis reported NVIDIA halved SOCAMM memory in Rubin racks due to supply shortages [7], corroborated by SK Hynix preparing SOCAMM2 [10] and a Micron stock decline [11]; and SemiAnalysis publicly disputed OpenAI's CFO's stated plan of a Fall 2026 Vera Rubin training run, arguing the clusters and software stack won't be ready in time [12]. NVIDIA Blackwell swept all seven MLPerf Training 6.0 benchmarks on June 16, with GB300 NVL72 delivering 1.6x faster training than GB200 NVL72 [13].

Why it matters

The SOCAMM memory cut, if confirmed, is a platform specification downgrade affecting CPU-side orchestration capacity at the moment NVIDIA is positioning Vera CPU as the core of its agentic AI offering. The SemiAnalysis dispute with OpenAI's CFO sets up the first concrete public test of whether Vera Rubin can support frontier-scale training by late 2026, with implications for how quickly the industry can move off Blackwell-class infrastructure.

Open questions

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

  • SemiAnalysis argues Vera Rubin NVL72 clusters won't be stable enough for a frontier-scale training run by Fall 2026 [12]; does NVIDIA or OpenAI provide a counter-claim or revised timeline?

  • NVIDIA reportedly halved SOCAMM memory in Rubin racks due to supply shortages [7]; is this a permanent architectural change or a stopgap, and does SOCAMM2 [10] resolve it for production racks?

  • AgentPerf is produced by Artificial Analysis under NVIDIA direction [14]; does any independent body validate the methodology or run it on non-NVIDIA hardware?

Narrative

NVIDIA's Vera Rubin NVL72 is a high-density AI compute rack integrating 72 Rubin GPUs via NVLink at the rack level, requiring 600kW per rack [1]. Validation follows a hierarchy from L10 (single-server firmware) through L11 (single-rack scale-up domain) to L12 (full compute cluster with scale-out networking) [2]. Dell delivered the first fully validated VR200 NVL72 rack to CoreWeave on May 31, clearing L11 [3], and Microsoft completed bring-up of its first Rubin VR200 NVL72 via Foxconn on June 1 [4]. Neither operator has reported clearing L12. The NVIDIA CEO confirmed in Seoul on June 5 that Samsung, SK Hynix, and Micron are all certified HBM4 suppliers, with SK Hynix holding roughly 60-70% share, Samsung 25-30%, and Micron the remainder [5][6].

A separate memory subsystem — SOCAMM, the module serving the Vera CPU rather than the Rubin GPU's HBM4 — became the subject of a disputed SemiAnalysis report claiming NVIDIA halved its capacity in Rubin racks due to supply shortages [7]. NVIDIA initially called the report fake news; SemiAnalysis defended it with physical evidence from the SK Hynix Computex booth [8][9]. Korea Herald subsequently reported SK Hynix is preparing SOCAMM2 for Vera Rubin [10], and a decline in Micron stock followed the report's wider circulation [11]. These signals suggest the SOCAMM reduction is real, affecting CPU-side orchestration capacity rather than GPU throughput.

On June 16, SemiAnalysis flagged a public statement by OpenAI's CFO that the company's next major training run will occur in Fall 2026 on Vera Rubin hardware [12]. SemiAnalysis argues this timeline is implausible: Vera Rubin NVL72 clusters are unlikely to be sufficiently stable by then, and the software stack will not be mature enough for a frontier-scale training run [12]. No response from NVIDIA or OpenAI has been reported. The same day, NVIDIA Blackwell (GB300 NVL72) swept all seven MLPerf Training 6.0 benchmarks — the first time any platform was submitted across every benchmark — with GB300 NVL72 delivering up to 1.6x faster training than GB200 NVL72 [13]. CoreWeave trained DeepSeek-V3 671B at 8,192 GPUs in 2.02 minutes [13]. MLPerf is run by an industry consortium, giving it more methodological independence than AgentPerf, which was designed by Artificial Analysis under NVIDIA direction and released June 12 claiming 20x more agents per megawatt for GB300 NVL72 over H200 [14].

Enterprise deployment of Vera CPU is on a 2027 timeline. HPE AI Factory with NVIDIA announced on June 16 that Vera CPU will be available with HPE Private Cloud AI in 2027, framed as a purpose-built agentic orchestration component with Confidential Computing and governance integration [15]. NVIDIA also demonstrated a co-packaged optics switch with Lambda, positioning power efficiency in inter-GPU communication as a competitive dimension alongside raw compute performance [16].

