121 companies that consume their foundational model at Native (own / self-hosted models) proximity — closest to the metal, cheapest tokens, most control. Avg Foundation Proximity Score 100/100.
Cybersecurity · AWS · Native (own / self-hosted weights)
Abnormal runs a cloud-native, API-first behavioral AI platform that integrates with cloud email and collaboration suites (e.g., Microsoft 365, Google Workspace) outside the mail flow, ingesting user, email, and identity telemetry into proprietary behavioral models and a data lake to drive detection, response, and human-risk scoring.[1][5][7][8][9] Its core is a proprietary behavioral foundation model and related ML pipelines hosted in its own cloud environment, exposed through its SaaS console and APIs rather than via public/open-source model distribution.[7][9]
Healthcare · AWS · Native (own / self-hosted weights)
Ambient clinical-documentation AI with its own medical models plus partner LLMs, grounded in clinician audio.
Creative · Multi-Cloud · Native (own / self-hosted weights)
Adobe runs a **multi-cloud** architecture primarily across AWS, Azure, and selected GCP sovereign regions, with Adobe Experience Platform and Experience Cloud services deployed on both AWS and Azure, and Firefly’s GPU-heavy training and inference infrastructure built on AWS. Core data and AI workloads rely on managed storage, compute, and networking primitives from these clouds, with Adobe layering proprietary platforms such as Experience Platform, Sensei, and Firefly on top.[1][2][3][5][6][7][12][14][15]
AI/LLM · Multi-Cloud · Native (own / self-hosted weights)
AI21 Labs develops and serves its own Jamba family of LLMs and related services via its AI21 Studio API and private deployments, while also distributing Jamba models through third‑party clouds such as Azure and upcoming NVIDIA APIs.[5][6][7] Public materials describe deployment options across public cloud and private/on‑prem environments but do not reveal a full production stack beyond this high‑level multi‑cloud posture.[3][5][6]
AI/LLM · Hybrid · Native (own / self-hosted weights)
Aleph Alpha’s current product layer appears to be PhariaAI, a sovereign enterprise stack that includes knowledge capture, development, operation, access control, and monitoring components. Public materials also show first-party API access and containerized deployment patterns, but the exact production datastore and hosting mix are not fully disclosed, so the infrastructure is best characterized as hybrid rather than purely on-prem or purely cloud.[2][5][10]
Healthcare AI · AWS · Native (own / self-hosted weights)
Ambient documentation with its own clinical models plus OpenAI, grounded in encounter audio.
AI/LLM · Multi-Cloud · Native (own / self-hosted weights)
Anthropic runs **Claude** and related services on a safety-first **multi-cloud** compute fabric spanning AWS Trainium2, Google TPUv7 and NVIDIA GPUs, fronted by Kubernetes‑based microservices (API gateways, orchestration, rate limiting, caching, and safety filters), with state held in PostgreSQL, vector stores, Redis, and cloud object storage.[1][10][13] Production offerings like Claude API and Managed Agents expose a fully managed orchestration and agent runtime, while emerging self‑hosted sandboxes move tool execution into customer infrastructure but keep Claude inference, routing, and session state on Anthropic’s cloud.[2][3][11]
Telecom · Hybrid · Native (own / self-hosted weights)
Telecom; in-house network ML plus partner LLMs (Ask AT&T on Azure OpenAI) for employees.
Design/Eng · AWS · Native (own / self-hosted weights)
Design/CAD cloud with its own generative-design models plus partner LLMs for assistants.
AI-infra · Multi-Cloud · Native (own / self-hosted weights)
Model-inference platform serving open models (Llama, etc.) on autoscaling GPU infra.
Banking (EU) · Hybrid · Native (own / self-hosted weights)
European bank; in-house risk ML plus Mistral for EU-sovereign generative AI.
Travel · Hybrid · Native (own / self-hosted weights)
Travel marketplace; huge own ranking/pricing ML plus OpenAI for the AI Trip Planner.
