← Insights
Verified Research · The AI Stack

Mapping the Enterprise AI Stack

Model choice is rarely the deciding factor. Integration, governance, and adoption decide most AI results; that is where to budget and sequence.

42 evidence clusters checked against primary and official sources · 2 industry stats caught as false and held out · contested figures flagged
Loading the stack…

The stack, layer by layer

Vendor snapshot August 2026 (targeted refresh from July 2026) · static rendition generated from the same data as the interactive map

07 Governance

trust & audit · lasting moat · key players: ServiceNow · Collibra · OneTrust · MS Purview

Permissions, audit trails, policy, lineage, model-release constraints, and observability: the constraint layer that makes AI safe and allowable to put on production data.

What happens here: Decides who and what is allowed to act, records what happened, and proves it to auditors and regulators. Production paths run through here; it is the chokepoint auditors test.

As frontier models expand: The durable top of the stack. Autonomous agents need guardrails more than people do: a person senses when they are out of bounds, an agent acts confidently past them, so more capable agents need more constraint, not less. The July 2026 refresh adds a sharper permission read: frontier model access, customer eligibility, energy/load obligations, and public-interest bargains can become deployment gates. Raw capability commoditizes; the version wrapped in permissions, audit, and predictable cost is what an enterprise can actually deploy, and what stays defensible.

06 Action & Orchestration

agents & workflows · emerging moat · key players: UiPath · ServiceNow · Agentforce · Copilot · n8n · Stripe

Where reasoning becomes execution: agents, workflows, RPA, and the integration glue that fires real actions in real systems.

What happens here: Turns a model's output into a committed transaction: opening a ticket, moving an order through quote-to-cash, routing an approval. The layer where value lands as intelligence gets cheap.

As frontier models expand: The hottest contested layer in the stack. Frontier labs push up from reasoning into agents; ServiceNow pushes down from governance; Salesforce pushes up from CRM. Stripe now reaches from payments and billing toward agentic commerce, while its agreed OpenRouter acquisition joins model routing, cost, and usage control to that economic layer. The question for any buyer: how much action, routing, and transaction evidence am I about to consolidate under one vendor?

05 Reasoning

inference & generation · going commodity · key players: Anthropic · OpenAI · Google · Meta · Mistral · OpenRouter

The intelligence itself: inference, generation, and reasoning over whatever context the layers below supply.

What happens here: Takes a well-specified question plus context and returns an answer or a plan. Capability per dollar is collapsing here faster than anywhere else in the stack.

As frontier models expand: The fastest-commoditizing layer in the stack, so the labs do not stay put; they expand up into action and down into tooling and semantics. A second control point is now explicit: gateways route across many models using price, speed, reliability, and task fit. Stripe's agreed OpenRouter acquisition makes that gateway part of billing and profitability infrastructure, but it does not prove durable neutrality, portability, or pricing power.

04 Semantic Layer

meaning & context · your moat · key players: Palantir · dbt · Databricks Unity Catalog

Maps raw data to business meaning: translates a cryptic field like status_cd = 04 into "active customer in good standing." This is how data actually moves through your specific organization.

What happens here: Holds the company-specific definitions, joins, and logic that make data interpretable. Standard industry dictionaries (FIBO, ACORD) commoditize; the logic unique to your business (the exceptions, the legacy quirks, the integration that exists only because two systems disagree on what "closed" means) does not. That specificity is the moat, and it cannot be bought off the shelf.

As frontier models expand: Frontier models try to skip this layer by deriving meaning on the fly (text-to-SQL), but that produces answers that look right and are not, where a real semantic layer returns nothing rather than guess. The company-specific wiring stays defensible; the standard definitions do not.

03 Compute

compute & runtime · commodity + capital control · key players: AWS · Azure · GCP · NVIDIA · CoreWeave · Meta Compute (reported)

Raw processing power: where transformation jobs and model inference actually run.

What happens here: Provides the metered capacity that everything above consumes. A commoditized utility in itself, but the hyperscalers use it as a foothold to sell up the stack.

As frontier models expand: Hyperscalers bundle AI services on top of compute to climb into the reasoning layer and capture more than utility margins. In August 2026, NVIDIA signed nonbinding memorandums with six financial institutions for platforms designed to mobilize more than $500 billion of third-party capital over time. That target is not committed capital or NVIDIA revenue, and final agreements plus project-level underwriting still decide what gets built. Meta Compute remains reported and provisional.

