A 12-layer analysis of the AI ecosystem — silicon to vertical applications — with competitive landscapes, market sizing, funding metrics, and strategic build opportunities across funding tiers. Refreshed against publicly available information as of August 17, 2026.
The AI stack remains a $2+ trillion ecosystem in 2026, spanning 12 layers from silicon to verticals. Value concentration sits at silicon (NVIDIA) and the model layer (Anthropic, OpenAI); the application layer is ascending.
Anthropic overtook OpenAI in annualized revenue (~$43B$47B May · ~$69B est. Jul[9] vs. ~$25B) with 40% enterprise API share (Menlo Ventures, Dec 2025) vs. 25%27%[4] for OpenAI. Targeting an H2 2026 IPO at a potential $900B+ valuation. Closed a $65B Series H at $965B post-money (May 28); confidentially filed for IPO June 1, targeting an October 2026 Nasdaq listing.[7,8]
SpaceX acquired xAI in an all-stock deal on Feb 2, 2026: $250B standalone within a $1.25T combined entity. xAI subsequently dissolved as a standalone company on May 6, 2026, rebranded as SpaceXAI[1]. SpaceX itself went public June 12, 2026 in the largest IPO in history ($75B raised, $2T+ valuation, NASDAQ: SPCX) and moved to acquire Cursor-maker Anysphere for $60B four days later[11,12]. A fundamental reshaping of the model layer landscape.
Inference workloads now exceed 55% of AI cloud infrastructure spend. Cerebras IPO'd May 14, 2026 at ~$56-66B FDV~$95B Day-1 market cap[5] (cooled to ~$60B by July 13; ~$64B Aug 9)[15][25]; OpenAI's $20B+ equity stake validates inference-specific silicon at scale. Fireworks' $1.5B Series D at $17.5B on >$1B annualized revenue re-rates the serving layer above it.[22]
Feb 2026 SaaSpocalypse erased ~$285B in software market cap following Anthropic's Cowork launch. Thomson Reuters dropped 15.83% in a single session (largest on record). Per-seat SaaS → outcome/fractional-FTE pricing transition is underway. By early July 2026 the public software basket was back to green for the year at the index level, though the recovery is uneven[19].
Six charts carrying the quantitative spine of this report. Each has a data table beneath it, and each names its basis. Where a bar is dashed, it is a projection rather than an actual.
| Stage | Share | Source |
|---|---|---|
| Use AI in ≥1 function | 88% | Stanford AI Index 2026 |
| Use generative AI in ≥1 function | 70% | Stanford AI Index 2026 |
| At least experimenting with agents | 62% | McKinsey State of AI |
| Scaling an agentic system | 23% | McKinsey State of AI |
| Fully scaled AI in any function | <10% | Stanford AI Index 2026 |
| EBIT impact >5% | ~6% | McKinsey State of AI |
| Period | Spend | Basis |
|---|---|---|
| Late 2024 | $3.5B | actual |
| Mid 2025 | $8.4B | actual |
| End 2026 | $15B | projected |
| Layer | Market size |
|---|---|
| 12 Verticals | $2.5T |
| 11 App Platforms | $315B |
| 04 Cloud | $1.1T |
| 02 Compute capex | $725B |
| 01 Silicon | $500B |
| 08 Inference | $106B |
| 06 Models (funding) | $80B |
| 05 Data Infra | $50B |
| 03 Networking | $20B |
| 10 Middleware | $8–10B |
| 09 Orchestration | $7.6B |
| 07 MLOps | $5–8B |
| Year | Consumption | Basis |
|---|---|---|
| 2024 | 415 TWh | actual |
| 2025 | 485 TWh | actual |
| 2026 | 1,000 TWh | superseded projection |
| 2030 | 945 TWh | IEA base case |
| Segment | Year | Size |
|---|---|---|
| AI applied to security | 2026 | $30–35B (bar uses midpoint) |
| AI applied to security | 2032 | ~$60B (extrapolated) |
| Securing AI itself | 2026 | $1.65B |
| Securing AI itself | 2032 | $13.5B |
Specialized chips—GPUs, TPUs, custom ASICs, and HBM—designed to accelerate the matrix math and parallel computation underpinning ML training and inference. The foundation of all AI capability.
AI chip revenue in 2026 (Deloitte). AMD CEO Lisa Su projects a $1T TAM by 2030. Broader semiconductor market surpassed $830B in 2025 (Omdia). NVIDIA FY26 data center revenue alone reached $193.7B; Q1 FY27 (May 2026): $81.6B total revenue (+85% YoY), $75.2B data center, networking $15B/qtr[14].
NVIDIA maintains an extraordinary moat through its CUDA software ecosystem (20+ years of developer investment). Custom silicon is eroding this at the hyperscaler level. The DeepSeek efficiency shock caused a $600B single-day drop in NVIDIA's market cap. Market is supply-constrained, not demand-constrained.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | NVIDIA | H200, B200, GB300, Vera Rubin | Full-stack: CUDA, NVLink, GPU-aware networking, software moat | FY26 total $215.9B (+65%); DC $193.7B | |
| 2 | AMD | MI300X, MI400, MI450 Instinct | Price-performance; ROCm open-source stack | Data center growing ~60% YoY | $90B OpenAI deal for 6GW MI450 (Oct 2025) |
| 3 | Google (TPU) | TPU v6 Trillium | GCP integration; Transformer-optimized | Internal + cloud resale | Anthropic 3.5GW + $40B Google investment (Apr 2026) |
| 4 | Broadcom | Custom AI ASICs (XPUs) | Bespoke silicon for Google, Meta, others | Networking + ASIC revenue surging | Major Anthropic TPU demand |
| 5 | Amazon (AWS) | Trainium 2/3, Inferentia | Cost-optimized training/inference; Bedrock integration | Internal + cloud resale | Trainium3 3× faster than T2 |
| 6 | Intel | Gaudi 3, Falcon Shores | x86 ecosystem leverage | New AI chips Oct 2025 | Restructuring underway |
| 7 | Cerebras | Wafer Scale Engine 3 | Entire wafer as single chip; ultra-low latency | $510M 2025 rev (+76%); $238M net income | IPO May 14, 2026: $5.55B raised @ |
| 8 | Groq | LPU / Groq 3 LPX | Deterministic latency; fastest tokens-per-sec | $20B | Groq Inc. nominally independent; not acquired |
| 9 | SK Hynix / Samsung / Micron | HBM3e, HBM4 | High Bandwidth Memory for all AI accelerators | HBM $100B+ by 2030 | $16B (2024) → $32.6B (2026) |
| 10 | Marvell | Custom AI accelerators | Custom silicon for hyperscalers | Strong data center growth | AI-exposed revenue accelerating |
Physical infrastructure — servers, racks, cooling, power delivery — that houses AI silicon and enables large-scale training and inference. Converts electricity and hardware into usable compute capacity.
Combined 2026 hyperscaler capex (Q1 2026 earningsQ2 2026 earnings, Jul 22–30[20]): AMZN $200B+~$220B, GOOGL $180–190B$195–205B, META $125–145B$130–145B[20], MSFT ~$120B (or ~$190B incl. off-BS leases)~$175B calendar 2026 incl. finance leases — a lease reclassification, not a spending cut[20]. Big-four total now ~$720–745B, up ~78% from ~$410B in 2025; Alphabet's Q2 free cash flow turned negative at −$5.9B, and Amazon attributed its $20B raise to higher memory costs. Global data center energy consumption projected to reach 1,000 TWh in 2026 (IEA).
The neocloud segment has exploded. CoreWeave carries $21B~$29B+[17] in long-term debt against $5B$12–13B guided 2026 (Q1: $2.08B, +112% YoY)[17] revenue — model works only while GPU demand outpaces cost of capital. Power availability is the primary bottleneck; hyperscalers pursuing nuclear, geothermal, and direct energy partnerships. UPDATE: Anthropic-SpaceX struck a deal for full capacity of Colossus One (~300 MW, 220k+ GPUs) on May 6, 2026; ~$15B annual run-rate, adding SpaceX as Anthropic's fourth major compute counterparty[2].
