NVIDIA GPU. is a massively parallel processor and surrounding platform used for graphics, AI training and inference, scientific computing, simulation, media, and data analytics. Data-center generations such as A100, H100 and B200 combine streaming multiprocessors, tensor cores, high-bandwidth memory, large caches, RAS, secure execution features, and high-speed scale-up links. The useful product is the GPU plus module, baseboard, network, system, firmware, CUDA stack, libraries, and deployment tooling—not a die in isolation. Semiconductor economics couple very large fixed commitments to uncertain product demand. Architecture, software, verification, masks, process qualification, factories, equipment, substrates, packaging capacity, test time, and inventory must be funded before lifetime volume is known. At the leading edge, design and mask nonrecurring expense can reach hundreds of millions of dollars, while a greenfield logic fab can require well above ten billion dollars and years to ramp. Mature nodes remain economically important because analog, RF, power, embedded memory, display, sensor, connectivity, and control functions do not automatically benefit from maximum transistor density. Revenue therefore depends on product mix, wafer starts, die area, yield, package complexity, utilization, pricing, customer concentration, and the timing of replacement cycles—not merely nominal node.

Business model, market position, and economics. NVIDIA’s strategic advantage combines silicon cadence with CUDA compatibility, optimized libraries, compilers, frameworks, networking, reference systems, and developer reach. That ecosystem reduces time to working performance and raises switching cost. GPU demand is mediated by foundry wafers, advanced packaging, HBM, substrates, networking, power, cooling, and datacenter construction. Accelerator price is therefore only one part of total cluster cost, and availability of complete systems can matter more than nominal chip production. Competitive advantage accumulates across reusable IP, talent, design methodology, process recipes, yield history, packaging know-how, developer tools, customer relationships, standards, and installed software. These assets reinforce one another but also create switching costs and concentration risk. A strong product can still lose if its toolchain is difficult, supply is constrained, total system cost is poor, or customers cannot qualify it in time. Conversely, an older node or architecture can remain attractive when it is stable, available, inexpensive, security-qualified, and supported for a decade. Roadmaps should be read as directional commitments; production readiness requires design kits, working silicon, repeatable yield, capacity, packaging, and customer shipments.

Technology, product architecture, and implementation. An SM schedules warps across scalar, vector, tensor, load/store, special-function, register, shared-memory, and cache resources. Tensor cores accelerate supported matrix types and sparsity modes; the memory hierarchy rewards coalescing, reuse, tiling, and overlap. HBM supplies enormous bandwidth but remains far slower than on-chip storage. NVLink and NVSwitch provide scale-up connectivity, while InfiniBand or Ethernet provides scale-out. Collective communication, topology, CPU and NIC placement, storage, and checkpointing determine distributed efficiency. A credible comparison starts at the workload and system boundary. Peak arithmetic, core count, transistor count, or process label alone says little about useful performance. Engineers examine sustained throughput, tail latency, memory capacity and bandwidth, cache behavior, interconnect topology, I/O, precision support, compiler maturity, power envelopes, cooling, reliability, security, serviceability, and software portability. For process and manufacturing choices they add density by circuit type, voltage range, SRAM scaling, analog behavior, design rules, IP readiness, yield learning, reticle limits, packaging, and qualification. Published specifications are usually conditional on product configuration and workload, so normalized measurements and clear test conditions matter.

Execution, supply chain, and engineering risk. Peak low-precision tensor numbers depend on data type, sparsity, clocks, and operation definition. H100 and B200 variants differ in form factor, memory capacity, power, interconnect, and cooling; compare the exact SKU and system. Models may be capacity-bound by weights and KV cache, bandwidth-bound by token generation, communication-bound during training, or compute-bound in dense matrix phases. Utilization, batching, precision, parallelism, kernel fusion, compiler support, and reliability recovery dominate economics. The operating system behind a shipped chip spans architecture, RTL, verification, physical design, signoff, tapeout, mask preparation, wafer fabrication, probe, assembly, final test, firmware, drivers, libraries, system validation, and field support. A schedule slip in one layer can idle investment elsewhere. Capacity reservations, long-lead equipment, substrate allocation, export controls, geographic concentration, single-source materials, and qualified second sources shape resilience. Quality systems must connect inline process data to wafer sort, package test, board behavior, and field returns. Change control is especially strict for automotive, industrial, medical, aerospace, infrastructure, and other products with long service lives.

