Cloud Data and AI Stacks

AWS, Databricks, Google Cloud, IBM, and Microsoft

The Shared Stack

Five vendors, one stack shape, different product names.

Five Vendors Sell the Same Stack

  • Same layers: Every vendor sells infrastructure, a lakehouse, operational data, context, agents, and apps.
  • Different names: The product names differ, and many changed in 2025 and 2026.
  • Real differences: The vendors differ most in table format, catalog, model choice, and where their products run.
  • Not only clouds: Databricks runs on other vendors' clouds. IBM runs on its own cloud, other clouds, and on premises.

Data Layers Store and Run the Company's Data

  • Infrastructure: Custom chips and NVIDIA GPUs train and run the models.
  • Lakehouse: Open tables in object storage hold the data once. Many query engines read the same tables.
  • Operational data: Databases and streams hold the present. The lakehouse holds the history.
  • Graph engine: Operational graphs sit next to the databases. Analytical graphs sit on the lakehouse.

AI Layers Turn Data into Answers and Actions

  • Context: The context layer stores what data means: metrics, business terms, and relationships.
  • Agents and models: The agent platform hosts models and connects agents to tools and memory.
  • Apps: Enterprise chat, coding assistants, and app builders put agents in front of people.
  • Agent governance: Identity, registry, gateway, and policy span every layer.

Where the Vendors Differ

Same layers, different defaults.

Four Questions Place Any Vendor on the Stack

  1. Find the data: Which table format and catalog hold the data?
  2. Find the meaning: Where do business definitions and relationships live?
  3. Find the agents: Which platform builds agents, and which models does it offer?
  4. Find the people: Which chat and coding tools do employees open every day?

Table Format Matters Less, the Catalog Matters More

  • Iceberg first: AWS, Google Cloud, and IBM store data in Apache Iceberg by default.
  • Delta first: Microsoft stores data in Delta Lake and translates metadata for Iceberg readers.
  • Converging: Databricks is moving Delta Lake 5.0 onto Iceberg's metadata design.
  • Common ground: Every vendor now reads Iceberg, so format choice locks customers in less.

The catalog now decides who can write and govern the tables.

Shifts Across All Five Vendors

The same moves show up at every vendor in 2026.

MCP and A2A Connect Agents Across Platforms

  • MCP: The Model Context Protocol lets an agent call an outside tool through one interface.
  • A2A: The Agent2Agent protocol lets one agent call another, even on another platform.
  • MCP support: All five vendors support MCP.
  • A2A support: AWS, Google Cloud, IBM, and Microsoft support A2A.

One MCP server plugs into all five agent platforms.

Every Vendor Is Building a Context Graph

  • Problem: Agents give wrong answers when they pick the wrong table or meaning.
  • AWS: AWS Context maps existing data into a knowledge graph for agents.
  • Databricks: Genie Ontology is a "living context graph" behind Genie One and Genie Code.
  • Google Cloud: Knowledge Catalog builds a "dynamic context graph" from all data.
  • Microsoft: Fabric IQ holds an ontology of entities, relationships, and rules.
  • IBM: Context in watsonx.data is a federated context layer.

Every Cloud Now Ships a Graph Engine

  • Google Cloud: Spanner Graph is the operational graph. BigQuery Graph is the analytical graph. Both are GA.
  • Microsoft: Graph in Fabric is the analytical graph over OneLake. It reached GA on June 2, 2026.
  • AWS: Neptune Database is the operational graph. Neptune Analytics is the analytical graph.
  • No native engine: Databricks and IBM ship no graph engine in their lakehouse.
  • GQL: Google and Microsoft query with ISO GQL. AWS queries with Gremlin, openCypher, and SPARQL.

The graph model is now standard. Graph deals now open with "why a separate graph database?"

Context Graphs Share Four Traits

  • Written and learned: Teams write some definitions by hand. The layer learns others by watching queries.
  • Permissions: The layer answers with the caller's own permissions.
  • MCP access: Agents reach the layer through MCP tools.
  • Open standard: Apache Ossie specifies semantic layers and ontologies.

Each context graph lives inside its vendor's platform. A customer on two clouds builds two.

Agents Get a Control Plane

  • Identity: Each agent gets its own identity.
  • Registry: A registry lists every agent, MCP server, and tool.
  • Gateway: A gateway routes all agent traffic.
  • Policy: A policy check runs on every tool call.
  • Cross-cloud reach: Microsoft and IBM control planes now discover agents on other clouds.

