NVDA

Analyst DD (Substack)

• 8Substack · damnang·BULLISH / EXECUTION·32 days agopost

  The post contends that SemiAnalysis framed a possible Kyber rack-format slip as a catastrophic generational delay, when NVIDIA's roadmap commitments are made at the chip-generation level (Rubin Ultra, 2H 2027), not by rack format. The author judges the midplane-yield concern unlikely to translate into a 12-month delay and views the episode as, at most, a rack-mix adjustment (Oberon vs Kyber) that keeps revenue inside NVIDIA.

  Evidence

  • Kyber's official schedule is 2H 2027; a '>12 month' delay would mean late 2028/2029, which the author says is conflated with a mere slip into early 2028
  • NVIDIA is the only company with production-proven rack-scale scale-up, having deployed NVL72 since late 2024
  • Per-rack power has climbed from ~40 kW (Hopper) to 120-130 kW (Blackwell) to 190-230 kW (Rubin), with Kyber at the 600 kW class fitting 144 GPU packages
  • The VR200 NVL72 shipping this year already uses a midplane-based, cable-free compute-tray design; Jensen Huang said at CES 2026 it cut tray assembly from two hours to five minutes
  • The 78-layer midplane (three laminated 26-layer boards) is described as a yield-defense design allowing known-good sub-boards to be tested and joined
  • Midplane is a passive component; ~50% yield only doubles board cost on a rack costing millions, so poor yield constrains volume/cost rather than the schedule
  • NVL72x2 was never officially announced by NVIDIA and existed only in supply-chain reporting
  • Precedent exists for copper crossing the rack boundary (GB200 NVL36x2), and NPO offers an intermediate optical path
  Catalysts

  • Taiwan high-layer-count PCB and CCL suppliers' monthly revenue disclosures (~10th of each month) in 2H 2026 as a diagnostic signal for the Kyber-vs-Oberon rack mix
  • Rubin ramping in 2H 2026; Kyber and Rubin Ultra slated for 2H 2027
  Risks

  • Midplane yield at ~1 square meter and 78 layers with sub-25-micrometer trace spacing remains unproven at scale
  • If Kyber slips, 2027 rack mix tilts toward Oberon-based Rubin/Rubin Ultra racks
  • Flagship NVL576 multi-rack optical scale-up could ship late or in small early volume
  • Roadmaps and specifications reflect announced plans and may change
• 7Substack · damnang·NEUTRAL / MOAT·38 days agopost

  Nvidia is the reference incumbent the 12 inference startups are trying to solve bottlenecks around more cheaply. Rather than being displaced, it is defending and absorbing, having taken a non-exclusive license to Groq's inference technology and brought over key people including founder Jonathan Ross.

  Evidence

  • Dec 2025: Nvidia secured a non-exclusive license to Groq's inference technology, reported at roughly $20B, and brought over founder Jonathan Ross and key people
  • Groq's latency optimization is now an element used inside Nvidia's own inference stack
  • Nvidia takes most CoWoS advanced-packaging capacity, constraining new entrants that rely on HBM
  • The contest is framed as 'who solves, more cheaply, the bottleneck Nvidia is solving expensively'
  Catalysts

  • Continued licensing/absorption of inference-startup technology and talent
  • How hyperscalers and startups press the inference market from above and below
  Risks

  • Low-cost DRAM camp (d-Matrix, Tenstorrent) deliberately routes around HBM and CoWoS where Nvidia dominates
  • Specialized designs target inference workloads Nvidia does not cover efficiently, eroding utilization advantages
• 6Substack · damnang·BEARISH / MOAT·69 days agopost

  The author frames NVIDIA's GPU architecture as structurally advantaged over Cerebras's wafer-scale design: hardware support for multiple precisions plus the CUDA software ecosystem lets rapid software innovation convert directly into performance, while Cerebras's fixed FP16 hardware is 'forever chasing one generation behind.' Even where Cerebras wins on raw decode speed, NVIDIA retains workload portability and dominates training and frontier-scale serving.

  Evidence

  • NVIDIA published 1,038 tokens/s for DGX B200 on Llama 4 Maverick vs Cerebras's 2,522 tokens/s, but under different conditions (batch size, precision, speculative decoding)
  • HBM bandwidth climbs each generation: H100 ~3.35 TB/s, H200 4.8 TB/s, B200 8 TB/s
  • HBM capacity scaling fast: H200 141GB, B200 ~180GB per GPU, projected past 1TB by 2028
  • NVIDIA added dedicated FP8 Tensor Cores starting with H100 and extended to FP4 with Blackwell, cutting compute and power
  • Major inference frameworks (vLLM, FlashAttention, Triton, CUTLASS, xformers) are all built on CUDA; any model/framework runs on NVIDIA
  • All major frontier labs train on NVIDIA GPUs, Google TPUs, or their own chips; the $20B OpenAI Cerebras contract is inference only
  • On AWS Bedrock, Trainium handles prefill and Cerebras is used only as a decode accelerator
  Catalysts

  • NVIDIA already moving to FP4 (Blackwell) while Cerebras would need 2-3 years to add FP8 in WSE-4
  • HBM capacity projected past 1TB by 2028, widening the capacity gap vs Cerebras's ~44GB SRAM
  Risks

  • Cerebras sits at the top tier for inference decode speed on public benchmarks (2,522 vs 1,038 tokens/s)
  • Cerebras signed a >$20B multi-year OpenAI compute deal and an AWS Bedrock deployment
  • Benchmark comparison conditions differ, so the speed gap may not be apples-to-apples