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | CoreWeave | GPU cloud (NVIDIA-focused) | AI-first neocloud; purpose-built for GPU workloads | $5.13B (2025); $12–13B guide (2026) | Backlog ~$88–95B post Meta+$21B & Anthropic deals; NVIDIA $2B strategic investment Jan 2026 |
| 2 | Equinix | Colocation & interconnection | 260+ data centers; carrier-neutral | $8B+ revenue | AI-driven demand accelerating |
| 3 | Digital Realty | Data center REIT | Large-scale campus deployments | $5B+ revenue | Multi-hundred MW AI campuses |
| 4 | Lambda Labs | GPU cloud | H100 at $2.99/GPU-hr; developer-friendly | ~$500M run rate; $1.5B Series E | 24MW AI factory in Kansas City |
| 5 | Crusoe | AI compute + energy | Clean energy-powered data centers | VC-backed | Sustainability resonating with enterprises |
| 6 | Nebius | GPU cloud (NVIDIA partner) | European neocloud; Yandex spinout | Growing in EU | Q2 2025 +625% YoY |
| 7 | RunPod | Serverless GPU cloud | Developer-first; low-cost inference | $120M ARR; 90% YoY growth | Scale on just $20M seed |
| 8 | Vultr | Cloud compute | Global edge; cost-competitive | Acquired by DigitalOcean (2025) | Expanding GPU capacity |
High-speed networking fabrics — InfiniBand, Ethernet, NVLink, optical transceivers — that connect GPUs within and across servers. Network bandwidth and latency directly constrain AI training speed and inference throughput.
AI networking in 2025 (650 Group). NVIDIA networking Q4 FY26 reached $11.0B (+263% YoY). Spectrum-X Ethernet surpassed a $10B annualized run rate. 1.6 Tbps switches enter volume deployment in 2026.
InfiniBand held ~80% of AI back-end networks in 2023; Ethernet surpassed it in 2025. UEC 1.0 (June 2025) provides InfiniBand-like RDMA on open-standard Ethernet. NVIDIA benefits either way. Data movement costs now exceed compute costs at scale.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | NVIDIA (Mellanox) | InfiniBand XDR, Spectrum-X, NVLink | Full networking stack; sub-2µs latency; GPU-aware routing | $11.0B/quarter networking (Q4 FY26) | +263% YoY; Spectrum-X at $10B+ run rate |
| 2 | Broadcom | Tomahawk/Jericho switch ASICs | Bandwidth doubling every 2 years; 51.2 Tbps switches | Dominant Ethernet switch ASIC share | Leading CPO deployment |
| 3 | Arista Networks | Data center Ethernet switches | Leading Ethernet DC switch vendor | Leading position in DC switching | AI back-end networking outperformance |
| 4 | Cisco | AI networking portfolio | Enterprise distribution; hybrid cloud | Large installed base | Expanding AI product line |
| 5 | Marvell | 800G/1.6T optical DSPs | Leading optical interconnect; custom PHY | Strong 800G lead | AI networking silicon growing rapidly |
| 6 | Celestica | White-box switches/servers | ODM for hyperscaler AI networking | Top 3 AI networking vendor | Hyperscaler custom designs |
| 7 | DriveNets | AI Fabric (Ethernet) | Fabric-scheduled Ethernet for AI; SDN | VC-backed growth stage | Open alternative to NVIDIA lock-in |
| 8 | Cornelis Networks | Omni-Path (CN5000) | Cost-competitive InfiniBand alternative at 400 Gbps | Niche but growing | Price-sensitive HPC/AI deployments |
Managed compute, storage, and AI-specific services that abstract infrastructure complexity. The primary distribution and monetization chokepoint for enterprise AI consumption.
Overall cloud market in 2026 (IaaS + PaaS + SaaS). Q4 2025 cloud infrastructure spending reached $119B (+29% YoY). GenAI-specific cloud services grew 140–180% in Q2 2025.
Big Three control 63% of the market. Google Cloud fastest-growing at 50%, driven by Gemini and Anthropic partnership. Multi-cloud adoption reached 89% of enterprises (Flexera 2026). Shift from training to inference favors elastic, serverless compute models.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | AWS | Bedrock, SageMaker, Trainium, Inferentia | Broadest catalog (200+); largest GPU fleet | ~$120B annualized; 30% market share | Q1 2026 grew 28%; |
| 2 | Microsoft Azure | Azure AI Foundry, Copilot Studio | Deepest enterprise ecosystem (M365, GitHub) | ~$100B+; 22% market share | 39% Q4 growth; Claude Opus 4.6 in Azure |
| 3 | Google Cloud | Vertex AI, Gemini API, TPU access | Best-in-class AI/ML platform; TPU infrastructure | ~$50B+; 14% market share | Backlog $460B+ Q1 2026; 50% Q4 growth |
| 4 | Oracle Cloud | OCI AI, GPU Supercluster | Aggressive pricing; database ecosystem | Growing from smaller base | xAI Colossus on OCI; Stargate partner |
| 5 | CoreWeave | AI-native GPU cloud | Purpose-built for AI; best GPU performance/$ | $5.13B (2025); $12–13B (2026) | IPO'd March 2025; ~$88–95B backlog |
| 6 | IBM | Watsonx | Enterprise AI governance; regulated industry focus | Acquired DataStax | Targeting compliance-heavy enterprises |
| 7 | Alibaba Cloud | PAI, Tongyi Qianwen | Dominant in China; competitive pricing | Largest cloud in APAC | Growing AI services in APAC |
Systems for storing, processing, indexing, and serving data to AI models — vector databases, data lakehouses, feature stores, labeling platforms, and synthetic data generators. Data quality is the primary determinant of AI application quality.
Broader AI data infrastructure market. Vector database market alone reaches $3.2B+ in 2026 (growing to $4.3B by 2028). Databricks at $134B valuation with $4.8B ARR (+55%); Snowflake FY26 revenue ~$4.2B.
Vector database market undergoing rapid commoditization. Traditional DBs (pgvector, MongoDB, Elasticsearch) adding native vector capabilities. 61% of enterprises delay AI initiatives due to lack of trusted data. Moat shifting from vector storage to unified data platforms.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Databricks | Lakehouse, Mosaic ML, Lakebase | Unified analytics + AI training on one platform | $134B valuation; $4.8B ARR (+55%); $4B Series L Dec 2025 | S-1 ready H2 2026; acquired Tecton, Antimatter, SiftD |
| 2 | Snowflake | Cortex AI, Arctic embeddings | Data cloud with native AI/ML; enterprise base | FY26 revenue ~$4.2B | AI features driving consumption growth |
| 3 | Pinecone | Serverless vector database | Fully managed; zero-ops vector search | $138M total raised at $750M valuation | Reportedly exploring sale; Notion churned to pgvector |
| 4 | MongoDB | Atlas Vector Search + Voyage AI | Vectors alongside operational data | $1.9B+ revenue | $220M Voyage AI acquisition |
| 5 | Weaviate | Open-source vector DB | Hybrid search (BM25 + vector); modular | $67M+ raised | HIPAA certified |
| 6 | Scale AI | Data labeling + evaluation | Human-in-the-loop; government contracts | $29B valuation after Meta $14.3B for 49% stake (Jun 2025) | Wang departed for Meta; Droege CEO |
| 7 | Qdrant | Open-source vector DB | Rust-native; performance-optimized | $38M+ raised | Fastest-growing open-source vector DB |
| 8 | Zilliz / Milvus | Enterprise vector DB | Billion-scale vector search; Kubernetes-native | $113M+ raised | Most scalable open-source option |
| 9 | Chroma | Lightweight vector DB | Developer-friendly; embeddable; simplest API | $20M+ raised | Default for prototyping |
| 10 | Fivetran / Airbyte | ELT for AI pipelines | Data connectors to unify enterprise data for AI | Fivetran: $5.6B valuation | Essential plumbing for RAG and fine-tuning |
Research labs and companies that train large foundation models — LLMs, multimodal models, diffusion models — which serve as the base intelligence layer for all downstream applications. Highest concentration of capital and talent in the AI stack.
