--- title: AI Market Reference Architecture document: ai-market-reference-architecture canonicalUrl: https://brianletort.ai/industry/architecture schemaVersion: 2026.05.02 lastUpdated: '2026-05-14' counts: categories: 7 vendors: 50 metrics: 21 modelCards: 10 positions: 36 sources: 32 categories: - ai_infrastructure - model_portfolio - data_management - ai_ready_data - developer_tools - agent_tools - commercial_tools --- # AI Market Reference Architecture *A living layer-cake taxonomy of AI infrastructure, models, data, developer tools, agent tools, commercial tools, metrics, and model cards. Last updated 2026-05-14.* **Counts.** 7 categories, 50 seeded vendors, 21 metrics, 36 position histories, 10 model cards, 32 sources. ## Thesis The LLM Evolutionary Tree explains where models came from. This reference architecture explains how AI market power moves through the layer cake: infrastructure, model portfolios, raw data management, AI-ready data, developer tools, agent tools, and commercial tools. The strategic question is whether SaaS keeps owning the data, or enterprises own their data and permission agents into it. ## Categories ### Layer 1 — AI Infrastructure (`ai_infrastructure`) Physical and cloud substrate for AI: GPUs, accelerators, networking, datacenters, power, storage, and managed compute capacity. **Boundary.** Place a vendor here when its primary role is supplying scarce compute, hardware, cloud capacity, or the physical systems that make AI workloads possible. **Buyer question.** Where does the compute, power, and hardware capacity come from? **Responsibility boundary.** Provider owns capacity, availability, hardware lifecycle, and physical resilience. Customer owns workload placement, demand forecasting, and utilization risk. **Seeded vendors.** - **Amazon Web Services** (leader, high confidence) — Full-stack AI cloud with capacity, managed model access, and enterprise procurement gravity. - **Broadcom** (leader, medium confidence) — Custom silicon and networking control point for hyperscale AI systems. - **Google** (leader, high confidence) — Frontier model, cloud platform, and consumer distribution leader with unusually strong first-party data and infrastructure. - **Microsoft** (leader, high confidence) — Enterprise distribution leader through Microsoft 365, GitHub, Azure, and identity surfaces; the strategic question is how much AI spend bundles into existing contracts. - **NVIDIA** (leader, high confidence) — The merchant accelerator and systems control point for most frontier AI buildouts; watch revenue mix, networking attach, and supply availability. - **Alibaba** (challenger, high confidence) — China-scale cloud and open-model contender; important for regional sovereignty and open-weight benchmark pressure. - **AMD** (challenger, medium confidence) — Challenger accelerator supplier; relevance rises when open software, cloud availability, and price-performance create credible substitution. - **Cloudflare** (challenger, medium confidence) — Edge-network runtime contender for distributed inference, routing, and security controls. - **Groq** (challenger, medium confidence) — Specialized inference hardware/runtime contender with latency as the primary differentiation. - **Together AI** (challenger, medium confidence) — Open-model serving and fine-tuning specialist; useful signal for open-weight production demand. **Metrics.** - **Data-center capex growth:** 53% YoY (Q1 2025, high confidence). AI infrastructure remains capital constrained; capex direction is the first-order signal for who can supply capacity. - **High-end accelerated servers:** > one-third of total data-center capex (2025 forecast, medium confidence). Accelerator-heavy systems are no longer a niche line item; they are a large share of the data-center investment stack. - **Data-center capex forecast:** 30% growth (2025 forecast, high confidence). Sustained infrastructure expansion keeps the AI substrate layer strategic, not commodity. ### Layer 2 — Model Portfolio (`model_portfolio`) Foundation, frontier, open-weight, specialist, and routed model portfolios that provide the intelligence primitives consumed by higher layers. **Boundary.** Place a vendor here when model capability, model choice, model routing, or model governance is the primary product decision. **Buyer question.** Which models should we use, for which workloads, at what cost and risk? **Responsibility boundary.** Model provider owns training, model behavior, release cadence, and safety defaults. Customer owns model selection, workload routing, evals, and fallback policy. **Seeded vendors.