Executive Summary: AWS AI Infrastructure

AWS is the first vendor in this assessment series that makes a credible claim across every layer of the 4+1 model, including Layer 2C. The structural difference between AWS and every on-prem vendor (Dell, HPE, VAST) is the direction of authority. On-prem vendors build upward from hardware, attempting to extend authority into orchestration and runtime layers. AWS builds downward from managed services, extending authority into custom silicon (Trainium, Inferentia, Graviton), custom networking (EFA/SRD, Nitro), and now on-prem infrastructure (AWS AI Factories).

The DAPM classification for AWS is structurally inverted compared to on-prem vendors. The enterprise architect using AWS retains less direct authority at every layer, but gains operational leverage that on-prem vendors cannot match. The question is not whether AWS has the capabilities. The question is whether the enterprise architect has made the authority delegation explicit, and whether they understand what borrowed judgment they inherit when they adopt AWS’s Reasoning Plane as their own.

AWS already has the control plane everyone else is trying to build. The problem is that customers do not always see where AWS’s control plane ends and their own authority begins. When Bedrock routes inference, SageMaker auto-scales, or Karpenter provisions nodes, those are Layer 2C functions operating invisibly inside managed services. This is the ‘DGX Realization’ that birthed the 4+1 model: the cloud operates an invisible Reasoning Plane that becomes visible only when you try to replicate it on bare metal.

The OpenAI partnership (2GW of Trainium capacity, Stateful Runtime on Bedrock) and NVIDIA deepened collaboration (1M+ GPUs including Blackwell and Rubin) demonstrate AWS positioning as the substrate on which multiple AI ecosystems converge, creating one of the broadest Layer 3 ecosystems and one of the most complex borrowed judgment landscapes.

Layer-by-layer status: Layer 0A (Custom Silicon Plus the Broadest NVIDIA Fleet, Rented by the Instance), Layer 0B (EFA and SRD, an AI Fabric AWS Designed), Layer 0C (FSx for Lustre and S3 Express One Zone Feed the Accelerators), Layer 1A (Glue, Lake Formation and SageMaker Catalog Over Open Tables), Layer 1B (Managed Retrieval Portfolio), Layer 1C (Managed Pipelines (Spark/Airflow Substrates)), Layer 2A (EKS + Capacity Management), Layer 2B (Bedrock + Open Paths), Layer 2C (Intelligence-2C Toolkit; Placement Implicit), Layer 3 (+1) (Broadest Ecosystem).

Assessment framework: 4+1 Layer AI Infrastructure Model. Scoring model: Decision Authority Placement Model (DAPM) — Retained, Delegated, or Ceded. Published by The CTO Advisor LLC (DBA The Advisor Bench). Author: Keith Townsend. Date assessed: October 27, 2026. Version: v1.40 - 4+1 v3: Layer 0 Sub-Layers.

AWS AI Infrastructure

Mapped to the 4+1 Layer AI Infrastructure Model

v1.40 - 4+1 v3: Layer 0 Sub-LayersAssessed October 27, 2026Sources & revision history
ACTIVE ASSESSMENT

Summary Finding

AWS is the first vendor in this assessment series that makes a credible claim across every layer of the 4+1 model, including Layer 2C. The structural difference between AWS and every on-prem vendor (Dell, HPE, VAST) is the direction of authority. On-prem vendors build upward from hardware, attempting to extend authority into orchestration and runtime layers. AWS builds downward from managed services, extending authority into custom silicon (Trainium, Inferentia, Graviton), custom networking (EFA/SRD, Nitro), and now on-prem infrastructure (AWS AI Factories).

The DAPM classification for AWS is structurally inverted compared to on-prem vendors. The enterprise architect using AWS retains less direct authority at every layer, but gains operational leverage that on-prem vendors cannot match. The question is not whether AWS has the capabilities. The question is whether the enterprise architect has made the authority delegation explicit, and whether they understand what borrowed judgment they inherit when they adopt AWS’s Reasoning Plane as their own.

AWS already has the control plane everyone else is trying to build. The problem is that customers do not always see where AWS’s control plane ends and their own authority begins. When Bedrock routes inference, SageMaker auto-scales, or Karpenter provisions nodes, those are Layer 2C functions operating invisibly inside managed services. This is the ‘DGX Realization’ that birthed the 4+1 model: the cloud operates an invisible Reasoning Plane that becomes visible only when you try to replicate it on bare metal.

The OpenAI partnership (2GW of Trainium capacity, Stateful Runtime on Bedrock) and NVIDIA deepened collaboration (1M+ GPUs including Blackwell and Rubin) demonstrate AWS positioning as the substrate on which multiple AI ecosystems converge, creating one of the broadest Layer 3 ecosystems and one of the most complex borrowed judgment landscapes.

Strength
Moderate
Gap
Partner
Layer 0A · ComputeCompute2 LABSCustom Silicon Plus the Broadest NVIDIA Fleet, Rented by the Instancedecides: vendor · Ceded▼

CPUs, accelerators, servers, integrated systems, OS and hypervisor, physical plant

Vendor-Provided

Graviton (Arm CPUs)Retained

AWS-designed Arm CPUs for general and host compute. Arm is a multi-vendor instruction set, so binaries built here run on other Arm CPUs (Graviton, Cobalt, Axion, Ampere, Arm OEM servers) without a rebuild: Retained, the commodity x86 reading. Who designs the core doesn't place it (Arm CPU ruling, October 4, 2026).

NVIDIA GPU Instances on AWS (P5, P5e, P5en, P6-B200, P6-B300)Retained

P5 (H100), P5e and P5en (H200), P6-B200 (B200) and P6-B300 (B300), up to eight GPUs per instance. Standard NVIDIA instances: a CUDA or PyTorch workload moves to another cloud's NVIDIA instances or an OEM server without a rebuild. Retained under the Layer 0 runtime rule, read from the customer's seat; the CUDA dependency is NVIDIA's and sits in the NVIDIA column.

Inferentia2 and Trainium2 Instances Through PyTorch on the Neuron SDK (Inf2, 1 to 12 Chips; Trn2, 1 or 16 Chips)Retained

Inf2 rents as single instances from inf2.xlarge (1 chip) to inf2.48xlarge (12 chips), and Trn2 as trn2.3xlarge (1 Trainium2 chip) or trn2.48xlarge (16). Workloads written in PyTorch and compiled through the Neuron SDK keep their model code portable to other hardware: Retained under the Layer 0 runtime rule.

Inferentia2 and Trainium2 Through Neuron-Specific Code (NKI Kernels, Neuron Compiler Tuning)Ceded

Kernels written in the Neuron Kernel Interface and compiler tuning for Inferentia and Trainium don't run elsewhere, and leaving means rewriting them: Ceded under the Layer 0 runtime rule. Recorded as a split because most Inf2 and Trn2 workloads hold both paths.

Trainium3 UltraServers (64 or 144 Chips on a Captive Chip-to-Chip Interconnect)Ceded

Trainium3 GA: EC2 Trn3 UltraServers, Gen1 at 64 chips and Gen2 at 144 chips, up to 4.4x higher performance, 3.9x higher memory bandwidth and 4x better performance per watt than Trn2 UltraServers, per AWS's GA announcement. Designed for agentic AI, MoE models, large-scale RL. Trn3 is sold only as UltraServers. Ceded on the integrated-system rule, the NVL72 reading: pod-scale capacity on a captive chip-to-chip interconnect can't be lifted as deployed. Read the same way as TPU pods on Google's row. Trainium4 expected 2027, not scored.