Foundation model funding in 2025 — 40% of all global AI investment. Anthropic crossed ~$43B (Apr)$47B (May) · Yipit est. ~$69B (Jul 10)[9] ARR; OpenAI at ~$25B ($2B/month)(growth plateaued Feb–Apr)[16]. Top two captured 14% of all global VC across all sectors in 2025.
Anthropic overtook OpenAI in revenue at 40% enterprise API share (Menlo Ventures, Dec 2025) vs. 25%27%[4] OpenAI. Claude Code holds 54% market share. SpaceX acquired xAI (Feb 2, 2026) in an all-stock deal. Open-source models (Llama, Mistral, DeepSeek) compressing pricing power. Capital intensity extreme.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Anthropic | Enterprise-first; safety; coding dominance | 40% enterprise share; | ||
| 2 | OpenAI | GPT-5, o3, ChatGPT, Codex | Consumer brand; broadest distribution | ~$25B ARR | Stargate $500B JV; AWS $138B; AMD 6GW MI450; confidential S-1 Jun 8, leaning 2027 listing[16] |
| 3 | Google DeepMind | Gemini 2.5/3, Gemini Flash | Full-stack Google integration; TPU advantage | Internal (Google subsidiary) | Powers Workspace, Android, Search |
| 4 | Meta AI | Llama 3/4 (open-weight) | Open-weight models; community ecosystem | Internal (Meta subsidiary) | $35.2B CoreWeave deal (Sept 25 + Apr 26); |
| 5 | Musk-backed; X data access | $230B valuation (Jan 2026 Series E, $20B raised) | Acquired by SpaceX Feb 2, 2026 (all-stock): $250B within $1.25T combined entity. Dissolved as standalone May 6, 2026; now a division of SpaceX. Parent SpaceX public Jun 12, 2026 (SPCX, $2T+)[11] | ||
| 6 | Mistral AI | Mistral Large 3, Pixtral | European champion; open-source roots; cost-efficient | €11.7B valuation (Sept 2025 Series C, €1.7B led by ASML for 11% stake) | $830M debt March 2026 for Paris DC; $400M+ ARR |
| 7 | DeepSeek | DeepSeek V3, R1 | Efficiency breakthroughs; open-weight | Government-backed | Shook industry compute assumptions |
| 8 | Cohere | Command R+, North, Tiny Aya | Enterprise + sovereign AI focus | $7B valuation; $240M ARR (Q4 2025); ~$1.7B total raised | Targeting 2026 IPO; 70% gross margins; Saab partnership |
| 9 | AI21 Labs | Jamba models | MoE + Mamba hybrid architectures | Rumored NVIDIA acquisition target | Differentiated architecture |
| 10 | Stability AI / Black Forest Labs | Stable Diffusion, FLUX | Open image/video generation | Restructured; FLUX gaining share | Leading open-source generative media |
Tools for fine-tuning, distilling, quantizing, evaluating, and managing models post-training. Bridges the gap between a raw foundation model and a production-ready model tailored to specific enterprise needs.
MLOps market in 2026, growing at 25–30% CAGR. Hugging Face at $4.5B valuation (1M+ models hosted). Weights & Biases acquired by CoreWeave (closed 2025), now operates as CoreWeave's MLOps software stack. LLMOps segment is nascent but growing explosively.
MLOps bifurcating: traditional (tracking, registry) is mature; LLMOps (RAG evaluation, guardrails) is nascent and fragmented. Biggest gap: evaluation — enterprises cannot reliably measure AI accuracy or safety. Serving a frontier model can cost $100M+/day in compute.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Hugging Face | Model Hub, Transformers, Inference Endpoints | De facto standard for open model distribution; 1M+ models | $4.5B valuation; $395M raised | Community lock-in; largest model registry |
| 2 | Weights & Biases | Experiment tracking, model registry, W&B Models | Best-in-class ML experiment tracking | Acquired by CoreWeave (closed 2025) | CoreWeave's MLOps stack |
| 3 | Databricks (Mosaic ML) | Model training + fine-tuning on lakehouse | Unified data + model lifecycle; governance | Part of $134B Databricks | Training integrated into data platform |
| 4 | Anyscale / Ray | Distributed compute framework for ML | Open-source; scales fine-tuning across GPUs | $1B+ valuation | Used by OpenAI, Anthropic, and others |
| 5 | NVIDIA NeMo | Model customization framework | NVIDIA hardware optimized; guardrails built-in | Part of NVIDIA ecosystem | Enterprise focus on customization |
| 6 | Predibase / Ludwig | Fine-tuning platform | Low-code fine-tuning on open-source models | $20M+ raised | Developer-friendly abstractions |
| 7 | Arize / Phoenix | ML observability + LLM evaluation | Production monitoring for model drift and quality | $78M+ raised | LLM evaluation becoming critical |
| 8 | Braintrust | LLM evaluation platform | Prompt engineering + evaluation in one platform | $60M+ raised | Growing developer adoption |
Systems that deploy trained models to serve predictions and generate outputs at production scale. This is where AI models meet real users and generate revenue — and the fastest-growing segment of the entire stack.
AI inference market 2025 → 2030 at 19.2% CAGR. Inference surpassed training in data center revenue late 2025. By early 2026, inference crossed 55% of AI cloud infrastructure spending ($37.5B).
Every AI query costs $0.0016–$0.10 in GPU compute; at billion queries/day that's $1.6M–$100M/day. This is why Meta signed a $35B CoreWeave deal focused on inference. Inference-specific silicon thesis validated by Cerebras IPO at ~$56-66B FDV~$95B Day-1 market cap[5] (May 14, 2026), ~$60B by Jul 13, ~$64B Aug 9[15][25] with OpenAI's $20B+ equity commitment, and NVIDIA's $20B Groq asset deal. On the serving side, Fireworks raised $1.5B at $17.5B on >$1B annualized revenue (Jul 16, 2026), and Anthropic's Opus 5 cut frontier-class agentic pricing in half at $5/$25 per M tokens.[22]
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | NVIDIA (TensorRT) | TensorRT-LLM, Triton Inference Server | Deepest GPU optimization; lowest latency on NVIDIA hardware | Part of NVIDIA software revenue | Standard for GPU inference optimization |
| 2 | vLLM (UC Berkeley) | Open-source LLM serving engine | PagedAttention; highest throughput open-source serving | Open-source; community-maintained | Adopted by most major AI companies |
| 3 | CoreWeave | Managed inference cloud | GPU-optimized serving; distributed observability | $12B+ projected 2026 | 9 of 10 leading model providers on platform |
| 4 | AWS (Inferentia/Bedrock) | Managed inference via Bedrock | Cost-optimized custom silicon; serverless inference | Part of AWS AI revenue | Trainium/Inferentia reducing costs 40%+ |
| 5 | Together AI | Inference API + open-source serving | Fast, cheap inference for open-source models | $1B annualized rev (Mar 2026); in talks for $1B at $7.5B | Developer community; strong pricing |
| 6 | Groq | LPU / Groq 3 LPX inference | Deterministic latency; fastest tokens-per-sec | $20B | Nominally independent |
| 7 | Fireworks AI | Inference optimization platform | FireAttention; compound AI system serving | ||
| 8 | Cerebras | Wafer-scale inference | Ultra-low latency; single-chip model execution | IPO May 14, 2026 @ | $510M 2025 rev (+76%) |
| 9 | Baseten | Model serving platform | GPU-optimized serving; Truss framework | $100M+ raised | Strong developer adoption |
| 10 | Modal | Serverless inference | Pay-per-second GPU; developer-friendly | $75M+ raised | Popular for batch and on-demand workloads |
Frameworks and tools that help developers build AI applications — chaining model calls, managing context windows, integrating retrieval, handling agent workflows, and connecting to external tools.
AI coding tools revenue in 2026 (more than 2× the $5.1B 2024 figure). 85% of devs use AI tools, 73% regularly (JetBrains April 2026). Average developer uses 2–4 simultaneously. AI agent market $7.6B (2025) → $50.3B (2030) at 45.8% CAGR.
