Executive Summary: Google Cloud AI Infrastructure

Google Cloud is the only vendor in this assessment series that owns a frontier foundation model, and that single fact restructures the entire 4+1 analysis. Google built Gemini, trains Gemini on its own TPUs, optimizes its silicon for Gemini’s training requirements, and weaves Gemini’s intelligence into every layer of its cloud platform. This creates a model-integrated stack: an architecture where the frontier model is not a component plugged into infrastructure but the intelligence that pervades the infrastructure.

No other vendor assessed possesses this vertical integration. Google owns every layer of the 4+1 model with proprietary IP: custom silicon (TPUs), custom networking (Virgo), proprietary storage (Colossus/Spanner/BigQuery), its own runtime and frameworks (JAX/Pathways), its own frontier models (Gemini), and a unified orchestration surface (Gemini Enterprise Agent Platform).

The DAPM implication is not merely that every layer is Ceded. It is that every layer is ceded to a unified intelligence. With AWS, authority is distributed across multiple vendors’ judgment (AWS infrastructure, Anthropic model reasoning, ISV application logic). That distribution creates complexity but also structural checks. With Google Cloud + Gemini, the enterprise concentrates authority in one vendor’s judgment across every layer, from silicon to application.

This is the deepest expression of vertical integration in enterprise technology since the mainframe era. The enterprise gains end-to-end optimization that no multi-vendor assembly can match. But the 4+1 framework makes visible what the integration obscures: the enterprise has no fallback position at any layer. Google Distributed Cloud (GDC) addresses data sovereignty without addressing judgment sovereignty: GDC still runs Google’s software stack and Google’s models.

The structural question: does concentrating all layers of authority and all layers of model judgment in a single vendor deliver enough value to justify the governance position, and has the enterprise made that concentration explicit rather than inheriting it by default?

Layer-by-layer status: Layer 0A (TPUs Plus the NVIDIA Fleet, Rented by the Slice and the Instance), Layer 0B (A RoCE GPU Network and the TPU Inter-Chip Interconnect, Fabrics Google Runs), Layer 0C (Managed Lustre and the Cloud Storage Rapid Tier Feed the Accelerators), Layer 1A (Knowledge Catalog Over Open Tables; Gemini Context Graph in Preview), Layer 1B (Model Prep & Managed Retrieval), Layer 1C (BigQuery-Native Pipelines), Layer 2A (GKE + Sovereign Extensions), Layer 2B (Model-Integrated Stack), Layer 2C (Complete Intelligence-2C Plane, No Placement), Layer 3 (+1) (Antigravity and Gemini Enterprise: First-Party Agents That Do the Work).

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: v1.46 - 4+1 v3: Layer 0 Sub-Layers.

Disclosure: The CTO Advisor has a commercial relationship with this vendor. Assessment findings are independent. See disclosure policy.

Google Cloud AI Infrastructure

Mapped to the 4+1 Layer AI Infrastructure Model

v1.46 - 4+1 v3: Layer 0 Sub-LayersAssessed October 27, 2026Sources & revision history
ACTIVE ASSESSMENT

Summary Finding

Google Cloud is the only vendor in this assessment series that owns a frontier foundation model, and that single fact restructures the entire 4+1 analysis. Google built Gemini, trains Gemini on its own TPUs, optimizes its silicon for Gemini’s training requirements, and weaves Gemini’s intelligence into every layer of its cloud platform. This creates a model-integrated stack: an architecture where the frontier model is not a component plugged into infrastructure but the intelligence that pervades the infrastructure.

No other vendor assessed possesses this vertical integration. Google owns every layer of the 4+1 model with proprietary IP: custom silicon (TPUs), custom networking (Virgo), proprietary storage (Colossus/Spanner/BigQuery), its own runtime and frameworks (JAX/Pathways), its own frontier models (Gemini), and a unified orchestration surface (Gemini Enterprise Agent Platform).

The DAPM implication is not merely that every layer is Ceded. It is that every layer is ceded to a unified intelligence. With AWS, authority is distributed across multiple vendors’ judgment (AWS infrastructure, Anthropic model reasoning, ISV application logic). That distribution creates complexity but also structural checks. With Google Cloud + Gemini, the enterprise concentrates authority in one vendor’s judgment across every layer, from silicon to application.

This is the deepest expression of vertical integration in enterprise technology since the mainframe era. The enterprise gains end-to-end optimization that no multi-vendor assembly can match. But the 4+1 framework makes visible what the integration obscures: the enterprise has no fallback position at any layer. Google Distributed Cloud (GDC) addresses data sovereignty without addressing judgment sovereignty: GDC still runs Google’s software stack and Google’s models.

The structural question: does concentrating all layers of authority and all layers of model judgment in a single vendor deliver enough value to justify the governance position, and has the enterprise made that concentration explicit rather than inheriting it by default?

Strength
Moderate
Gap
Partner
Layer 0A · ComputeCompute3 LABSTPUs Plus the NVIDIA Fleet, Rented by the Slice and the Instancedecides: vendor · Ceded▼

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

Vendor-Provided

TPU Single-Host Slices Through an Open Runtime (TPU7x Ironwood, v6e Trillium, v5p VMs Under JAX, PyTorch/XLA, vLLM)Retained

Generally available TPUs rent as single-host VMs: v6e from 1 chip (ct6e-standard-1t, 4t, 8t), Ironwood and v5p at 4 chips per VM. Workloads written in JAX or in PyTorch through XLA also run on NVIDIA GPUs, so the opinions lift without a rebuild: Retained under the Layer 0 runtime rule.

TPU Single-Host Slices Through TPU-Specific Code (Pallas Kernels Written for TPU, Sharding Tuned to TPU Memory)Ceded

Kernels and sharding written for the TPU's memory hierarchy don't run elsewhere, and leaving means rewriting them: Ceded under the Layer 0 runtime rule. Most TPU workloads hold both paths, which is why the slices are recorded as a split.

TPU Pod-Scale Capacity (Multi-Host Slices on the Inter-Chip Interconnect)Ceded

Multi-host slices up to full pods, wired by Google's inter-chip interconnect. 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. Google designs TPUs to train Gemini, so the enterprise doesn't direct the optimization priorities; that's narrative, not what places the chip. Read the same way as Trainium3 UltraServers on AWS's row.

NVIDIA GPU Instances on Google Cloud (A4, A3 Ultra, A3 Mega, A3 High; G4 Fractional GPUs)Retained

A4 (eight B200), A3 Ultra (eight H200), A3 Mega and A3 High (H100), and G4 on RTX PRO 6000 in fractions down to 1/8 of a GPU. 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.

NVIDIA NVL-Class Rack-Scale Capacity on Google Cloud (A4X on GB300 and GB200 NVL72)Ceded

A4X Max and A4X instances on GB300 and GB200 NVL72 racks, 18 instances and 72 GPUs to an NVLink domain. 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. Not scored within this chip under the GA evidence table: A4X Max ("To get started with the A4X Max preview").

Google Distributed Cloud (GDC)Ceded

On-prem deployment of Google Cloud services. Connected and air-gapped configs. 4 racks to hundreds. NVIDIA A100, H100, H200 and B300 machine types (the air-gapped docs list a4-ultragpu-8g-gdc at eight B300), and Gemini models on-prem. Managed GDC Provider initiative (Clarence, Gulf Energy, T-Systems, WWT). NATO deployment. Customer provides facility; Google provides and operates HW+SW. Addresses data sovereignty but not judgment sovereignty. Ceded under the Layer 0 portability acid test (Keith, October 16, 2026): can the services that support Layers 1 and 2 be decoupled from the vendor's Layer 0? Here they can't. Gemini runs on NVIDIA Blackwell inside GDC, so the silicon isn't what holds it; the service is. Gemini, Agent Platform and the managed services don't run on another vendor's NVIDIA GPUs, and the hardware comes bound to them. This matches AWS AI Factories and OCI Dedicated Region, the same shape, which read Ceded. Who operates the region (Google, or a Managed GDC Provider) is a prose fact and doesn't move the grade (a local cloud region is scored like the vendor's other regions, October 16, 2026).

Axion Arm CPUs (C4A, N4A VMs; C4A Bare Metal)Retained

Google-designed Arm CPUs rented as VMs: C4A on Arm Neoverse V2 up to 72 vCPUs (GA October 30, 2024, Titanium SSD variants GA January 2025, bare metal GA May 2026) and N4A on Neoverse N3.