NVIDIA NVL-Class Rack-Scale Capacity on AWS (P6e-GB200 and P6e-GB300 UltraServers)Ceded

P6e-GB200 and P6e-GB300 UltraServers on NVIDIA GB200 and GB300 NVL72. P6e-GB200 instances "are only available in UltraServers," and P6e-GB300 has been generally available since December 2, 2025, through the AWS sales team. Ceded on the integrated-system rule, the Lenovo NVL-class reading: a rack-scale system built around a captive NVLink fabric can't be lifted as deployed.

Nitro System (Hypervisor and Hardware Isolation)Ceded

AWS's Nitro System: the Nitro card, which carries networking, local NVMe storage, management and monitoring; the Nitro security chip, integrated into the motherboard; and the Nitro hypervisor, "a lightweight hypervisor that manages memory and CPU allocation." Bare metal instances run on the same system without the hypervisor. EFA, SRD and UltraClusters are read at 0B.

AWS AI Factories: NVIDIA Compute (On-Prem)Ceded

Dedicated on-prem environments as private AWS Region. Customer provides space and power; AWS deploys and manages NVIDIA GPUs on P6-B200 and P6e-GB200 UltraServers, networking, storage, and full managed services (Bedrock, SageMaker). Inverted vs Dell/HPE: AWS operates infrastructure the customer houses. AWS's FAQ also lists P6-B300 and P6e-GB300, which its product page still calls "upcoming," so they're noted, not scored.

Intel Gaudi Instances on AWS (DL1, 8x Habana Gaudi HL-205)Retained

dl1.24xlarge: eight first-generation Gaudi accelerators, 256 GiB accelerator memory, listed on the current accelerated computing specifications page.

AMD Radeon Pro Instances on AWS (G4ad, 1 to 4x Radeon Pro V520)Retained

G4ad instances with AMD Radeon Pro V520 GPUs on AMD EPYC hosts, listed on the accelerated computing specifications page. A graphics-class GPU, not an Instinct part.

AWS AI Factories: Trainium Compute (On-Prem)Ceded

The same on-prem private AWS Region with AWS Trainium ("Trn2, Trn3" per AWS's FAQ), which notes "You can deploy multiple accelerator types within a single AWS AI Factory." AWS deploys and manages it in space the customer houses. Ceded, the AI Factories reading.

NVIDIA-Provided

NVIDIA GPU Silicon

P5 through P6e: H100, H200, B200, B300, GB200 and GB300. The CUDA dependency is read on NVIDIA's row.

◆ Gap Analysis

AWS rents the widest accelerator menu on the instrument: Trainium3 UltraServers, Trainium2 and Inferentia2, and NVIDIA from P5 through P6e-GB300. Graviton is the host CPU, and Annapurna Labs designs the custom silicon. The buyer gets per-workload silicon choice with no procurement cycle. That fluidity is the point and the problem. On premises, accelerator choice is a capital decision made once. In the cloud it recurs with every workload, and that matching is itself a Layer 2C function. The architect's concern is the runtime underneath. A workload written in PyTorch keeps its exit; one built on Neuron-specific kernels stays on Trainium and Inferentia. The UltraServers and NVL-class racks are integrated systems on captive interconnects, and AWS AI Factories put the same AWS-run stack inside the customer's building. Calibration: GCP and Azure read strong on the same rented shape; Dell and HPE strong on hardware the enterprise owns.

◆ Borrowed Judgment

Rented, so the runtime call is AWS's. AWS decides which host an instance lands on and when maintenance moves it, and Dedicated Hosts, the one pin, don't support the P5 family that carries the grade: vendor decides, visible, not overridable, Ceded. Portability splits by runtime: NVIDIA instances and PyTorch on Neuron read Retained; Neuron-specific code, the UltraServers and AI Factories read Ceded.

◆ Working Notes

Not a chip (G2, October 16, 2026): EC2 Dedicated Hosts and placement groups are a placement mechanism, read as an authority fact. Dedicated Hosts don't support the P5 family, so they pin nothing on the fleet that carries the grade. Named, not scored (October 22, 2026): Trn2 UltraServers, which AWS's Trn2 page lists as "available in preview"; NVIDIA Rubin and Vera Rubin on AWS, with no instance type in the accelerated computing docs (the AI Factories FAQ says AWS plans to offer them "once generally available"); and Trainium4.

Layer 0B · FabricNetwork FabricEFA and SRD, an AI Fabric AWS Designeddecides: vendor · Ceded▼

Interconnects, switching and routing, network virtualization

Vendor-Provided

Trainium3 UltraServers: Captive Chip-to-Chip InterconnectCeded

The captive chip-to-chip interconnect inside the UltraServer: AWS's Neuron docs put 64 Trainium3 devices in one scale-up domain on NeuronLink-v4 and NeuronSwitch-v1, and 144 in the second generation. Ceded with the system on the integrated-system rule.

P6e GB200 UltraServers: NVLink FabricCeded

The captive NVLink fabric inside the rack-scale system. Ceded with the system on the integrated-system rule.

EFA + SRD (OS-Bypass Transport and Multipath Fabric)Ceded

EFA OS-bypass networking at 3,200 Gbps and SRD, AWS's multipath, fault-tolerant transport, wiring EC2 UltraClusters. The transport decisions sit at 0B. EFA also carries storage: on EFA-enabled FSx for Lustre Persistent 2 file systems, GPUDirect Storage moves data between the file system and GPU memory, bypassing the CPU, for P5, P5e, P5en, P6-B200 and P6-B300 clients.

AWS AI Factories: Networking (On-Prem)Ceded

The networking AWS deploys and manages inside the customer's facility, read like AWS's other regions.

Amazon VPC (Subnets, Route Tables, Security Groups, Gateways, Peering, Transit Gateway)Ceded

A logically isolated virtual network the enterprise defines: subnets pinned to one Availability Zone, route tables per subnet or gateway, security groups, internet and NAT gateways, VPC peering, Transit Gateway, and flow logs. AWS's network virtualization in AWS's regions.

NVIDIA-Provided

NVLink in P6e UltraServers

The GB200 rack's NVLink domain is NVIDIA's, captive inside the integrated system.

NIXL over EFA

NVIDIA's transfer library runs over EFA for disaggregated inference.

◆ Gap Analysis

EFA gives instances OS-bypass networking at up to 3,200 Gbps, and SRD, AWS's own multipath transport, carries it across EC2 UltraClusters of 20,000 GPUs. EFA also carries FSx for Lustre traffic, with GPUDirect Storage straight into GPU memory. Trainium3 and GB200 UltraServers add captive chip-to-chip fabrics inside the rack, and VPC wraps it all in networks, route tables and security groups the enterprise writes. The buyer gets a scale-out fabric built for collective traffic without buying a switch. The architect's concern is that every piece is AWS's design. SRD and EFA run nowhere else, and the opinions built against them don't leave. Calibration: GCP's RoCE GPU network and Azure's InfiniBand read strong on the same shape; HPE's Slingshot and NVIDIA's Spectrum-X strong as fabrics the enterprise buys.

◆ Borrowed Judgment

Ceded where the grade rests. The route tables and security groups are the enterprise's code, but the grade rests on the transport, and SRD makes the multipath and congestion calls inside AWS's fabric with no per-decision override: vendor decides, visible, not overridable, Ceded. Portability: EFA, SRD and VPC are single-vendor surfaces, Ceded.

