Upgrade to Pro — share decks privately, control downloads, hide ads and more …

From Stack to Fluid: A Hypothesis for Architect...

Sponsored · Your Podcast. Everywhere. Effortlessly. Share. Educate. Inspire. Entertain. You do you. We'll handle the rest.
Avatar for ICHICHI ICHICHI
August 04, 2026

From Stack to Fluid: A Hypothesis for Architecture in the AI Era from a NW Perspective

A Note from an Architect specialized in Cloud & AI

Avatar for ICHICHI

ICHICHI

August 04, 2026

More Decks by ICHICHI

Other Decks in Research

Transcript

  1. JAWS-NW LT From Stack to Fluid : A Hypothesis for

    Architecture in the AI Era from a NW Perspective A Note from an Architect specialized in Cloud & AI 2026.08.04 | 叶 奕池 JAWS-NW LT 2026.08.04
  2. 01 About the presenter 叶 奕池/ Yichi.YE Architect l Enterprise

    Cloud & Applied AI l キンドリルジャパン株式会社 ME:Generate a Pic of me in animal figure ME:Generate a Pic of me in your image GPT: Claude: - Cloud SME specialized in Enterprise Hybrid Cloud & Applied AI - AI Architecting Research Lead l Kyndryl Japan & Global Architect Community - Talent Development Program Core l Kyndryl Japan AWS Community - 2025-2026 Japan AWS Top Engineer - JAWS-NW Core Member Contact : - LinkedIn:Yichi.YE - X:YICHI @ichichi_0417 Company Official Shoot Above the Noise:Think Deep. Build Long. JAWS-NW LT 2026.08.04 01
  3. About Today 02 The Scope & Logic 02 Stack vs

    Flow 構造から振る舞いへ 04 05 06 07 Missing Question Capacity Fungibility Agent-Driven Flow Diagnostic Question Fluid Architecture 問いの出発点 AWS Researchから借りた視点 AIで変わるFlow Capacity Designの問い A Working Hypothesis 01 03 Hub-Spoke 予測可能性の前提 本日お話する内容は私個人の見解であり、私が所属する会社・組織の立場、戦略、意見を代表するものではありません。 JAWS-NW LT 2026.08.04 02
  4. 03 Practical Insight “ AI Systemが変わり、 新たなAI Capabilityが求められるとき、 Network Architectureは本当に今までのままでよいのだろうか。

    The Missing Question: AWS re:InventやAWS Summit では、AI関連の新しいサービスが毎年発表される。 ネットワーク領域も例外ではない。 しかし、こうしたアップデートを個別に追う議論からは、一つの問いが抜け落ちがちである。 JAWS-NW LT 2026.08.04 03
  5. 04 Architecture Lens Stack vs Flow & Hub-Spoke STACK Hub-Spoke

    FLOW Application VPC A Data Database SHIFT Storage Context Feedback Compute Action Inference VPC B Transit Gateway VPC C VPC D Network Structure Behavior Traffic Pattern is reasonably predictable. Layer / Component / Dependency Movement / Runtime / Circulation Application構成から、必要な経路とCapacityを事前に見積もれる。 Question What happens when system behavior becomes more dynamic than ever before. SystemのBehaviorが、これまで以上に動的になるとき、従来のArchitectureの前提は、そのままでよいのだろうか。 JAWS-NW LT 2026.08.04 04
  6. 05 Research Lens Capacity Fungibility: Evidence from Research SOURCE DESIGN

    PRINCIPLE RNGはAI System向けに設計されたものではない。ここでは設計原則のみを借りる。 “ The root problem is that fat trees Capacity should be fungible. lack capacity fungibility※1. FIGURE 1 | topology contrast (recreated) 必要な場所に、 Fat-tree Flat Expander 利用可能なCapacityを柔軟に届くこと。 T1 T10 S1 T2 S2 Pre-Research Core Thought T9 A1 A2 T3 Traffic Patternの予測が難しくなるほど、 A3 small cut T8 T1 T2 T3 T4 T4 Capacityを固定経路に閉じ込めない設計の価値が高まる。 T5 T7 T5 T6 capacity can be stranded many alternate paths JAWS-NW LT 2026.08.04 ※1:RNG: Flat Datacenter Networks at Scale, Bernardi et al. | Amazon Web Services | arXiv:2604.15261 05
  7. 06 Agent-Driven Flow From Predictable Traffic to Agent-Driven Flow INFERENCE

    Agent selects the Flow at Runtime. EXAMPLE Sakana Fugu “A multi-agent system delivered as one model※2” ONE LOGICAL INTERFACE Model A 従来のClient–Server通信とは異なり、 Agentは状況に応じてツールを選択し、Modelを切り替え、 複数のSub-agentを並行して呼び出す。 OpenAI-compatible Endpoint Fugu Orchestrator Agent B Sub-agent NETWORK IMPLICATION Tool C runtime selection / switching / parallel invocation 誰が、どこへ、どれだけ通信するかを、 設計時に固定的に見積もることが難しくなる。 STABLE OUTSIDE Logical Interface JAWS-NW LT 2026.08.04 DYNAMIC INSIDE ≠ Internal Flow Topology ※2:Sakana AI, "Fugu," GitHub repository, 2026. Available: https://github.com/SakanaAI/fugu (product page: https://sakana.ai/fugu/) 06
  8. 07 From Observation to Conclusion From Observation to Conclusion OBSERVATION

    01 OBSERVATION 02 OBSERVATION 03 Stack explains Structure. Capacity should be fungible. Agent selects Flow at Runtime. AI Systemでは、Stack Structureだけでなく、 Traffic Patternの予測が難しいほど、 Logical Interfaceが安定していても、 Flow/Behaviorを捉える必要がある。 Capacityを固定経路に閉じ込めない価値が高まる。 Internal Flow Topologyは動的に変化する。 CURRENT CONCLUSION Network may need to evolve from a static communication layer to a resilient Flow Fabric. Networkは、Stack最下層の静的な通信レイヤーから、System境界を越えるFlowを支えるResilientなFlow Fabricへと進化する必要があるのではないか。 CURRENT CONCLUSION ≠ FINAL ANSWER 現時点の結論であり、検証・修正・棄却され得るWorking Hypothesis。 JAWS-NW LT 2026.08.04 07
  9. 08 Working Hypothesis Fluid Architecture CENTRALIZED GOVERNANCE GUARDRAILS + AUTONOMY

    CONTROL TOWER Centralizedな統制と Policy / Security / Trust / Boundary Distributedな判断を両立する。 Centralized Guardrails RESILIENT FLOW FABRIC FLOW OVER LAYERS Structureだけでなく、 RESILIENCE THROUGH ADAPTABILITY Data Context Inference Action Feedback 固定Patternではなく、 Circulationを設計する。 変化するFlowを受け止める。 continuous circulation across system boundaries Actor A Actor B Actor C Agent / Region Model / Account Tool / Service DISTRIBUTED DECISION-MAKING JAWS-NW LT 2026.08.04 08
  10. Thank you 叶 奕池 l Architect l Enterprise Cloud &

    Applied AI From Stack to Fluid: A Hypothesis for Architecture in the AI Era from NW Perspective A Note from an Architect specialized in Cloud & AI