** - **Amazon Web Services** (leader, high confidence) — Full-stack AI cloud with capacity, managed model access, and enterprise procurement gravity. - **Anthropic** (leader, high confidence) — Frontier model provider moving upward through Claude Code, computer use, enterprise deployments, and consulting-style implementation services. - **Databricks** (leader, high confidence) — Data-platform control point for enterprises that want AI close to governed data and ML operations. - **Google** (leader, high confidence) — Frontier model, cloud platform, and consumer distribution leader with unusually strong first-party data and infrastructure. - **Hugging Face** (leader, high confidence) — Open-model registry and developer distribution layer; the clearest public signal for open ecosystem pull. - **Meta** (leader, high confidence) — Open-weight ecosystem anchor with massive consumer distribution; track whether open model gravity converts into developer and enterprise control. - **Microsoft** (leader, high confidence) — Enterprise distribution leader through Microsoft 365, GitHub, Azure, and identity surfaces; the strategic question is how much AI spend bundles into existing contracts. - **OpenAI** (leader, high confidence) — Frontier model portfolio leader now pushing upward into developer tools, agents, and commercial work through Codex, ChatGPT, and consulting-style enterprise services. - **Alibaba** (challenger, high confidence) — China-scale cloud and open-model contender; important for regional sovereignty and open-weight benchmark pressure. - **DeepSeek AI** (challenger, high confidence) — Efficiency and open-weight pressure point that forces the frontier to justify price, latency, and training-capital intensity. - **HeyGen** (challenger, medium confidence) — Video generation specialist; relevant as multimodal models pressure point-solution durability. - **Mistral AI** (challenger, high confidence) — European frontier and open-weight contender; relevance rises with sovereignty, portability, and enterprise deployment concerns. - **Together AI** (challenger, medium confidence) — Open-model serving and fine-tuning specialist; useful signal for open-weight production demand. - **Vercel** (challenger, medium confidence) — Frontend and AI runtime contender where application deployment, model routing, and developer workflow meet. - **Writer** (challenger, medium confidence) — Enterprise writing and knowledge-work platform; a test case for whether specialist commercial AI survives model-provider expansion. - **xAI** (challenger, medium confidence) — Frontier challenger with distribution through X and fast model iteration. - **Adept** (specialist, low confidence) — Computer-use and action-model precedent; retained because it shaped the category even as market structure keeps moving. - **Cohere** (specialist, medium confidence) — Enterprise and retrieval-oriented model specialist. - **ElevenLabs** (specialist, medium confidence) — Voice AI specialist; useful marker for modality-specific application markets outside general chat. **Metrics.** - **SWE-bench performance jump:** 71.7% (2024 vs 2023, high confidence). Software-engineering capability moved from toy benchmark to enterprise relevance, changing the model layer's practical value. - **Open-weight vs closed-weight performance gap:** 1.70% (2024 vs prior year, high confidence). Narrowing gaps increase buyer leverage and make openness a strategic dimension, not only a developer preference. - **Live model price, latency, speed, and context comparisons:** available (live, high confidence). Model leadership should be benchmark-adjusted by cost and latency, not read from capability leaderboards alone. ### Layer 3 — Data Management (`data_management`) Raw systems of record, warehouses, lakehouses, operational databases, and document stores that hold enterprise data before it is shaped for AI use. **Boundary.** Place a vendor here when the product primarily stores, governs, or manages source data that still needs processing, wrapping, indexing, or context compilation before an AI system can consume it directly. **Buyer question.** Who owns the system of record, and how do we keep AI from being trapped inside SaaS silos? **Responsibility boundary.** Data platform owns storage, lineage, access control, and governance surfaces. Customer owns data quality, semantic modeling, and agent-access policy. **Seeded vendors.