Three coding tool leaders: Claude Code (complex tasks), Cursor (IDE experience; $60B SpaceX acquisition pending[12], closing by end Aug 2026[27]), GitHub Copilot (enterprise deployment). 73% of developers use AI coding tools regularly. MCP (Model Context Protocol) crossed 97M installs by March 2026, becoming the standard for AI-tool connectivity.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | Anthropic | Claude Code | Terminal-native; 82% SWE-bench; 54% market share | $2.5B+ ARR from Claude Code | 46% most-loved (JetBrains Apr 2026) |
| 2 | Anysphere | Cursor | AI-native IDE; 72% autocomplete acceptance | $2B ARR; | 70% Fortune 1000; |
| 3 | GitHub (Microsoft) | Copilot | Broadest IDE support; 4.7M paid subscribers (+75% YoY) | Microsoft Copilot suite | Moving to token-based billing June 1, 2026 |
| 4 | LangChain | LangChain, LangGraph, LangSmith | Largest orchestration ecosystem; 500+ integrations | $130M+ raised | LangGraph → production standard for agents |
| 5 | LlamaIndex | LlamaIndex, LlamaCloud | Best-in-class RAG; 300+ data connectors | $50M+ raised | 44K GitHub stars; purpose-built retrieval |
| 6 | Microsoft | Agent Framework (AutoGen + Semantic Kernel) | Enterprise-grade; .NET/Python/Java; Azure integration | Part of Microsoft | GA Q1 2026; Azure-heavy enterprises |
| 7 | CrewAI | Multi-agent orchestration | Role-based multi-agent collaboration | $40M+ raised | Scaling enterprise customers |
| 8 | Vercel | v0 (AI app generator) | Frontend-focused; Next.js ecosystem | $250M+ raised; $3.5B valuation | Pioneered "vibe coding" |
| 9 | Replit | Replit Agent | Browser-based AI coding; full-stack app generation | $200M+ raised; $1.16B valuation | Agent 3 builds entire web apps |
| 10 | Windsurf (Codeium) | AI code editor | Free tier; broader model support | Acquired by OpenAI (~$3B) | Competition intensifies |
API gateways, model routers, authentication, rate limiting, caching, and cost management systems that sit between AI models and applications. The "plumbing" that makes AI reliable, observable, and cost-effective in production.
API management market broadly. AI-specific middleware nascent at $1–2B. AI observability market projected to reach $3B+ by 2028.
40% of companies spend $10M+/year on AI, but cost visibility tooling is primitive. AI gateways routing queries to cheapest model gaining adoption as enterprises use 2–4 models simultaneously. LLM observability being absorbed by APM vendors. Middleware layer has weak moats — with one exception now forming: the shared context layer, where governance, lineage, and accumulated company knowledge create switching costs that gateways and routers never had.
| # | Company | Product | Differentiator | Revenue / Funding | Signal |
|---|---|---|---|---|---|
| 1 | OpenAI | API platform (direct) | Largest model API by volume; developer ecosystem | Part of OpenAI revenue | Default API for many applications |
| 2 | Anthropic | Claude API (direct + reseller) | Highest enterprise API share (40%); strongest reasoning | Part of Anthropic revenue | 40% enterprise share (Menlo Dec 2025) |
| 3 | OpenRouter | Multi-model API gateway | Single API for 100+ models; auto routing and fallback | VC-backed | Default aggregator for model comparison |
| 4 | Portkey | AI gateway + observability | Unified API gateway with monitoring; guardrails | $20M+ raised | Growing in enterprise AI ops |
| 5 | Helicone | LLM observability | Open-source LLM monitoring; cost tracking | $15M+ raised | Developer-friendly; strong community |
| 6 | LiteLLM | Open-source proxy | Unified API across 100+ LLM providers; load balancing | Open-source | De facto standard for multi-model orchestration |
| 7 | Datadog | AI observability | LLM monitoring integrated into existing APM | Part of $2.2B Datadog revenue | Leveraging enterprise relationships |
| 8 | Mem0 / Zep | AI memory management | Persistent memory for AI agents across sessions | $10–20M raised each | Solving context management at scale |
| 9 | Coconut | Shared context layer for AI agents | Model-agnostic, governed, versioned company context; every fact carries its source, and gaps are flagged rather than filled with a guess | Private (Coconut AI Inc.) | Targets the context-governance gap named in Part 4. Competitive set is platform-native, not peer: Claude Projects/Cowork, ChatGPT memory, and M365 Copilot can each bundle this away, with Glean ($300M ARR) the comparison every enterprise buyer makes |
Horizontal platforms that enable businesses to build, deploy, and manage AI-powered applications — including no-code/low-code AI builders, enterprise AI platforms, and agent deployment platforms.
Global SaaS market in early 2026, growing at 20% CAGR toward $1.1T by 2032. AI-native SaaS spending increased 108% YoY. Gartner: 80% of enterprises deploy GenAI-enabled apps by end of 2026.
The Feb 3–5, 2026 "SaaSpocalypse" wiped ~$285B in market cap following Anthropic's Cowork launch. Thomson Reuters dropped 15.83% in a single session (largest on record); LegalZoom -19.68%; software ETFs ~-20% YTD by March. Publicis Sapient reducing SaaS licenses by ~50%. Shift from per-seat to outcome-based pricing. Vertical SaaS has stronger defensibility than horizontal.
| # | Company | Product | Differentiator | Revenue | Signal |
|---|---|---|---|---|---|
| 1 | Salesforce | Einstein AI, Agentforce | CRM-native AI; massive enterprise install base | $37B+ | AI agents across sales, service, marketing |
| 2 | ServiceNow | Now Assist, AI Agents | ITSM/workflow automation with embedded AI | $10B+ | Agent-to-agent communication protocols |
| 3 | Microsoft | Copilot (M365, Dynamics, Power Platform) | Deepest enterprise software integration | Part of $240B+ Microsoft | Copilot across entire product suite |
| 4 | Palantir | AIP (Artificial Intelligence Platform) | Ontology-based AI; defense/intelligence pedigree | $3B+; $250B+ market cap | Fastest-growing enterprise AI platform |
| 5 | Workspace AI (Gemini) | AI across Docs, Sheets, Gmail, Meet | Part of Google Cloud | 3B+ Workspace users | |
| 6 | Notion AI | AI-augmented productivity | Consumer-friendly; horizontal knowledge management | $150M+ ARR | Strong developer/startup adoption |
| 7 | Zapier / Make | AI-powered workflow automation | No-code automation with AI steps; 7,000+ integrations | $230M+ ARR (Zapier) | AI automation becoming core use case |
| 8 | Relevance AI / Lindy | Agent builder platforms | No-code AI agent deployment for business users | $10–20M raised each | Democratizing agent creation |
AI applications built for specific industries — legal AI, healthcare AI, financial AI, sales AI — that deliver complete solutions to end users. This is where AI value is ultimately captured.
Global spending on AI-enabled applications projected in 2026, growing 44% YoY. Vertical AI software grows from $133.5B (2025) to $194B (2029) — the largest segment by revenue.