Layer 0C · StorageStorage SubstrateFSx for Lustre and S3 Express One Zone Feed the Acceleratorsdecides: vendor · Delegated▼

Arrays, storage servers and software, front-end protocols: where the data sits

Vendor-Provided

Amazon S3 (Object Storage, Including S3 Express One Zone)Ceded

De facto object storage standard. S3 Tables (re:Invent 2025): Iceberg-native table storage. S3 Express One Zone: single-digit ms latency. The S3 API is its own chip. This is the service AWS runs behind it: storage classes, lifecycle rules, replication, access policies and the tooling around them, which don't lift to another S3 store without a rebuild. Ceded.

AWS AI Factories: Storage (On-Prem)Ceded

The storage AWS deploys and manages inside the customer's facility, read like AWS's other regions.

Amazon EBS (Block Volumes and Snapshots; io2 Block Express, gp3, HDD Types)Ceded

Block volumes attached to EC2 instances, SSD- and HDD-backed types, Elastic Volumes resizing, snapshots that copy across accounts, Regions and Availability Zones, and a Snapshots Archive tier. Volume data is replicated automatically across servers inside one Availability Zone.

Amazon FSx for Lustre (Scratch and Persistent File Systems; SSD, HDD, Intelligent-Tiering Classes; S3 Data Repositories)Ceded

Managed Lustre file systems with metadata (MDT) and object storage targets (OSTs), provisioned metadata IOPS on Persistent 2, EFA and GPUDirect Storage for EFA-enabled clients, and import from S3. The Lustre client is its own chip. This is the service AWS runs: deployment types, throughput tiers, S3 data repository links and backups, which don't lift to another Lustre. Ceded.

Amazon FSx for NetApp ONTAP (NFS, SMB, iSCSI and NVMe; Single-AZ and Multi-AZ; SnapMirror, FlexCache, Capacity-Pool Tiering, SnapLock)Ceded

A fully managed AWS service built on NetApp's ONTAP file system: multiprotocol access over NFS, SMB, iSCSI and NVMe, automatic tiering to a lower-cost capacity pool, snapshots, clones, SnapMirror replication and FlexCache, managed through AWS's tools or NetApp's ONTAP CLI, REST API and Console. AWS sells it and runs it, and it decides at runtime. The opinions accumulate in ONTAP, so it reads Ceded to NetApp on portability, as ONTAP does through any seller; they lift to NetApp's own systems and the other clouds' ONTAP services, with an operating seam the customer owns.

Nitro System: NVMe Front End for EBSCeded

Nitro presents EBS volumes to the instance as NVMe block devices and runs EBS encryption on the host servers. Facet of the Nitro System chip.

S3 API on Amazon S3 (Buckets and Objects Behind the S3 Standard)Delegated

The S3 API, which independent object stores implement. Workloads confined to the supported common S3 operations move by changing endpoints, credentials and provider-specific configuration; the data, policies and unsupported service features migrate or adapt. Delegated. The service behind it is the Amazon S3 chip, Ceded.

Lustre Client Protocol on FSx for LustreDelegated

Mounted with the open Lustre client, which DDN, HPE and the other clouds' managed Lustre services share: the client and its tooling stay usable against another Lustre, while service topology, stored layouts, backups and integrations are revalidated or recreated. Delegated. The managed service is the FSx for Lustre chip, Ceded.

Front-End Protocols on FSx for NetApp ONTAP (NFS, SMB, iSCSI, NVMe/TCP, S3 Through Access Points)Delegated

NFS v3 to v4.2, SMB and iSCSI, NVMe/TCP on second-generation file systems with six or fewer HA pairs, and S3 through Amazon S3 access points attached to volumes, per AWS's FSx for ONTAP guide, which independent implementations share: hosts and applications re-point to another server or target speaking the same protocols, and the data and provider-specific policies migrate with it. Delegated. The S3 access points cover a subset of S3 operations. Ceded behind them: ONTAP's data management is NetApp's, and the managed service's control plane is AWS's.

NVIDIA-Provided

GPUDirect Storage

The GPU-direct read path into FSx for Lustre uses NVIDIA's GDS stack.

◆ Gap Analysis

AWS covers the substrate end to end. S3 is the object foundation, S3 Express One Zone serves single-digit-millisecond reads, EBS handles block, FSx for NetApp ONTAP serves file and block on ONTAP as a managed service, and FSx for Lustre is the parallel file system that feeds training clusters. The buyer gets AI-throughput storage sized to the workload, not to an array purchase. The cell rests on FSx for Lustre, and that's where the enterprise keeps a handle. It sets a file's stripe layout with lfs setstripe and re-lays an existing file with lfs migrate. Underneath, S3 decides durability placement across Availability Zones with no override, and EBS replicates inside the zone on its own. Calibration: Azure and GCP read strong on managed Lustre too; DDN and WEKA strong on the same function sold as software or appliances.

◆ Borrowed Judgment

Delegated where the grade rests. FSx for Lustre places data across its storage targets, and the enterprise can override a specific file's layout: vendor decides, visible, overridable, Delegated. Portability: S3 and the Lustre client are multi-vendor standards, Delegated; EBS is AWS's own API, Ceded; FSx for NetApp ONTAP reads Ceded to NetApp, as ONTAP does through any seller.

Layer 1A · DataData Foundation & Governance1 LABGlue, Lake Formation and SageMaker Catalog Over Open Tablesdecides: code · Retained▼

The data itself, organized and governed: the Governance Catalog that Layer 2C queries

Vendor-Provided

Amazon S3 + S3 Tables (S3 Tables half: Iceberg-native tables)Ceded

S3 Tables (re:Invent 2025): Iceberg-native table storage in table buckets, with an Iceberg REST endpoint for standard table operations and table maintenance through S3 Tables' own API. The Apache Iceberg surface is its own chip. The service behind it, its operations, management and configuration, is AWS's and doesn't lift: Ceded.

AWS Glue Data Catalog + Lake FormationCeded

Centralized metadata with catalog federation to remote Iceberg catalogs. Fine-grained access control, cross-account sharing, column/row-level security. SageMaker and Bedrock inherit IAM/Lake Formation context. The primitives for 1A→2C exist; the composition is customer-built. Single-vendor API: the litmus names this class Ceded: the opinions accumulate against a surface only AWS implements, with nowhere to take them. Re-lensed July 2026 to match the Azure/GCP/OCI treatment of the architecturally identical service.

SageMaker CatalogCeded

Discovery, subscription, governed sharing of data assets within SageMaker Unified Studio. Single-vendor API: the litmus names this class Ceded: the opinions accumulate against a surface only AWS implements, with nowhere to take them. Re-lensed July 2026 to match the Azure/GCP/OCI treatment of the architecturally identical service.

Apache Iceberg REST Endpoint on Amazon S3 Tables (Standard Table Operations)Delegated

Apache Iceberg has independent implementations: code and definitions written to it move to another compatible implementation within the features both support, the data moves by export, replication or rewrite, and the provider's control-plane configuration and operational state don't lift. Delegated. AWS documents the S3 Tables Iceberg REST endpoint as a set of the standardized Iceberg REST operations for standard table metadata; table maintenance (snapshots, compaction) runs through S3 Tables' own API and doesn't lift. The service behind it is its own chip, Ceded.