02 Storage

storage & retrieval · pure commodity · key players: Snowflake · Databricks · Redshift · BigQuery

Where data lands and persists at scale: warehouses, lakes, and lakehouses.

What happens here: Holds the organization's data in queryable form. Commoditized as raw storage, so the players are racing upward to where margin lives.

As frontier models expand: Snowflake (Cortex) and Databricks (Unity Catalog, Mosaic) push up into the semantic and reasoning layers, pressuring anyone who assumed the meaning layer was theirs alone to own.

01 Data Origination

records & signals · proprietary data · key players: SAP · Oracle · Salesforce · Workday · SCADA / IoT

Systems of record where business data is first created: transactions, customer records, sensor and telemetry streams.

What happens here: Generates the raw material the whole stack runs on. Largely insulated from frontier models (they don't create your transactions), so the systems themselves stay durable.

As frontier models expand: Mostly durable. The interesting move is Salesforce reaching up from its CRM origination base into orchestration via Agentforce: origination players wanting to own the action layer above them.

Expansion plays

AWS

operates in: 02 Storage · 03 Compute · expanding into: 05 Reasoning · 06 Action & Orchestration

Turning cloud infrastructure into the operating system for enterprise AI. Runs storage (2) and compute (3) as the underlying utility. Pushing up into reasoning (5) via Bedrock and the Nova model family, and into action (6) via AgentCore and Managed Agents, with IAM, Guardrails, and CloudTrail bundled in, so the governance travels with the agent.

What it means for buyers: If your data and compute already live on AWS, Bedrock and AgentCore are the frictionless next step; agent spend even rolls into your existing cloud commitment. Convenient, and a quiet way to single-source three layers at once.

Microsoft

operates in: 03 Compute · 05 Reasoning · expanding into: 01 Data Origination · 06 Action & Orchestration · 07 Governance

Bundling AI across the entire stack, from cloud to inbox to governance. Runs compute (3, Azure) and reasoning (5, via Azure AI Foundry and Copilot, multi-model). Reaching into origination (1, Dynamics and M365), action (6, Copilot Studio and Agent Service), and governance (7, Purview and Agent 365). Few layers Microsoft is not touching.

What it means for buyers: The broadest bundle in enterprise software. For most organizations Microsoft is already everywhere, so the AI expansion rides existing licenses, which is exactly what makes it hard to evaluate, or resist, layer by layer.

Google

operates in: 02 Storage · 03 Compute · 05 Reasoning · expanding into: 04 Semantic Layer · 06 Action & Orchestration

The clearest case of one company running storage, compute, and reasoning at once, and reaching into the rest. Runs compute (3, GCP), storage (2, BigQuery), and reasoning (5, Gemini 3.1). Pushing into the semantic layer (4) via BigQuery's data-to-meaning features and into action (6) via the Gemini Enterprise Agent Platform (formerly Vertex AI) and Project Mariner. Kurian's pitch: own the full stack from chip to inbox.

What it means for buyers: When one vendor runs three layers and is reaching into two more, "diversified stack" gets hard to claim. The convenience is real; so is the concentration. The buyer's job is to know which layers they have effectively single-sourced.

OpenAI

operates in: 05 Reasoning · expanding into: 06 Action & Orchestration · 01 Data Origination · 07 Governance

Extending from the model into agents, end-user apps, and the plane that governs them. Runs reasoning (5) with GPT-5 and the o-series. Pushing into action (6) via AgentKit, Workspace Agents, and Codex; toward the user (1) via ChatGPT Enterprise and apps in ChatGPT; and into governance (7) via Frontier, its platform for managing AI coworkers with shared context, permissions, and evaluation. Building a stack, not just an endpoint.

What it means for buyers: Same escape-the-commodity logic as the other labs, now reaching all the way to the governance plane: Frontier puts OpenAI in the same contest as ServiceNow's Control Tower and Microsoft's Agent 365. Worth watching whether your model provider becomes your application vendor and your agent-governance layer at once.

Stripe + OpenRouter

operates in: 01 Data Origination · expanding into: 05 Reasoning · 06 Action & Orchestration

Joining economic infrastructure with model routing and agentic transactions. Stripe operates transaction and billing infrastructure at the data-origination boundary. Its agreed OpenRouter acquisition reaches toward reasoning through routing across 400+ models and more than 80 providers, while agentic commerce products reach into action. Neither party disclosed a transaction price, and closing plus product integration still matter.