** - **Databricks** (leader, high confidence) — Data-platform control point for enterprises that want AI close to governed data and ML operations. - **Google** (leader, high confidence) — Frontier model, cloud platform, and consumer distribution leader with unusually strong first-party data and infrastructure. - **Palantir** (leader, medium confidence) — Enterprise operating-layer contender where AI is embedded into governed workflows and decision processes. - **Salesforce** (leader, high confidence) — Incumbent workflow distributor; the strategic question is whether AI agents reinforce or unbundle CRM process ownership. - **Snowflake** (leader, high confidence) — Governed data-cloud control point for AI workloads that start with enterprise data access and policy. - **Wiz** (leader, medium confidence) — Cloud security control point that can extend into AI posture management as models and agents become part of the attack surface. - **Alibaba** (challenger, high confidence) — China-scale cloud and open-model contender; important for regional sovereignty and open-weight benchmark pressure. - **Elastic** (challenger, medium confidence) — Search incumbent with hybrid retrieval and observability roots; relevant when AI-ready data starts from search, logs, and operational documents. - **Glean** (challenger, medium confidence) — Enterprise search and assistant layer that sits near the boundary between AI-ready data and commercial workflow. - **MongoDB** (challenger, medium confidence) — Operational data incumbent moving upward by adding vector search and AI application data patterns. - **Credo AI** (specialist, medium confidence) — AI governance specialist; useful pure-play signal for policy, audit, and risk-management demand. **Metrics.** - **Organizations using AI:** 78% (2024 vs 2023, high confidence). More organizational AI use increases pressure to expose governed systems of record to agents without copying everything into SaaS silos. - **Generative AI in at least one business function:** 71% (2024 vs 2023, high confidence). Business-function adoption makes source-data access, lineage, and permissions first-class AI architecture concerns. - **Enterprise GenAI spend:** $37B (2025 estimate, medium confidence). Enterprise AI spend eventually lands against data platforms, not only model APIs and point tools. ### Layer 4 — AI-Ready Data (`ai_ready_data`) Processed, indexed, embedded, retrievable, policy-wrapped, and API-exposed data that can be consumed directly by models and agents. **Boundary.** Place a vendor here when the product turns raw enterprise data into retrieval, context, embeddings, vector search, API access, or model-ready memory. **Buyer question.** How do we make our data usable by AI without surrendering it to every SaaS vendor? **Responsibility boundary.** AI-ready data provider owns retrieval quality, indexing, wrappers, and access surfaces. Customer owns source-data truth, freshness, permissions, and which agents may consume which context. **Seeded vendors.** - **Amazon Web Services** (leader, high confidence) — Full-stack AI cloud with capacity, managed model access, and enterprise procurement gravity. - **Databricks** (leader, high confidence) — Data-platform control point for enterprises that want AI close to governed data and ML operations. - **Hugging Face** (leader, high confidence) — Open-model registry and developer distribution layer; the clearest public signal for open ecosystem pull. - **Pinecone** (leader, high confidence) — Managed vector database leader; important because it turns raw enterprise content into retrieval surfaces agents can consume directly. - **Snowflake** (leader, high confidence) — Governed data-cloud control point for AI workloads that start with enterprise data access and policy. - **Cloudflare** (challenger, medium confidence) — Edge-network runtime contender for distributed inference, routing, and security controls. - **Elastic** (challenger, medium confidence) — Search incumbent with hybrid retrieval and observability roots; relevant when AI-ready data starts from search, logs, and operational documents. - **Glean** (challenger, medium confidence) — Enterprise search and assistant layer that sits near the boundary between AI-ready data and commercial workflow. - **Groq** (challenger, medium confidence) — Specialized inference hardware/runtime contender with latency as the primary differentiation. - **MongoDB** (challenger, medium confidence) — Operational data incumbent moving upward by adding vector search and AI application data patterns. - **Perplexity** (challenger, medium confidence) — AI answer/search application that competes on user habit, citation UX, and model-routing quality. - **Together AI** (challenger, medium confidence) — Open-model serving and fine-tuning specialist; useful signal for open-weight production demand. - **Vercel** (challenger, medium confidence) — Frontend and AI runtime contender where application deployment, model routing, and developer workflow meet. - **Weaviate** (challenger, high confidence) — Open-source vector and hybrid-search contender for teams that want more control over the AI-ready data layer. - **Zilliz / Milvus** (challenger, medium confidence) — Milvus ecosystem signal for open vector infrastructure and high-scale retrieval workloads. - **Cohere** (specialist, medium confidence) — Enterprise and retrieval-oriented model specialist. - **LangChain** (specialist, medium confidence) — Developer orchestration and observability specialist for multi-model, tool-using applications. **Metrics.