The control plane holder sees every agent and every tool call.

Memory and Evaluation Become Managed Services

  • Agent memory: Memory stores what an agent learned across sessions.
  • Memory status: AWS and Google Cloud memory are GA. The other three are in Beta or preview.
  • Evaluation: Evaluation scores agent answers against test cases before and after release.
  • Evaluation status: AWS, Google Cloud, and Microsoft evaluation is GA.

The memory layer is still open at three of five vendors.

Operational Data Moves Next to the Lakehouse

  • Why: The business runs on transactions, and the lakehouse needs them within seconds, not overnight.
  • Postgres: Databricks, AWS, Google Cloud, and Microsoft each offer managed Postgres.
  • Vectors: Every vendor's operational database now adds vector search.
  • Streams: Streams now feed the context layer as well as the lakehouse.

Vector search alone no longer sets a product apart.

App Builders and Custom Chips Round Out the Stack

  • App builders: Business users describe an app, and the platform builds it on company data.
  • App builder status: Apps in Amazon Quick is GA. The others are in Beta or preview.
  • Custom chips: AWS, Google Cloud, Microsoft, and IBM build their own chips.
  • NVIDIA too: The same four also rent NVIDIA GPUs next to their own chips.

Agents Now Run Inside the Services

  • First wave: Built-in agents handle incidents, failed pipelines, and cost spikes.
  • Turned on, not built: The customer turns the agent on instead of building one.
  • Examples: AWS DevOps Agent, Azure SRE Agent, Gemini Cloud Assist, Genie ZeroOps, and IBM Concert.
  • Needs a map: Each agent needs a map of how services, code, and deployments connect.

AWS DevOps Agent builds an application topology. Genie ZeroOps reads Unity Catalog lineage.

Where Neo4j Fits

A graph adds connected context to each layer.

A Customer-Owned Graph Works Across Every Vendor

  • One graph, every cloud: A customer on two clouds keeps one context graph instead of two.
  • Traceable answers: The graph returns multi-hop answers that trace back to their source.
  • A map for built-in agents: Built-in agents need a map of how services connect. A graph is that map.
  • Reached through MCP: Every agent platform reaches the graph through the Neo4j MCP Server.
  • Built from the lakehouse: The lakehouse keeps the tables. The graph adds the links.
  • Graph depth: Graph Data Science runs more than 65 algorithms. Cypher 25 supports most mandatory GQL features.

Takeaways

  • One stack, five vendors: The layers match. The names, formats, and catalogs differ.
  • Iceberg is common ground: Format locks customers in less. The catalog decides more.
  • Every cloud now ships a graph engine: Each one stays inside its own platform.
  • MCP is the shared interface: One MCP server reaches every agent platform.
  • Memory is still open: Agent memory is in Beta or preview at three of five vendors.
  • Neo4j spans the clouds: A customer-owned graph serves agents on any vendor.

Overview talk. The arc runs from the stack all five vendors now sell, through where the vendors differ, into the shifts that show up across all of them in 2026, and ends with where Neo4j fits. Product names and status were checked against vendor pages on September 27, 2026. Graph engine facts were checked on September 29, 2026. Names changed often in 2025 and 2026, so check again before reuse. Source material: cloud-integration/hyperscaler.md.

Databricks and IBM are not hyperscalers in the strict sense, but they follow the same stack shape, so the comparison includes them.

Six layers in two groups. The three data layers store and run the company's data. The three AI layers turn that data into answers and actions. Agent governance runs across every layer. Vendors now also build agents into their own services, such as an agent that finds the cause of an outage. Each built-in agent belongs to the layer of the service it runs in, so the stack keeps its six layers.

An operational database stores the current state that an app changes many times per second, such as an order. A stream carries events as they happen. A graph engine stores entities and the relationships between them. It follows a chain of relationships in one query. Spanner Graph and Amazon Neptune Database are operational graphs. BigQuery Graph, Graph in Microsoft Fabric, and Amazon Neptune Analytics are analytical graphs.

Agents read the context layer before they query, so they pick the right table and the right meaning. Vendors call it a context graph, an ontology, or a semantic layer. Vendors call the governance pieces an agent control plane.