Social/Consumer · Hybrid · Native (own / self-hosted weights)
ByteDance runs a large-scale hybrid architecture with proprietary cloud-native data and streaming systems (ByteHouse, ByteHTAP, ByteMQ/Flux) atop commodity storage and compute, tightly integrated with recommendation and multimodal AI stacks for products like TikTok/Douyin and robotics.[1][3][6][7][8][9] Real-time streaming (Kafka/ByteMQ), HTAP, and on-device AI infrastructure (Pitaya) form the core substrate for online training, inference, and global content delivery.[1][2][3][8][10][11]
Design · AWS · Native (own / self-hosted weights)
Massive consumer design platform on AWS with MySQL + DynamoDB and a WASM-accelerated rendering pipeline. Magic Studio blends in-house generative models with partners (incl. acquired Leonardo.ai).
AI-hardware · On-Prem · Native (own / self-hosted weights)
Designs wafer-scale AI chips and serves open models (Llama) at record speed on its own systems.
Consumer AI · GCP · Native (own / self-hosted weights)
Character.AI runs a proprietary end‑to‑end stack on Google Cloud, training and serving its own Kaiju/C‑series LLMs on GPU and TPU infrastructure (including H100 clusters) with heavy inference optimization, backed by managed databases (AlloyDB, Spanner), Redis, and BigQuery on GCP.[1][2][3][4][7][10][11][15] The production platform uses dense transformer models (Kaiju Small/Medium/Large) trained and hosted on GCP, with int8-optimized inference and custom serving infrastructure for low‑latency conversational workloads.[3][7][11]
EdTech · AWS · Native (own / self-hosted weights)
Study platform blending its own content ML with OpenAI for its learning assistant.
Infrastructure · On-Prem · Native (own / self-hosted weights)
Cloudflare runs its own Linux-based servers and private backbone as a globally anycasted edge and core network, with most products (CDN, Zero Trust, R2, KV, Queues, Agents, Containers) built on an internal proxy chain and the Workers/Durable Objects developer platform.[3][9][10] Control planes and schedulers (for Containers, Workers AI, agents, etc.) are themselves implemented on Cloudflare’s stack, indicating a predominantly self-hosted, on-prem architecture rather than dependence on hyperscale public clouds.[7][8][10]
AI/LLM · Multi-Cloud · Native (own / self-hosted weights)
Enterprise-focused LLM lab training its own Command/Embed/Rerank models, served across clouds (Google Cloud, Oracle, AWS) and deployable in-VPC for data-sensitive customers. North platform targets RAG.
Telecom/Media · Hybrid · Native (own / self-hosted weights)
Cable/media; own voice-remote + network ML plus partner LLMs for assistants.
AI/LLM · Multi-Cloud · Native (own / self-hosted weights)
RAG-native enterprise platform training its own grounded language models.
Robotics · AWS · Native (own / self-hosted weights)
Warehouse robotics on RFM-1, its own robotics foundation model (team since joined Amazon).
CX AI · GCP · Native (own / self-hosted weights)
Contact-center AI with its own real-time coaching models plus partner LLMs over call data.
Autonomous · GCP · Native (own / self-hosted weights)
GM self-driving running its own perception/planning ML; restructured toward personal-vehicle autonomy.
DevTools · Multi-Cloud · Native (own / self-hosted weights)
Cursor runs a multi-cloud, event-driven architecture with most CPU/backend services on AWS, major GPU inference clusters on Azure plus newer GPU clouds, and specialized AI workloads on GCP, fronted by stateless gateways, queues, and GPU worker fleets for LLM inference and agents.[6][10] Workspace data and embeddings flow through turbopuffer, Pinecone, and Kafka-compatible streams, with Temporal-based orchestration for agent workflows and large multi-GPU clusters (including Fireworks AI-hosted proprietary fine-tuned models) serving code-focused LLMs.[4][6][7][8][10]
Cybersecurity · Hybrid · Native (own / self-hosted weights)
Self-Learning AI built in-house; models the network rather than calling an LLM.