** - **Vector and hybrid retrieval availability:** available across Pinecone, Weaviate, Milvus, MongoDB, Elastic (current, high confidence). Retrieval infrastructure is now broad enough to treat AI-ready data as a separate layer between raw datastores and agent execution. - **Model and context gateway pattern:** emerging (2026, medium confidence). API gateways, model routers, and context wrappers become the control plane for which agents can consume which data. - **Open-model serving demand proxy:** downloads and model-card activity (live, medium confidence). Open ecosystem activity is a maintainable proxy for AI-ready data and serving demand where vendors do not disclose token volume. ### Layer 5 — Developer Tools (`developer_tools`) Build tools for humans and agents creating software: IDE copilots, coding agents, CLI agents, context management, orchestration, evals, and tool harnesses. **Boundary.** Place a vendor here when the primary workflow is building, modifying, testing, or operating software with AI assistance. **Buyer question.** Which tools turn developers into agent supervisors and software editors? **Responsibility boundary.** Tool provider owns coding UX, context packaging, model/tool invocation, and review affordances. Customer owns repository permissions, acceptance criteria, and production release governance. **Seeded vendors.** - **Amazon Web Services** (leader, high confidence) — Full-stack AI cloud with capacity, managed model access, and enterprise procurement gravity. - **Anthropic** (leader, high confidence) — Frontier model provider moving upward through Claude Code, computer use, enterprise deployments, and consulting-style implementation services. - **GitHub** (leader, high confidence) — Developer workflow distribution leader; coding is one of the clearest paid AI application categories. - **Hugging Face** (leader, high confidence) — Open-model registry and developer distribution layer; the clearest public signal for open ecosystem pull. - **Microsoft** (leader, high confidence) — Enterprise distribution leader through Microsoft 365, GitHub, Azure, and identity surfaces; the strategic question is how much AI spend bundles into existing contracts. - **NVIDIA** (leader, high confidence) — The merchant accelerator and systems control point for most frontier AI buildouts; watch revenue mix, networking attach, and supply availability. - **OpenAI** (leader, high confidence) — Frontier model portfolio leader now pushing upward into developer tools, agents, and commercial work through Codex, ChatGPT, and consulting-style enterprise services. - **Cloudflare** (challenger, medium confidence) — Edge-network runtime contender for distributed inference, routing, and security controls. - **Cursor** (challenger, medium confidence) — AI-native coding application where agentic development behavior is visible earlier than in many enterprise categories. - **Vercel** (challenger, medium confidence) — Frontend and AI runtime contender where application deployment, model routing, and developer workflow meet. - **Hermes** (specialist, medium confidence) — Personal agent stack signal: scheduled operations, memory, tool use, and exception escalation over owned infrastructure. - **LangChain** (specialist, medium confidence) — Developer orchestration and observability specialist for multi-model, tool-using applications. - **OpenCode** (specialist, low confidence) — Open developer-agent signal; tracked because open coding agents may become the portability layer between models, repos, and local tools. - **Zeroclaw** (specialist, medium confidence) — Lightweight agent runtime signal for owned-agent execution rather than SaaS-owned workflow automation. - **OpenClaw** (watch, low confidence) — Emerging open agent-tooling signal; tracked as a future-stack example, not a settled public market leader. **Metrics.** - **Coding tools as leading GenAI application category:** top enterprise spend category (2025 estimate, medium confidence). Developer tools are the first place where AI changes production work, not only information retrieval. - **Coding agents and CLI tools:** Cursor / Codex / Claude Code / OpenCode (2026, medium confidence). Build tools are moving from autocomplete to agent supervision, with context management and tool harnesses becoming the real differentiators. - **Context management as build substrate:** emerging (2026, medium confidence). The durable build-tool layer may be less about the IDE and more about how context, tools, repo state, and review loops are packaged for agents. ### Layer 6 — Agent Tools (`agent_tools`) Agent runtimes and operational tools that run scheduled, triggered, multi-agent, tool-using, and exception-escalating work. **Boundary.