NVIDIA-Provided

Assessment pending

◆ Gap Analysis

One of the most mature governance catalogs on the instrument. Glue Data Catalog and Lake Formation hold the metadata and access policy, SageMaker Catalog adds discovery and governed sharing, and S3 Tables organize the data as Iceberg. Catalog federation reaches remote Iceberg catalogs, which no other row matches. The buyer gets one governed catalog that higher layers can query through an API. The architect's concern is the join. A reasoning plane asking the catalog what the data is before it places work isn't a product; the primitives exist, and the composition is the enterprise's. Federation also stops at proprietary on-prem storage metadata. S3's storage function reads at 0C, and the grade rested on the catalog, so it doesn't move. The grade rested on the catalog, so it doesn't move.

◆ Borrowed Judgment

Lake Formation evaluates the LF-Tags and grants the enterprise's administrators assign, deterministically, and CloudTrail logs each data access: code decides, visible, overridable, Retained. Portability: S3 Tables' Iceberg REST endpoint is a Delegated door over the Ceded table-bucket service; Glue, Lake Formation and SageMaker Catalog are single-vendor APIs, Ceded.

Layer 1B · RetrievalContext Management & Retrieval1 LABManaged Retrieval Portfoliodecides: vendor · Delegated▼

Low-latency retrieval for RAG — vector/hybrid search, context windows

Vendor-Provided

Amazon Bedrock Knowledge BasesCeded

Managed RAG: ingest → chunk → embed → index → retrieve. Supports OpenSearch, Aurora/pgvector, Pinecone, Redis. Vertically integrates 1B+1C within managed boundary. Single-vendor API: the litmus names this class Ceded: the opinions accumulate against a surface only AWS implements, with nowhere to take them. Re-lensed July 2026 to match the Azure/GCP/OCI treatment of the architecturally identical service.

Amazon OpenSearch ServerlessCeded

Vector search with HNSW/FAISS. Default Bedrock Knowledge Bases backend. Serverless scaling. The OpenSearch API surface is its own chip. The service behind it, its operations, management and configuration, is AWS's and doesn't lift: Ceded.

Amazon NeptuneDelegated

Graph database for relationship-aware retrieval. Multi-hop reasoning for agentic workloads. Delegated survives on the consumed interface: Gremlin and openCypher are multi-vendor query standards. The weakest survivor of the July 2026 re-lens, flagged as such.

Amazon S3 VectorsCeded

Native vector storage in S3 buckets. GA Dec 2025. Up to 2 billion vectors per index, sub-100ms on frequent queries, roughly 90 percent cheaper than specialized vector databases. A metadata filter on query-vectors gives hybrid semantic-plus-structured retrieval in one call. Usable standalone or as a Bedrock Knowledge Bases storage engine. The put/query interface is single-vendor AWS with no open implementation elsewhere, so the store is captive even though the vectors re-embed cheaply: Ceded. The cheapest 1B option is also the most captive interface.

OpenSearch API on Amazon OpenSearch ServerlessDelegated

The OpenSearch API has independent implementations: code and definitions written to it move to another compatible implementation within the features both support, the data moves by export, replication or rewrite, and the provider's control-plane configuration and operational state don't lift. Delegated. Code written to the supported subset of OpenSearch indexing and search operations moves to another OpenSearch; Serverless collections, security policies and service configuration don't. The service behind it is its own chip, Ceded.

NVIDIA-Provided

Assessment pending

◆ Gap Analysis

Bedrock Knowledge Bases erases the 1B/1C boundary within its managed surface: borrowed judgment, not a gap. AWS makes chunking, embedding, retrieval strategy decisions on the customer’s behalf. The enterprise should ask whether defaults suit their domain. The governance fork mirrors 2A and 2B: consume Bedrock Knowledge Bases and the chunk, embed, and retrieve decisions are AWS's (Delegated), or self-orchestrate the pipeline on primitives (InvokeModel embeddings into S3 Vectors or OpenSearch) and Retain the retrieval-strategy opinions. Self-orchestration Retains the pipeline logic, not the store: OpenSearch stays portable, S3 Vectors is captive to a single-vendor API. Interoperability gap: no unified retrieval abstraction across Bedrock Knowledge Bases + self-hosted Weaviate + Neptune. Routing logic between backends is a Layer 2C function living in application code.

◆ Borrowed Judgment

Moderate, with a Retain path. Self-orchestrating the embed-and-index pipeline on primitives (InvokeModel embeddings into S3 Vectors or OpenSearch) Retains the retrieval-strategy opinions, though the chosen store keeps its own authority. Consume Bedrock Knowledge Bases and AWS makes retrieval quality decisions the enterprise inherits without explicit governance. Compare to VAST (InsightEngine, tighter but VAST-controlled) or Dell (Elastic, separate ISV). Decision authority under the override rule (October 5, 2026): Bedrock decides the search strategy by default, but each Retrieve request can set numberOfResults, filter, overrideSearchType (hybrid or semantic), and a reranking configuration, and results return relevance scores. Vendor decides, visible, overridable, Delegated, the Snowflake and Databricks 1B reading.

◆ Working Notes

Neptune graph-based retrieval is increasingly relevant for agentic workloads needing relationship-aware context, a pattern neither Dell’s Elastic nor VAST’s InsightEngine natively provides.

Layer 1C · PipelinesData Movement & PipelinesManaged Pipelines (Spark/Airflow Substrates)decides: code · Retained▼

Move/transform data — ETL/ELT, lineage, cost-aware movement, KV cache tiering

Vendor-Provided

AWS Glue ETL (Spark Substrate)Ceded

Glue: serverless ETL with Spark 3.5.6, Iceberg 1.10. SageMaker Unified Studio: horizontal integration across 1A/1C/2B with one-click onboarding. Single governed environment collapsing organizational boundaries across data engineering, data science, ML engineering. The Spark API surface is its own chip. The service behind it, its operations, management and configuration, is AWS's and doesn't lift: Ceded.

SageMaker Unified StudioCeded

Single-vendor development and pipeline console: workspace and pipeline-definition opinions accumulate against an AWS-only surface. Ceded, split from the Spark-substrate ETL chip.

Amazon MWAA (Managed Airflow)Ceded

Managed Airflow for complex DAGs. Step Functions for serverless multi-step workflows in ML pipeline reference architectures. Chip scoped to MWAA: managed open-source Airflow. DAGs port to any Airflow. Step Functions split below. The Airflow DAGs surface is its own chip. The service behind it, its operations, management and configuration, is AWS's and doesn't lift: Ceded.

AWS Step FunctionsCeded

Single-vendor state-machine language: workflow definitions have nowhere to run but AWS. Ceded, split from the Airflow chip.

Spark API on AWS Glue ETLDelegated

The Spark API has independent implementations: code and definitions written to it move to another compatible implementation within the features both support, the data moves by export, replication or rewrite, and the provider's control-plane configuration and operational state don't lift. Delegated. The service behind it is its own chip, Ceded.

Airflow DAGs on Amazon MWAADelegated

Airflow DAGs have independent implementations: code and definitions written to them move to another compatible implementation within the features both support, the data moves by export, replication or rewrite, and the provider's control-plane configuration and operational state don't lift. Delegated. The service behind it is its own chip, Ceded.

NVIDIA-Provided

Assessment pending

◆ Gap Analysis

AWS horizontal integration vs VAST vertical integration: same functional coverage, different authority models. VAST = one authority boundary, fewer choices, fewer seams. AWS = many services sharing governance via Lake Formation/IAM, more policy control, more operational complexity. SageMaker Unified Studio addresses fragmentation but underlying services remain distinct. Decision authority: Glue Spark ETL carries the grade and runs the enterprise's own transformation code; auto scaling changes worker count, not the output, and MWAA and Step Functions run the DAGs and state machines the enterprise defined. Run history is documented per job. Code decides, visible, overridable, Retained, the judgment test (October 4, 2026).