Timeline

  • 2026-01: Jensen Huang announces at CES that Vera Rubin NVL72 is in full production; Rubin chip debuts with 336 billion transistors and 50 petaflops AI performance. [33][34][35]
  • 2026-02: SK Hynix begins HBM4 mass production shipments to NVIDIA, holding approximately 70% of NVIDIA's HBM4 orders. [36][27]
  • 2026-03-17: Nscale acquires 8GW Monarch Compute Campus in West Virginia; Microsoft signs 1.35GW LOI co-announced with NVIDIA. [37][38][24][39]
  • 2026-05: Samsung sells out entire 2026 HBM4 supply; rack prices reach $8.8M; shortage projected until 2028. [40][29][41][28]
  • 2026-05: NVL72's 600kW per-rack power requirement documented as incompatible with existing data centers, requiring greenfield construction. [1][30][31]
  • 2026-05: NVIDIA equity stake in Nscale confirmed at approximately $674M; Microsoft Rubin GPU deployment revised to 130,000 units. [42][32][25]
  • 2026-05-18: First Vera CPUs hand-delivered to OpenAI, Anthropic, and other leading AI labs. [43][44][45]
  • 2026-05-21: NVIDIA reports Q1 2026 earnings: $81.6B revenue, up 85% year-over-year. [17][46]
  • 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. [47][48][49][50][23][51]
  • 2026-05-31: Dell delivers world's first fully validated Vera Rubin NVL72 rack to CoreWeave; L11 scale-up domain confirmed operational. [26][52][3][2][53][54][55]
  • 2026-06-01: Microsoft completes first Rubin VR200 NVL72 bring-up via Foxconn at COMPUTEX; Jensen states wafer-level production started, rack-level mass production not yet begun. [4]
  • 2026-06-05: NVIDIA CEO confirms three HBM4 suppliers certified in Seoul; describes Vera Rubin as 'in full production'; NVSwitch Tray BoM open-sourced, disclosing nine AMD EPYC 3151 CPUs per VR200 rack. [18][5][21]
  • 2026-06-10: NVIDIA demonstrates co-packaged optics switch with Lambda for power-efficient inter-GPU data transfer. [16]
  • 2026-06-12: NVIDIA publishes AgentPerf benchmark results via Artificial Analysis: GB300 NVL72 runs up to 20x more agents per megawatt than H200. [14][56]
  • 2026-06: SemiAnalysis reports NVIDIA halved SOCAMM memory in Rubin racks due to supply shortages; SK Hynix readies SOCAMM2; Micron stock declines on the report. [7][10][11]
  • 2026-06-16: SemiAnalysis disputes OpenAI CFO's claim of a Fall 2026 Vera Rubin training run, arguing clusters and software stack won't be ready. [12]
  • 2026-06-16: NVIDIA Blackwell sweeps all seven MLPerf Training 6.0 benchmarks; GB300 NVL72 is 1.6x faster than GB200 NVL72; CoreWeave trains DeepSeek-V3 671B at 8,192-GPU scale in 2.02 minutes. [13]
  • 2026-06-16: HPE AI Factory with NVIDIA announces Vera CPU availability in HPE Private Cloud AI for 2027, with Confidential Computing and Agent Toolkit integration. [15]

Perspectives

NVIDIA / Jensen Huang

Q1 2026 earnings ($81.6B, +85% YoY) validate AI demand; CEO confirmed three-vendor HBM4 certification; Vera Rubin described as 'in full production'; AgentPerf frames GB300 as 20x more agent-efficient than H200; MLPerf Training 6.0 confirms GB300 NVL72 as the fastest training platform across all seven benchmarks.

Evolution: Strengthened: MLPerf Training 6.0 adds consortium-run independent validation alongside the NVIDIA-directed AgentPerf result.

[17][4][18][5][14][15][13]

SemiAnalysis

Vera Rubin FP16 gains are only ~1.6x over GB200 while HBM capacity is flat; COMPUTEX keynote rated F tier; NVSwitch BoM discloses AMD EPYC 3151; SOCAMM memory was halved due to supply shortages; OpenAI CFO's Fall 2026 Vera Rubin training claim 'doesn't add up' given cluster stability and software maturity.

Evolution: Extended: the SOCAMM cut report has gained corroborating signals, and the OpenAI CFO dispute extends SemiAnalysis's critical stance to deployment timelines.