Vertical AI has the strongest moat potential in the entire stack — domain-specific data flywheels, regulatory certification, workflow lock-in. Winning formula: frontier model + proprietary domain data + purpose-built workflow automation. Harvey's leap from $5B to $11B and into talks at $15.5B on $350M annualized revenue by August 2026[24] is the clearest signal of moat strength. The pattern now extends down-market: against Ironclad at $200M ARR and a $3.2B mark, SpotDraft reached 700+ customers and ~$380M by selling contract lifecycle management to mid-market legal teams the enterprise vendors price out, and differentiating on an architecture rather than a model — the document is processed on the endpoint NPU and never reaches a server. Pricing evolving toward outcome-based and fractional-FTE models.
| # | Company | Product | Differentiator | Funding / Valuation | Signal |
|---|---|---|---|---|---|
| 1 | Harvey | AI for legal | Purpose-built for law firms; trusted by top firms | 100K+ lawyers; 25K+ custom agents; Hexus acquired Jan 2026 | |
| 2 | Perplexity | AI search engine | Answer engine with citations; consumer + enterprise | $20–22B valuation (early 2026 Series E-6); ~$200M ARR | Snapchat partnership Jan 2026; $750M Azure commit |
| 3 | Glean | Enterprise AI search | Permissions-aware knowledge graph; connects all enterprise tools | $7.2B valuation (Jun 2025); $200M ARR (Dec 2025, doubled in 9 months); $765M raised | 100M+ agent actions/yr; targeting 1B |
| 4 | Writer | Enterprise AI platform | Brand-safe content generation; governed AI | $1.9B valuation | Fortune 500 customers |
| 5 | Jasper | AI marketing platform | Marketing content generation at scale | $1.7B valuation | Pivoting to enterprise workflows |
| 6 | Abridge | AI clinical documentation | Real-time clinical conversation summarization | ~$5.3B valuation (mid-2025 rounds) | Healthcare-specific compliance |
| 7 | EvenUp | AI for personal injury law | Automated demand letter generation | $1B+ valuation | Deep vertical specialization |
| 8 | Viz.ai | AI for radiology/stroke detection | FDA-cleared AI diagnostic imaging | $500M+ valuation | Regulatory moat |
| 9 | Suki AI | AI medical documentation | Voice-enabled clinical documentation assistant | $500M+ valuation | Physician adoption growing |
| 10 | Hebbia | AI for knowledge work | Enterprise document analysis; financial/legal focus | $700M+ valuation | Matrix reasoning architecture |
| 11 | Ironclad | Enterprise contract lifecycle management | Category leader for enterprise legal; deepest workflow and integration surface | $200M ARR (Jan 2026, +34% YoY); $3.2B valuation; $334M raised | No new round since the Jan 2022 Series E; crossed $100M ARR in 2024 and doubled it in two years |
| 12 | SpotDraft | AI-native contract lifecycle management (VerifAI) | Mid-market CLM sold beneath Ironclad and Icertis. Deliberate split: contract review, clause extraction, risk scoring, redlining, and playbook application run on the Snapdragon X Elite NPU inside Word, so the document never reaches a server; login, licensing, collaboration, and analytics stay cloud-side | ~$92M raised; ~$380M post-money (Jan 2026 Series B extension, Qualcomm Ventures) | 700+ customers, up from ~400 a year earlier; 1M+ contracts/yr, volume +173% YoY; ~100% revenue growth guided for 2026 |
Part 1 is supply-side: who builds what, layer by layer. It does not answer the question an investment committee
asks first, which is whether any of this is working for the people
paying for it. This part is the demand side, and it is deliberately
less flattering than the rest of the document.
The headline is a gap.
Enterprise adoption of AI is close to universal and enterprise value
capture from AI is rare. Both facts are well evidenced, they are not in
tension, and the distance between them is the single most important number in
this report.
| Measure | Value | Source |
|---|---|---|
| Anticipate positive agentic ROI | 171% (expected) | PagerDuty exec survey |
| Attribute any EBIT impact | 39% | McKinsey |
| EBIT impact >5% | ~6% | McKinsey |
McKinsey's high performers are three times more likely to have fundamentally redesigned workflows around AI, and three times more likely to be scaling agents in a function. The differentiator is not model choice, spend, or technical sophistication. It is willingness to change the process rather than bolt AI onto the existing one. This is the single most actionable finding in the report for an operator, and it argues that the binding constraint on enterprise AI value is organisational, not technological.
Gartner projects that more than 40% of agentic AI projects will be cancelled by 2027, citing unclear ROI and weak governance rather than capability shortfalls. Read alongside the 171% ROI executives say they anticipate, this is an expectations problem: buyers are underwriting returns the deployed systems are not yet producing. For vendors this is a near-term revenue risk that does not show up in any ARR chart, and it is the demand-side reason the governance opportunities in Part 3 are rated as highly as they are.
| Measure | Value |
|---|---|
| AI-skilled role growth | +69% |
| Total jobs market | +9% |
| Wage premium 2026 | +62% |
| Wage premium 2025 | +57% |
The supply side of this stack is priced for broad enterprise value capture that has not yet occurred. That does not invalidate the infrastructure thesis, because inference demand is real and power-constrained regardless of whether buyers can attribute EBIT. But it does mean the application and vertical layers carry more execution risk than their funding multiples imply, and it makes the workflow-redesign and governance layers structurally more attractive than a pure capability play. The 6% who capture value did so by changing how work happens. Whoever sells that change profitably is positioned better than whoever sells the model.
Gap: No turnkey solution for real-time AI spend optimization across multi-model environments. Enterprises using 2–4 LLMs simultaneously cannot intelligently route queries to the cheapest model, forecast inference costs, or identify prompt-level waste.
Pain: 40% of companies spend $10M+/year on AI; Cloud Efficiency Rate dropped from 80% → 65%. 500–1,000% cost underestimation at scale (Gartner). Q2 2026 sharpened this: big-four capex reached ~$720–745B, Amazon raised its guide $20B on memory costs alone, and Alphabet posted negative Q2 free cash flow[20]. That cost base gets passed through in token prices. The countervailing force is Opus 5 at half of Fable 5 pricing[21] — per-token costs are falling while total spend rises, which is precisely the condition that makes FinOps tooling necessary rather than optional. TAM: ~$2–4B within 3 years.
3-person team with FinOps background builds "Cloudflare Workers meets Apptio for LLMs." Monetize via % of savings identified. 5-year outcome: $500M+ ARR acquired by Datadog or ServiceNow.
Gap: No turnkey RAG evaluation solution across hybrid retrieval strategies. Enterprises cannot systematically measure retrieval quality, chunk relevance, or end-to-end answer accuracy.
Pain: 87% cite data readiness as primary AI impediment. Arize, Braintrust, and Ragas offer partial solutions only. Watch the ceiling: retrieval evaluation is being absorbed into context-layer platforms that own the corpus and can score grounding natively. The standalone window is narrowing — build for the evaluation harness, not the retrieval stack. TAM: ~$1–2B within 3 years.
Small ML team builds "Datadog for RAG" — automated quality scoring, regression testing, alerting. $500–$5,000/month per pipeline. 5-year: $200M ARR, acquired by Databricks or Datadog.
Gap: No standardized solution for logging, auditing, and proving compliance of AI agent actions in regulated industries. Current frameworks (LangGraph, CrewAI) have minimal built-in governance.
Pain: Gartner: 40% of enterprise apps will embed AI agents by end of 2026. GRC software is a $15B market. Upgraded this cycle: the June 2, 2026 US Executive Order created a voluntary pre-release framework with government early access to frontier models and August 1 benchmarking deliverables, while the EU's July 2026 Cybersecurity and AI action plan pushes the opposite, prescriptive direction[29]. Any vendor selling into both now needs dual-regime evidence, and nobody produces it off the shelf. Regulation moved this from a nice-to-have to a procurement gate. TAM: ~$1–3B~$2–5B within 3 years.
4–6 engineers with GRC or audit background build the evidence layer: immutable action logs, policy attestation, and export packs mapped to both the US voluntary framework and EU AI Act Article 12 record-keeping. Sell to the compliance buyer, not the platform team. 5-year: $150–300M ARR, acquired by a GRC incumbent (OneTrust, Drata) or ServiceNow.
Gap: MCP is becoming the standard for connecting AI models to enterprise tools, but no dominant platform simplifies building, testing, and deploying MCP servers securely.
Pain: By end of 2026, successful SaaS launches will advertise "MCP-native." Building MCP servers requires custom engineering per tool. TAM: ~$500M–1.5B within 3 years.
Gap: Legal, healthcare, and financial workflows handle text that is privileged, PHI, or MNPI. The default AI architecture ships that text to a third-party API. There is no standard toolkit for running extraction, classification, and risk scoring entirely on the endpoint while keeping quality near frontier-model levels.