** Place a product here when the defining behavior is autonomous or semi-autonomous execution: agents talking to agents, invoking tools, maintaining memory, and paging humans on exception. **Buyer question.** Which agent systems can we safely let act on our behalf? **Responsibility boundary.** Agent tool owns execution loop, tool-use guardrails, memory, logging, and escalation mechanics. Customer owns delegated authority, permissions, rollback, and exception policy. **Seeded vendors.** - **Anthropic** (leader, high confidence) — Frontier model provider moving upward through Claude Code, computer use, enterprise deployments, and consulting-style implementation services. - **CrowdStrike** (leader, medium confidence) — Security incumbent where AI changes both defense workflows and governance requirements. - **GitHub** (leader, high confidence) — Developer workflow distribution leader; coding is one of the clearest paid AI application categories. - **OpenAI** (leader, high confidence) — Frontier model portfolio leader now pushing upward into developer tools, agents, and commercial work through Codex, ChatGPT, and consulting-style enterprise services. - **Palantir** (leader, medium confidence) — Enterprise operating-layer contender where AI is embedded into governed workflows and decision processes. - **Salesforce** (leader, high confidence) — Incumbent workflow distributor; the strategic question is whether AI agents reinforce or unbundle CRM process ownership. - **ServiceNow** (leader, high confidence) — Enterprise workflow incumbent positioned for governed AI task execution inside service-management processes. - **Cursor** (challenger, medium confidence) — AI-native coding application where agentic development behavior is visible earlier than in many enterprise categories. - **Harvey** (challenger, medium confidence) — Legal AI specialist with domain workflow depth; a clean example of the commercial-tool layer above models and agent infrastructure. - **Sierra** (challenger, medium confidence) — Customer-facing agentic workflow contender; watch for production deployments and accountability controls. - **Adept** (specialist, low confidence) — Computer-use and action-model precedent; retained because it shaped the category even as market structure keeps moving. - **Agent Zero** (specialist, medium confidence) — General-purpose agent runtime signal for tool use, memory, and autonomous task loops. - **Cognosys** (specialist, low confidence) — Early agentic workflow specialist; tracked as a category signal, not a settled leader. - **Hermes** (specialist, medium confidence) — Personal agent stack signal: scheduled operations, memory, tool use, and exception escalation over owned infrastructure. - **LangChain** (specialist, medium confidence) — Developer orchestration and observability specialist for multi-model, tool-using applications. - **OpenCode** (specialist, low confidence) — Open developer-agent signal; tracked because open coding agents may become the portability layer between models, repos, and local tools. - **Zeroclaw** (specialist, medium confidence) — Lightweight agent runtime signal for owned-agent execution rather than SaaS-owned workflow automation. - **OpenClaw** (watch, low confidence) — Emerging open agent-tooling signal; tracked as a future-stack example, not a settled public market leader. **Metrics.** - **Agent startups tracked:** 170+ (March 2025, medium confidence). Agentic workflows are a distinct market structure, not just another app feature. - **Agent startup funding:** $3.8B (2024, medium confidence). Funding growth signals that investors view delegated work as a separate value pool. - **Public production-readiness disclosure:** uneven (2026, medium confidence). Human-in-loop controls, tool permissions, and audit logs should carry as much weight as demos. ### Layer 7 — Commercial Tools (`commercial_tools`) Finished point solutions and vertical AI tools: video, voice, legal, sales, research, support, writing, and other specialist business outcomes. **Boundary.** Place a vendor here when buyers purchase a packaged business outcome, not a reusable model, data layer, developer tool, or agent runtime. **Buyer question.** Which specialist tools survive when agents can work directly over our own data? **Responsibility boundary.** Tool provider owns workflow UX, packaged domain behavior, and product controls. Customer owns data portability, agent access, workflow redesign, and whether point tools become durable or get absorbed by agent systems. **Seeded vendors.