Data Platform · Multi-Cloud · Native (own / self-hosted weights)
Databricks runs a two-layer architecture where a Databricks-managed control plane hosts the UI, APIs, metadata, and orchestration services, while customer workloads execute in a compute/data plane inside the customer’s AWS, Azure, or GCP account (or Databricks’ serverless account) against cloud object storage using Spark and Delta Lake.[3][4][9][11] Around this core, Databricks positions an open-core lakehouse stack (Delta Lake, Spark, MLflow) with proprietary governance, serverless, and AI platform services, integrated into medallion-style production patterns.[1][2][6][12]
AI/Translation · Hybrid · Native (own / self-hosted weights)
European MT leader training its own translation + writing models on its own EU infrastructure.
FoodTech · AWS · Native (own / self-hosted weights)
Food delivery; own dispatch/ETA ML plus OpenAI for support.
Hardware/Enterprise · Hybrid · Native (own / self-hosted weights)
Dell AI Factory packages on-prem open models (Llama) + partners (Cohere) for enterprise/sovereign deployment.
Media · Hybrid · Native (own / self-hosted weights)
Media/parks; large in-house recommendation/VFX ML plus partner LLMs for assistants; cautious GenAI posture.
Search/Data · Multi-Cloud · Native (own / self-hosted weights)
Ships ELSER, its own retrieval model, and plugs partner LLMs for RAG.
Gaming · Hybrid · Native (own / self-hosted weights)
Game publisher; large in-house animation/matchmaking ML plus partner LLMs for tooling.
AI/Voice · Multi-Cloud · Native (own / self-hosted weights)
ElevenLabs exposes proprietary speech, audio, and agent models via HTTPS REST, SSE streaming, and WebSocket APIs behind api.elevenlabs.io, backed by credit-based authentication and multi-tenant SaaS infrastructure designed for high concurrency and enterprise scalability.[1][7][9] Core value is provided by in-house deep learning voice and audio models rather than a foundational LLM as the primary engine.[2][5]
Gaming · Hybrid · Native (own / self-hosted weights)
Unreal Engine + Fortnite; in-house generative/animation ML (MetaHuman) on its own infra.
Biotech · AWS · Native (own / self-hosted weights)
EvolutionaryScale operates ESM3 as a proprietary, GPU-accelerated protein LLM stack on AWS, exposing commercial access primarily via AWS SageMaker, AWS HealthOmics, and soon Amazon Bedrock, while distributing a smaller open version of ESM3 and earlier ESM models with code and some weights on GitHub and Hugging Face for non-commercial and open use.[3][5][7][11] Internal architecture details beyond this cloud/HPC and PyTorch-centric GPU cluster setup are not publicly documented.[3][10][12]
Wholesale · AWS · Native (own / self-hosted weights)
Wholesale marketplace; own recommendation ML plus OpenAI for merchandising assistants.
Robotics · Hybrid · Native (own / self-hosted weights)
Figure’s production architecture is centered on its in-house **Helix** vision-language-action stack running on vertically integrated humanoid hardware and a proprietary BotQ manufacturing/software infrastructure, with robots streaming high-bandwidth data for continuous training and fleet improvement[1][2][7][9]. Core control and reasoning have moved off external LLMs to fully in-house Helix models, with cloud services used as supporting infrastructure rather than as foundational AI engines[7][9].
AI Infra · Multi-Cloud · Native (own / self-hosted weights)
Fast inference for open models on its own serving stack.
Logistics · AWS · Native (own / self-hosted weights)
Freight forwarding; in-house logistics ML plus OpenAI for ops assistants.
E-Commerce (India) · GCP · Native (own / self-hosted weights)
Indian commerce giant (Walmart-owned); own recommendation ML plus Gemini-based shopping assistants.
FinTech (Africa) · AWS · Native (own / self-hosted weights)
Pan-African payments; in-house fraud ML plus OpenAI for support.
SaaS (India) · AWS · Native (own / self-hosted weights)
CX/ITSM suite; Freddy AI blends its own models with OpenAI over customer data.