Titanium (Offload Adapters and Titanium Offload Processors)Ceded

Google's offload system: security microcontrollers, adapters and offload processors that run network and storage functions off the host, sanitize NVMe commands, and encrypt data in transit and at rest. All third- and fourth-generation general-purpose VMs use it.

NVIDIA-Provided

NVIDIA GPU Silicon

A3 through A4X Max: H100, H200, B200, GB200 and GB300. The CUDA dependency is read on NVIDIA's row.

◆ Gap Analysis

Google rents two accelerator families. TPU7x Ironwood, v6e Trillium and v5p come as single-host slices or multi-host pods on the inter-chip interconnect. NVIDIA comes as A3, A4, fractional G4, and A4X Max and A4X on GB300 and GB200 NVL72. Axion is Google's Arm CPU, and Titanium offloads network and storage work from every current host. The buyer gets per-workload silicon choice. On premises that match is made once at procurement; here it recurs, and it's a Layer 2C function. The architect's concern is the runtime. JAX or PyTorch through XLA keeps its exit; Pallas kernels and sharding tuned to TPU memory stay on TPU. Pods and NVL72 racks are integrated systems on captive interconnects, and Google Distributed Cloud (GDC) puts Google-operated services in the customer's building that run on no one else's hardware. Calibration: AWS 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 Google's. Google decides which host a VM or TPU slice lands on, when maintenance moves it and which generation it stocks, and the enterprise reverses none of it: vendor decides, visible, not overridable, Ceded. Portability splits by runtime: NVIDIA instances, Axion and TPU slices under an open runtime read Retained; TPU-specific code, pods, NVL72, Titanium and GDC read Ceded.

◆ Working Notes

Named, not scored (October 5, 2026 GA gate): TPU 8t and TPU 8i, listed as Coming Soon; at GA they read like the GA generations. Lab-measured (July 2026): the gemma4-tpu-inference lab found Trillium v6e capacity dry in all three attempted zones with quota clean, so the constraint was capacity, not policy. A5X bare metal on Vera Rubin NVL72 is announced and isn't in the accelerator-optimized machine docs (read October 22, 2026), so it's named, not scored. Out of the NVIDIA NVL-Class Rack-Scale Capacity on Google Cloud chip under the GA evidence table (October 24, 2026): A4X Max ("To get started with the A4X Max preview").

Validated in Labs

Hands-on builds on real hardware that test where authority actually holds at this layer.

Lab 012Buy the harness, not the tier

Open weights on capacity I already carry clear 17 of 22 repairs for nothing, and the gate names the five they miss. Renting more hardware for the rest is slower and dearer than the API. The minimum that finishes the job is a mini model in an agentic harness, at $3.14. Two Opus generations finish the same 22 for four times that.

Lab 008Renting the chip was the easy part

Renting the TPU is fast; serving a bring-your-own model on it means adopting Google’s stack or quantizing off-box. The friction was the finding.

Lab 007The CPU exit is a batch lane, not a serving lane

Across 22 measured configurations, zero met the interactive latency bar; the Xeon lane delivers throughput or interactive latency, not both.

Layer 0B · FabricNetwork FabricA RoCE GPU Network and the TPU Inter-Chip Interconnect, Fabrics Google Runsdecides: vendor · Ceded▼

Interconnects, switching and routing, network virtualization

Vendor-Provided

GPU RDMA Network (RoCE on A4X; Rail-Aligned, Separate From Host Networking)Ceded

The RDMA network that joins A4X Max and A4X instances: GPUDirect RDMA over RoCE on a hierarchical, rail-aligned architecture, on interfaces separate from general-purpose host traffic. Google runs the fabric and makes its routing and congestion calls. Ceded, the EFA and Azure InfiniBand reading. Not scored within this chip under the GA evidence table: A4X Max ("To get started with the A4X Max preview, please contact your Google Cloud sales representative.").

TPU Pods: Inter-Chip InterconnectCeded

Google's captive inter-chip interconnect wiring multi-host slices. Ceded with the pod on the integrated-system rule.

A4X (GB300 and GB200 NVL72): NVLink FabricCeded

The captive NVLink fabric inside the NVL72 racks, 18 instances and 72 GPUs to an NVLink domain. Ceded with the system on the integrated-system rule. Not scored within this chip under the GA evidence table: A4X Max ("To get started with the A4X Max preview, please contact your Google Cloud sales representative.").

Google Cloud VPC (Global Networks, Regional Subnets, Routes, Firewall Rules and Hierarchical Policies; Andromeda)Ceded

A virtual network inside Google's production network on Andromeda, Google's network virtualization stack. VPC networks, routes and firewall rules are global; subnets are regional. Firewall rules are enforced on the VMs.

Titanium: Network OffloadCeded

Network processing offloaded from the host CPU to the Titanium adapter, which the bandwidth docs tie to higher egress bandwidth. Facet of the Titanium chip.

Google Distributed Cloud (GDC): NetworkingCeded

Each organization gets its own isolated virtual network: a Default VPC for workloads, a system Infra VPC, project network policies as the segmentation boundary, and interconnect resources to external private networks. Facet of the GDC chip.

NVIDIA-Provided

NVLink in A5X Vera Rubin NVL72

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

NIXL, ConnectX and BlueField

NIXL for disaggregated inference, and ConnectX and BlueField for GPU networking. Google manages the NVIDIA integration layer.

◆ Gap Analysis

Google joins its A4X Max and A4X GPU instances on a rail-aligned RDMA network over RoCE, kept on interfaces apart from host traffic. The TPU inter-chip interconnect wires multi-host slices inside a pod, and the GB300 and GB200 NVL72 racks carry a captive NVLink fabric inside. Titanium offloads network processing from the host. VPC runs on Andromeda, with global networks, routes and firewall rules the enterprise writes. The buyer gets a scale-out fabric built for collective traffic without buying a switch. The architect's concern is visibility as well as design. The enterprise can't audit how Google's fabric shares bandwidth or prioritizes traffic. Virgo, the next fabric, was announced for TPU 8t and isn't generally available. Calibration: AWS's EFA 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 firewall rules and routes are the enterprise's code, but the RDMA GPU network and the inter-chip interconnect make the routing and bandwidth calls, and the enterprise can't see how they share bandwidth: vendor decides, not visible, not overridable, Ceded. Portability: the GPU network, the interconnect, VPC and Titanium are single-vendor surfaces, Ceded.

◆ Working Notes

Named, not scored (October 22, 2026): Virgo, announced April 22, 2026 for TPU 8t, which is Coming Soon; TPUDirect RDMA to storage and multisite clusters, announced alongside it; A5X on Vera Rubin NVL72, which isn't in the accelerator-optimized machine docs. Out of the GPU RDMA Network chip under the GA evidence table (October 24, 2026): A4X Max ("To get started with the A4X Max preview, please contact your Google Cloud sales representative."). Out of the A4X Max and A4X chip under the GA evidence table (October 24, 2026): A4X Max ("To get started with the A4X Max preview, please contact your Google Cloud sales representative.").

Layer 0C · StorageStorage SubstrateManaged Lustre and the Cloud Storage Rapid Tier Feed the Acceleratorsdecides: vendor · Delegated▼

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

Vendor-Provided

Managed Lustre (Parallel File System for Training Data)Ceded

10 TB/s bandwidth (10x YoY, claimed 20x faster than other hyperscalers), 80 PB capacity. RDMA-enabled. Storage substrate feeding TPU and GPU clusters, GA June 30, 2025; moved from 1C on October 5, 2026. The Lustre client is its own chip. This is the service Google runs, its performance tiers and the data movement around it, which don't lift to another Lustre. Ceded.

Cloud Storage XML API With S3 Interoperability (HMAC Keys)Delegated

The XML API with HMAC keys, which Google documents as interoperable with some S3 tools and libraries. 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, Rapid tier on Colossus, storage classes, lifecycle and replication, is the Cloud Storage chip, Ceded.

Cloud Storage + Rapid Tier (Colossus), through the JSON API and client libraries (GCS-native)Ceded

Standard object storage at planetary scale. Rapid tier uses Colossus, Google’s internal distributed storage platform (previously powering Search, Gmail, YouTube, Gemini training). Sub-millisecond read/write. The enterprise gets the same storage engine that holds Google’s training data. The service, its tiers, lifecycle and replication, and the JSON API and client libraries are Google's: Ceded. The XML API with S3 interoperability is its own chip.

Persistent Disk and Hyperdisk (Zonal Disks; Regional PD and Hyperdisk Balanced High Availability)Ceded

Block storage for Compute Engine. Zonal disks, plus synchronously replicated regional disks that hold one replica in the instance's zone and one in a secondary zone the enterprise picks; Asynchronous Replication to a second region.