Ask them in order. The next grid answers the data, agents, and people questions for each vendor. The meaning question is the context layer, which gets its own slides in the next section. Who governs the agents is the control plane, also in the next section. For the people question, the usual answers are Office, Slack, a browser, or an IDE.

The rows follow the four questions: data first, then agents, then people, then where it all runs. - Table format: AWS, Google Cloud, and IBM default to Apache Iceberg. Microsoft defaults to Delta Lake and translates the metadata for outside Iceberg readers. Databricks manages both as equals. - Catalog: Unity Catalog is the only one that governs data, models, agents, and MCP tools in one place. Google's Knowledge Catalog can federate to AWS Glue, Unity Catalog, and Snowflake, in Preview. - Agent platform: Gemini Enterprise Agent Platform replaced Vertex AI on April 22, 2026. Microsoft Foundry puts agents, models, and tools in one Azure resource with one set of access controls. - Models: Microsoft Foundry lists more than 10,000 models. Google Model Garden lists more than 200. - Chat: Microsoft 365 Copilot lives inside the Office apps most companies already use. Gemini Enterprise also connects to Microsoft 365. watsonx Orchestrate targets IT leaders more than end users. - Coding: Kiro, Antigravity, IBM Bob, and GitHub Copilot help developers build any software. Genie Code works only inside Databricks. It reads Unity Catalog tables, columns, and lineage before it writes code. - Where it runs: AWS, Google Cloud, and Microsoft run on their own cloud. Databricks is multi-cloud. IBM is the only vendor here with an on-premises lakehouse edition.

The last point is our inference from the vendor pages. Delta Lake 5.0 was presented at Data + AI Summit 2026. It adopts the Iceberg v4 metadata tree, so Delta and Iceberg clients will read and write one on-disk format. It has not shipped yet: the Databricks Iceberg docs still list Iceberg v1 to v3. OneLake translation limits: it writes Iceberg V2 only, handles Parquet only, and skips tables with 5,000 or more commits. Bridge to the next section: the catalog is also growing up. It no longer just lists tables. Every vendor is turning it into a context layer that stores what the data means.

One view of where each shift stands. The next slides start with MCP, the protocol most of these shifts depend on. Then they follow the grid's rows from top to bottom. "Announced" means the vendor announced the product, but the source does not state a release status. Google rebuilt Gemini Cloud Assist around proactive agents at Next '26 on April 22, 2026. IBM announced IBM Concert on May 5, 2026. The graph engine row counts native engines only. Databricks offers the GraphFrames library on Spark. IBM sells DataStax Graph in DataStax Enterprise.

AgentCore Gateway turns APIs and Lambda functions into MCP tools. Databricks offers managed MCP servers and a GA Genie One MCP server. Google runs managed MCP servers for its own services. watsonx Orchestrate and Foundry agents both call MCP servers. Google supports A2A 1.0, and IBM supports 0.3.0. Databricks documents A2A in a community blog post.

An example of a wrong meaning is the term "revenue." AWS, Databricks, Google Cloud, and Microsoft all describe their context layer as a graph. Only Microsoft documents the graph engine behind it: Graph in Fabric stores the Fabric IQ instance graph. - AWS Context is coming soon. Pieces available now: Glue Data Catalog business context is in preview, and S3 annotations are GA. - Genie Ontology is in Public Preview and on by default since August 6, 2026. - Knowledge Catalog is GA. Its Context API is in preview. - The Fabric IQ workload is GA. The ontology item is in preview. Work IQ and Foundry IQ round out Microsoft's context layer. - Context in watsonx.data entered private preview on May 5, 2026. - Snowflake Horizon Context competes for the same layer.

This is the most important slide for a Neo4j audience. A context graph records what data means. A graph engine stores the entities and relationships themselves. It runs traversals and graph algorithms over them. - Spanner Graph has been GA since December 2, 2024. BigQuery Graph reached GA on August 31, 2026. Google says the two share one graph schema and query language. - Graph in Fabric stores the instance graph of the Fabric IQ ontology. Microsoft says the ontology "declares what connects and why." - AWS Context and the Amazon Quick knowledge graph name no graph store. The open-source Context Ontology Accelerator uses Neptune. - GQL is ISO/IEC 39075:2024. Neptune supports neither GQL nor SQL/PGQ. Neo4j Cypher 25 supports most mandatory GQL features. Each engine lives inside one vendor's platform and serves that vendor's agents. That validates the graph model. It also means the difference is now a graph the customer owns, queries with Cypher, and uses across every cloud. The "why a separate graph database" point is our inference from the vendor pages.

