AI Software

Hyperscaler Cloud Revenue

Quarterly cloud revenue from each hyperscaler's 10-Q. AWS, Microsoft Intelligent Cloud (Azure is not separately disclosed but called out via growth rate), Google Cloud, and Oracle OCI. AI-attributable contribution is captioned where management has disclosed it.

Milestone Disclosures

  • 2024-01-30Microsoft. First quarter Microsoft attributed a specific point contribution from AI services to Azure growth (approximately 6 points). (MSFT FY2024-Q2 earnings call)
  • 2025-01-29Microsoft. Total Microsoft AI business reached $13B annual run-rate, up 175% year over year. (MSFT FY2025-Q2 earnings call (Satya Nadella))
  • 2024-07-23Alphabet. Google Cloud crossed $10B in quarterly revenue for the first time. (Alphabet Q2-2024 earnings release)
  • 2025-09-09Oracle. RPO step-up to approximately $455B on signed AI compute contracts including OpenAI. (ORCL FY2026-Q1 earnings call)

Quarterly cloud revenue, USD billions, calendar 2023-Q1 forward. AWS is the AWS reporting segment from Amazon. Azure is the Intelligent Cloud segment (Microsoft does not report Azure as a standalone line). Google Cloud is the Alphabet segment. OCI is Oracle Cloud Infrastructure, mapped from Oracle’s May fiscal year to the closest calendar quarter. Hover a point for management-disclosed AI contribution annotations where available. Click legend pills to toggle series.

Hyperscaler AI Capex

Annual capex 2020 through 2026E from cash flow statements. AI-attributable share is annotated where management has disclosed it (Meta and Microsoft routinely; others sparsely). Estimates marked at reduced opacity.

Faded bars are 2026E sell-side consensus. AI-attributable share appears in tooltip where management disclosed it.

Management AI-Capex Disclosures

  • 2025-01-29Meta Platforms (Mark Zuckerberg). The majority of our capex going forward will be on AI infrastructure. (META Q4-2024 earnings call)
  • 2025-02-06Amazon (Andy Jassy). The vast majority of capex spend is on AI for AWS. (AMZN Q4-2024 earnings call)
  • 2025-02-04Alphabet (Sundar Pichai, Ruth Porat). 2025 capex approximately $75B, principally driven by AI infrastructure. (Alphabet Q4-2024 earnings call)
  • 2025-01-29Microsoft (Amy Hood). Capex is dominated by long-lived assets for AI and cloud workloads. (MSFT FY2025-Q2 earnings call)
  • 2025-06-11Oracle (Larry Ellison). OCI capex is being driven essentially entirely by AI training compute demand. (ORCL FY2025-Q4 earnings call)

Annual capital expenditure, USD billions, calendar 2020 through 2026E. Cash-flow capex from each company’s 10-K; 2026E reflects sell-side consensus (JPMorgan, Morgan Stanley, Goldman Sachs). Amazon capex includes fulfillment plus transportation, not AWS-only. Microsoft additionally records material finance-lease right-of-use additions for data centers, not included in cash-flow capex. AI-attributable share is management-disclosed only; companies use qualitative language (“majority”, “substantial majority”) rather than a dollar split.

Neocloud and Inference Platform Comparison

GPU-rental neoclouds (CoreWeave, Nebius, Lambda, Crusoe, etc.) and inference platforms (Together, Fireworks, Modal, Replicate). Per-row status, estimated GPU count, contracted revenue or RPO, FY24 revenue, and major customers.

GPU Rental

Companies that own fleets of NVIDIA GPUs and lease compute capacity directly.

CompanyStatusGPU count (est.)RPO / contractedFY24 revenueMajor customers

CoreWeave

PublicCRWV

IPO 2025-03

approximately 250,000$26B$1.9BMicrosoft (62% of FY2024 revenue per CRWV S-1), Meta, NVIDIA, IBM, Cohere, Mistral, Anthropic (announced)

Nebius

PublicNBIS

IPO 2024-10

approximately 22,000n/d$220MAI labs and enterprises (specific named customers not disclosed at customer level)

Applied Digital

PublicAPLD

IPO 2022-04

approximately 10,000 (est.)n/d$165MCoreWeave (announced lease arrangement), one large hyperscaler (named publicly in Sept 2024 press release)

IREN (formerly Iris Energy)

PublicIREN

IPO 2021-11

approximately 4,300 H100 / H200 (est.)n/d$188MAI customers not disclosed; bitcoin mining historically the dominant business

Lambda

Series C $480M April 2024 at approximately $1.5B valuation; Series D $480M February 2025 reportedly at $4B (Reuters).