◆ Borrowed Judgment

Delegated with customer-retained configuration. AWS provides pipeline services; customer defines transformations and flows. The same fork as 1B applies to the embed-and-index pipeline: consume Bedrock Knowledge Bases and the chunk, embed, and index movement is collapsed and managed (Delegated), or self-orchestrate it on primitives (InvokeModel embeddings feeding S3 Vectors or OpenSearch) and Retain the pipeline opinions, with the chosen store carrying its own authority. Operational complexity of maintaining consistent governance across many AWS accounts is the trade-off for flexibility.

Layer 2A · OrchestrationInfrastructure OrchestrationEKS + Capacity Managementdecides: vendor · Ceded▼

GPU scheduling, quotas, RBAC, fair-share scheduling, utilization optimization

Vendor-Provided

Amazon EKS Auto Mode + KarpenterDelegated

Managed K8s with GPU-aware scheduling. EKS Auto Mode automates cluster/compute management. Karpenter: open-source autoscaler provisioning exact instance types. Mixed compute (NVIDIA, Trainium, Inferentia, Graviton). Note: no DRA support, Capacity Blocks negate scale-to-zero. EKS is a managed service behind the standard Kubernetes API: manifests lift to another conformant cluster, so Delegated (operation delegated to AWS; the standard interface keeps the opinions portable).

Capacity ManagementCeded

Capacity Block Reservations, Flex Start, Savings Plans, Spot. All AWS-controlled allocation. These are capacity acquisition mechanisms, not workload placement reasoning. Not scored within this chip under the GA evidence table: Flex Start (I found no AWS product named Flex Start; Flex-start VMs are a Google Cloud provisioning model, so this looks like a misattribution.).

NVIDIA-Provided

NVIDIA GPU Operator (on EKS)

Available but optional. AWS controls GPU scheduling through EKS Auto Mode and Karpenter. Run:ai available but not required.

◆ Gap Analysis

For Dell/HPE, Layer 2A is where authority slips to NVIDIA. For AWS, Layer 2A is where AWS retains authority through managed services while integrating NVIDIA optionally. GPU scheduling primitives are AWS-controlled. Governance choice: Retain 2A by running self-managed EKS, or Cede 2A by consuming Bedrock (no EKS, no Karpenter: AWS handles 2A invisibly). Both legitimate; the choice is a governance decision with DAPM implications. Capability and deployment are separate readings. The grade states what an architect can deploy on this vendor's paper; it does not assert that a given customer consumes it. The common alternative shape here is concrete and first-party: Red Hat OpenShift Service on AWS is jointly engineered and supported by Red Hat and AWS, so a customer running it consumes AWS for Layer 0 and capacity while Red Hat owns the 2A control plane, and EKS is licensed past rather than consumed. Self-managed Kubernetes and Slurm on EC2 are the other shape, already noted above as the Retain-2A path. In both, the 2A judgment the enterprise inherits is not AWS's.

◆ Borrowed Judgment

Low to moderate depending on path. EKS (Kubernetes interface): Delegated, managed K8s, manifests lift to another conformant cluster. Bedrock consumption: Ceded. NVIDIA dependency is optional, structurally different from Dell/HPE where Run:ai is the primary GPU scheduler.

◆ Working Notes

Capacity Blocks sometimes described as proto-2C but cost-optimized capacity acquisition is not multi-objective placement reasoning. Out of the Capacity Management chip under the GA evidence table (October 24, 2026): Flex Start (I found no AWS product named Flex Start; Flex-start VMs are a Google Cloud provisioning model, so this looks like a misattribution.).

Layer 2B · RuntimeApplication Runtime & Execution1 LABBedrock + Open Pathsdecides: model · Delegated▼

Model serving, agent execution, inference APIs, distributed inference

Vendor-Provided

Model Serving via Standard Interfaces (Bedrock OpenAI-compatible endpoints; Claude Platform on AWS)Delegated

Foundation model access consumed through genuine multi-vendor standard interfaces: Bedrock exposes OpenAI-compatible endpoints, and the native Claude Platform on AWS (April 2026) brings Anthropic’s Messages API onto AWS paper. Both are interfaces independent vendors implement, so the integration opinions written against them lift to any compatible endpoint. A managed service behind a standard interface is Delegated. The residual switching cost is behavioral recalibration, graduated by how much application logic sits in deterministic code outside the model.

Customer-Created and Managed Tools (Function Calling / Client-Side)Retained

GA through Bedrock's OpenAI-compatible endpoints and the Converse API's tool use. The enterprise owns and operates the business logic, APIs, commands and deterministic validators invoked through tool calls: the seam where control passes from instructing the model to executing code outside it. The opinions are the enterprise's and run against a multi-vendor interface, so they lift to another provider with an adapter and re-evaluation, not a rebuild. What the enterprise cedes at this seam is the decision to act: the model chooses whether, when and with which arguments to call, and that judgment is inherited under the serving component above. Added instrument-wide on September 2, 2026 (M4): the seam exists on every runtime that hands control to portable customer code, and a real Retained position is scored, not narrated.

Bedrock Platform Surface (Converse/InvokeModel, Fine-Tuning, Guardrails)Ceded

The capture that graduates beyond the standard interface. The native Converse and InvokeModel surface is implemented only by AWS, and fine-tuning including RFT produces adapters that do not export. Opinions accumulated here have nowhere to go. Model breadth (Claude, Nova, Llama, OpenAI, Mistral, Cohere, Nemotron) is a catalog fact, not an authority fact.

Bedrock AgentCore RuntimeCeded

Serverless agent runtime. Framework-agnostic (Strands, LangChain, CrewAI). Protocol-agnostic (MCP, A2A). Model-agnostic. MicroVM session isolation. 2M+ SDK downloads in 5 months. Single-vendor API: the litmus names this class Ceded: the opinions accumulate against a surface only AWS implements, with nowhere to take them. Re-lensed July 2026 to match the Azure/GCP/OCI treatment of the architecturally identical service.

SageMaker AI (Managed)Ceded

Training, fine-tuning, inference endpoints. LMI containers with vLLM. Multi-LoRA. Supports Trainium + NVIDIA. vLLM on EKS and Ray on EKS for fully self-hosted (Retained). Chip scoped to the managed platform: SageMaker training/serving APIs are single-vendor. Ceded, matching Vertex on the GCP row. The self-hosted path is split below.

Self-Hosted OSS Runtimes on EC2/EKSRetained

vLLM, Ray, KServe-class serving the customer operates on rented compute: open substrates; the opinions lift to any infrastructure. The Retained escape valve of the AWS runtime story.

Strands Agents SDKRetained

AWS open-source agentic framework. Model-first, native AgentCore/Guardrails/OpenTelemetry integration. Multi-agent patterns with A2A.

NVIDIA-Provided

NVIDIA GPU Instances + NIM

P5/P6 instances. NVIDIA Nemotron via Bedrock. NVIDIA dependency optional: Trainium-only inference is architecturally possible.