** - **Adobe** (leader, high confidence) — Creative-workflow incumbent with strong distribution and provenance needs. - **Anthropic** (leader, high confidence) — Frontier model provider moving upward through Claude Code, computer use, enterprise deployments, and consulting-style implementation services. - **CrowdStrike** (leader, medium confidence) — Security incumbent where AI changes both defense workflows and governance requirements. - **Google** (leader, high confidence) — Frontier model, cloud platform, and consumer distribution leader with unusually strong first-party data and infrastructure. - **Meta** (leader, high confidence) — Open-weight ecosystem anchor with massive consumer distribution; track whether open model gravity converts into developer and enterprise control. - **Microsoft** (leader, high confidence) — Enterprise distribution leader through Microsoft 365, GitHub, Azure, and identity surfaces; the strategic question is how much AI spend bundles into existing contracts. - **OpenAI** (leader, high confidence) — Frontier model portfolio leader now pushing upward into developer tools, agents, and commercial work through Codex, ChatGPT, and consulting-style enterprise services. - **Palantir** (leader, medium confidence) — Enterprise operating-layer contender where AI is embedded into governed workflows and decision processes. - **Salesforce** (leader, high confidence) — Incumbent workflow distributor; the strategic question is whether AI agents reinforce or unbundle CRM process ownership. - **ServiceNow** (leader, high confidence) — Enterprise workflow incumbent positioned for governed AI task execution inside service-management processes. - **Wiz** (leader, medium confidence) — Cloud security control point that can extend into AI posture management as models and agents become part of the attack surface. - **Glean** (challenger, medium confidence) — Enterprise search and assistant layer that sits near the boundary between AI-ready data and commercial workflow. - **Harvey** (challenger, medium confidence) — Legal AI specialist with domain workflow depth; a clean example of the commercial-tool layer above models and agent infrastructure. - **HeyGen** (challenger, medium confidence) — Video generation specialist; relevant as multimodal models pressure point-solution durability. - **Perplexity** (challenger, medium confidence) — AI answer/search application that competes on user habit, citation UX, and model-routing quality. - **Sierra** (challenger, medium confidence) — Customer-facing agentic workflow contender; watch for production deployments and accountability controls. - **Writer** (challenger, medium confidence) — Enterprise writing and knowledge-work platform; a test case for whether specialist commercial AI survives model-provider expansion. - **xAI** (challenger, medium confidence) — Frontier challenger with distribution through X and fast model iteration. - **Cognosys** (specialist, low confidence) — Early agentic workflow specialist; tracked as a category signal, not a settled leader. - **Credo AI** (specialist, medium confidence) — AI governance specialist; useful pure-play signal for policy, audit, and risk-management demand. - **ElevenLabs** (specialist, medium confidence) — Voice AI specialist; useful marker for modality-specific application markets outside general chat. **Metrics.** - **GenAI application spend:** $19B+ (2025 estimate, medium confidence). Commercial tools are already a major value-capture zone, but the durable winners need workflow depth that survives model-provider expansion. - **Microsoft 365 Copilot paid enterprise seats:** 20M+ (April 2026, high confidence). Seat adoption is the cleanest public signal for commercial AI distribution through incumbent software suites. - **Specialist tool pressure:** model providers moving upward (2026, medium confidence). OpenAI and Anthropic moving into implementation services raises the bar for point solutions: they need proprietary workflow depth, data access, or domain-specific trust. ## Top model cards - `gpt-5-5` — Frontier closed reasoning default. Best fit: High-stakes reasoning, agentic coding, multimodal work, and workloads where frontier quality outweighs portability. - `claude-opus-4-7` — Frontier enterprise reasoning and coding model. Best fit: Long-form reasoning, coding, tool-use workflows, and enterprise contexts that weight safety and trust posture heavily. - `gemini-2-5-pro` — Frontier multimodal reasoning model. Best fit: Multimodal and long-context workloads, especially where Google Cloud or Google product distribution is already strategic. - `deepseek-r1` — Open-weight reasoning pressure point. Best fit: Reasoning workloads where cost, portability, and self-hosting leverage matter more than first-party product polish. - `llama-4-maverick` — Open ecosystem frontier family. Best fit: Enterprise and developer contexts that value ecosystem breadth, inspectability, and deployment flexibility. - `qwen-3-thinking-2507` — Open-weight reasoning and