FP4 / ternary-weight architectures decouple AI capability from raw transistor density — embargoed nations maintain competitive development.

FP4 / ternary-weight architectures decouple AI capability from raw transistor density — embargoed nations maintain competitive development. | Ternary/FP4/sub-bit model production deployments

Key catalyst: Ternary/FP4/sub-bit model production deployments

Watch events: BitNet-class ternary deployment scale; FP2 / 1-bit model research milestones

FP4 training + INT8/INT4 inference now standard at Chinese labs. DeepSeek, Qwen, MiniMax all using sub-8-bit quantization in production.

Beneficiaries (24)

Adverse (6)

{
  "nia": false,
  "mode": "THESIS",
  "role": "Host-VC",
  "context": "Shift from FP32 to ternary weights reduces hardware requirements by >1 order of magnitude; proves AI progression not purely hardware-limited.",
  "to_year": 2029,
  "conv_cues": "mathematically decouple; proving",
  "direction": "HAPPEN",
  "from_year": 2026,
  "timeframe": "2026+",
  "conv_level": "HIGH",
  "milestones": [
    {
      "kind": "llm_pre_event",
      "label": "FP4 inference lands in mainstream open-source frameworks",
      "notes": "HIT — FP4 (NVFP4 + MXFP4) confirmed in llama.cpp by April 2026 per InsiderLLM. Decouples capability from FP16 transistor density.",
      "source": "https://insiderllm.com/guides/fp4-inference-llamacpp-nvfp4-mxfp4/ — FP4 already landed in llama.cpp by April 2026",
      "status": "hit",
      "weight": 0.4,
      "ordinal": -6,
      "source_id": null,
      "confidence": 0.99,
      "source_url": "https://insiderllm.com/guides/fp4-inference-llamacpp-nvfp4-mxfp4/",
      "expected_date": "2026-04-01",
      "observed_date": "2026-04-15",
      "research_origin": "deep_research",
      "expected_date_range": {
        "to": "2026-06-30",
        "from": "2026-01-01"
      },
      "measurement_criterion": "FP4 (NVFP4 or MXFP4) inference support added to llama.cpp, vLLM, or TensorRT-LLM with public benchmarks showing minimal accuracy loss vs FP16"
    },
    {
      "kind": "llm_pre_event",
      "label": "DeepSeek V4 KV cache quantization confirms architecture decoupling",
      "source": "https://dev.to/soytuber/deepseek-v4-flash-gemmaqwen-kv-cache-quantization-384k-context-2m0",
      "status": "hit",
      "weight": 0.4,
      "ordinal": -5,
      "source_id": null,
      "confidence": 0.95,
      "source_url": "https://dev.to/soytuber/deepseek-v4-flash-gemmaqwen-kv-cache-quantization-384k-context-2m0",
      "expected_date": "2026-04-24",
      "observed_date": "2026-04-24",
      "research_origin": "deep_research",
      "measurement_criterion": "DeepSeek V4-Flash demonstrates KV cache quantization at 384K context window with material memory reduction"
    },
    {
      "kind": "prereq",
      "label": "Nvidia became the world's first $5 trillion company (late 2025), operating a near-monopoly on advanced AI chips.",
      "status": "hit",
      "weight": 0.5,
      "ordinal": -4,
      "source_id": "SEM_011",
      "expected_date": "2026-04-29",
      "observed_date": "2026-04-29"
    },
    {
      "kind": "prereq",
      "label": "Nvidia Data Center revenue +66% YoY, contributing ~90% of $57B fiscal Q3 revenue; >$4.5T market cap entirely underpinned by AI silicon.",
      "status": "hit",
      "weight": 0.5,
      "ordinal": -3,
      "source_id": "SEM_027",
      "expected_date": "2026-04-29",
      "observed_date": "2026-04-29"
    },
    {
      "kind": "prereq",
      "label": "Nvidia's Arizona-based TSMC factory successfully fabricated cutting-edge semiconductors on US soil for first time in decades (October 2025).",
      "status": "hit",
      "weight": 0.5,
      "ordinal": -2,
      "source_id": "SEM_014",
      "expected_date": "2026-04-29",
      "observed_date": "2026-04-29"
    },
    {
      "kind": "prereq",
      "label": "Nvidia quadrupled chip production output while only doubling human headcount — achieved by deploying AI coding tools (Cursor, Claude Code) a",
      "status": "hit",
      "weight": 0.5,
      "ordinal": -1,
      "source_id": "SEM_012",
      "expected_date": "2026-04-29",
      "observed_date": "2026-04-29"
    },
    {
      "kind": "event",
      "label": "FP4 / ternary-weight architectures decouple AI capability from raw transistor density — embargoed nations maintain competitive development.",
      "status": "partial",
      "weight": 1,
      "ordinal": 0,
      "source_id": "SEM_022",
      "expected_date": "2026-05-01",
      "observed_date": "2026-05-01"
    },
    {
      "kind": "llm_pre_event",
      "label": "Ternary-weight production model released by 
... (truncated)