◆ Gap Analysis

AWS owns multiple 2B surfaces (Bedrock, SageMaker, AgentCore, EKS). NVIDIA dependency is optional in a way it’s not for Dell/HPE. Agent frameworks blur 2B/2C/3 boundaries: AgentCore bundles Runtime (2B) + Policy (2C) + agent logic (3). Product boundary ≠ architectural boundary. Borrowed judgment: using Bedrock to access Anthropic Claude or Meta Llama means the model provider’s alignment decisions become part of the enterprise’s AI system. Guardrails constrain output but reasoning in model weights is not customer-configurable.

◆ Borrowed Judgment

Varies by path. Bedrock through its OpenAI-compatible endpoints and Claude Platform on AWS: Delegated, plus the model provider's borrowed judgment. Bedrock's native Converse and InvokeModel surface, fine-tuning, and Guardrails: Ceded. AgentCore Runtime: Ceded, a single-vendor API matching the Azure, GCP, and OCI agent runtimes. SageMaker AI (managed): Ceded, matching Vertex. Self-hosted vLLM, Ray, or KServe on EC2 or EKS: Retained, as are the tools the enterprise writes and the Strands Agents SDK. Runtime proliferation is itself a 2C decision AWS doesn't automate.

◆ Working Notes

Product boundary (AgentCore = Runtime + Policy + Evaluations + Memory + Registry) doesn’t align with 4+1 architectural boundary (Runtime = 2B, Policy = 2C, agent logic = 3). Same cross-layer bundling seen in Google’s and VAST’s products.

Layer 2C · ReasoningAgentic Infrastructure — The Reasoning Plane1 LABIntelligence-2C Toolkit; Placement Implicitdecides: vendor · Delegated▼

Policy-driven placement and resource coordination — the Autonomy Layer

Vendor-Provided

Bedrock GuardrailsCeded

Content filtering, PII protection, topic blocking. ApplyGuardrail API works with any model. Cross-account organizational safeguards (GA Apr 2026). Single-vendor API: the litmus names this class Ceded: the opinions accumulate against a surface only AWS implements, with nowhere to take them. Re-lensed July 2026 to match the Azure/GCP/OCI treatment of the architecturally identical service.

AgentCore Policy (GA Mar 2026)Ceded

Centralized governance outside agent code. Natural language → Cedar policy. Intercepts every tool call before execution. 13 AWS regions. Single-vendor API: the litmus names this class Ceded: the opinions accumulate against a surface only AWS implements, with nowhere to take them. Re-lensed July 2026 to match the Azure/GCP/OCI treatment of the architecturally identical service.

AgentCore Evaluations + MemoryCeded

Built-in evaluators for correctness, safety, adherence, consistency. Episodic Memory for stateful reasoning across sessions. Single-vendor API: the litmus names this class Ceded: the opinions accumulate against a surface only AWS implements, with nowhere to take them. Re-lensed July 2026 to match the Azure/GCP/OCI treatment of the architecturally identical service.

NVIDIA-Provided

No NVIDIA Layer 2C Dependency

All Layer 2C components are AWS IP. NVIDIA does not control governance, policy, or reasoning in the AWS stack.

◆ Gap Analysis

Intelligence Layer 2C (partially present): AgentCore Policy + Guardrails + Evaluations govern agent behavior. Real and productized. Natural-language-to-Cedar conversion is the most accessible policy authoring in this assessment. Infrastructure Layer 2C (not built): No service answers ‘given data residency, cost, latency, and compliance, should this run on Trainium in us-east-1 or NVIDIA in eu-west-1?’ Capacity primitives are building blocks, not a policy-driven placement engine querying 1A governance metadata. The structural insight: AWS already has the control plane everyone else is trying to build, but it’s implicit. Dell’s 2C gap is product absence. AWS’s 2C gap is visibility and authority: the capability exists but is implicit, managed, and Ceded. Read across the instrument by shape rather than by roster. Three positions exist at 2C. One cohort ships nothing and the enterprise owns the function by default. A second and now larger cohort ships productized Intelligence-2C (agent identity, gateways, registries, runtime policy enforcement, audit) and reads moderate or strong on it, with authority Ceded to whoever built the control plane. Nobody ships Infrastructure-2C as a configurable product. Naming which vendor sits where is what the heat map is for; repeating it inside a cell only guarantees the cell goes stale the next time a peer moves. AWS sits in the second cohort with a distinguishing wrinkle worth stating plainly. Its Infrastructure-2C decisions are real and they are made continuously, inside managed services, on the enterprise's behalf. They are simply never surfaced as a control the architect can see or set. That is a different thing from not having the capability, and it is the reason this cell reads the way it does. The question is not ‘Does AWS have Layer 2C?’ but ‘How much can the enterprise configure, audit, and control, and how much has been Ceded without explicit classification?’

◆ Borrowed Judgment

Intelligence 2C: Low. AgentCore Policy, Guardrails, Evaluations are AWS IP. Customer defines policies; AWS enforces. Infrastructure 2C: Ceded (implicit). Placement decisions inside managed services without explicit customer policy input. When SageMaker auto-scales or Bedrock routes, those are 2C functions the enterprise has Ceded without classification. DAPM discipline demands: for every managed service placement decision, classify as Delegated (customer sets policy) or Ceded (AWS decides).

◆ Working Notes

The re:Invent 2025 and AgentCore announcements are the strongest vendor validation of the 4+1 model’s Layer 2C thesis. The ‘invisible Reasoning Plane’ observation is the conceptual foundation of the 4+1 model itself. Watch-list (descored July 12, 2026, /reconcile GA-gate): AWS Agent Registry, in preview; a preview is never a scored component on any row. Scores at doc-confirmed GA. Instrument note: live Infrastructure-2C (a policy engine deciding at request time where inference runs relative to the data, which model serves the request, and how cost, latency, and compliance are arbitrated) does not ship from any vendor on this instrument. It is recorded here as a universal finding about the market, not as a charge against this vendor. The sibling finding applies here as well: you cannot prompt your way to deterministic output. The controls credited in this cell answer whether an action is legal (permissions, schema and policy checks, rate and token limits, approval gates), and none of them validates whether the outcome was right under a graduated escalation policy. Controls that work by instructing the model shift the output distribution without pinning it. That deterministic outcome-validator is missing instrument-wide, and like live placement it is noted as universal rather than charged to one vendor.

Layer 3 (+1) · ApplicationsAI Application Layer — The Value PlaneBroadest Ecosystemdecides: model · Delegated▼

AI-powered business capabilities — business logic, workflow automation

Vendor-Provided

Model + ISV EcosystemDelegated

Anthropic Claude, Amazon Nova, Meta Llama, OpenAI, Mistral, Cohere, NVIDIA Nemotron. Thousands of ISV applications. 11,000+ government agencies.

AWS Kiro (Agentic IDE Platform)Ceded

Spec-driven agentic development platform replacing Amazon Q Developer (new signups ended May 2026). Three surfaces: VS Code-compatible IDE, CLI, and autonomous cloud agent. Spec-driven development generates requirements.md, design.md, and tasks.md before code: specs are source-of-truth, code is build artifact. Hooks system: 17 automated quality gates (security, linting, testing, validation) firing on file save and PR events. Multi-model routing: Claude Sonnet for reasoning-heavy specs, Amazon Nova for high-throughput code generation, Bedrock as unified model plane. 50+ Powers (MCP integrations: Figma, Terraform, Stripe, Datadog). Autonomous agent executes backlog tasks and opens PRs without developer in the loop. Deep AWS context: native Powers for AWS pricing, docs, Well-Architected, cost analysis. The most opinionated developer AI surface from any cloud vendor. It enforces structured development discipline rather than freeform 'vibe coding.' Single-vendor API: the litmus names this class Ceded: the opinions accumulate against a surface only AWS implements, with nowhere to take them. Re-lensed July 2026 to match the Azure/GCP/OCI treatment of the architecturally identical service. Not scored within this chip under the GA evidence table: autonomous cloud agent ("Kiro autonomous agent is now available on Kiro Web (Preview) as a new autonomous mode . Kiro Web is currently in preview").