What it means for buyers: The gateway can become an economic control point: it sees which model ran, what it cost, how it performed, and what gets billed. Preserve exportable usage records, routing policy, evaluation history, and model/provider portability so one convenience layer does not quietly become both your AI control plane and your commercial ledger.

Anthropic

operates in: 05 Reasoning · expanding into: 06 Action & Orchestration · 04 Semantic Layer

Escaping the commoditizing model layer by moving into agents and the tooling around context. Runs reasoning (5) with the Claude family. Pushing into action (6) via Claude Managed Agents, Claude Code, and Cowork, and toward the semantic layer (4) via the Model Context Protocol (MCP), now an industry-wide standard for wiring models to enterprise context. Capturing more than a metered model call.

What it means for buyers: The model vendor does not intend to stay a model vendor. Do not architect on the assumption that your reasoning provider stays in its lane; today's API is tomorrow's agent and context layer. MCP becoming a shared standard is the tell: own the wiring, not just the model.

Salesforce

operates in: 01 Data Origination · expanding into: 06 Action & Orchestration · 04 Semantic Layer

Reaching up from the customer record into the actions taken on it. Runs customer-data origination (1) with Sales and Service Cloud. Pushing into action (6) via Agentforce and toward the semantic layer (4) via Data 360: wanting the workflow and the meaning, not just the record. The move is partly defensive: data platforms pull the records downward while productivity bundles encroach on the entry point.

What it means for buyers: An origination player climbing into action to stay durable. For a Salesforce-heavy shop, Agentforce is the easy yes, but it deepens dependence at a second layer and bets your workflow on a vendor being squeezed from two directions.

ServiceNow

operates in: 07 Governance · expanding into: 06 Action & Orchestration · 05 Reasoning

Using a governance and workflow foothold as the wedge to expand down into agent orchestration and embedded reasoning. Runs governance (7): the AI Control Tower governs every agent, model, and action in the enterprise. Pushing into action (6) via the AI Agent Orchestrator, Now Assist, and Otto, and toward reasoning (5) via the purpose-built Now LLM. Nvidia's Huang called it "the operating system of enterprise AI agents."

What it means for buyers: A governance incumbent becoming an agent-orchestration platform. The question is not whether the agents work; it is how much of your action layer runs through one vendor's control plane, and the exit cost in three years.

Snowflake

operates in: 02 Storage · expanding into: 04 Semantic Layer · 05 Reasoning

Climbing out of commoditized storage into the meaning and reasoning layers above. Runs storage (2). Pushing into the semantic layer (4) via Cortex Analyst and Semantic Views, and into reasoning (5) via Cortex AI: turning a data warehouse into a place where meaning and inference happen. The catch: these only see data that already lives inside Snowflake.

What it means for buyers: If your data already lives in Snowflake, letting it own the semantic layer too is the path of least resistance: convenient, but a consolidation decision dressed up as a feature upgrade, and one that stops at the warehouse boundary.

Databricks

operates in: 02 Storage · expanding into: 04 Semantic Layer · 05 Reasoning

Turning the lakehouse into the governed semantic and AI layer, not just storage. Runs storage (2). Pushing into the semantic layer (4) via Unity Catalog business semantics and Metric Views, and into reasoning (5) via Mosaic AI and Agent Bricks, pressuring from the storage layer anyone who assumed the meaning layer was theirs alone.

What it means for buyers: Same consolidation logic as Snowflake, from the lakehouse side. The pull is to let your storage vendor own everything from raw data to model output: efficient, until you want to move.

Meta

operates in: 01 Data Origination · 05 Reasoning · expanding into: 03 Compute · 06 Action & Orchestration

Using consumer distribution, open-weight reasoning, and potentially excess AI infrastructure as adjacent routes out of pure model commoditization. Runs massive consumer/data-origination surfaces (1) and reasoning/model products (5). Reported July 2026 Meta Compute plans would push into compute (3) by renting excess AI capacity or hosted model access; consumer AI experiments push toward action/distribution (6).

What it means for buyers: Do not treat Meta as a confirmed enterprise cloud yet. Treat it as a watch-list vendor whose consumer reach, open-weight model strategy, and possible compute resale could reprice neocloud assumptions and force buyers to ask exactly which layer a vendor is selling.