regional sovereignty signal. Best fit: Reasoning workloads where open weights, regional availability, and China-scale ecosystem signals matter. - `mistral-large-3` — European frontier and sovereignty model. Best fit: European enterprise and sovereignty-sensitive workloads that need strong performance without a single US frontier dependency. - `command-a` — Enterprise retrieval and workflow specialist. Best fit: Retrieval-heavy enterprise workflows, knowledge applications, and workloads that value business-context fit over raw frontier rank. - `grok-4` — Frontier challenger with social distribution. Best fit: Buyers exploring frontier alternatives and consumer/social-context distribution signals. - `gpt-oss-120b` — Open-weight strategic hedge from a closed-model leader. Best fit: Portability-sensitive teams that want open-weight leverage without leaving the OpenAI model family narrative entirely. ## Monthly position history ### Layer 1 — AI Infrastructure - **NVIDIA** — leader (96/100, up). Merchant accelerator leadership remains the infrastructure benchmark; score rises with continued capex pull-through and networking attach. - **Microsoft** — leader (89/100, up). Infrastructure position strengthens when Azure capacity and Copilot distribution reinforce each other. - **Amazon Web Services** — leader (88/100, flat). Cloud capacity and procurement breadth keep AWS in the leader band. - **AMD** — challenger (70/100, up). Challenger score improves when substitution pressure and cloud availability become more credible. ### Layer 2 — Model Portfolio - **OpenAI** — leader (94/100, flat). Frontier capability plus product habit keeps OpenAI in the leader band; open-weight pressure shows up as efficiency pressure, not displacement. - **Anthropic** — leader (91/100, up). Enterprise trust and agentic coding strength keep Anthropic gaining within the model layer. - **Google** — leader (89/100, up). Google benefits from simultaneous model, cloud, and consumer distribution signals. - **Hugging Face** — leader (82/100, up). Open-model distribution and model-card gravity keep Hugging Face in the platform leader band. ### Layer 3 — Data Management - **Databricks** — leader (88/100, up). Governed data gravity keeps Databricks central as enterprise AI shifts from experiments to operated systems. - **Snowflake** — leader (84/100, up). Snowflake remains a leader when AI starts from governed enterprise data. ### Layer 4 — AI-Ready Data - **Pinecone** — leader (84/100, up). Pinecone anchors the vector-store portion of the AI-ready data layer. - **Weaviate** — challenger (74/100, up). Weaviate gains when teams want open, hybrid, and self-controlled retrieval infrastructure. - **Vercel** — challenger (71/100, up). Runtime position improves as model routing becomes part of application deployment rather than a separate platform decision. - **Cloudflare** — challenger (69/100, up). Edge distribution and security posture keep Cloudflare relevant as inference gets closer to users and policies. ### Layer 5 — Developer Tools - **Anthropic** — leader (85/100, up). Claude Code makes Anthropic one of the most important movers in build tooling, not only model APIs. - **GitHub** — leader (84/100, up). GitHub owns a high-frequency execution surface where agentic behavior can become daily workflow rather than demo. - **OpenAI** — leader (82/100, up). Codex and implementation services move OpenAI upward from model portfolio into the developer-tools layer. - **Cursor** — challenger (77/100, up). Cursor gains as coding remains the cleanest category for visible AI productivity and agentic workflow adoption. ### Layer 6 — Agent Tools - **Anthropic** — leader (87/100, up). Claude's coding and tool-use posture keeps Anthropic central to early production agentic workflows. - **ServiceNow** — leader (80/100, up). ServiceNow is positioned where delegated action meets governed enterprise process. - **Hermes** — specialist (70/100, up). Hermes represents owned-agent operations over personal infrastructure and owned data, with humans paged only on exception. - **Sierra** — challenger (69/100, up). Sierra is a useful challenger signal for customer-facing production agents. ### Layer 7 — Commercial Tools - **Microsoft** — leader (92/100, up). Paid Copilot seats and enterprise distribution make Microsoft the application-layer benchmark. - **Google** — leader (88/100, up). Gemini consumer adoption keeps Google in the leader band even when enterprise monetization is harder to isolate. - **Salesforce** — leader (81/100, up). Salesforce stays strong where AI rides existing CRM workflow ownership. - **Wiz** — leader (80/100, up). Cloud security posture extends naturally into AI posture as models and agents become part of the attack surface. --- Compact JSON: . Raw taxonomy YAML: . Canonical HTML: .