Filestore (Zonal, Regional, Enterprise Multishares; NFSv3 and NFSv4.1)Ceded

Managed NFS file shares. Zonal tiers for HPC up to 100 TiB, regional tiers resilient to zone outages, and multishares for GKE; Basic tiers are legacy. NFS is its own chip. This is the service Google runs, its tiers, regional replication and multishares, which don't lift to another NFS server. Ceded.

Google Cloud NetApp Volumes (NFSv3, NFSv4.1, NFSv4.2, SMB, iSCSI, NVMe/TCP; Flex Unified ONTAP-Mode; External Replication Over SnapMirror)Ceded

Google's fully managed storage service, generally available since August 24, 2023: NFS, SMB, iSCSI and NVMe/TCP volumes across several service levels, with snapshots, backups and asynchronous volume replication. ONTAP-mode for Flex Unified pools went GA on April 17, 2026, and external replication, which uses bidirectional SnapMirror between ONTAP-based systems and NetApp Volumes, went GA for Flex Unified volumes on May 18, 2026. Google 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.

Google Distributed Cloud (GDC): Block StorageCeded

Network-based SSD block storage managed through Kubernetes APIs, with dual-zone persistent volumes replicated asynchronously. Facet of the GDC chip. Ceded, with Persistent Disk and Hyperdisk: the opinions accumulate in Google's block API.

Google Distributed Cloud (GDC): Object Storage (S3-Compatible)Ceded

S3-compatible object storage with dual-zone buckets replicated asynchronously. Facet of the GDC chip. The S3 API is its own chip; the dual-zone buckets and their replication are Google's: Ceded, with Cloud Storage.

Lustre Client Protocol on Managed 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 Managed Lustre chip, Ceded.

Front-End Protocols on Filestore (NFSv3, NFSv4.1)Delegated

NFSv3 and NFSv4.1, 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 managed service is the Filestore chip, Ceded.

Front-End Protocols on Google Cloud NetApp Volumes (NFSv3, NFSv4.1, NFSv4.2, SMB, iSCSI, NVMe/TCP, S3 in ONTAP-Mode)Delegated

NFSv3, NFSv4.1, NFSv4.2, SMB, iSCSI and NVMe/TCP, with S3 multiprotocol in ONTAP-mode only, per Google's NetApp Volumes overview; availability varies by service level and mode. Independent implementations share them: 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. Ceded behind them: ONTAP's data management is NetApp's, and the managed service's control plane is Google's.

S3-Compatible API on Google Distributed Cloud (GDC) Object StorageDelegated

The S3-compatible API on GDC's object storage, 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 dual-zone buckets and their replication are the GDC object storage chip, Ceded.

NVIDIA-Provided

Assessment pending

◆ Gap Analysis

Google covers the substrate end to end. Cloud Storage is the object foundation, and its Rapid tier runs on Colossus, the storage platform behind Search, YouTube and Gemini training, with sub-millisecond reads and writes. Persistent Disk and Hyperdisk handle block, Filestore serves NFS, and NetApp Volumes serves file and block on ONTAP as a managed service. Managed Lustre is the parallel file system that feeds TPU and GPU clusters, at up to 10 TB/s and 80 PB. The buyer gets AI-throughput storage sized to the workload. The cell rests on Managed Lustre, and that's where the enterprise keeps a handle. It sets striping explicitly with lfs setstripe and lfs setdirstripe across the instance's storage targets. Underneath, Cloud Storage decides durability placement with no per-object override. Calibration: AWS and Azure 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. Managed Lustre places data across its storage targets, and the enterprise can override a file's or a directory's layout: vendor decides, visible, overridable, Delegated. Portability: Filestore's NFS and the Managed Lustre client are multi-vendor standards, Delegated, and Cloud Storage splits by interface, Delegated through the S3-compatible XML API and Ceded through the JSON API; Persistent Disk, Hyperdisk and GDC's block storage read Ceded, GDC's S3-compatible object storage reads Delegated, and NetApp Volumes reads Ceded to NetApp, as ONTAP does through any seller.

Layer 1A · DataData Foundation & Governance1 LABKnowledge Catalog Over Open Tables; Gemini Context Graph in Previewdecides: vendor · Delegated▼

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

Vendor-Provided

Smart Storage Object ContextsCeded

Structured, mutable, IAM-governed context on every Cloud Storage object (GA): tags, classifications, and extracted entities the enterprise writes or annotation pipelines attach, discoverable by downstream systems. Automated annotation at write time with Gemini is preview and not scored. Not scored within this chip under the GA evidence table: Custom contexts with extended data (structured payloads Google services attach during automated analysis) ("Custom contexts with extended data ( Preview ): Contexts that consist of a key, a string value, timestamps, and an optional structured JSON payload.").

Knowledge Catalog (Formerly Dataplex Universal Catalog)Ceded

Governance catalog across BigQuery and Cloud Storage: automatic discovery (GA), column-level lineage (GA), data profiling and quality scans, data products (GA), Iceberg REST cataloging (GA), dbt metadata import (GA), and policy tags with row and column security. Gemini context-graph features (unstructured insights, lookupContext, relationships discovery) are preview and not scored.

BigQuery StorageCeded

Serverless columnar storage for structured/semi-structured. Managed Iceberg tables. Separates storage and compute. BigQuery spans storage, analytics, ML, and governance in a single service.

Lakehouse Runtime Catalog and Governance (Formerly BigLake)Ceded

Unified fabric across Cloud Storage and BigQuery: single schema, multiple engines, row and column governance through the BigQuery Storage API across access paths including open-source engines, multi-cloud reach through BigQuery Omni. The governance and runtime catalog are Google's own surface. BigQuery writes and automatic table management for managed Iceberg tables are preview and not scored. Moved from 1C October 5, 2026: a catalog grades at 1A.

Apache Iceberg REST Catalog Interface (Lakehouse)Delegated

Iceberg tables served through the Apache Iceberg REST catalog, a multi-vendor standard interface Spark, Trino, and other engines speak; Lakehouse namespace and table commands GA April 21, 2026, automated cataloging in Knowledge Catalog GA April 16, 2026. A Google-managed implementation behind a standard interface: the tables and their catalog contract lift to another Iceberg REST catalog.

NVIDIA-Provided

Assessment pending

◆ Gap Analysis

Knowledge Catalog, formerly Dataplex Universal Catalog, answers what the data is, generally available: discovery, column-level lineage, profiling and quality scans, data products, and automated cataloging of the Lakehouse Iceberg REST catalog. Policy tags and row and column security are enforced across BigQuery. Sensitive Data Protection publishes classification as aspects for BigQuery, Cloud SQL and Vertex AI datasets. The buyer gets a catalog a reasoning plane can query for lineage, quality and the classification of structured data. The architect's concern is where classification stops. It doesn't reach Cloud Storage buckets, so object classification is what the enterprise writes into object contexts. The Gemini half that would build the context graph agents ground on is mostly preview. Cloud Storage holds the bytes, a 0C function. The grade rested on the catalog, so it doesn't move.

◆ Borrowed Judgment

Policy tags, row and column security and quality rules are the enterprise's, enforced as written; discovery, profiling and classification are Google's judgment, and the enterprise can override each result per object. The grade rests on the catalog's description of the data: vendor decides, visible, overridable, Delegated, the Purview, Horizon and Unity Catalog reading. Portability: the Iceberg REST interface is Delegated; Knowledge Catalog, the Lakehouse runtime catalog and BigQuery Storage are Google's, Ceded.

◆ Working Notes

Named, not scored (preview): unstructured data insights, data relationships discovery, lookupContext, unstructured data profiling, third-party metadata connectors and Smart Storage automated annotations. At GA, read whether the catalog ranks and routes context (2C) or describes data (1A). Out of the Smart Storage Object Contexts chip under the GA evidence table (October 24, 2026): Custom contexts with extended data (structured payloads Google services attach during automated analysis) ("Custom contexts with extended data ( Preview ): Contexts that consist of a key, a string value, timestamps, and an optional structured JSON payload.").

Layer 1B · RetrievalContext Management & Retrieval1 LABModel Prep & Managed Retrievaldecides: vendor · Delegated▼

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

Vendor-Provided

BigQuery ML Embeddings for Vector Search (ML.GENERATE_EMBEDDING, AI.GENERATE_EMBEDDING)Ceded

BigQuery ML generates embeddings in the warehouse with Google's models, and Google documents how to "directly use the VECTOR_SEARCH function on the ML.GENERATE_EMBEDDING output" for semantic search and retrieval-augmented generation. Vectors are useful only with the model that made them: Ceded to the model owner. BigQuery ML's training and inference over tabular models is read at 2B.