Privateapproximately 100,000 (est.)n/d$500MMicrosoft (announced multi-year contract 2024), AI labs

Crusoe Energy

Series D $686M December 2024 at approximately $2.8B; Series E reportedly raised 2025 at higher valuation.

Privateapproximately 50,000 (est.)$10Bn/dOpenAI (Stargate JV), Oracle (Stargate JV), Microsoft (announced)

Voltage Park

Funded by Jed McCaleb's Navigation Fund (philanthropic vehicle); reportedly approximately $500M invested in compute capacity.

Privateapproximately 24,000 (est.)n/dn/dSpot-rental marketplace customers; not named

Inference Platform

Companies that resell or layer software over third-party GPU capacity to offer model serving.

CompanyStatusGPU count (est.)RPO / contractedFY24 revenueMajor customers

RunPod

Series A $20M May 2024 led by Intel Capital.

Privaten/a (resells)n/dn/dDevelopers and small-to-midmarket inference customers

Together AI

Series B $106M March 2024 at approximately $1.25B; Series C $305M February 2025 at approximately $3.3B (Reuters).

Privaten/a (primarily inference and fine-tuning over leased fleets)n/dn/dOpen-source AI developers and enterprises running Llama, DeepSeek, Mistral, Qwen models

Fireworks AI

Series B $52M July 2024 at $552M post-money (Sequoia, Benchmark).

Privaten/a (resells)n/dn/dDoordash, Quora Poe, Cresta, Cursor (announced inference partners)

Replicate

Series B $40M December 2023 led by a16z; valuation approximately $350M.

Privaten/a (resells)n/dn/dDevelopers using SDXL, Llama, FLUX, Whisper variants

Modal

Series A $16M Aug 2023 led by Redpoint; Series B reportedly closed in 2025 (specific figure not public).

Privaten/a (resells, serverless)n/dn/dAI startups and ML teams running batch inference and training jobs

GPU counts are point-in-time estimates compiled from S-1 disclosures, 10-Q narrative, announced data-center power capacity, and reported procurement orders. RPO (remaining performance obligation) is reported by CoreWeave; private peers report contracted-revenue or multi-year contract value in the same field with a label note. Hover any figure for the specific source citation. n/d = not disclosed.

Open Weights Inference Pricing

Effective dollars per million tokens across providers for open-weights models in three size classes (small around 7B, medium around 70B, frontier reasoning). Demonstrates commoditization in the inference layer once a model is openly released.

Llama 3 / 3.1 / 3.3 70B (medium open-weights generalist). Released 2024-04-18 (Llama 3); 2024-07-23 (3.1); 2024-12-06 (3.3).

Effective USD per million tokens (input and output blended at 3:1, matching the Artificial Analysis convention). Quarterly snapshots, calendar 2024-Q2 through 2026-Q1. Null cells mean the model was not yet released or the provider serves only distilled variants. Groq is a throughput-tier outlier (latency-optimized); its $/MTok is competitive but the headline value is tokens per second, not price. Hover the chart for per-provider values; click legend pills to toggle. (Sources: Artificial Analysis, provider pricing pages.)

Data and ML Platforms

Databricks, Snowflake, MongoDB, HuggingFace, Scale AI. Per row: FY24 revenue, AI-attributable revenue contribution where disclosed, notable products. The platform layer benefits from AI workloads without the capex burden of building models.
CompanyStatusRevenue (FY24)AI-attributableNotable AI products

Databricks

Series J $10B December 2024 at $62B post-money; subsequent extension reported 2025 at higher valuation per Bloomberg.