Pain: SpotDraft proved the demand shape, and the useful detail is the split rather than the slogan. Its VerifAI runs embeddings, clause extraction, risk scoring, redlining, and playbook application on the Snapdragon X Elite NPU end-to-end inside Microsoft Word, so the document never reaches a server; login, licensing, collaboration, orchestration, and fleet analytics stay in the cloud. That is the replicable pattern — keep the privileged artifact on the endpoint, keep the control plane central — and Qualcomm Ventures funded the company to scale it. Two honest constraints: it currently lands on a narrow hardware footprint (Snapdragon X Elite laptops), and the tooling around it is hand-rolled. Quantization, offline evaluation, model update distribution, and proving to an auditor that inference stayed local are all bespoke today. The gap is that toolchain, not the silicon. TAM: ~$1–2B within 3 years.
Gap: Enterprises wanting to fine-tune models for specific domains (legal, medical, financial) lack curated, high-quality training datasets. Options are expensive custom labeling or noisy web scrapes.
Pain: Regulated industries cannot use general web data. Synthetic data tools lack domain depth. TAM: ~$500M–1B within 3 years.
Gap: Context engineering tools exist (LangChain, Mem0, Zep) but no unified platform governs where that context comes from. Result: 66% of enterprises get biased or misleading AI insights; 57% duplicate AI efforts across departments.
Pain: 88% have a formal context strategy, but 87% cite data readiness as a significant impediment. 61% frequently delay AI initiatives. TAM: ~$5–10B within 5 years.
Status change — the category is now contested. Coconut is building exactly this: a shared, living, model-agnostic context layer where every fact carries its source, answers are identical across Claude, ChatGPT, Copilot, and Slack, and missing context is flagged rather than guessed. Qontext raised $2.7M pre-seed in Berlin on the same thesis (HV Capital, with the founders of n8n, Neo4j, Celonis, and make.com participating). The thesis is validated; the open window is narrower than it was in July. Differentiation now has to come from governance depth and lineage, not from the idea.
15–30 engineers build "Okta for AI context" — governance platform sitting between enterprise data systems and AI apps. $50–200/user/month. 5-year: $500M+ ARR, IPO or acquired by Databricks/Snowflake for $5–10B.
Gap: Regulated industries (healthcare, financial services, legal, government) need complete, compliant AI solutions combining frontier model capabilities with domain-specific workflow automation and regulatory certification.
Pain: Vertical SaaS growing at 23.9% CAGR — nearly double the broader SaaS market. Harvey, Abridge, EvenUp prove the pattern. Most verticals remain underserved. Strengthened this cycle: Harvey is in talks at $15.5B on $350M annualized revenue (~$300M ARR), a 41% valuation step in five months[24]. The more useful signal for a new entrant is below the top: SpotDraft reached 700+ customers and ~$380M valuation by taking contract lifecycle management to mid-market legal teams that Ironclad and Icertis price out, and won on deployment speed and on-device privacy rather than model quality. The frontier lab is not the competitor; the incumbent's implementation cycle is. TAM: ~$10–30B across verticals within 5 years.
20–40 people with domain expertise build "Harvey for Insurance" — AI automating claims processing from intake to settlement. State-by-state regulation = enormous barrier to entry. 5-year: $1B+ ARR, 60%+ gross margins, pursuing IPO.
Gap: No platform manages the complete agent lifecycle — deployment, monitoring, debugging, cost optimization, safety testing, rollback — as enterprises deploy autonomous agents at scale.
Pain: 66% of engineering teams report AI outputs look correct but fail during testing. Current tools (LangSmith, Arize) offer observability but not full lifecycle management. TAM: ~$3–8B within 5 years.
15–25 engineers with APM/DevOps background build "Datadog for AI Agents." $0.01 per agent action monitored. 5-year: $300M–500M ARR; standard infrastructure for enterprise agent operations.
Gap: Physical industries (manufacturing, construction, agriculture, logistics) lack AI systems integrating vision, sensor data, robotics control, and language models into unified operational platforms.
Pain: Manufacturing AI adoption <15% despite $14T output. Edge inference hardware (Qualcomm, NVIDIA Jetson) is now capable. Capital is arriving: Atoms, Travis Kalanick's physical-AI company, raised $1.7B led by a16z in July 2026 — one of the largest private rounds of the month and a signal that the category is moving from research to buildout. TAM: ~$15–25B within 5 years.
Gap: As AI handles sensitive data and makes autonomous decisions, new attack surfaces emerge — prompt injection, data poisoning, model extraction. No comprehensive AI-specific cybersecurity solution exists.
Pain: Critical injection defense is a top enterprise concern. Defense tooling is primitive. The June 2, 2026 Executive Order put AI-enabled cyber defense on the federal procurement path and directed criminal enforcement against AI-enabled attacks; the EU's July action plan targets the same surface from the resilience side[29]. Federal demand signal now exists where it did not in July, which shortens the enterprise sales cycle for anyone with a credible product. TAM: ~$3–5B within 5 years.
Gap: GPU-based compute is hitting power efficiency walls. Photonic computing, neuromorphic chips, and wafer-scale integration could deliver 10–100× efficiency gains, breaking the power constraints limiting data center scaling.
Pain: Data centre energy projected to reach 1,000 TWh in 2026reached ~485 TWh in 2025; IEA base case ~945 TWh by 2030[30]. Power availability is the #1 bottleneck, and the binding limit is interconnection queues rather than aggregate consumption. Cerebras IPO at ~$56–66B FDV (May 2026)Cerebras trades at ~$64B (Aug 9, 2026) after a ~$95B Day-1 close, with Q2 reported Aug 12: $210M core revenue against ~$194M expected and FY26 guidance raised to $880–890M, yet the stock fell ~14% after hours — technology thesis validated, multiple still compressing[31][25] and OpenAI's $20B+ equity stake validate the inference-specific silicon thesis at scale. SambaNova's $1B Series F at $11B (Jul 8) confirms private capital is still funding non-NVIDIA architectures[23]. Caution: the public comp has round-tripped ~27% off its Day-1 close, so the thesis is validated at the technology level but the exit multiple is no longer the one that priced in May. TAM: ~$50–100B within 7 years.
Deep-tech startup develops photonic AI accelerators — 50× energy efficiency over GPUs for inference. Requires $200M+ for tape-out. 5-year: acquired by NVIDIA/AMD/Intel for $10B+, or becomes "TSMC of photonic computing."
Gap: Nations outside the US (EU, India, Saudi Arabia, Japan, South Korea) need full-stack sovereign AI: data centers, custom silicon, foundation models, and apps tailored to local languages and regulations.
Pain: EU AI Act requires data residency; Saudi Arabia invested $100B+ in AI infrastructure; Mistral raised €1.7B from ASML for European AI sovereignty. TAM: ~$100B+ within 5 years.
Consortium partners with sovereign wealth fund to build full-stack AI platform for Middle East or Southeast Asia. Combines local data centers with Llama/Mistral models + vertical AI apps. 5-year: $5B+ revenue with government-backed demand certainty.
Gap: No company has built a comprehensive AI-first enterprise suite that replaces Salesforce + ServiceNow + SAP with an agent-native architecture from the ground up.
Pain: $285B market rout when Anthropic's Cowork plugins launched. Publicis Sapient reducing SaaS licenses by ~50%. The $315B SaaS market is being restructured. TAM: ~$50–100B within 7 years.
Palantir-scale ambition — AI-native enterprise OS that replaces point-solution SaaS. Agents handle CRM, ITSM, HR, finance, legal through a single intelligence layer. $500M+ capital required. 5-year: "next Salesforce" at $50B+.
Gap: AI has shown breakthrough potential (AlphaFold), but no company has built a full-stack platform taking a drug target from identification through lead optimization, clinical trial design, and regulatory submission.
Pain: Drug development costs $2.6B per approved drug; clinical trial failure rates exceed 90%. AI could reduce timelines 30–50% and costs 50%+. TAM: ~$30B+ within 7 years.
Gap: AI data centre power consumption projected to reach 1,000 TWh by 2026~485 TWh in 2025 rising to ~945 TWh by 2030 (IEA base case)[30]. Grid capacity is the binding constraint: ERCOT alone received 198GW of large-load interconnection applications in Q1 2026 against a 438GW queue, with 4–7 year energization timelines in major hubs. Dedicated clean energy (SMRs, geothermal, grid-scale storage) for AI compute creates massive value.