Google Cloud Remote MCP Servers (BigQuery, AlloyDB, Spanner, Cloud SQL, Bigtable, Cloud Storage, Knowledge Catalog)Delegated

Managed MCP servers that "Run on the service's infrastructure and offer an HTTP endpoint to AI applications," enabled with each product's API. The BigQuery server (https://bigquery.googleapis.com/mcp) works with "AI applications including Antigravity, ChatGPT, Claude, and custom applications," exposing table discovery and SQL tools. A Google-run service behind the multi-vendor MCP standard: Delegated. Data Agent Kit, the IDE extension and coding-agent plugin that consume these servers, is developer tooling, noted rather than scored here.

Feature Store on Gemini Enterprise Agent Platform (Formerly Vertex AI Feature Store)Ceded

Managed feature serving for online/offline ML models and agents. Consistent feature serving across training and inference. 1B→2B bridge.

Agent Search (Formerly Vertex AI Search; Discovery Engine API)Ceded

Turnkey managed RAG and enterprise search via the Discovery Engine API (formerly Vertex AI Agent Builder, now part of Gemini Enterprise Agent Platform). Create a data store over Cloud Storage, BigQuery, or websites and Google handles parsing, chunking, embedding, indexing, ranking, and grounded retrieval end to end. The direct analog to Bedrock Knowledge Bases, but on a proprietary managed datastore with no swappable open backend, so more captive than Bedrock KB: the retrieval opinions have no open exit.

BigQuery Vector SearchCeded

VECTOR_SEARCH over an IVF vector index, GA (September 27, 2024). Hybrid keyword-plus-vector VECTOR_SEARCH and the ScaNN index type are in preview and sit in the notes. A self-orchestrate primitive: the enterprise builds the retrieval pipeline, but the store is BigQuery, a single-vendor API with no open exit.

Firestore Vector SearchCeded

K-nearest-neighbor vector search on transactional Firestore data, GA, with composite pre-filters and LangChain and LlamaIndex integrations. A self-orchestrate primitive on a single-vendor Firestore API: convenient for operational RAG, captive at the store layer.

NVIDIA-Provided

Assessment pending

◆ Gap Analysis

No meaningful capability gap. Most mature Layer 1B in the assessment series. BigQuery ML eliminates data-to-model handoff. LookML Agent automates semantic model construction. Data Agent Kit enables agent-driven pipeline development. The gap is governance over model-generated data artifacts. When LookML Agent generates a semantic model, when Data Agent Kit writes a pipeline, when BigQuery ML trains a model, who reviews the output for correctness? These are AI-generated artifact governance problems. Google provides no productized capability for governing model-generated data artifacts at scale. Data Agent Kit’s explicit support for Claude Code and non-Google tooling is strategically significant: the one point in the stack where third-party model access is genuinely equal. Retrieval is the surface this layer is named for, and Google covers it comprehensively but captively. Agent Search (formerly Vertex AI Search; the Discovery Engine API) is the turnkey managed RAG product, the direct analog to Bedrock Knowledge Bases: ingest, chunk, embed, index, and grounded retrieval handled end to end. BigQuery VECTOR_SEARCH and Firestore Vector Search are the self-orchestrate primitives. The governance fork mirrors AWS: consume Agent Search and Google owns the retrieval decisions (Ceded), or self-orchestrate on BigQuery or Firestore vector and Retain the retrieval-strategy opinions in code. The twist against AWS: every native Google vector store is single-vendor (BigQuery, Firestore, and Vector Search on Gemini Enterprise Agent Platform are all Ceded), so self-orchestration Retains the pipeline logic but still lands on a captive store unless the enterprise deliberately chooses AlloyDB or Cloud SQL pgvector, the one portable escape. Even the turnkey surface is more captive than AWS: Bedrock Knowledge Bases is Delegated because it orchestrates over swappable open backends, while Agent Search runs a proprietary managed datastore with no open exit.

◆ Borrowed Judgment

The semantic model as borrowed judgment: when LookML Agent generates definitions, every analytics query and agent interaction using those definitions inherits Gemini’s interpretation of business logic. Powerful (automates weeks of manual semantic modeling) and risky (embeds model judgment in the analytical foundation). The pipeline-building agent as borrowed judgment: Data Agent Kit agents write Dataflow jobs and BigQuery transformations. The enterprise inherits the agent’s data engineering judgment: join strategies, filter logic, null handling. Previously human expertise, now model-generated. The retrieval surfaces deepen the capture: Agent Search, BigQuery vector, and Firestore vector are all single-vendor APIs, so the enterprise inherits Google opinions on chunking, ranking, and grounding with no portable exit short of pgvector. Decision authority (October 5, 2026): Agent Search (formerly Vertex AI Search) parses, chunks, and ranks, but each request can set boostSpec, filter, and a relevance filter that overrides the global threshold. Vendor decides, visible, overridable, Delegated. Self-orchestrated BigQuery or Firestore vector would read code and Retain the retrieval logic.

◆ Working Notes

The comparison to AWS SageMaker Unified Studio: AWS provides a single governed environment across services (service-wide integration). Google achieves integration through BigQuery spanning storage, analytics, ML, and governance (service-deep integration). Google = tighter integration at cost of BigQuery lock-in. AWS = service diversity at cost of integration complexity. Watch-list (preview, not scored): LookML Agent (Gemini-powered semantic-model generation) - moved off the scored components per the GA-gate until it reaches GA; it remains a notable forward signal for model-generated semantic models. Dataflow and Managed Service for Apache Spark are processing engines, not retrieval, so they're not scored at 1B. GA evidence table (October 24, 2026), per named item: hybrid keyword-plus-vector VECTOR_SEARCH is marked Preview in BigQuery's release notes and comes out of the vector search chip until Google affirms it GA.

Layer 1C · PipelinesData Movement & PipelinesBigQuery-Native Pipelinesdecides: code · Retained▼

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

Vendor-Provided

BigQuery-Native Pipelines (Data Transfer Service, Dataform SQL Workflows, BigQuery Pipelines and Data Preparation, Continuous Queries)Ceded

Scheduled ingest from S3, Azure Blob, Redshift, Teradata, Snowflake, Salesforce, and Oracle through the Data Transfer Service; SQL transform workflows with dependencies, schedules, and assertions in Dataform; BigQuery pipelines with tables, views, data sources, and data quality tests as tasks (GA July 30, 2026); Gemini-assisted data preparation over Cloud Storage and Drive files (GA March 23, 2026); continuous queries streaming results to Pub/Sub, Bigtable, and Spanner (stateful operations are Preview and not scored). The workflows are written against BigQuery and don't lift to another warehouse without a rebuild: Ceded.

AI Data Prep in BigQuery (AI.GENERATE_EMBEDDING; Autonomous Embedding Generation)Ceded

Embeddings generated in SQL against Google or open embedding models (GA March 6, 2026), and autonomous embedding generation that keeps a declared embedding column current as rows change (GA June 17, 2026). Ceded on the embeddings carve-out: the vectors are useless without the same model at query time.

Dataflow (Apache Beam Batch and Streaming; Dataflow ML With RunInference and MLTransform)Ceded

Managed runner for the enterprise's Beam pipelines, with inference, embedding, and enrichment steps inside the same pipeline. The Apache Beam pipelines surface is its own chip. The service behind it, its operations, management and configuration, is Google's and doesn't lift: Ceded.

Datastream (Change Data Capture to BigQuery, Cloud Storage, and Iceberg)Ceded

Change data capture from operational databases with per-object and partial backfill. Datastream detects new source columns and adds them to the destination on its own. Sources marked Preview (Salesforce Marketing Cloud, Dataverse, ServiceNow) aren't scored.

Managed Service for Apache Airflow (Formerly Cloud Composer 3)Ceded

Managed Airflow running the enterprise's DAGs (GA December 16, 2024; Airflow 3 GA April 15, 2026). The Airflow DAGs surface is its own chip. The service behind it, its operations, management and configuration, is Google's and doesn't lift: Ceded.

Pub/Sub (BigQuery and Cloud Storage Subscriptions; AI Inference Single Message Transform)Ceded

Streaming ingest with subscriptions that write straight to BigQuery and Cloud Storage, and per-message model enrichment through the AI Inference transform (GA April 6, 2026). Import topics are named, not scored, pending per-source badges.