AWS Context learns join paths from agent queries. Genie Ontology ranks snippets partly by how often they are used. AWS Context inherits IAM and Lake Formation permissions. Genie Ontology gates every snippet by Unity Catalog permissions. MCP entry points: AWS Context MCP tools, the Genie One MCP server, and the Foundry IQ MCP server. Apache Ossie was called Open Semantic Interchange until it entered the Apache Incubator. Snowflake founded it, and AWS and Databricks joined. The platform-bound point is our inference from the vendor pages.

- AWS: AgentCore Identity, AgentCore Gateway, Policy in AgentCore, and AWS Agent Registry are all GA. - Databricks: Unity Gateway reached GA on August 4, 2026. - Google Cloud: Agent Identity, Agent Registry, and Agent Gateway are GA. - IBM: The watsonx Orchestrate Agentic Control Plane is in private preview. - Microsoft: Microsoft Agent 365 reached GA on May 1, 2026. Entra Agent ID gives each agent an Entra identity. Microsoft Agent 365 syncs registries with Amazon Bedrock and Google Cloud in public preview. watsonx Orchestrate discovers agents in Google's agent platform and Microsoft Foundry.

Memory examples: user preferences and past decisions. AgentCore Memory has been GA since October 2025. Databricks managed agent memory is in Beta and stores long-term memory in Lakebase. Google Memory Bank features reached GA in June and July 2026. Databricks evaluates agents with MLflow 3. IBM monitors agents with watsonx.governance.

Postgres products: Databricks Lakebase, Amazon Aurora DSQL and Aurora PostgreSQL, AlloyDB for PostgreSQL, and Azure HorizonDB. IBM offers Db2 with DiskANN vector indexing and DataStax Astra DB. IBM completed its purchase of Confluent on March 17, 2026. IBM pairs the Confluent Real-Time Context Engine with Context in watsonx.data. The vector point is our inference from the vendor pages.

Apps in Amazon Quick reached GA on September 1, 2026. Genie App Builder is in Beta. Google's Agent Studio app builder and Apps in Copilot Studio are in preview. IBM's closest product is IBM Bob, which targets developers. Chips: Trainium3, Inferentia2, and Graviton5 at AWS. Ironwood TPU and Axion at Google, with TPU 8t and 8i coming soon. Maia 200 and Cobalt 200 at Microsoft. Telum II and the Spyre Accelerator at IBM. Databricks runs on the hardware of its host clouds.

AWS DevOps Agent reached GA on March 31, 2026. It covers AWS, Azure, and on-premises systems, and reaches on-premises tools through MCP. Azure SRE Agent reached GA in March 2026 and can restart, scale, or roll back within policy guardrails. Genie ZeroOps is in private preview and tests a fix in a sandbox before a human applies it. Built-in agents will likely spread from operations to other routine tasks. That point is our inference from the vendor pages. Bridge to the Neo4j section: a map of how things connect is a graph.

The same stack from the start of the deck. Neo4j adds a role on four layers: the context layer, the agent layer, the apps layer, and the lakehouse. The vendor keeps the chips, the operational databases, and agent governance. AgentCore Gateway, Databricks Agent Bricks, and IBM watsonx Orchestrate all connect to MCP servers. The Neo4j Agent Memory library ships an integration for Amazon Bedrock AgentCore.

Each vendor builds a context graph inside its own platform. Neo4j adds one that the customer owns, queries with Cypher, and uses across every cloud. Graph deals on Google Cloud and Microsoft will open with "why a separate graph database?" Neo4j answers with depth and reach. Teams that learn GQL on Google Cloud or Microsoft can carry those skills to Neo4j. - Neo4j Graph Intelligence for Microsoft Fabric reached GA on October 23, 2025. It brings Graph Data Science algorithms into Fabric, so it adds to Graph in Fabric on Microsoft accounts. - Neo4j Graph Analytics runs inside Snowflake as a Native App. - Neo4j Aura is listed on AWS Marketplace and Google Cloud Marketplace. AuraDB Professional is listed on Microsoft Marketplace. The graph stores entities and the relationships between them, so an agent can follow a chain of links in one query. The built-in agents point is our inference. AWS DevOps Agent builds an application topology, and Genie ZeroOps reads lineage. Both are maps of connected things, which is what a graph stores.