NVIDIA-Provided

NVIDIA NIM on Bedrock

NVIDIA models via Bedrock API alongside all other providers.

◆ Gap Analysis

Broadest Layer 3 in this assessment, different category than Dell ISV partnerships, HPE Unleash AI, or VAST Cosmos. Each Layer 3 application brings its own governance domain. AgentCore Policy and Guardrails provide cross-agent governance primitives; whether they compose into enterprise-wide agent governance remains an implementation question. The Retained/Delegated boundary is not uniform. Custom Strands on self-hosted EKS: fully Retained. Bedrock Agents / Q / partner apps: substantially Delegated. Same enterprise may have both patterns simultaneously. Kiro represents AWS's strongest Layer 3 opinion: spec-driven development enforces structured requirements before code generation. This is an opinionated development methodology embedded in tooling: the enterprise Delegates development workflow decisions to AWS's architectural opinions about how AI-assisted software should be built. The autonomous agent (cloud agent executing tasks and opening PRs without human in the loop) creates a new DAPM question: when Kiro's agent writes and ships code autonomously, who owns the judgment embedded in that code? The developer who assigned the task, or Kiro's multi-model routing logic that chose which model to apply? Compare to Google Antigravity 2.0 (agent orchestration platform, multi-agent parallel execution, Gemini-native) and GitHub Copilot (IDE-embedded coding agent with cloud agent for autonomous PR creation). All three clouds now have agentic developer surfaces that span Layer 2B (execution) and Layer 3 (application). The competitive dynamics are shifting from 'which cloud has the best models' to 'which cloud has the most productive developer surface.'

◆ Borrowed Judgment

Distributed and complex. Model providers bring training data, alignment, safety decisions as inherited borrowed judgment. AWS platform defaults shape application behavior. DAPM Action 3 applies with force: when you move off AWS, what judgment doesn’t move with you? Answer: almost everything above Layer 0. Decision authority (October 5, 2026): Amazon Quick carries the grade. Write actions default to Always Ask: every invocation shows a confirmation prompt with Allow, Trust, or Deny, and a connector author can lock an action to Always Ask (Quick tool-permissions docs). Model decides, visible, overridable, Delegated.

◆ Working Notes

OpenAI partnership (Stateful Runtime on Bedrock) + NVIDIA (1M+ GPUs) position AWS as convergence substrate for multiple AI ecosystems. Broadest possibilities, most complex borrowed judgment landscape. The Q Developer → Kiro transition (new Q Developer signups ended May 2026) is a significant strategic signal. AWS is consolidating its developer AI surface into a single opinionated platform rather than maintaining parallel tools. Kiro's spec-driven approach is the inverse of 'vibe coding': it imposes engineering discipline through AI tooling. Whether enterprise development teams accept this opinionated workflow or prefer the freeform approach of Cursor/Copilot is the adoption question. Red Hat Summit 2026 announced OpenShift Dev Spaces support for Kiro alongside Microsoft Copilot, Claude CLI, Cline, Continue, and Roo, meaning Kiro can run inside Red Hat's governed platform. This cross-vendor interoperability matters for the 4+1 model: the developer tool (Layer 3) can be decoupled from the infrastructure platform (Layer 2A). An enterprise could run Kiro on OpenShift on Dell hardware: three vendors' authority at three different layers. Named, not scored at Layer 3 (October 19, 2026: a builder whose only output is software the enterprise runs earns no Layer 3 grade): Strands SDK applications (an SDK whose output is the enterprise's own agents; read at 2B). To notes under the GA evidence table (October 24, 2026): Bedrock Agents (Managed) (The owner closed the managed service to new customers and points them to AgentCore, so it is not offered today; existing customers keep running (maintenance mode). ("Amazon Bedrock Agents (now Amazon Bedrock Agents Classic) is no longer open to new customers.")). To notes under the GA evidence table (October 24, 2026): Amazon Q (Q Business is closed to new customers and Q Developer IDE plugins reach end of support April 30, 2027 (docs point to Kiro), so the owner no longer offers Amazon Q to new buyers. ("Amazon Q Business is no longer open to new customers. For capabilities similar to Q Business, explore Amazon Quick.")). Out of the AWS Kiro chip under the GA evidence table (October 24, 2026): autonomous cloud agent ("Kiro autonomous agent is now available on Kiro Web (Preview) as a new autonomous mode . Kiro Web is currently in preview").