Knowledge Catalog Lineage and Data Quality on Pipelines (Formerly Dataplex Universal Catalog)Ceded

Lineage recorded automatically from BigQuery, Dataflow, Airflow, Spark, and pipeline jobs (GA; Looker lineage is Preview), and scheduled data quality scans with alerts on pipeline output. The day-two surface; the catalog itself is scored at 1A.

Airflow DAGs on Google Cloud's Managed Service for Apache AirflowDelegated

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.

Apache Beam Pipelines on DataflowDelegated

Apache Beam pipelines 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

BigQuery carries the strong on its own. It ingests on a schedule, transforms with dependencies and assertions, runs real-time SQL, and prepares data for AI inside the engine, with embeddings kept current as rows change. That's the Snowflake and Databricks shape, where the warehouse ships the pipeline. Dataflow, Datastream, managed Airflow, and Pub/Sub add depth around it, and lineage and data quality on the pipelines pass the day-two test. The grade doesn't need them. The cell was rebuilt October 5, 2026. It used to rest on five chips that couldn't carry it: two storage tiers, a catalog, a Preview federation feature, and an enrichment feature with no product documentation. The products that actually move data on Google Cloud appeared only in prose. Capability didn't change. The evidence did. Decision authority: BigQuery executes what the enterprise configured. Data preparation's Gemini suggestions happen at design time, and at runtime the accepted steps run as written; autonomous embedding maintains a column the enterprise declared. Code decides, visible, overridable, Retained, the judgment test and its defaults clause (October 4 and 5, 2026). Datastream is the exception: it reads the source and adds columns on its own. It adds depth but doesn't carry the grade.

◆ Borrowed Judgment

Low at runtime, higher at design time. The pipelines run what the enterprise wrote, but the enterprise writes them in BigQuery's dialect, and Gemini's preparation suggestions shape what gets written. The judgment Google keeps is in Datastream's schema handling and in the embedding models the AI prep calls. The same embed-and-index fork as 1B applies to the RAG ingest pipeline: consume Agent Search (formerly Vertex AI Search) and the ingest, chunk, and embed movement is managed for you (Ceded), or self-orchestrate it in Dataflow or BigQuery and Retain the pipeline logic, though the chosen store stays captive.

◆ Working Notes

Named, not scored (October 5, 2026): Cross-Cloud Lakehouse is Preview under Pre-GA terms, with catalog federation and intelligent caching also Preview; Smart Storage auto-annotate at ingest appears in a Google blog with no product documentation. RAG Engine ingestion is GA only in europe-west3 and europe-west4 (allowlist in the US regions, Preview elsewhere) and isn't cited as a carrier. Re-homed under the M2 storage ruling: Managed Lustre to Layer 0, BigLake (renamed Lakehouse, April 2026) to 1A; Rapid Tier and Smart Storage were already scored at 1A. The asymmetry between data plane federation and control plane federation (absent at 2C) is a structural finding. Google invests in making data accessible across clouds but not in making agent governance portable across clouds. Data accessibility without orchestration portability draws workloads toward GCP as the governance center.

Layer 2A · OrchestrationInfrastructure OrchestrationGKE + Sovereign Extensionsdecides: vendor · Ceded▼

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

Vendor-Provided

GKE + GKE Agent SandboxDelegated

Managed K8s with AI-era extensions. Agent Sandbox: gVisor-based secure isolation, 300 sandboxes/second/cluster with sub-second time to first instruction. Infrastructure built for the agentic era, not retrofitted. GKE is a managed service behind the standard Kubernetes API: manifests lift to another conformant cluster, so Delegated (operation delegated to Google; the standard interface keeps the opinions portable).

Fluid ComputeCeded

GCE + GKE dynamically shifting workloads in real-time. CPUs for branchy agent logic, secure sandboxes, RL, SLM inference, RAG. GPU/TPU for training and large-model inference. Proto-Layer 2C: routes based on workload characteristics, not business context.

GDC (Sovereignty Analysis)Delegated

On-prem Google Cloud services: GKE, Agent Platform, managed storage, Gemini Flash, Blackwell GPUs. Air-gapped for sensitive workloads. Addresses data sovereignty (where computation happens) but NOT judgment sovereignty (whose model drives computation). GDC runs Gemini on-prem: same recursive dependency, inside the enterprise perimeter.

Capacity ManagementCeded

CUDs (1yr/3yr), on-demand, preemptible/spot, Dynamic Workload Scheduler. Same pattern as AWS: capacity acquisition (2A), not workload placement (2C).

NVIDIA-Provided

NVIDIA GPU Operator (on GKE)

Available for NVIDIA instances on GKE. Google manages the GPU integration layer.

◆ Gap Analysis

GKE is the most mature managed K8s for AI workloads. GKE Agent Sandbox has no equivalent in Dell, HPE, or AWS portfolios: 300 sandboxes/second is built for agentic workload density. Fluid Compute sits at the 2A/2C boundary. Its dynamic workload shifting is more than capacity acquisition: runtime decisions about which compute type serves which workload. But less than full 2C: routes on workload characteristics, not business context (data residency, compliance tags, cost targets). The Fluid Compute → Knowledge Catalog connection does not exist: workload placement does not consult governance metadata. Same Infrastructure Layer 2C gap every vendor has. GDC: the full sovereignty analysis reveals that data sovereignty ≠ judgment sovereignty. Knowledge Catalog on GDC uses Gemini. Smart Storage on GDC uses Gemini. Agent Platform on GDC uses Google’s runtime. The enterprise gains physical sovereignty but retains the same judgment concentration. The ‘self-driving cloud’ narrative implies Gemini-powered infrastructure operations: autonomous root-cause analysis on infrastructure telemetry. If the Reasoning Plane is itself Gemini-powered, the operational intelligence and application intelligence are the same intelligence. 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. Customers routinely run self-managed Kubernetes, Slurm, or a third-party distribution on Compute Engine rather than adopting GKE, in which case Google is bought as substrate and the 2A judgment belongs to whoever runs the control plane. GKE's maturity is a statement about the option on offer, not about occupancy of the layer.

◆ Borrowed Judgment

GKE consumed via the standard Kubernetes interface: the enterprise's manifests and operators lift to any Kubernetes (Delegated: managed K8s service; the enterprise could switch without rebuilding). Google's optional Gemini-driven scheduling optimization is the only borrowed-judgment layer, and only if the enterprise opts in. GDC as borrowed judgment in sovereign packaging: physical control over facility, Google’s judgment in software, model, governance, and operations. Data sovereignty with judgment concentration. Fluid Compute as proto-2C borrowed judgment: when Fluid Compute routes agent work to CPU vs GPU, that routing is Google’s judgment about optimal compute matching. Enterprise doesn’t configure the routing policy.

◆ Working Notes

Data sovereignty vs judgment sovereignty: the 4+1 framework should distinguish between where data resides (which GDC addresses) and whose model’s reasoning shapes the AI system (which GDC does not address). An enterprise running GDC air-gapped has data sovereignty while fully ceding judgment sovereignty.

Layer 2B · RuntimeApplication Runtime & Execution2 LABSModel-Integrated Stackdecides: model · Delegated▼

Model serving, agent execution, inference APIs, distributed inference

Vendor-Provided

Gemini Enterprise Agent PlatformCeded

Unified platform for building, scaling, governing, optimizing agents. Subsumes all Vertex AI services. Agent Studio (low-code), ADK (code-first, Python/Go/Java/TypeScript, model-agnostic, open-source), Model Garden (200+ models incl Gemini, Claude, Llama, Gemma), Agent Runtime, Agent-to-Agent Orchestration, Agent Identity (GA), Agent Gateway, Agent Observability, Agent Registry, Memory Bank, Antigravity (desktop app + CLI). Not scored within this chip under the GA evidence table: Agent Gateway, Antigravity (desktop app + CLI) ("Release Agent Gateway for Private Preview.").

Model Serving via the OpenAI-Compatible Vertex EndpointDelegated

Vertex documents an OpenAI-compatible endpoint for Gemini, and serves third-party models (Claude through Anthropic’s Messages API, Llama) alongside it. Consumption through those interfaces is a managed service behind a multi-vendor standard, so the integration opinions lift to any compatible endpoint and the serving interface is Delegated. The model itself remains Google’s; what is portable is the interface, not Gemini.

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

GA through the OpenAI-compatible Vertex endpoint and Gemini function calling. 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.