<svg viewBox="0 0 760 470" xmlns="http://www.w3.org/2000/svg" font-family="-apple-system,Segoe UI,Roboto,sans-serif">
  <rect x="0" y="0" width="760" height="470" fill="#0d1117"/>
  <text x="380" y="28" fill="#e6edf3" font-size="21" font-weight="700" text-anchor="middle">NVIDIA — The AI Compute Stack, Chip to Datacenter</text>
  <text x="380" y="48" fill="#8b98a5" font-size="12" text-anchor="middle">GPU silicon + interconnect + systems + software = the dominant AI training platform</text>

  <!-- === LEFT: GPU Die Generations (stacked view) === -->
  <rect x="30" y="62" width="220" height="280" rx="6" fill="#080d14" stroke="#233043" stroke-width="1.2"/>
  <text x="140" y="80" fill="#e6edf3" font-size="11" text-anchor="middle" font-weight="600">GPU Generations</text>

  <!-- Blackwell B200 -->
  <rect x="45" y="90" width="190" height="42" rx="4" fill="#122040" stroke="#76b900" stroke-width="1.4"/>
  <text x="140" y="108" fill="#76b900" font-size="10" text-anchor="middle" font-weight="600">Blackwell B200 (2024)</text>
  <text x="140" y="122" fill="#8b98a5" font-size="8.5" text-anchor="middle">208B transistors, 2 dies (TSMC 4NP), 4.5 PFLOPS FP8</text>

  <!-- Hopper H100 -->
  <rect x="45" y="138" width="190" height="42" rx="4" fill="#122040" stroke="#60a5fa" stroke-width="1.2"/>
  <text x="140" y="156" fill="#93c5fd" font-size="10" text-anchor="middle" font-weight="600">Hopper H100 (2022)</text>
  <text x="140" y="170" fill="#8b98a5" font-size="8.5" text-anchor="middle">80B transistors, TSMC 4N, 989 TF FP16</text>

  <!-- Ampere A100 -->
  <rect x="45" y="186" width="190" height="38" rx="4" fill="#14202c" stroke="#8b98a5" stroke-width="1"/>
  <text x="140" y="202" fill="#8b98a5" font-size="9.5" text-anchor="middle">Ampere A100 (2020)</text>
  <text x="140" y="216" fill="#6b7684" font-size="8" text-anchor="middle">54B, TSMC 7N, 312 TF FP16</text>

  <!-- Volta V100 -->
  <rect x="45" y="230" width="190" height="34" rx="4" fill="#14202c" stroke="#6b7684" stroke-width="0.8"/>
  <text x="140" y="246" fill="#6b7684" font-size="9" text-anchor="middle">Volta V100 (2017)</text>
  <text x="140" y="258" fill="#6b7684" font-size="8" text-anchor="middle">21B, TSMC 12nm, first Tensor Cores</text>

  <!-- Pascal P100 -->
  <rect x="45" y="270" width="190" height="30" rx="4" fill="#14202c" stroke="#6b7684" stroke-width="0.6"/>
  <text x="140" y="288" fill="#6b7684" font-size="8.5" text-anchor="middle">Pascal P100 (2016) — HBM2 debut</text>

  <!-- Scaling arrow -->
  <line x1="240" y1="110" x2="240" y2="280" stroke="#76b900" stroke-width="0.8" stroke-dasharray="3,2"/>
  <text x="246" y="200" fill="#76b900" font-size="8" transform="rotate(90,246,200)">~4x perf/gen</text>

  <!-- Die photo style indication -->
  <text x="140" y="320" fill="#6b7684" font-size="8.5" text-anchor="middle">each gen: +transistors, +HBM BW, +TF</text>
  <text x="140" y="335" fill="#6b7684" font-size="8" text-anchor="middle">die size: 400→800+ mm²</text>

  <!-- === CENTER: System scale-up === -->
  <rect x="265" y="62" width="235" height="280" rx="6" fill="#080d14" stroke="#233043" stroke-width="1.2"/>
  <text x="382" y="80" fill="#e6edf3" font-size="11" text-anchor="middle" font-weight="600">System Hierarchy</text>