IT (Japan) · Hybrid · Native (own / self-hosted weights)
Builds Takane, its own enterprise Japanese LLM (with Cohere), for regulated/government use.
Biotech · AWS · Native (own / self-hosted weights)
Cell-programming foundry training its own biological/sequence models over massive lab data.
Super-app · AWS · Native (own / self-hosted weights)
SEA super-app; large in-house logistics/fraud ML plus OpenAI for assistants.
Productivity · AWS · Native (own / self-hosted weights)
Long-standing in-house NLP models plus GPT for generative rewriting.
AI Hardware · On-Prem · Native (own / self-hosted weights)
Groq builds and operates its own on-premise inference infrastructure around custom Language Processing Units (LPUs) and GroqChip hardware, exposing this via stateless, multi-region HTTP/gRPC APIs used by customers’ gateways, queues, and orchestrators.[1][3][6][7] The production patterns described in public materials place Groq as a dedicated low-latency inference backend behind customer-managed proxies, Redis queues, and Kubernetes-based orchestrators, rather than as a full cloud stack provider.[1][3][4][6]
Healthcare · AWS · Native (own / self-hosted weights)
Hippocratic AI’s production architecture centers on the Polaris constellation: a core healthcare conversation LLM surrounded by 30+ specialized supervisor, verifier, and tool-call models, many built on open-source foundations but trained and operated as proprietary healthcare agents in AWS.[1][3][5][8] The system uses online and offline LLM judges plus task-specific engines (e.g., overdose, labs/vitals, benefits, scheduling) to monitor and govern voice interactions and workflow actions at clinical scale.[1][4][8]
AI/LLM · AWS · Native (own / self-hosted weights)
The model/dataset hub on AWS, storing weights as Git LFS repos over S3 and running Inference Endpoints/Spaces. Stewards open libraries (Transformers, Diffusers) — the open-source center of gravity for ML.
AI/LLM · Azure · Native (own / self-hosted weights)
Trains its own Inflection models (now enterprise-focused) on Azure supercompute.
Biotech · Multi-Cloud · Native (own / self-hosted weights)
Generative-chemistry drug discovery on its own target-discovery + molecule-generation models.
FinTech · AWS · Native (own / self-hosted weights)
Intuit Assist runs on its own GenOS platform with partner LLMs underneath.
Biotech · GCP · Native (own / self-hosted weights)
Public materials indicate Isomorphic Labs runs its production AI/drug-design engine on Google Cloud, using Google Cloud AI Hypercomputer with TPU/GPU fleets orchestrated by Google Kubernetes Engine for autoscaling inference workloads. The company publicly emphasizes its own drug-design engine and AlphaFold-derived models rather than exposing a foundational LLM-centric stack.
AgTech · AWS · Native (own / self-hosted weights)
Precision-ag + autonomy on large in-house computer-vision/ML models running on equipment edge.
Marketing · AWS · Native (own / self-hosted weights)
Marketing automation; own predictive ML plus OpenAI for content over its customer-data platform.
Electronics (Korea) · On-Prem · Native (own / self-hosted weights)
Trains EXAONE, its own foundation models, for products/enterprise on its own infrastructure.
AI Infra · Multi-Cloud · Native (own / self-hosted weights)
Builds PyTorch Lightning + a studio to train/serve open models (Llama, etc.) on cloud GPUs.
AI/LLM · Hybrid · Native (own / self-hosted weights)
Liquid AI builds and serves its own Liquid Foundation Models (LFM and LFM2) with a custom hybrid liquid/convolution/attention architecture, optimized for both data-center and fully on-device deployment across CPUs, GPUs, and NPUs.[1][3][4][9][11] Public materials emphasize hardware-in-the-loop training and edge/PC deployment, but do not expose a full production cloud stack, suggesting a mix of self-hosted/model-serving infrastructure plus OEM/partner integrations rather than a single public hyperscaler.[9][11]
Retail · GCP · Native (own / self-hosted weights)
Home-improvement retailer on Google Cloud; Gemini-based associate/customer assistants + own ML.