The House Model AdvantageCeded

Gemini on TPU: silicon designed for this model, networking designed for its topology, distributed runtime (Pathways) built for its coordination, inference optimized for its architecture, governance (Knowledge Catalog) powered by it, orchestration defaults to it. No other vendor achieves this degree of vertical optimization. Third-party models (Claude, Llama) run on NVIDIA GPUs, supported but not co-optimized. The playing field is structurally tilted. The standard-interface consumption path is scored separately; what remains here is the platform surface beyond it, where the vertical co-optimization and the governance defaults accumulate.

Pathways (Distributed Runtime)Ceded

Google’s distributed runtime for superpod-scale training and inference coordination. Proprietary to Google Cloud with no implementation elsewhere, so coordination and scaling opinions written against it have nowhere to go. This is the part of the framework stack that actually captures.

JAX (Apache 2.0; GPU, CPU and TPU Through XLA)Delegated

Apache 2.0 and runs on GPU and CPU through XLA as well as TPU. The consumed interface is the framework API, which independent vendors implement; model code written against it lifts to other hardware. TPU optimization is a performance fact, not an authority fact.

TorchTPU (PyTorch on TPU)Delegated

Full PyTorch on TPU; PyTorch code is hardware-agnostic, so it lifts to other accelerators without rebuilding. The owner’s wording hedges: the launch post calls TorchTPU "in active development" while saying it’s "used in production" by some of the largest Google Cloud customers. That’s no preview marker, so it reads GA, but it’s the softest GA evidence in this cell.

vLLM on TPU (vLLM JAX)Delegated

vLLM optimized for TPU through the vLLM project’s TPU backend, now vLLM JAX. Open source with real alternatives; serving configuration lifts to vLLM on other hardware.

llm-d (Kubernetes-Native Distributed Inference; CNCF Sandbox)Delegated

Open-source Kubernetes-native distributed inference serving with multi-vendor participation, on GKE with TPU and GPU. The deployment and routing configuration lifts to any conformant Kubernetes.

NVIDIA-Provided

NVIDIA GPU Instances

A3/A3 Mega/A3 Ultra/A5X for third-party model inference. CUDA ecosystem required for non-Gemini models.

◆ Gap Analysis

Layer 2B is the center of gravity for the model-integrated stack. The model provider, runtime provider, and infrastructure provider are the same company. When the enterprise runs Gemini on Agent Platform on TPU, it borrows Google’s judgment at the model layer, runtime layer, framework layer, and silicon layer simultaneously. A single entity’s priorities shape the entire execution path. The Agent Platform collapses Layers 2B, 2C, and 3 into a single product surface: Agent Runtime (2B infrastructure), Agent Identity/Gateway/Registry/Orchestration/Observability (2C governance), Agent Studio/ADK/Antigravity (Layer 3 development). The product boundary does not align with the architectural boundary. AWS separates these: Bedrock (model access) is distinct from AgentCore Runtime (agent execution) is distinct from AgentCore Policy (governance). AWS’s separation preserves architectural boundaries the enterprise can independently govern. Google’s collapse optimizes integration but prevents swapping the governance layer (2C) while keeping the runtime (2B). The NVIDIA dependency at 2B is optional for Gemini (TPU-native) but required for third-party models. The enterprise using Claude on Google Cloud pays a structural performance tax: Claude runs on NVIDIA GPUs through a runtime designed for Gemini. Model Garden’s 200+ models are API-equal but not silicon-equal.

◆ Borrowed Judgment

The model-integrated runtime: Gemini on TPU inherits Google’s judgment at model layer (training data, alignment, safety), runtime layer (scheduling, scaling, session management), framework layer (JAX/Pathways optimization), and silicon layer (TPU architecture). Most concentrated borrowed judgment in the assessment series. The open-source hedge: llm-d, TorchTPU, vLLM, ADK provide genuine open alternatives. Google opens components that reduce adoption friction (frameworks, SDKs) while keeping authority-concentrating components closed (Agent Runtime infrastructure, Agent Gateway, Pathways). Production deployment pulls open tools into Google’s managed surface where authority shifts from Retained to Ceded.

◆ Working Notes

The 2B/2C collapse prevents the enterprise from independently governing the orchestration layer. An enterprise that wants Google’s Agent Registry and Agent Identity but AWS’s Bedrock for model access and its own governance engine for policy enforcement cannot compose that architecture. The components are bundled. Lab-measured (July 2026, reconcile v1.6): the gemma4-tpu-inference lab measured the bring-your-own serving path on the self-serve TPU lane, using a model that is itself TPU-native. The constraint it hit is common to mixture-of-experts models sharded across accelerators on any substrate: two global key-value heads cap pure tensor parallelism at 2, and at that width bf16 does not fit v5e's per-chip memory budget. The single-device contrast in the lab program, 128GB of unified memory on the GB10 serving the same model whole, is a memory-topology difference, not a runtime-openness difference. The lane ruling stands on its own terms: on reachable self-serve silicon the deployable paths were quantizing off-box or adopting Google's serving stack, and that adoption cost is this cell's Ceded reading in practice. Whether Google's own managed stack serves this model well was not the question and was not measured. Out of the Gemini Enterprise Agent Platform chip under the GA evidence table (October 24, 2026): Agent Gateway, Antigravity (desktop app + CLI) ("Release Agent Gateway for Private Preview.").

Layer 2C · ReasoningAgentic Infrastructure — The Reasoning Plane2 LABSComplete Intelligence-2C Plane, No Placementdecides: model · Ceded▼

Policy-driven placement and resource coordination — the Autonomy Layer

Vendor-Provided

Agent Identity (GA)Ceded

Agents as identity principals with authentication, authorization, audit. Control plane function: determines which agents exist as governed entities.

Agent GatewayCeded

Protocol-level governance for MCP and A2A communications. Security partner integrations (Broadcom, Check Point, Cisco, CrowdStrike, F5, Netskope, Okta, Palo Alto, Zscaler). Spans Layer 0 (networking), 2B (runtime), and 2C (orchestration).

Agent-to-Agent OrchestrationCeded

Deterministic multi-agent workflow routing. Control plane function: determines which agent handles which subtask. Not scored within this chip under the GA evidence table: A2A agents on Agent Runtime ("Preview This feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms .").

Agent RegistryCeded

Catalog of agents with ownership, capabilities, protocols, invocation details. Administrator-controlled discoverability. Control plane function: determines which agents are available and who can use them.

Agent ObservabilityCeded

Monitoring, tracing, debugging across production agent populations. Feedback loop for detecting faulty reasoning and intervening.

NVIDIA-Provided

No NVIDIA Layer 2C Dependency

All Layer 2C components are Google IP. NVIDIA does not control governance, policy, or reasoning in Google’s stack.

◆ Gap Analysis

Google’s Intelligence Layer 2C is the most complete productized offering in the assessment series: Agent Identity + Gateway + Registry + Orchestration + Observability + Memory Bank. Together they constitute a genuine control plane for agent governance. Infrastructure Layer 2C (the autonomous placement engine) is NOT built as a customer-configurable product. The capacity primitives (Fluid Compute, CUDs, DWS) are building blocks, but they don’t compose into a policy-driven placement engine querying Knowledge Catalog governance metadata. Same gap as AWS and every other vendor. Google’s implicit Layer 2C is the most sophisticated in the assessment: managed services make autonomous placement, scaling, routing, and capacity decisions invisibly. The enterprise cannot see, configure, audit, or override these decisions. The model-integrated Reasoning Plane: if Google’s ‘self-driving cloud’ uses Gemini for infrastructure decisions, then the model powering the enterprise’s agents (Layer 3) is the same model governing agent orchestration (Intelligence 2C) is the same model deciding where agents run (Infrastructure 2C). One model’s judgment pervades every decision surface. Cross-cloud orchestration gap: Google federates the data plane (Cross-Cloud Lakehouse) but NOT the control plane. Agent Platform governs GCP agents only. Enterprise running agents across multiple clouds has no cross-platform agent governance surface, unless all agents route through Google’s Agent Gateway, which cedes cross-cloud governance to Google. The captive-but-best dilemma: this is evidence the control plane CAN be built as a coherent capability. The enterprise architect who wants it has one option: adopt Google Cloud. The federated alternative does not exist. Calibration (September 2, 2026): strong on the completeness criterion, not on placement. All five legs of an Intelligence-2C plane are productized and GA here: first-class agent identity, a request-time gateway, a registry, cross-agent orchestration, and observability. Salesforce lacks the identity leg, VMware the registry and orchestration legs, Nutanix everything but the gateway, and each reads moderate for that stated reason. No reasoning mechanism ships; placement stays the universal gap.

