QUANTA · EVIDENCE WALL

The world is moving toward Noeton

We don't need to convince you. We archive industry events against the five principles — and let the trend speak. Click a badge to filter.

  1. 2026-01 P2 · Dataflow compute

    HBM4 ramps with custom base dies

    Logic dies fabbed on advanced nodes become standard in HBM4 — "compute moves toward memory" enters the DRAM product roadmap itself.

  2. 2025-12 P1 · Memory-centric

    Mooncake merges into vLLM v1

    Its Transfer Engine becomes the official KV connector for P/D disaggregation, then joins the PyTorch ecosystem — KV pooling becomes open infrastructure.

  3. 2025-12 P3 · Model-OS

    MCP donated to the Linux Foundation (AAIF)

    Private protocol to neutral governance in 13 months — the fastest standardization of a syscall surface in computing history.

  4. 2025-11 P1 · Memory-centric

    Samsung & SK hynix standardize LPDDR6-PIM

    The two DRAM giants jointly standardize low-power PIM; SK's roadmap pairs PIM with CXL by ~2028 — the exhibit-to-product divide gets paved.

  5. 2025-09 P5 · Intent interface

    Computer-use agents become products

    Anthropic/OpenAI computer-use agents reach public beta; HarmonyOS Intents Kit registers app capabilities as system-level intents.

  6. 2025-09 P2 · Dataflow compute

    NVIDIA Rubin CPX: a prefill-only GPU

    30 PFLOPS deliberately paired with GDDR7 instead of HBM, decode left to its HBM sibling — the incumbent productizes "two physics, two chips".

  7. 2025-08 P1 · Memory-centric

    Montage samples CXL 3.1 memory controller

    PCIe 6.2 ×8, dual-channel DDR5-8000; endorsed by Samsung/AMD/Intel — a Chinese company in the first tier of the CXL ecosystem.

  8. 2025-06 P1 · Memory-centric

    Huawei CloudMatrix 384: UnifiedBus supernode

    NPUs/CPUs/DRAM/SSDs peer-connected without CPU mediation; ~3.6× memory capacity and ~2.1× bandwidth vs a comparable GPU rack. Rack-scale pooled memory, in production.

  9. 2025-03 P3 · Model-OS

    OpenAI adopts MCP; Google follows

    The biggest rival adopts a competitor's protocol — N×M adapter pain beats ecosystem moats.

  10. 2025-03 P4 · Tiered cognition

    NVIDIA ships Dynamo

    GTC 2025: the vendor's own "disaggregated inference OS" — KV-aware routing across prefill/decode pools. Phase splitting becomes first-party.

  11. 2025-02 P1 · Memory-centric

    Mooncake wins FAST 2025 Best Paper

    The top storage venue crowns "organize the cluster around the cache" — the academy endorses the memory-centric route.

  12. 2024-11 P3 · Model-OS

    Anthropic releases MCP

    Typed schemas unify the model × tool surface — "schema is the ABI" goes from analogy to open protocol.

  13. 2024-09 P0 · Premise

    Reasoning models move compute to decode

    o1 opens the test-time-compute era (R1 open-sources it): token budgets shift massively to decode — enlarging the bandwidth-bound share of inference.

  14. 2024-06 P1 · Memory-centric

    Mooncake: KVCache-centric architecture published

    Tsinghua & Moonshot publish Kimi's serving platform: the cluster reorganized around the KV cache, with P/D split and pooled idle DRAM/SSD.

  15. 2024-06 P4 · Tiered cognition

    Apple Intelligence + Private Cloud Compute

    An ~3B on-device floor with escalation to Apple-silicon cloud — a two-tier degraded-mode design shipped to hundreds of millions of devices.

The paper version of this wall is §12 "Industry Trajectory", new in v2.

Inclusion bar: publicly verifiable product launches, awarded papers, standardization events — no rumors, no roadmap slideware. Continuously updated.