AI-coding · GCP · Native (own / self-hosted weights)
Magic’s public evidence points to a proprietary, long-context model stack centered on its own LTM models rather than a packaged third-party LLM. The clearest infrastructure signal is a Google Cloud partnership for AI supercomputers, but the public record does not expose a full production data stack or compliance posture.
E-Commerce · AWS · Native (own / self-hosted weights)
LatAm commerce/fintech giant; large in-house recommendation/fraud ML plus OpenAI for assistants.
Automotive · Multi-Cloud · Native (own / self-hosted weights)
In-car MBUX assistant on Google Cloud Gemini; large in-house ADAS/vehicle ML.
Platform · Azure · Native (own / self-hosted weights)
Microsoft operates primarily as a cloud-first organization running its internal and external services on Azure, backed by globally distributed, Microsoft-operated datacenters and a hybrid-cloud capable architecture.[1][4][6] Core business, productivity, and analytics workloads are delivered via Azure IaaS/PaaS (VMs, app services, databases) plus newer unified analytics platforms like Microsoft Fabric/OneLake, integrated with shared Microsoft services such as Microsoft 365 and identity.[1][5][7]
Image AI · Hybrid · Native (own / self-hosted weights)
Midjourney runs a proprietary latent diffusion image model pipeline behind a Discord bot and web UI, with prompts flowing through a request-handling layer into GPU clusters for model inference, then storing generated images in large-scale object storage fronted by a CDN.[5][9] Public information focuses on model architecture and system-design patterns; specific cloud vendor, database products, and formal compliance regimes are not disclosed.[1][5][9]
AI/LLM · Multi-Cloud · Native (own / self-hosted weights)
Mistral runs a multi-cloud architecture where its hosted La Plateforme and Studio offerings sit atop partner clouds (Google Cloud, AWS, Azure, SAP, IBM, Snowflake, NVIDIA, Outscale), while open-weight models are also deployable on‑prem via standard inference containers.[10][11] The production stack is therefore split between first‑party EU‑centric hosting for regulated workloads and distribution of models through major cloud marketplaces and managed services.
Internet · Hybrid · Native (own / self-hosted weights)
NAVER runs its core portal and AI services primarily on **on-premise data centers** such as the GAK Sejong facility, increasingly augmented by NAVER Cloud’s sovereign AI infrastructure built with NVIDIA DSX and AMD EPYC/Instinct platforms for large-scale training and inference.[2][4][5][6][7] The production stack is vertically integrated—NAVER controls the data centers, cloud layer (NAVER Cloud), and proprietary foundation models like HyperCLOVA X, exposed as services to internal products and external partners.[2][5][7]
AI/LLM · Hybrid · Native (own / self-hosted weights)
OpenAI runs large GPU superclusters and Kubernetes-based orchestration across Azure and its own data centers in a hybrid setup, with Azure as the primary cloud and custom HPC clusters (e.g., Stargate) for training and serving frontier models.[15][16][19] Core application and API workloads use relational stores like PostgreSQL for accounts/settings and globally scalable databases such as Azure Cosmos DB plus Kafka streams for high-volume conversation, analytics, and event data.[1][8][12]
Healthcare AI · Multi-Cloud · Native (own / self-hosted weights)
Medical-evidence answer engine on its own models trained over peer-reviewed literature.
Telecom (EU) · Hybrid · Native (own / self-hosted weights)
European telecom partnering with Mistral for sovereign LLMs over its own network ML.
Cybersecurity · Multi-Cloud · Native (own / self-hosted weights)
Palo Alto Networks runs a multi-cloud SaaS architecture with AWS as the primary runtime for major platforms like Prisma Cloud, using services such as RDS, Redshift, Neptune, OpenSearch, Redis, and S3 alongside Kubernetes-based microservices for data ingestion, analytics, and Infinity Graph. GCP (including Vertex AI) and Azure are used selectively for AI/ML workloads and cloud-native firewall offerings, with proprietary security engines and data stores layered on top of these managed services.