◆ Borrowed Judgment

The captive control plane: enterprise inherits Google’s orchestration model: deterministic routing, Google-managed identity, Google-governed protocols. Well-engineered but unchallengeable: cannot substitute alternative orchestration logic within the Agent Platform boundary. The model-powered control plane: if the Reasoning Plane uses Gemini for infrastructure decisions, a model judgment error at the control plane layer is invisible to the enterprise, with no fallback to human decision-making. Intelligence 2C: Low borrowed judgment in the sense that the components are productized and configurable. High borrowed judgment in the sense that the governance logic itself (Agent Gateway protocol decisions, Agent Identity authentication model, Orchestration routing patterns) is Google’s, not the enterprise’s.

◆ Working Notes

Google’s 2C proves the Control Plane Working Notes thesis: the control plane can be built. The question is whether it can be liberated from the vendor boundary, and whether the model-integrated dimension (control plane powered by the same model it governs) is a pattern to replicate or to avoid. The asymmetry: data plane federates (Cross-Cloud Lakehouse), control plane does not. This serves Google’s strategic interest: data accessibility without orchestration portability draws workloads toward GCP as governance center. 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. Out of the Agent-to-Agent Orchestration chip under the GA evidence table (October 24, 2026): A2A agents on Agent Runtime ("Preview This feature is subject to the "Pre-GA Offerings Terms" in the General Service Terms section of the Service Specific Terms .").

Layer 3 (+1) · ApplicationsAI Application Layer — The Value PlaneAntigravity and Gemini Enterprise: First-Party Agents That Do the Workdecides: model · Delegated▼

AI-powered business capabilities — business logic, workflow automation

Vendor-Provided

Google Antigravity 2.0 (App, CLI and IDE Extensions; Enterprise Through Gemini Enterprise)Ceded

Google's agent that does the work: the Antigravity App (v2.5+), the agy CLI (v1.2+) and IDE extensions run autonomous multi-agent tasks in the enterprise's repositories and workspaces, with goals run to completion, scheduled background tasks, browser research and parallel subagents. GA product documentation includes an enterprise path: Google Cloud SSO or Workforce Identity Federation, IAM roles, VPC Service Controls, data residency for first-party Gemini models, and Gemini Enterprise Standard or Plus seat licenses, with Claude models available on administrator opt-in (JetBrains, Zed and Xcode extensions are Preview). Projects, plugins, schedules and agent configuration are Antigravity objects: Ceded. Not scored within this chip under the GA evidence table: JetBrains extension, Zed extension, Xcode extension ("(Enterprise support in Preview)").

Gemini Enterprise Deep Research Agent and Agent DiscoveryCeded

Google's own agents inside the Gemini Enterprise app that carry out research tasks for business users and surface the agents an enterprise has published. They do the work themselves: Ceded.

NVIDIA-Provided

NVIDIA Models via Model Garden

NVIDIA Nemotron and other NVIDIA models available alongside all other providers.

◆ Gap Analysis

Google's Layer 3 grade rests on first-party agents that do the work, and Antigravity carries it on its own. Antigravity 2.0 (launched at I/O, May 19, 2026) is an agent-first work surface: the desktop App, the agy CLI that replaced Gemini CLI, IDE extensions, and an SDK. Its agents take a goal and run it to completion, work in parallel as subagents, run on a schedule in the background, and research in the browser, the same shape as Claude Code, Kiro and GitHub Copilot's agent mode. An agent that does the work is a business application even when it also writes software. Its documentation is GA with an enterprise path: Gemini Enterprise seat licenses, Google Cloud SSO and Workforce Identity Federation, IAM, VPC Service Controls, and regional data residency for first-party Gemini models, with Anthropic's Claude models available on administrator opt-in. Gemini Enterprise adds Google's own Deep Research agent and agent discovery for business users. That stands with the best shipping implementation on the instrument: strong. Not scored at Layer 3: the Gemini model family (models earn no Layer 3 grade; read at 2B), the Agent Development Kit and the Agent Studio and Agent Designer builders (software-only builders earn no Layer 3 grade), and the consumer-enterprise feedback loop, which is a strategic observation rather than a product. Model Garden reads at 2B, inside Agent Platform: more than 200 models, Claude, Llama and Gemma among them, procured on Google's paper. The architect's concern is where the agents land. Antigravity's projects, plugins and schedules are Antigravity objects, and enterprise use runs through Gemini Enterprise and Agent Platform's identity, gateway, registry and observability (2B and 2C). Google keeps the model choice open at Layer 3 and concentrates the commitment at 2B and 2C. Calibration: AWS reads strong on Amazon Q and Kiro, Azure on Microsoft 365 Copilot and GitHub Copilot, Anthropic and OpenAI on their own agents.

◆ Borrowed Judgment

The enterprise borrows the model's judgment inside Antigravity's agents: the model decides which commands to run and which changes to make. Antigravity documents permissions, sandboxing and execution gates that stand between the agent's decision and its effect, so the reading is Delegated at the model boundary, as on AWS and Azure. Model choice is open (Gemini, or Claude on opt-in), and Model Garden widens it to more than 200 models procured on Google's paper, a relationship that reads at 2B; the work surface, its projects and plugins, and the enterprise plumbing under Gemini Enterprise are Google's.

◆ Working Notes

Named, not scored at Layer 3: the Gemini model family (read at 2B), the Agent Development Kit, Agent Studio and Agent Designer (software-only builders earn no Layer 3 grade), the consumer-enterprise feedback loop (Gemini across Search, Gmail, Docs and Android improves enterprise models; consumer-driven alignment can also constrain enterprise use). The Gemini CLI to Antigravity CLI migration (June 18, 2026) consolidated Google's developer agent surface. JetBrains, Zed and Xcode extensions are Preview. Named, not scored at Layer 3 (October 22, 2026): Model Garden. Model catalogs earn no Layer 3 chip; it reads at 2B, inside Agent Platform. This supersedes the earlier relationship-chip reading. Out of the Google Antigravity 2.0 chip under the GA evidence table (October 24, 2026): JetBrains extension, Zed extension, Xcode extension ("(Enterprise support in Preview)").