Pain: Microsoft revived Three Mile Island; Google signed geothermal deals; power is the #1 bottleneck cited by data center operators. Upgraded this cycle on direct operator testimony: Andy Jassy said on July 30 that even at ~$220B of 2026 capex Amazon will not have enough capacity to meet demand in 2026, and expects the same in 2027[20]. When the largest buyer states publicly that money is not the constraint, the constraint is power and shells. Roughly 60% of hyperscaler capex already goes to power and data center shells rather than chips. TAM: ~$50B+ within 7 years.
Developer secures interconnect queue positions and firm generation (SMR, geothermal, or gas-plus-storage bridge) and sells power-plus-shell as a bundled product to hyperscalers and neoclouds on 15-year contracts. Capital-intensive and permitting-bound, but demand risk is close to zero through 2030.
| # | Opportunity | Tier | Rating | Key Rationale |
|---|---|---|---|---|
| 1 | AI-Native Vertical SaaS for Regulated Industries | Venture ($5–50M) | HIGH | |
| 2 | Enterprise Context & Data Governance Layer | Venture ($5–50M) | HIGH | Databricks' Tecton/Antimatter/SiftD acquisitions confirm consolidation; Coconut and Qontext now building it directly — validated but contested |
| 3 | AI Cost Management / FinOps for LLMs | Bootstrappable (<$5M) | HIGH | Copilot on token billing; ~$720–745B big-four capex meets Opus 5 at half of Fable 5 pricing — spend up, unit cost down[20] |
| 4 | AI-First Enterprise Software Suite | Deep-Pocketed ($100M+) | HIGH | $285B SaaSpocalypse proved replacement thesis at market scale |
| 5 | Next-Gen Compute Infrastructure (Post-GPU) | Deep-Pocketed ($100M+) | HIGH |
Both are horizontal, model-agnostic, and reach revenue quickly. AI FinOps has the largest addressable market and weakest incumbent coverage. Newly competitive alternative: Agent Compliance & Audit Trails, upgraded to HIGH this cycle — the June 2026 Executive Order and the EU July action plan turned governance evidence into a procurement gate with no off-the-shelf answer.[29]
Moat from regulatory certification, proprietary data, and workflow integration is the strongest in the stack. Go one tier below the marquee names: SpotDraft won mid-market legal on deployment speed and on-device privacy, not model quality. Second priority: Enterprise Context Governance — the "missing layer" every enterprise AI deployment needsstill the missing layer, but no longer unoccupied — Coconut and Qontext are building it now, so enter on governance depth and lineage rather than on the idea.
Largest market restructuring since the cloud transition. Requires world-class execution and 5–7 year horizon. Focused bet alternative: Sovereign AI infrastructure — near-certain demand backed by government budgets. Lower-variance alternative added this cycle: AI Energy Infrastructure, upgraded to HIGH after Amazon stated that ~$220B of 2026 capex still will not meet demand through 2027. Power, not capital, is the binding constraint.[20]
Pure-play model development — capital requirements are prohibitive; Anthropic and OpenAI have insurmountable leads. · Pure-play vector databases — commoditizing rapidly. · Undifferentiated AI middleware — thin wrappers with no moat.
Part 1 reads the stack by altitude: each layer consumes the one below and supplies the one above. Three forces do not respect that structure. Security, regulation and energy cut vertically through all twelve layers, and none of them can be located at a single altitude. Treating any of them as a thirteenth layer would be a category error, because a layer is defined by what it consumes and what it supplies, and these consume and supply at every level at once.
Each spine below is the stack transposed: one row per layer, read top to bottom. The pattern each one traces is the finding. Security runs as a defensive concern from L01 to L11 and then inverts at L12, where AI is sold as security rather than secured. Regulation concentrates at the two ends and thins in the middle. Energy binds hard at the bottom four layers and is close to irrelevant above them, which is itself worth seeing at a glance rather than asserting.
Two markets, not one, and conflating them is the standard error. AI applied to security is mature at roughly $30–35B in 2026. Securing AI itself is far smaller and far faster: MarketsandMarkets puts agentic AI security at $1.65B in 2026 reaching $13.5B by 2032 (42% CAGR), while Dell'Oro sizes AI systems security at close to $8B by 2030 from a standing start in 2024.
Consolidation is running ahead of category formation. Palo Alto absorbed Protect AI, Check Point took Lakera, Cisco took both Robust Intelligence and Astrix, and Cyera signed a $1B letter of intent for Oasis. Independents are being acquired before reaching scale, which implies the durable position in most of this spine is a feature inside a platform rather than a standalone product. The exception is the agent layer, where Zenity's $125M Series C in August 2026 shows capital still funding independents. For an investor the read is straightforward: at L07 and L08 you are underwriting an acqui-exit, at L09 and L12 you are still underwriting a company.
| Layer | Exposure at this altitude | Who sells here | Signal |
|---|---|---|---|
| 01 | Hardware root of trust, TEEs on-accelerator, die and package attestation | NVIDIA Confidential Computing, Intel TDX, AMD SEV-SNP, Arm CCA | Blackwell is the first TEE-I/O capable GPU; confidential-mode throughput penalty fell from 30–40% to near parity |
| 02 | Tenant isolation on shared GPU fleets, memory residue, BMC and firmware | Fortanix, Edgeless Systems, operator-native controls | Neocloud tenancy is less hardened than hyperscaler; audit evidence gap widening |
| 03 | East-west encryption at RDMA line rate, fabric segmentation | NVIDIA BlueField DPUs, Cisco Hypershield, Arista | Encryption at 800G without throughput loss remains unsolved cheaply |
| 04 | Non-human and agent identity, KMS, endpoint access control | Astrix (Cisco), Oasis (Cyera $1B LOI), Entro, GitGuardian, Wiz | Fastest-consolidating category; NHI rounds passed $340M in the last year |
| 05 | Training-data poisoning, provenance and lineage, PII and PHI leakage | Cyera, Varonis, BigID, Unity Catalog | Underpins the Enterprise Context Governance opportunity in Part 3 |
| 06 | Weight exfiltration, dangerous-capability evals, pre-release red teaming | Gray Swan, Irregular, Haize Labs, lab-internal | Gray Swan $40M Series A (May 2026); cited in 11 frontier model system cards |
| 07 | Model artifact supply chain, unsafe deserialization, malicious hub models | HiddenLayer, Prisma AIRS (ex-Protect AI), JFrog ML | Protect AI absorbed by Palo Alto Jul 2025; HiddenLayer independent on $56M |
| 08 | Runtime prompt injection, output filtering, redaction, token flooding | Prompt Security, Lakera (Check Point), Robust Intelligence (Cisco) | Two of the three leading independents acquired; this layer is now platform-owned |
| 09 | Agent permissions, tool-call sandboxing, egress control, human-in-loop gates | Zenity, Invariant Labs, E2B, Browserbase | Zenity $125M Series C (Aug 3, 2026); capital still backing independents here |
| 10 | MCP server trust and authn, gateway policy, spend limits as a control | Cloudflare AI Gateway, Portkey, Kong | OX Security disclosed ~200,000 vulnerable MCP instances (Apr 2026); NSA guidance since issued |
| 11 | DLP for copilots, oversharing via enterprise search, shadow AI | Microsoft Purview, Netskope, Island, Witness AI | Oversharing is the most-cited blocker to copilot rollout; a permissions problem, not a model problem |
| 12 | AI applied to security operations — detection, triage, response | CrowdStrike Charlotte, Microsoft Security Copilot, Dropzone, Prophet, Torq | Dropzone $57.4M / 100+ enterprise customers; Prophet $41M; all three majors shipped agentic SOC at RSAC 2026 |
L04, L09 and L12. Identity (L04) and agent control (L09) are the two surfaces where enterprises are actively writing cheques and where no incumbent yet owns the category. L12 is a genuine vertical with real revenue. Everything else is either already absorbed into platform vendors (L07, L08) or inherited from general infrastructure security (L02, L03), and should be priced as a feature rather than a market.