AI/Search · Multi-Cloud · Native (own / self-hosted weights)
Perplexity operates a multi-cloud architecture: it historically built and hosted its Sonar models on AWS (including Bedrock for Anthropic/Claude) while later signing a large Azure agreement to deploy models via Microsoft Foundry, indicating active use of both providers.[1][3][8][9][15] Its core production stack centers on proprietary Sonar LLMs (built on and fine‑tuned from open-source bases like Mistral and later Llama 3.x) combined with partner frontier models (GPT‑5/4-class, Claude, Gemini, Grok) selected per query.[1][2][5][6][10][11][13][14]
Robotics · Hybrid · Native (own / self-hosted weights)
Physical Intelligence runs a proprietary robotics foundation model stack with large-scale training and data infrastructure, using WEKA for high-performance data and Oracle Cloud Infrastructure for hosted compute, complemented by Kubernetes-based CI/CD and GPU-heavy cloud environments.[1][2][7][11][12] The architecture centers on Python/C++ robotics services, PyTorch/JAX-based model training on cloud GPUs, and data pipelines from robot teleoperation and simulation into foundation model training and evaluation.[11][12]
Social · AWS · Native (own / self-hosted weights)
Built Navigator-1, its own multimodal foundation model (fine-tuned from open-source bases incl. Qwen) powering the Pinterest Assistant at ~10% the cost of frontier models; big in-house recommendation ML.
AI-coding · AWS · Native (own / self-hosted weights)
Poolside builds and trains proprietary foundation models via its internal **Model Factory** (Titan training stack on a ~10K GPU cluster, Apache Iceberg data layer, Kubernetes orchestration) and then deploys those models and agentic systems fully inside customer VPCs/on‑prem through the Poolside Platform, with additional distribution via AWS Bedrock and Trainium-backed inference.[1][2][7][8][9][5][6][11] The production architecture is therefore a proprietary, Kubernetes-based ML/agent stack running on AWS (and customer infrastructure) with an Iceberg-centered data plane and tightly integrated model training, evaluation, and deployment pipelines.[1][2][7][8][9][11]
E-Commerce (Japan) · Hybrid · Native (own / self-hosted weights)
Japanese commerce/fintech ecosystem training its own Japanese-first Rakuten AI models.
FinTech (India) · AWS · Native (own / self-hosted weights)
Indian payments; in-house fraud/risk ML plus OpenAI for support assistants.
Banking · On-Prem · Native (own / self-hosted weights)
Canadian bank running large in-house risk ML; Cohere models deployed in-VPC for sovereign generative AI (Borealis AI).
Biotech · GCP · Native (own / self-hosted weights)
Drug discovery on its own phenomics foundation models trained over petabytes of cellular imaging.
AI/LLM · Multi-Cloud · Native (own / self-hosted weights)
Reka trains and serves its own multimodal encoder–decoder frontier models (Core, Flash, Edge) on custom Kubernetes-based GPU clusters spanning multiple vendors, using PyTorch on large H100/A100 fleets and a separate A10/A100 inference stack.[2][5] Public deployment is via Reka’s own web app and API endpoints (chat.reka.ai, platform.reka.ai, showcase.reka.ai), with an OpenAI-compatible API server for Edge provided through vLLM and Hugging Face artifacts.[2][7][12]
Telecom (India) · On-Prem · Native (own / self-hosted weights)
Telecom/digital giant building sovereign Indian-language models (Jio Brain / BharatGPT) on its own infrastructure.
AI Infra · GCP · Native (own / self-hosted weights)
Runs open models as one-click APIs on its own GPU fleet.
Gaming · Hybrid · Native (own / self-hosted weights)
League/Valorant maker; in-house anti-cheat + matchmaking ML on its own infra; off the foundational-LLM map at the core.
Automotive · AWS · Native (own / self-hosted weights)
EV maker training its own driver-assist/vehicle ML; off the foundational-LLM map at the core of the product.