[19][20][21][9][8][7][12]

OpenAI

CFO publicly stated the next major training run will occur in Fall 2026 on Vera Rubin hardware [12]; no response to SemiAnalysis's rebuttal has been reported.

Evolution: New voice in the thread; position is contested but undefended publicly.

[12][22]

Microsoft / Foxconn

First hyperscaler to complete Rubin VR200 NVL72 bring-up via Foxconn as ODM; also anchor customer via Nscale (130,000 GPUs, 1.35GW LOI for West Virginia).

Evolution: Consistent.

[23][24][25][4]

Dell / CoreWeave

Dell delivered the world's first fully validated Vera Rubin NVL72 rack to CoreWeave on May 31, clearing L11; CoreWeave subsequently achieved the fastest DeepSeek-V3 671B training time in MLPerf Training 6.0 at 8,192-GPU scale.

Evolution: Extended: CoreWeave's MLPerf result positions it as the leading operator for Blackwell-class infrastructure validation.

[26][3][13]

Memory and supply chain analysts

HBM4 supply is three-vendor with SK Hynix ~60-70%, Samsung ~25-30%, and Micron the remainder; SOCAMM supply appears constrained enough to have forced a memory cut in Rubin racks.

Evolution: Split: HBM4 supply picture improved with CEO-level confirmation; SOCAMM supply picture has worsened with the reported cut.

[27][28][29][6][5][7][11]

Data center infrastructure analysts

NVL72's 600kW per-rack power requirement is incompatible with existing data center infrastructure, establishing greenfield construction as a structural bottleneck independent of memory supply.

Evolution: Consistent.

[1][30][31]

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][18] — the two characterizations leave actual rack delivery timelines ambiguous. [4][18][5]
  • OpenAI's CFO states the next major training run will happen on Vera Rubin in Fall 2026 [12]; SemiAnalysis argues NVL72 clusters won't be sufficiently stable and the software stack won't be mature enough by then [12] — neither NVIDIA nor OpenAI has responded to the dispute. [12]
  • SemiAnalysis reports NVIDIA halved SOCAMM memory in Rubin racks due to supply shortages [7], a claim NVIDIA called fake news; SK Hynix preparing SOCAMM2 [10] and a Micron stock decline [11] provide corroborating signals, but NVIDIA has not confirmed the cut. [7][10][11]
  • NVIDIA promotes AgentPerf — produced by Artificial Analysis under NVIDIA direction — claiming 20x more agents per megawatt for GB300 NVL72 over H200 [14]; MLPerf Training 6.0, run by an industry consortium, placed the same platform at 1.6x faster training than GB200 NVL72 [13], showing how metric choice and benchmark provenance shape the apparent scale of generational gains. [14][13]
  • NVIDIA markets Vera Rubin on a 10x cost-per-token reduction; SemiAnalysis documents FP4/FP8 gains of ~3.5x over GB200 while FP16 gains are ~1.6x and HBM capacity is flat [19], making the efficiency claim workload-specific rather than uniform. [19]
  • NVIDIA holds approximately $674M in Nscale equity [32] while describing Nscale publicly as a commercial partner, making large deployment announcements co-publicized by both companies non-arm's-length transactions [24]. [32][24]