NVIDIA

operates in: 03 Compute · expanding into: no adjacent push

Turning compute control into a financed infrastructure platform. NVIDIA operates at compute. Its August 2026 memorandums with six financial institutions are designed to create independently underwritten financing platforms and mobilize more than $500 billion of third-party capital over time. The memorandums are not final funding commitments, and NVIDIA revenue or project creditworthiness cannot be inferred from the headline target.

What it means for buyers: Compute procurement now includes a capital-structure question. Ask who provides the outside cash flow, what support is contingent, which asset lives back the financing, and who bears loss if utilization, delivery, refinancing, or residual value misses the model.

As AI turns intelligence into a commodity, the advantage shifts to what can't be bought off the shelf: the execution and governance layers that are hard to run well, the semantic wiring specific to your business, and the permission/compute-control layer that determines what can actually run in production. Everything else, you and your competitor buy from the same vendors.

Claim ledger

The verification record behind the map: the August 2026 pass checked 42 evidence clusters across the retained ledger plus the July and August trigger evidence. The eighteen load-bearing clusters are shown, the two false claims included; contested figures stay flagged on the map itself.

42/42
Checked
evidence clusters traced to primary or official sources, August 2026
2
False
industry stats caught as misattributed and held in the avoid bucket
11
Corrected
figures, scopes, and attributions tightened against primaries
8
Demoted
preprint, reported-only, or single-source claims held out of stronger findings
ConfirmedCemri et al. (NeurIPS 2025): why multi-agent LLM systems fail; the coordination-failure taxonomy behind the orchestration layer's read.arxiv.org
ConfirmedLiu et al. (TACL 2024): Lost in the Middle; long-context degradation that keeps the context layer load-bearing.arxiv.org
DemotedSingle-agent parity under equal token budgets: preprint, not peer-reviewed; held as a contested signal, never asserted.arxiv.org
ConfirmedNVIDIA FY2025 results (SEC 8-K): $115.2B data-center revenue; the silicon layer's margin evidence.sec.gov
ConfirmedNVIDIA Q1 FY2027 (May 2026): $75.2B data-center revenue on $81.6B total; durable value still concentrating at compute.nvidia.com
ConfirmedNVIDIA compute-financing memorandums (Aug 2026): more than $500B with six financial institutions; confirmed with transaction-structure limits: nonbinding until final agreements.nvidia.com
ConfirmedStripe agrees to acquire OpenRouter (Aug 19, 2026): a 400+ model gateway moving into a payments company; the routing-and-transaction control point, limits stated until close.stripe.com
DemotedMeta Compute: reported exploration of compute sales, not company-confirmed; carried as provisional on the map.bloomberg.com
CorrectedOpenAI figures: ~$852B valuation confirmed at the March 2026 round; the run-rate de-specified to the official $20B+ (late 2025); the climb above $40B by Aug 2026 attributed as reported.bloomberg.com
ConfirmedMCP adoption and donation: OpenAI and Google adoption in 2025, donated to the Linux Foundation's agentic AI foundation that December; the interoperability layer's standardization evidence.anthropic.com
CorrectedMCP security claims: MCPoison and CurXecute are Cursor-IDE flaws, not general MCP vulnerabilities; the security-lag read rests on scan findings, not confirmed exploits.checkpoint.com
DemotedMIT NANDA "95% of pilots fail": preliminary, non-peer-reviewed, contested denominator; flagged on the map, never asserted.fortune.com
ConfirmedGartner (June 2025): over 40% of agentic-AI projects predicted canceled by end-2027; an analyst prediction, attributed as one.gartner.com
ConfirmedS&P Global Voice of the Enterprise (n=1,006): companies scrapping most AI initiatives rose from 17% to 42% year over year.spglobal.com
ConfirmedShadow AI: employees at more than 90% of surveyed firms use personal AI for work while about 40% of firms hold official subscriptions.fortune.com
False"McKinsey: 98% expect reskilling within 5 years": misattributed; McKinsey's actual finding is ~75% expecting role shifts. Held in the avoid bucket.mckinsey.com
False"Narrow scope ships on time 65% vs 16%": misattributed; the figures are task-accuracy results from a clinical study, not delivery rates. Held in the avoid bucket.nature.com
DemotedAI-equity bubble diagnostics: preprint stochastic-volatility analysis; held as a valuation caution, not stack evidence.arxiv.org