Not a market so much as a procurement gate and a compliance cost. The adjacent GRC software market is roughly $15B. The material fact for 2026 is divergence in kind: the US June 2026 Executive Order is voluntary and access-based, while the EU is prescriptive. Any vendor selling into both now needs dual-regime evidence, and nobody produces it off the shelf.
The EU timeline just moved, and it moved in the direction that reduces near-term pressure. On June 16, 2026 the European Parliament deferred most high-risk obligations: Annex III use-based duties slip from August 2026 to December 2027, and Annex I product-regulated duties from August 2027 to August 2028. What did take effect on schedule on August 2, 2026 is Article 50 transparency and AI Office enforcement over general-purpose models. The practical consequence is that near-term obligations are about disclosure and GPAI, not high-risk certification, which pushes the compliance-tooling opportunity in Part 3 further out than the July reading implied.
| Layer | Exposure at this altitude | What applies | Signal |
|---|---|---|---|
| 01 | Export controls, entity lists, fab and tooling restrictions | Jan 2026 Commerce rule reopened H200 / MI325X to China; late-May 2026 rules re-tightened on Blackwell and closed the offshore-subsidiary loophole | Chinese vendors took ~41% of China's AI accelerator server market in 2025, up from NVIDIA's ~95% share in 2022 (IDC) |
| 02 | Siting, permitting, water disclosure, large-load interconnection | 200+ state data-centre bills in 2025, 40+ enacted; DOE proposed federal jurisdiction over loads above 20MW | Regulatory risk is now local and political, not federal |
| 03 | Export controls on high-end networking equipment | Inherited from L01 and L02 | Low direct exposure |
| 04 | Data residency and sovereignty | EU residency requirements enforced at the cloud boundary | The main transmission path for EU obligations into US vendors |
| 05 | Copyright, lawful basis for training data, residency | Active litigation across jurisdictions | Exposure attaches to the corpus, not the weights, so it concentrates here rather than at L06 |
| 06 | GPAI obligations, transparency, pre-release evaluation | EU Article 50 and GPAI enforcement took effect Aug 2, 2026 as scheduled. High-risk obligations deferred: Annex III to Dec 2027, Annex I to Aug 2028 (Parliament vote, Jun 16, 2026) | US June 2026 EO is voluntary and access-based; the two regimes are diverging in kind, not just in degree |
| 07 | Documentation, record-keeping, third-party assessment | Lands here first when high-risk rules apply | Deferred to Dec 2027, which removes near-term urgency |
| 08 | Content marking, deepfake labelling, output disclosure | Article 50 enforced at the serving boundary, live now | The one EU obligation with immediate teeth |
| 09 | Audit trails, action logging, human oversight | June 2026 US EO and EU record-keeping both land here | Turned agent governance evidence into a procurement gate |
| 10 | Policy enforcement and evidence generation | Natural enforcement point for both regimes | The gateway is the only component that sees every call |
| 11 | Workplace surveillance, accessibility, employment law | Attaches to the deploying enterprise, not the vendor | Underrated; rarely appears in vendor risk registers |
| 12 | Sectoral certification per vertical | State-by-state insurance, HIPAA, FINRA, FDA | The moat named in the top-ranked opportunity in Part 3 |
L01 and L06, with a long tail at L12. Export controls at silicon are the highest-consequence regulatory force in the stack and the one this report has historically under-covered. L06 carries the live EU obligations. L12 carries sectoral certification, which is a moat rather than a cost. The middle of the stack (L03, L07, L10) is largely regulated by inheritance.
Global data-centre electricity consumption was approximately 485 TWh in 2025, and the IEA base case projects roughly 945 TWh by 2030, about 3% of global electricity, up from 1.5% in 2024. Accelerated servers grow near 30% annually against 9% for conventional servers and account for close to half the net increase. This corrects the 1,000 TWh-in-2026 figure carried in prior editions, which was a 2024-vintage projection that did not materialise.
The constraint is interconnection, not generation. There is no shortage of prospective electrons; there is a shortage of approved connections and delivery timelines. ERCOT alone received 198GW of large-load applications in Q1 2026 against a 438GW queue, energization runs 4–7 years in major hubs, and DOE has moved to assert federal jurisdiction over loads above 20MW specifically to compress that. This is why Amazon can state that $220B of 2026 capex still will not meet demand: capital is not the binding input.
| Layer | Exposure at this altitude | Evidence | Signal |
|---|---|---|---|
| 01 | Wafer and advanced-packaging capacity; perf-per-watt | CoWoS allocation, HBM supply | Every efficiency gain here compounds through all 11 layers above |
| 02 | Grid interconnection, firm generation, water, shells | ERCOT took 198GW of large-load applications in Q1 2026 against a 438GW queue | Energization runs 4–7 years in major hubs, up to a decade elsewhere. ~60% of hyperscaler capex is power and shells |
| 03 | Optics and switching power draw | Rising share of rack power as clusters scale | One of the few remaining cheap efficiency levers |
| 04 | PUE, siting economics, cooling | Hyperscaler PUE gains largely exhausted; long tail still has headroom | Sets the floor under every price above |
| 05 | Storage and ETL draw | Small relative to L02 | Not a binding constraint |
| 06 | Training-run power events | Largest single events, but inference has overtaken training in aggregate | Training is the headline; inference is the curve |
| 07 | Quantization and distillation as demand-side levers | The only relief that does not require new generation | Lever, not a constraint |
| 08 | Steady-state token-serving load | Now over 55% of AI infrastructure spend | Inference efficiency, not training efficiency, determines the 2030 power curve |
| 09 | Multi-step agent traces multiply L08 draw | A single agent trace can consume 10–100× the tokens of one completion | Indirect multiplier |
| 10 | Routing and caching as efficiency levers | Semantic caching and cheap-model routing | Lever, not a constraint |
| 11 | None material | — | Not a binding constraint |
| 12 | None material | — | Not a binding constraint |
L01 through L04 only. The thinness of the top eight rows is the finding, not a gap in the research. Physical constraints bind at the bottom of the stack and are close to irrelevant above L04, which is precisely why application-layer companies can scale without capital intensity while infrastructure companies cannot. The one exception worth watching is L08 and L09: inference efficiency and agent-trace multiplication are demand-side levers on L02 power draw, and they are the only levers available on a shorter horizon than grid interconnection.
A landscape report is only as good as its willingness to say where it is weak. This appendix states the method, the confidence tiers, and the known defects that survived this cycle.
Every quantified claim is checked against public primary sources on each refresh, prioritised by volatility: valuations, ARR, IPO status and corporate structure are re-verified every cycle; silicon roadmaps and sovereign policy less often. Corrections are logged in the Updates banner with the superseded value struck through inline and a superscript link back to the entry. No figure is carried forward on the strength of having appeared in a prior edition, which is exactly how the 1,000 TWh energy error survived three cycles before entry 30 caught it.
A Filed or disclosed: SEC filings,
earnings calls, official press releases. Treat as fact.
B Credible third-party estimate:
Sacra, Menlo, Dell'Oro, IEA, analyst surveys. Directionally reliable,
point values soft.
C Market-sizing projection or
extrapolation. Useful for order of magnitude only. Every TAM in Part 3 is
tier C, and the 2032 figure in the AI-security chart is the softest number
in the document.
Survey-based demand data is self-reported. The 88%, 39%
and 6% figures in Part 2 come from executive surveys, which systematically
over-report adoption and under-report failure. Read them as upper bounds on
adoption and lower bounds on disappointment.
Market sizings use mixed bases. The layer chart places TAM,
annual capex and annual funding on one axis because that is how the
underlying sources report them. It is legitimate for rank order and
misleading for arithmetic. Do not sum it.
Newly added companies carry structural data only. The
companies added this cycle to close the Layer 03 and search-coverage gaps
have verified layer placement, taglines and relationships, but financial
stats only where a figure was verified this cycle. An empty stat is an
absence of verified data, never an implied zero.
Private-company revenue is estimated. ARR figures for
private companies derive from third-party estimators and should be treated as
tier B throughout.