Private$3.0B

ARR

n/d

Mosaic AI plus DBRX plus Genie cited qualitatively as the principal AI bundle inside Lakehouse.

Mosaic AI (training and fine-tuning platform from the MosaicML acquisition), DBRX (open-weights model), AI Functions in SQL, Genie (natural-language analytics), Unity Catalog vector search.

Snowflake

n/a (public)

PublicSNOW$3.6B

FY2025 revenue (fiscal year ended 31 Jan 2025)

n/d

Cortex AI revenue contribution is not separately broken out as of the FY2025 10-K.

Cortex AI (LLM functions inside the data warehouse), Cortex Search (vector and lexical retrieval), Cortex Agents, Document AI, Snowpark Container Services hosting.

MongoDB

n/a (public)

PublicMDB$2.0B

FY2025 revenue (fiscal year ended 31 Jan 2025)

n/d

Voyage AI (embedding models) acquired February 2025 for an undisclosed sum; integrated into Atlas.

Atlas Vector Search (native vector search inside MongoDB Atlas), Voyage AI embeddings (acquired Feb 2025), Atlas Stream Processing for real-time AI workloads.

Hugging Face

Series D $235M August 2023 at $4.5B post-money; lead investors Google, Amazon, NVIDIA, Salesforce, Sound Ventures.

Private$70M

Reported ARR

$70M

Revenue mix: enterprise hub subscriptions, inference endpoints, AutoTrain, compute reseller margin.

Hub (the dominant model and dataset registry), Inference Endpoints, AutoTrain, Spaces, transformers and diffusers libraries.

Scale AI

Meta acquired a 49% stake for $14.3B in June 2025; Alexandr Wang became Meta's Chief AI Officer concurrent with the deal. Scale's pre-deal valuation was $29B (Series F May 2024).

Private$870M

Reported revenue (calendar 2024)

$870M

Includes federal Donovan defense AI revenue (Scale Federal) plus commercial RLHF and evaluation contracts plus Outlier consumer labeling brand.

Outlier (consumer labeling), Donovan (defense Gen AI), Scale Studio (model evaluations), Scale Data Engine.

Latest fiscal-year revenue for the data and ML platform layer. Public-company revenue is from the 10-K; private-company revenue is from reported figures (Reuters, Bloomberg, The Information) and marked n/d where the company has not disclosed. AI-attributable revenue is management-disclosed only. Snowflake and MongoDB have not split AI revenue; Databricks characterizes AI as the majority of new pipeline but has not separately disclosed Mosaic AI / DBRX / Genie revenue. Hover any figure for the specific source citation.

Vector Database Market

Pure-play vector databases (Pinecone, Weaviate, Chroma, Qdrant, Milvus / Zilliz) versus embedded options (Postgres pgvector, MongoDB Atlas Vector Search, Snowflake Cortex Search, Elastic, Redis). Includes type, market share where measured, public versus private, and latest funding.

Pure-play vector databases

Vector-native databases built specifically for embeddings and approximate-nearest-neighbor search.

VendorStatusGitHub starsMarket share signalLatest funding

Pinecone

Privaten/aRetool 2024 reports Pinecone as the leading pure-play vector DB by usage among surveyed AI builders (approximately one third of respondents using a pure-play indicated Pinecone). Pinecone is closed-source so GitHub-star ranking does not apply.Series B $100M April 2023 at $750M post-money led by Andreessen Horowitz. Subsequent reports suggest no priced round since; revenue trajectory characterized as 'plateauing' by a16z and ICONIQ surveys in 2025.

Weaviate

Private14kSecond-tier pure-play in Retool 2024; smaller share than Pinecone but actively used in enterprise RAG. Open-source repository commands a meaningful developer following.Series B $50M April 2023 at approximately $200M post-money led by Index Ventures; subsequent extension reported 2024.

Chroma

Private18kDominant choice for prototype and developer-local RAG per a16z survey; lower share in production deployments than Pinecone or Weaviate. Strong open-source repository momentum.Series A $18M April 2023 led by Quiet Capital; Series B $48M April 2025 led by Lightspeed at approximately $300M valuation per The Information.