Status: active and growing

Sources

  1. [1] The Data Center Isn't Ready. NVIDIA's Vera Rubin platform ships in… — reactive:nvidia-vera-computex-launch
  2. [2] At L10 your Firmware/BIOS and OS works on a single server, at L11 a single rack or scale-up domain works, and then at L1… — SemiAnalysis Twitter (2026-05-31)
  3. [3] Notably, passing L11 diags means that this rack is up and running, including the IMEX channels on the NVL72 scale-up dom… — SemiAnalysis Twitter (2026-05-31)
  4. [4] BREAKING NEWS: JENSEN JUST ANNOUNCED MICROSOFT HAS FINISHED BRING UP ON THEIR FIRST RUBIN VR200 NVL72 RACK with their OD… — SemiAnalysis Twitter (2026-06-01)
  5. [5] Nvidia Clears Memory's Big Three for Vera Rubin HBM4 Supply — reactive:nvidia-vera-computex-launch
  6. [6] Nvidia just cleared the memory bottleneck significantly for Vera Rubin by qualifying HBM4 from Samsung, SK Hynix, and Mi… — Rohan Paul Twitter (2026-06-08)
  7. [7] NVIDIA Halves SOCAMM Memory in Rubin Racks Amid Supply Shortages / X — reactive:nvidia-vera-computex-launch
  8. [8] SemiAnalysis responds to Vera SOCAMM controversy: Critics haven't visited SK Hynix's Computex booth — reactive:nvidia-vera-computex-launch
  9. [9] Our Vera SOCAMM note is causing a bit of a stir. As always some folks are jumping to the wrong conclusions. Those saying… — SemiAnalysis Twitter (2026-06-08)
  10. [10] SK hynix readies SOCAMM2 for Nvidia's Vera Rubin — reactive:nvidia-vera-computex-launch
  11. [11] NVIDIA Rubin SOCAMM Memory Cut Triggers Market Panic and MU Decline | KuCoin — reactive:nvidia-vera-computex-launch
  12. [12] ALERT: OpenAI's CFO claims their next big training run will happen in Fall 2026 on Vera Rubin but that doesn't add up. R… — SemiAnalysis Twitter (2026-06-16)
  13. [13] Fastest, Largest, Strongest: NVIDIA Blackwell Sweeps MLPerf Training 6.0 — NVIDIA Blog (2026-06-16)
  14. [14] NVIDIA Blackwell Leads on First Agentic AI Infrastructure Benchmark — NVIDIA Blog (2026-06-12)
  15. [15] HPE AI Factory With NVIDIA Expands for the Era of Agents — NVIDIA Blog (2026-06-16)
  16. [16] Nvidia released this video of its photonics co-packaged optics (CPO) switch with Lambda. — Rohan Paul Twitter (2026-06-10)
  17. [17] NVIDIA just dropped $81.6B in Q1 revenue up 85% YoY 🤯 — reactive:nvidia-vera-computex-launch (2026-05-21)
  18. [18] Seoul Purpose: How NVIDIA and South Korea Are Building the Future of AI — NVIDIA Blog (2026-06-05)
  19. [19] for more details on Nvidia's VR NVL72 Oberon and future roadmap, check out our article from February: — SemiAnalysis Twitter (2026-05-31)
  20. [20] F TIER KEYNOTEMAX: Jensen ComputeX presentation was one of the worst keynotes he has done. He announced nothing new on t… — SemiAnalysis Twitter (2026-06-01)
  21. [21] BREAKING NEWS: NVIDIA HAS JUST OPEN SOURCED THEIR RUBIN NVSWITCH TRAY BoM & DIAGRAM & IT INCLUDES AMD EYPC 3151 … — SemiAnalysis Twitter (2026-06-05)
  22. [22] Tips Excel on X: "OpenAI's Next AI Training Run Powered by NVIDIA Vera Rubin In a June 2026 episode, Friar announced OpenAI's major fall AI model training will use NVIDIA's Vera Rubin platform, now in full production with Rubin GPUs and advanced networking for complex AI tasks. She highlighted" / X — reactive:nvidia-vera-computex-launch
  23. [23] Microsoft's strategic AI datacenter planning enables seamless, large ... — reactive:nvidia-vera-computex-launch
  24. [24] Nscale acquires 8GW Monarch Compute Campus, Microsoft signs on for 1.35GW of compute - DCD — reactive:nvidia-vera-computex-launch
  25. [25] 130,000 Rubin GPUs Are Being Deployed at Nscale For Microsoft, Further Showing Massive Interest In NVIDIA's Next-Gen AI Chips — reactive:nvidia-vera-computex-launch
  26. [26] BREAKING NEWS: COREWEAVE & DELL IS THE FIRST CLOUD TO ANNOUNCE THAT THEY HAVE RUBIN VR200 NVL72 WITH FULLY PASSING L… — SemiAnalysis Twitter (2026-05-31)
  27. [27] SK Hynix Secures 70% of Nvidia's HBM4 Orders - Semicon — reactive:nvidia-vera-computex-launch