Sources & revision history · v1.40 - 4+1 v3: Layer 0 Sub-Layers

re:Invent 2025, GTC 2026, Bedrock AgentCore GA, AgentCore Policy GA (Mar 2026), SageMaker Unified Studio, AWS/NVIDIA collaboration, OpenAI/AWS partnership, analyst coverage. v1.2 (instrument reconciliation): 1A S3 and 2A EKS Retained→Delegated: a managed service behind a multi-vendor standard interface is Delegated; Retained is reserved for an open substrate the enterprise operates. Corrects the prior over-application of Retained. Also aligned the VAST Layer 2C cross-reference to VAST's current gap status (PolicyEngine GA end 2026). v1.3 (lab-validated against the TFD corpus RAG build): added Amazon S3 Vectors (GA Dec 2025) to Layer 1B as a Ceded component (single-vendor vector API, no open exit), and added the self-orchestrate-versus-Bedrock-KB Retain/Delegate fork to the 1B and 1C narration, mirroring 2A/2B. v1.4 (July 12, 2026, /reconcile): the hyperscaler lens drift resolved: AWS's single-vendor managed services re-lensed under the litmus's own rule (single-vendor API = Ceded; Delegated requires a multi-vendor standard or open substrate), aligning the row with the Azure/GCP/OCI treatment of architecturally identical services. Chips moved D→C: Glue Catalog + Lake Formation, SageMaker Catalog (1A); Bedrock Knowledge Bases (1B); Bedrock, AgentCore Runtime (2B); Guardrails, AgentCore Policy, Evaluations + Memory (2C); Amazon Q, Kiro, Bedrock Agents (L3). Invalid dual-enum chips split into per-mode components (SageMaker managed Ceded / self-hosted Retained; Bedrock Agents Ceded / Strands apps Retained; Glue ETL Delegated / Unified Studio Ceded; MWAA Delegated / Step Functions Ceded). Agent Registry (Preview) descored to the watch-list per the GA-gate. statusLabels moved to capability vocabulary. No capability grades changed. Inference-interface reconciliation (July 23, 2026): the completions-interface ruling (OpenAI v1.0) and its Messages-API extension (Anthropic v1.0) applied: model access consumed through a genuine multi-vendor standard inference interface is Delegated; proprietary surfaces beyond the interface remain Ceded. Bedrock stands Ceded (single-vendor Converse/InvokeModel); standard-interface paths on AWS paper noted with a breadth fact-check pending. /reconcile (September 1, 2026): inference-interface rule applied as a scored facet. Model access consumed through a multi-vendor standard interface (OpenAI-compatible completions, Anthropic Messages) split out as Delegated; the platform surface beyond the interface stays Ceded. Matches the Azure v1.4 treatment. /reconcile (September 1, 2026): retired status vocabulary replaced with gap. /reconcile (September 1, 2026): shared 2C findings stated explicitly (live-placement universality, outcome-validation finding). /reconcile (September 1, 2026): the hard-coded cross-vendor 2C roster replaced with a shape-based reading. The roster had drifted out of agreement with the rows it described (HPE 2C scored three Ceded, not Retained plus Delegated) and went stale on every peer promotion. /reconcile (September 1, 2026): the availability-not-occupancy reading applied at 2A: the grade states deployable capability on this vendor's paper and no longer reads as though the vendor's layer is automatically the customer's. /reconcile (October 5, 2026): IBM and Red Hat are now separate rows; citation retargeted to the Red Hat row, no grade moves. /reconcile (October 5, 2026): Layer 0 runtime rule (ruled September 28 and October 4, 2026). NVIDIA GPUs on AWS split: P5/P5e/P6 instances Ceded to Retained from the customer's seat, P6e GB200 Ceded on the integrated-system rule. Custom silicon holds Ceded, re-reasoned on the integrated-system rule (Trn3 UltraServers) in place of silicon ownership. Grade unchanged. /reconcile (October 5, 2026, same pass): custom silicon split. Trainium3 UltraServers Ceded on the integrated-system rule, matching TPU pods; Inferentia2 split by runtime; Graviton Ceded to Retained on the Arm CPU ruling. /reconcile (October 5, 2026): 1C authority column (rulings of September 15 and October 4, 2026). 1C authority vendor / Ceded to code / Retained, and visible null to true: Glue runs the enterprise's Spark code, and auto scaling changes worker count, not output (docs.aws.amazon.com/glue/latest/dg/auto-scaling.html); run history is documented (view-job-runs). Grade unchanged. /reconcile (October 5, 2026): authority re-read against product docs under the judgment test, override rule, and Layer 3 ruling: layer1a to code / Retained; layer3 to model / Delegated. Grades and components unchanged. /reconcile (October 5, 2026): authority read against product docs: layer0 visible true; layer1b visible true; layer2a visible true; layer1b Ceded to Delegated on per-request retrieval overrides. Grades and components unchanged. v1.17 (Layer 0 sub-layers, October 16, 2026): Layer 0 split into 0A Compute, 0B Network Fabric and 0C Storage Substrate by runtime decision (Keith, October 16, 2026); storage left 1A for 0C, and 1A, now Data Foundation & Governance, is graded on the catalog that classifies the data. 0A strong, 0B strong, 0C strong. v1.18 (October 19, 2026, /reconcile) Layer 3: Strands SDK applications (an SDK whose output is the enterprise's own agents; read at 2B) moved out of the chip list under the Layer 3 exclusion for software-only builders (Keith, October 19, 2026). Grade and authority unchanged. (Same pass) Plain-language pass: methodology rule numbers in the prose and audit notes replaced with the reasoning they stood for; no grade, chip or authority reading changed. v1.19 (October 19, 2026, /reconcile voice pass): em-dashes in the prose replaced with periods, colons, commas or parentheses; wording, grades, chips and authority readings unchanged. v1.20 (October 21, 2026, /reconcile): 0C chip added for Amazon FSx for NetApp ONTAP, from AWS's FSx for ONTAP user guide (read October 21, 2026). AWS sells and runs it, so it's AWS's chip; Ceded to NetApp on portability with AWS deciding at runtime (Keith, October 21, 2026). Grade and cell authority unchanged; the cell rests on FSx for Lustre. v1.21 (October 22, 2026, /reconcile): the 0B calibration line cites GCP's RoCE GPU network in place of Virgo, which GCP v1.26 moved to a named, unscored note. Grade and authority unchanged. v1.22 (October 22, 2026, /reconcile): the 0A NVL-class chip cites Lenovo's NVL-class racks as its precedent in place of Dell's XE9812, which Dell v2.22 moved to a named, unscored note. Grade and authority unchanged. v1.23 (October 22, 2026, /reconcile): 0B EFA chip records the storage path (FSx for Lustre over EFA with GPUDirect Storage, per the FSx for Lustre guide), and the Trainium3 chip drops a "sub-10µs latency" figure the Neuron docs don't carry, in favor of the documented 64- and 144-device scale-up domains. Found by the 0B fabric research. Grade and authority unchanged. v1.24 (October 22, 2026, /reconcile): the 0A NVIDIA chip lists P5en and P6-B300 and drops an undocumented Rubin claim and a market-position line; the NVL-class chip adds P6e-GB300 (GA December 2, 2025); the Neuron chips add Trn2 instances; the Trainium3 chip carries AWS's own figures (3.9x memory bandwidth, not 4x) and drops the latency figure, as 0B did in v1.23; the Nitro chip's detail moves from fabric facts to the Nitro System's 0A components. Trn2 UltraServers (preview) and Rubin on AWS go to named, unscored notes. Found by the 0A compute research. Grade and authority unchanged. v1.25 (October 23, 2026, /reconcile): 0A: AWS documents Trainium and NVIDIA as separate accelerator options. Found by the Layer 0 data pass after the sniff read. Grade and cell authority unchanged. v1.26 (October 23, 2026, storage door pass): The S3 API, the Lustre client and FSx for NetApp ONTAP's protocols are doors; the managed services read Ceded. Grade unchanged. v1.27 (October 23, 2026, Layer 1 managed-services pass): Amazon S3 + S3 Tables splits; Amazon OpenSearch Serverless splits; AWS Glue ETL splits; Amazon MWAA splits. Grade unchanged. v1.28 (October 24, 2026, /reconcile): GA evidence recorded on 5 of 5 Layer 1B chips under the GA evidence table. Grade unchanged. v1.29 (October 24, 2026, /reconcile): Layer 1B GA evidence migrated to the GA evidence table version 2. Grade unchanged. v1.30 (October 24, 2026, /reconcile): GA evidence recorded on 4 of 4 Layer 1A chips under the GA evidence table version 2. Grade unchanged. v1.31 (October 24, 2026, /reconcile): GA evidence recorded on 6 of 6 Layer 1C chips under the GA evidence table version 2. Grade unchanged. v1.32 (October 24, 2026, /reconcile): GA evidence recorded on 5 of 5 Layer 0B chips under the GA evidence table version 2. Grade unchanged. v1.33 (October 24, 2026, /reconcile): GA evidence recorded on 11 of 11 Layer 0A chips under the GA evidence table version 2. Grade unchanged. v1.34 (October 24, 2026, /reconcile): GA evidence recorded on 9 of 9 Layer 0C chips under the GA evidence table version 2. Grade unchanged. v1.35 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 2 Layer 2A chips under the GA evidence table version 2. Grade unchanged. v1.36 (October 24, 2026, /reconcile): GA evidence recorded on 7 of 7 Layer 2B chips under the GA evidence table version 2. Grade unchanged. v1.37 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 4 Layer 3 chips under the GA evidence table version 2. Grade unchanged. v1.38 (October 24, 2026, /reconcile): GA evidence recorded on 3 of 3 Layer 2C chips under the GA evidence table version 2. Grade unchanged. v1.39 (October 24, 2026, /reconcile): GA evidence table version 2.1 applied: 2 chips recorded or narrowed, 2 to notes. Grade unchanged. v1.40 (October 27, 2026): each component now carries a durable offering id; nothing else changed.

4+1 Layer AI Infrastructure Model · Vendor Assessment Series · The CTO Advisor LLC (DBA The Advisor Bench) · thectoadvisor.com