  <!-- Single GPU module -->
  <rect x="280" y="92" width="70" height="32" rx="3" fill="#122040" stroke="#60a5fa" stroke-width="1"/>
  <text x="315" y="112" fill="#93c5fd" font-size="8" text-anchor="middle">GPU + HBM</text>
  <text x="360" y="108" fill="#8b98a5" font-size="8">SXM module</text>

  <!-- NVLink connecting 8 GPUs -->
  <rect x="280" y="132" width="200" height="55" rx="4" fill="#0b1524" stroke="#76b900" stroke-width="1.2"/>
  <text x="380" y="148" fill="#76b900" font-size="9" text-anchor="middle" font-weight="600">DGX / HGX (8 GPUs)</text>
  <!-- 8 small GPU blocks -->
  <g fill="#122040" stroke="#3b82f6" stroke-width="0.5">
    <rect x="288" y="154" width="18" height="14" rx="1"/><rect x="310" y="154" width="18" height="14" rx="1"/>
    <rect x="332" y="154" width="18" height="14" rx="1"/><rect x="354" y="154" width="18" height="14" rx="1"/>
    <rect x="376" y="154" width="18" height="14" rx="1"/><rect x="398" y="154" width="18" height="14" rx="1"/>
    <rect x="420" y="154" width="18" height="14" rx="1"/><rect x="442" y="154" width="18" height="14" rx="1"/>
  </g>
  <!-- NVSwitch -->
  <rect x="330" y="172" width="80" height="12" rx="2" fill="#1a2a0a" stroke="#76b900" stroke-width="0.6"/>
  <text x="370" y="181" fill="#76b900" font-size="7" text-anchor="middle">NVSwitch (900 GB/s per GPU)</text>

  <!-- DGX SuperPOD -->
  <rect x="280" y="196" width="200" height="40" rx="4" fill="#0b1524" stroke="#a78bfa" stroke-width="1"/>
  <text x="380" y="212" fill="#c4b5fd" font-size="9" text-anchor="middle" font-weight="600">SuperPOD (32 DGX = 256 GPUs)</text>
  <text x="380" y="226" fill="#8b98a5" font-size="8" text-anchor="middle">InfiniBand 400G interconnect</text>

  <!-- Full cluster -->
  <rect x="280" y="244" width="200" height="40" rx="4" fill="#0b1524" stroke="#fbbf24" stroke-width="1"/>
  <text x="380" y="260" fill="#fbbf24" font-size="9" text-anchor="middle" font-weight="600">AI Supercomputer (100k+ GPUs)</text>
  <text x="380" y="274" fill="#8b98a5" font-size="8" text-anchor="middle">Meta, xAI, OpenAI clusters</text>

  <!-- Scale labels -->
  <text x="280" y="300" fill="#6b7684" font-size="8">1 GPU: 4.5 PF (B200)</text>
  <text x="280" y="312" fill="#6b7684" font-size="8">8 GPUs: 36 PF (one node)</text>
  <text x="280" y="324" fill="#6b7684" font-size="8">256 GPUs: 1.1 EF (SuperPOD)</text>
  <text x="280" y="336" fill="#6b7684" font-size="8">100k GPUs: ~450 EF (frontier training)</text>

  <!-- === RIGHT: Software stack === -->
  <rect x="515" y="62" width="215" height="195" rx="6" fill="#0b1220" stroke="#233043" stroke-width="1"/>
  <text x="622" y="80" fill="#e6edf3" font-size="11" text-anchor="middle" font-weight="600">Software Moat</text>

  <!-- Stack layers -->
  <rect x="530" y="90" width="185" height="18" rx="2" fill="#14261f" stroke="#34d399" stroke-width="0.8"/>
  <text x="622" y="103" fill="#6ee7b7" font-size="9" text-anchor="middle">PyTorch / JAX / TensorFlow</text>

  <rect x="530" y="114" width="185" height="18" rx="2" fill="#1a1520" stroke="#a78bfa" stroke-width="0.8"/>
  <text x="622" y="127" fill="#c4b5fd" font-size="9" text-anchor="middle">cuDNN / cuBLAS / NCCL</text>