  28. [28] SK Hynix Surges 15% to New High: HBM Shortage Until 2028, How Much Longer Can AI Memory King Rise? — reactive:nvidia-vera-computex-launch
  29. [29] Nvidia's memory costs soar 485%, latest AI systems now cost $7.8 ... — reactive:nvidia-vera-computex-launch
  30. [30] NVIDIA Vera Rubin: 600kW Racks by 2027 | Introl Blog — reactive:nvidia-vera-computex-launch
  31. [31] Nvidia's Vera Rubin GPU: Redesigning Data Centres for 600kW Racks — reactive:nvidia-vera-computex-launch
  32. [32] UK AI Infrastructure Startup Nscale Receives $674 Million (£500 ... — reactive:nvidia-vera-computex-launch
  33. [33] Nvidia CEO confirms Vera Rubin NVL72 is now in production — reactive:nvidia-vera-computex-launch
  34. [34] NVIDIA Vera Rubin AI Platform Hits Full Production CES 2026 ... — reactive:nvidia-vera-computex-launch
  35. [35] Nvidia debuts Rubin chip with 336B transistors and 50 petaflops of AI performance - SiliconANGLE — reactive:nvidia-vera-computex-launch
  36. [36] SK Hynix set to ship HBM4 for Nvidia's Vera Rubin this month — reactive:nvidia-vera-computex-launch
  37. [37] Nscale and Microsoft Announce Collaboration with NVIDIA and Caterpillar to Deliver 1.35GW of NVIDIA Vera Rubin NVL72 GPUs at Flagship AI Factory Campus in West Virginia — reactive:nvidia-vera-computex-launch
  38. [38] Nscale and Microsoft Announce Collaboration with NVIDIA and Caterpillar to Deliver 1.35GW of NVIDIA Vera Rubin NVL72 GPUs at Flagship AI Factory Campus in West Virginia — reactive:nvidia-vera-computex-launch
  39. [39] Nscale acquisition includes plan to build AI facility in Mason County — reactive:nvidia-vera-computex-launch
  40. [40] Samsung sells out of 2026 HBM4 supply as memory resurgence ... — reactive:aws-garman-a100-demand
  41. [41] Price of Nvidia's Vera Rubin NVL72 racks skyrockets to as much as $8.8 million apiece, but server makers' margins will be tight — Nvidia is moving closer to shipping entire full-scale systems — reactive:nvidia-vera-computex-launch
  42. [42] Nvidia-backed UK AI firm Nscale raises $1.1 billion funding round — reactive:nvidia-vera-computex-launch
  43. [43] Vera Arrives: NVIDIA’s First CPU Built for Agents Lands at Top AI Labs — NVIDIA Blog (2026-05-18)
  44. [44] NVIDIA hand-delivers first 1.2 TB/s Vera CPUs to OpenAI, Anthropic ... — reactive:nvidia-vera-computex-launch
  45. [45] Nvidia unveils details of new 88-core Vera CPUs positioned to compete with AMD and Intel – new Vera CPU rack features 256 liquid-cooled chips that deliver up to a 6X gain in CPU throughput | Tom's Hardware — reactive:nvidia-vera-computex-launch
  46. [46] "Demand has gone parabolic. The reason is simple: Agentic AI has arrived." — reactive:nvidia-vera-computex-launch (2026-05-21)
  47. [47] NVIDIA GTC Taipei at COMPUTEX: Live Updates on What’s Next in AI — NVIDIA Blog (2026-05-21)
  48. [48] NVIDIA Vera Rubin NVL72 wins Computex 2026 awards for AI ... — reactive:nvidia-vera-computex-launch
  49. [49] Meta Builds AI Infrastructure With NVIDIA — reactive:nvidia-vera-computex-launch
  50. [50] NVIDIA GTC 2026: Google Cloud Deepens Partnership for AI ... — reactive:nvidia-vera-computex-launch
  51. [51] The CEO of NVIDIA, looked at Matt Murphy and said "The next trillion dollar company, ladies and gentlemen." (Save this). — Milk Road AI Twitter (2026-06-02)
  52. [52] Dell just made history this weekend and it is the culmination of an execution streak that no other company in enterprise… — Milk Road AI Twitter (2026-05-31)
  53. [53] CoreWeave Completes Industry-First Bring-Up And Validation Of NVIDIA Vera Rubin NVL72 — reactive:nvidia-vera-computex-launch
  54. [54] HPCwire - Since 1987 – Covering the Fastest Computers in the World and the People Who Run Them — reactive:nvidia-vera-computex-launch
  55. [55] CoreWeave Completes Industry-First Bring-Up and Validation of NVIDIA Vera Rubin NVL72 - Las Vegas Sun News — reactive:nvidia-vera-computex-launch
  56. [56] NVIDIA just posted the first agentic AI benchmark results where GB300 NVL72 runs up to 20x more coding agents per megawa… — Rohan Paul Twitter (2026-06-12)