Sources & revision history · v1.46 - 4+1 v3: Layer 0 Sub-Layers

Google Cloud Next 2026 (Apr 22–24), GTC 2026, NVIDIA partnership, Forrester, SiliconANGLE, The New Stack, analyst coverage. v1.2 (instrument reconciliation): 2A GKE Retained→Delegated: managed K8s behind a standard interface is Delegated, not Retained. Cloud Storage remains Ceded (GCS-native API). v1.3 (lab-validated against the TFD corpus / vCTOA RAG build): added the retrieval surface to Layer 1B - Vertex AI Search (Discovery Engine, the Bedrock Knowledge Bases analog), BigQuery Vector Search, and Firestore Vector Search, all Ceded (single-vendor APIs); demoted the preview LookML Agent from a scored component to a notes watch-line per the GA-gate; and added the self-orchestrate-versus-managed Retain/Delegate fork to 1B/1C, mirroring the AWS row. 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. Note added at 2B (Vertex OpenAI-compatible endpoint); no chips moved: the scored components are platform-altitude. /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): the Frameworks component bundled open-source JAX, TorchTPU, vLLM, and llm-d with Google-proprietary Pathways under a single Ceded call. Split: Pathways stays Ceded, the open frameworks score Delegated on the open-substrate rule. Dataflow/Dataproc moved Ceded to Delegated: managed OSS behind a standard interface, matching aws 1C Glue ETL, databricks 1C Spark, and ibm 1C ML Pipeline Stack. /reconcile (September 1, 2026): shared 2C findings stated explicitly (live-placement universality, outcome-validation finding). /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): Layer 0 runtime rule (ruled September 28 and October 4, 2026). TPUs split by adopted runtime (JAX and PyTorch/XLA Retained, TPU-specific code Ceded); the ownership reasoning retired. NVIDIA GPUs on Google Cloud split: standard instances Ceded to Retained from the customer's seat, NVL72 capacity Ceded on the integrated-system rule. Grade unchanged. /reconcile (October 5, 2026, same pass): TPUs re-read on the generally available generations (TPU7x, v6e, v5p); TPU 8t and 8i descored to named-not-scored under the GA-gate. Single-host slices split by runtime; pod-scale capacity Ceded on the integrated-system rule, matching AWS Trainium3 UltraServers. /reconcile (October 5, 2026): Layer 1C rebuilt on GA evidence (docs.cloud.google.com BigQuery, Datastream, Pub/Sub, Composer, and Knowledge Catalog docs and release notes). Strong holds on BigQuery-native pipelines, which carry the grade alone (Keith, October 5, 2026). The five prior chips couldn't carry it: Managed Lustre moved to Layer 0 and BigLake to 1A under the M2 storage ruling, Rapid Tier and Smart Storage were already at 1A, Cross-Cloud Lakehouse (Preview) and Smart Storage auto-annotate (undocumented) named, not scored. 1C authority vendor / Ceded to code / Retained. Layer 0, 1A, and 1C grades unchanged. /reconcile (October 5, 2026): authority re-read against product docs under the judgment test, override rule, and Layer 3 ruling: layer1b to vendor / Delegated. Grades and components unchanged. /reconcile (October 5, 2026): product names updated to Google's April 2026 renames (docs.cloud.google.com Gemini Enterprise Agent Platform name changes): Vertex AI Search to Agent Search, Vertex AI Feature Store and Vector Search to Gemini Enterprise Agent Platform, Dataproc to Managed Service for Apache Spark; the first use in each chip keeps the former name. Earlier entries in this log keep the names in use at the time. Grades, DAPM, and authority unchanged. /reconcile (October 5, 2026): authority read against product docs: layer2a visible true; layer0 visible false to true (physicalHostTopology documents cluster, block, sub-block, and host placement, the AWS instance-topology reading). Grades and components unchanged. Layer 1A re-read (October 5, 2026): public-source check found the Gemini context-graph features in preview (Knowledge Catalog release notes) and Smart Storage automated annotation in preview. Grade holds strong on the GA catalog core, which classifies the data; Smart Storage rescoped to object contexts (GA); Knowledge Catalog credited on discovery, lineage, profiling, quality, data products, and structured-data classification, with the Sensitive Data Protection aspect integration's Cloud Storage limit stated; authority vendor / not overridable / Ceded to vendor / visible / overridable / Delegated on per-object override of aspects, object contexts, and sensitivity tags; BigLake split into the Lakehouse runtime catalog (Ceded) and the Iceberg REST catalog interface (Delegated). ChatGPT peer check (reviews/gcp-1a-chatgpt.json): 8 findings upheld, 1 rejected under the watch-list GA-date rule. v1.20 (October 16, 2026, Layer 0 portability acid test): Google Distributed Cloud's Layer 0 chip Delegated to Ceded. Keith's test asks whether the services that support Layers 1 and 2 decouple from the vendor's Layer 0, read on the service being measured. Gemini runs on NVIDIA Blackwell in GDC and still can't move to another vendor's NVIDIA GPUs, so the hardware reads with the service. Settles the reconcile item that had GDC Delegated against AWS AI Factories and OCI Dedicated Region Ceded. Layer 0 grade and authority unchanged; the layer2a GDC chip is untouched. v1.21 (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. v1.22 (October 19, 2026, /reconcile) Layer 3: the Agent Development Kit (an SDK; read at 2B), and Agent Studio and Agent Designer (low-code and no-code builders) 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) Layer 3 rebuilt on Antigravity 2.0, documented GA with a Gemini Enterprise path (Keith: "Antigravity on its own would make GCP Strong"), plus the Gemini Enterprise Deep Research agent; the Gemini model family chip left Layer 3 (models earn no Layer 3 grade) and the consumer-enterprise feedback loop moved to notes; Model Garden stays as a relationship chip. Authority vendor and Ceded to model and Delegated on Antigravity's documented execution gates, the AWS and Azure reading. Grade holds strong. (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. v1.23 (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. v1.24 (October 21, 2026, /reconcile): 0C chip added for Google Cloud NetApp Volumes, from Google's NetApp Volumes overview and release notes (read October 21, 2026: GA August 24, 2023; Flex Unified ONTAP-mode GA April 17, 2026; external SnapMirror replication GA May 18, 2026). Google sells and runs it, so it's Google's chip; Ceded to NetApp on portability with Google deciding at runtime (Keith, October 21, 2026). Grade and cell authority unchanged; the cell rests on Managed Lustre. v1.25 (October 21, 2026, /reconcile): 0C GDC storage facet split into block (Ceded) and S3-compatible object (Delegated), from reconcile's peer calibration: the object half read Ceded only because it shared a chip with block, while Cloud Storage reads Delegated through its S3-compatible API. Grade and cell authority unchanged; the cell rests on Managed Lustre. v1.26 (October 22, 2026, /reconcile): Virgo (announced April 22, 2026 for TPU 8t, with no availability statement) and A5X on Vera Rubin NVL72 (not in the accelerator-optimized machine docs) move to named, unscored notes at 0B and 0A. In their place go the GA RoCE GPU network on A4X Max and A4X (AI Hypercomputer networking overview) and the A4X Max and A4X NVL72 racks (machine docs: 18 instances and 72 GPUs to an NVLink domain). Found by the 0B fabric research. Grades and authority unchanged; the 0B cell rests on the RoCE GPU network and the inter-chip interconnect. v1.27 (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. v1.28 (October 22, 2026, /reconcile): Model Garden comes off Layer 3 under the model-catalog ruling (catalogs belong to 2B, where Agent Platform already lists it), superseding the earlier relationship-chip reading, and the relationship point moves to the Layer 3 DAPM prose. Grade and authority unchanged. v1.29 (October 22, 2026, /reconcile): the 0A NVIDIA chip adds A4 (eight B200) and drops an "industry-first" market-position phrase and an unsourced line on where third-party models run; GDC's GPU list follows the air-gapped machine types docs (A100 through B300); the NVIDIA column drops A5X and Vera Rubin, which the notes already hold as unscored; a stale A4 Ultra preview note comes out. Found by the 0A compute research. Grade and authority unchanged. v1.30 (October 23, 2026, /reconcile): Layer 1 research pass. 1B: rewritten (the chip described in-warehouse ML training; the 1B function is embedding generation for vector search). 1B: rewritten (the managed MCP servers are the capability, usable without the agent kit; the kit is a wrapper). 1B: "Dataflow + Managed Service for Apache Spark (Formerly Datapr" removed (Dataflow and Spark are processing engines, not retrieval). Found by the Layer 1 slot cut. Grade and cell authority unchanged. v1.31 (October 23, 2026, storage door pass): The Lustre client, the S3-compatible XML API, NFS on Filestore and NetApp Volumes' protocols are doors; the managed services read Ceded. Grade unchanged. v1.32 (October 23, 2026, Layer 1 managed-services pass): Managed Service for Apache Airflow splits; Dataflow splits. Grade unchanged. v1.33 (October 24, 2026, /reconcile): GA evidence recorded on 5 of 6 Layer 1B chips under the GA evidence table. Grade unchanged. v1.34 (October 24, 2026, /reconcile): Layer 1B under the GA evidence table version 2: GA evidence on every chip, preview items and unaffirmed chips to notes. Grade unchanged. v1.35 (October 24, 2026, /reconcile): GA evidence recorded on 4 of 5 Layer 1A chips under the GA evidence table version 2. Grade unchanged. v1.36 (October 24, 2026, /reconcile): GA evidence recorded on 9 of 9 Layer 1C chips under the GA evidence table version 2. Grade unchanged. v1.37 (October 24, 2026, /reconcile): GA evidence recorded on 4 of 6 Layer 0B chips under the GA evidence table version 2. Grade unchanged. v1.38 (October 24, 2026, /reconcile): GA evidence recorded on 7 of 8 Layer 0A chips under the GA evidence table version 2. Grade unchanged. v1.39 (October 24, 2026, /reconcile): GA evidence recorded on 12 of 12 Layer 0C chips under the GA evidence table version 2. Grade unchanged. v1.40 (October 24, 2026, /reconcile): GA evidence recorded on 4 of 4 Layer 2A chips under the GA evidence table version 2. Grade unchanged. v1.41 (October 24, 2026, /reconcile): GA evidence recorded on 5 of 6 Layer 2B chips under the GA evidence table version 2. Grade unchanged. v1.42 (October 24, 2026, /reconcile): GA evidence recorded on 1 of 2 Layer 3 chips under the GA evidence table version 2. Grade unchanged. v1.43 (October 24, 2026, /reconcile): GA evidence recorded on 4 of 5 Layer 2C chips under the GA evidence table version 2. Grade unchanged. v1.44 (October 24, 2026, /reconcile): GA evidence table version 2.1 applied: 7 chips recorded or narrowed. Grade unchanged. v1.45 (October 24, 2026, /reconcile): Layer 2 bundle split: Open Frameworks on TPU split by project. Grade unchanged. v1.46 (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