  <rect x="530" y="138" width="185" height="18" rx="2" fill="#122040" stroke="#60a5fa" stroke-width="0.8"/>
  <text x="622" y="151" fill="#93c5fd" font-size="9" text-anchor="middle">CUDA (parallel runtime)</text>

  <rect x="530" y="162" width="185" height="18" rx="2" fill="#2a1a0a" stroke="#f59e0b" stroke-width="0.8"/>
  <text x="622" y="175" fill="#fbbf24" font-size="9" text-anchor="middle">GPU driver + PTX ISA</text>

  <rect x="530" y="186" width="185" height="18" rx="2" fill="#122040" stroke="#76b900" stroke-width="1"/>
  <text x="622" y="199" fill="#76b900" font-size="9" text-anchor="middle">GPU hardware (SM + Tensor Cores)</text>

  <text x="622" y="222" fill="#8b98a5" font-size="8.5" text-anchor="middle">15+ years of CUDA ecosystem</text>
  <text x="622" y="236" fill="#8b98a5" font-size="8.5" text-anchor="middle">= billions of lines of GPU code</text>
  <text x="622" y="250" fill="#6b7684" font-size="8" text-anchor="middle">moat: developers locked in</text>

  <!-- Business numbers -->
  <rect x="515" y="265" width="215" height="80" rx="6" fill="#0b1220" stroke="#233043" stroke-width="1"/>
  <text x="622" y="283" fill="#e6edf3" font-size="10" text-anchor="middle" font-weight="600">By the Numbers (2025)</text>
  <text x="535" y="300" fill="#76b900" font-size="9.5">Market cap: ~3T USD</text>
  <text x="535" y="314" fill="#8b98a5" font-size="9">Data center rev: ~100B USD/yr</text>
  <text x="535" y="328" fill="#8b98a5" font-size="9">AI training GPU share: ~90%</text>
  <text x="535" y="342" fill="#6b7684" font-size="8.5">Fab: TSMC (100% outsourced)</text>

  <!-- Bottom -->
  <text x="380" y="378" fill="#e6edf3" font-size="10" text-anchor="middle" font-weight="600">Key insight: NVIDIA sells systems, not just chips — GPU + NVLink + NVSwitch + CUDA + libraries = full stack lock-in</text>
  <text x="380" y="396" fill="#8b98a5" font-size="9.5" text-anchor="middle">Competitors: AMD (MI300X), Intel (Gaudi), Google (TPU), startups (Cerebras, Groq, SambaNova) — all fighting CUDA moat</text>

  <!-- Interconnect detail -->
  <text x="380" y="425" fill="#76b900" font-size="9.5" text-anchor="middle">NVLink 5 (Blackwell): 1.8 TB/s GPU-to-GPU — faster than any CPU memory bus</text>
  <text x="380" y="440" fill="#6b7684" font-size="9" text-anchor="middle">Every large AI model trained since 2020 ran on NVIDIA GPUs — the picks-and-shovels of the AI gold rush</text>

  <text x="380" y="462" fill="#6b7684" font-size="11" text-anchor="middle">NVIDIA made GPUs the CPU of AI — then built the entire system around them so nobody can leave.</text>
</svg>

Evaluation, roadmap discipline, and CFS connection. Procurement should benchmark representative models at target sequence length, batch, quality, precision, and latency service level. Include tokens per second, time to train, energy, rack density, network, memory headroom, checkpoint time, failure recovery, software licensing, support, and expected model evolution. Roadmap names such as Rubin are forward-looking until exact products, configurations, availability, and measured workloads are established. Due diligence separates measured facts from marketing categories and forward-looking plans. Check the date, product form factor, memory configuration, power limit, software release, process variant, package, and whether a number is peak, typical, estimated, or independently reproduced. Company revenue rankings and foundry shares move with cycles, currency, reporting boundaries, and whether wafer manufacturing or end-product sales are counted. Procurement adds total landed cost, supply assurance, licensing terms, support, lifecycle, compliance, and exit options. Engineering teams should preserve traceable assumptions and revisit them when a roadmap, regulation, yield curve, or workload changes. CFS connects this topic to semiconductor architecture, implementation, verification, manufacturing, packaging, test, and deployed AI-system tradeoffs across the platform.