Qdrant

Private21kEuropean-headquartered (Berlin) pure-play; strong open-source uptake, particularly in cost-sensitive deployments. Smaller US enterprise footprint than Pinecone.Series A $28M January 2024 led by Spark Capital at approximately $200M post-money.

Milvus / Zilliz

Private30kMilvus is the highest-starred open-source vector database. Zilliz is the commercial vendor behind Milvus. Strong adoption in Asia and among LF AI graduates.Series B $60M January 2022 led by Hillhouse and Prosperity7; subsequent extension reported 2023.

Embedded vector search (inside existing databases)

Vector search added on top of an existing database or search engine. Compresses pure-play share by removing the need for a second store.

VendorStatusGitHub starsMarket share signalLatest funding

pgvector (PostgreSQL extension)

Open-source12kpgvector is the single largest compressor of pure-play vector DB market share per Retool 2024 and a16z 2024. Used by anyone running Postgres for transactional data who adds vector search rather than introducing a second database. Supported natively by AWS RDS, Aurora, Azure Database for PostgreSQL, GCP Cloud SQL, and Neon.n/a (open-source extension; primary author is Andrew Kane).

MongoDB Atlas Vector Search

PublicMDBn/aAtlas Vector Search is bundled inside MongoDB Atlas (the hosted MongoDB product). Adoption tied to broader MongoDB Atlas customer base; the Voyage AI acquisition (Feb 2025) added owned embedding models.n/a (public; MDB)

Snowflake Cortex Search

PublicSNOWn/aCortex Search is the Snowflake-native vector and lexical retrieval offering inside the data warehouse. Usage tied to Snowflake customer base for unstructured-data analytics.n/a (public; SNOW)

Elastic (Elasticsearch / Lucene)

PublicESTC70kElasticsearch added dense vector indexes in 7.3 and continued to improve through 8.x; Elastic positions itself for hybrid lexical-plus-vector retrieval at enterprise scale. Strong in customers who already standardized on Elastic for log search.n/a (public; ESTC)

Redis (Redis Vector Search)

Private67kRedis (now Redis Inc, post-2024 license shift) offers vector search inside Redis Enterprise. Strong in latency-sensitive caching plus retrieval workloads. Reverted to a more permissive license late 2024 to restore open-source mindshare.Last announced primary funding Series G $110M April 2021 at $2B post-money. Private; reported revenue approximately $200M ARR.

Market Dynamics

  • Embedded vector search inside existing databases (pgvector, Atlas Vector Search, Cortex Search) is compressing the addressable market for pure-play vector DBs. Most enterprise teams choose the path that does not add a new database. (Retool State of AI Report 2024; a16z LLM developer survey 2024.)
  • Pure-plays differentiate on workload-tier features: large-scale ANN performance, hybrid search, multi-tenancy, namespace isolation, and managed-service ergonomics. Pinecone, Weaviate, Qdrant, Zilliz all serve high-throughput RAG and agent workloads at scale. (Vendor product pages; benchmark studies (Qdrant benchmarks, ANN-Benchmarks).)
  • Open-source distribution dominates pure-play adoption: Milvus, Qdrant, Weaviate, Chroma all lead with open-source. Pinecone is the notable closed-source exception and faced commercial-traction questions through 2024 and 2025. (GitHub data; The Information reporting on Pinecone.)

No third-party tracker publishes a clean revenue-share table for this market. Signals captured: Retool State of AI 2024 (vector database usage among AI builders), a16z LLM developer survey 2024, and GitHub stars on each vendor’s primary open-source repository. Embedded vector search inside existing databases (pgvector, Atlas Vector Search, Cortex Search) is the principal compressor of pure-play addressable market. GitHub star counts are point-in-time and rounded; treat as ordinal not cardinal. Hover any cell for the specific source citation.

Ask Sterling

Register for a premium account to gain access to Sterling AI.

Get Started

Things you can ask Sterling:

Summarize Tesla's latest earnings reportWhy did NVIDIA's margins expand?Compare Apple vs Microsoft's cash flowWhat's driving the EV sector growth?
AI Software - Infrastructure | Sterling