Gaming · Hybrid · Native (own / self-hosted weights)
UGC gaming platform on its own data centers; trains its own generative models for 3D/code creation assistants.
Video AI · AWS · Native (own / self-hosted weights)
Runway’s public product and developer docs show a proprietary, cloud-hosted AI platform with first-party APIs and an internal model catalog rather than exposed third-party foundational model dependencies. The available evidence supports AWS as the most likely cloud footprint, but the exact database technologies and compliance attestations are not publicly confirmed in the sources provided.
AI-hardware · On-Prem · Native (own / self-hosted weights)
Own RDU AI chips plus its Samba-1 model and hosted open models for enterprise/government.
Enterprise · Hybrid · Native (own / self-hosted weights)
Joule is multi-model: SAP’s own ABAP LLM plus a Generative AI Hub fronting OpenAI, Gemini, Anthropic (Claude powers Joule agents), Llama, and sovereign Mistral/Cohere on SAP cloud.
AI/Data · AWS · Native (own / self-hosted weights)
Beyond data-labeling, Scale trains its own Defense Llama (fine-tuned Llama 3) for U.S. national security on its Donovan federal platform; evaluates frontier models for the DoD CDAO.
Cybersecurity · AWS · Native (own / self-hosted weights)
Purple AI combines SentinelOne’s own Ultraviolet models with Anthropic Claude + OpenAI GPT over its Singularity data lake; exposes a Purple MCP server for agentic SOC investigation.
Enterprise · Multi-Cloud · Native (own / self-hosted weights)
ServiceNow runs a proprietary SaaS platform across hyperscaler regions on AWS, Azure, and GCP, plus sovereign-cloud deployments, with a custom database layer (RaptorDB on PostgreSQL) replacing MariaDB/MySQL for most instances.[7][8] Its architecture is tightly integrated with NVIDIA NeMo microservices and GPU infrastructure for training and inference of its Apriel/Nemotron family of open models, while exposing workflow, observability, and discovery services over these multi-cloud foundations.[2][4][5][8]
Defense · Hybrid · Native (own / self-hosted weights)
Builds Hivemind, its own autonomy foundation for GPS/comms-denied flight.
Industrial · Azure · Native (own / self-hosted weights)
Industrial Copilot built with Microsoft on Azure OpenAI over its own engineering/automation ML.
Data Platform · Multi-Cloud · Native (own / self-hosted weights)
Snowflake runs a proprietary three-layer, multi-cluster shared-data architecture on AWS, Azure, and GCP, separating compressed columnar storage on cloud object stores from MPP compute warehouses and a distributed cloud services control plane.[1][16][19] Internally it is a service-oriented system with independently scalable storage, compute, and metadata/transaction services built for OLAP workloads.[16][19]
Cybersecurity · GCP · Native (own / self-hosted weights)
Developer-security platform on GCP/AWS scanning code and dependencies; DeepCode AI combines symbolic analysis with ML for fix suggestions. Backed by a curated open-source vulnerability database.
Conglomerate (Japan) · Hybrid · Native (own / self-hosted weights)
Telecom/investment group building its own Japanese LLMs (Sarashina) plus partner models; major AI-infra investor.
Automotive (EU) · Multi-Cloud · Native (own / self-hosted weights)
Automaker (Jeep/Peugeot/Fiat) embedding Mistral in-vehicle plus its own ADAS ML.
Healthcare AI · Multi-Cloud · Native (own / self-hosted weights)
Ambient clinical assistant blending its own medical models with partner LLMs.
Audio AI · AWS · Native (own / self-hosted weights)
Trains + serves its own music-generation models.
AI-coding · Multi-Cloud · Native (own / self-hosted weights)
Privacy-first code assistant with its own models plus optional partner LLMs, deployable air-gapped.
Retail · GCP · Native (own / self-hosted weights)
Retailer on Google Cloud; store-companion + guest AI on Gemini over its own merchandising ML.
Healthcare · GCP · Native (own / self-hosted weights)
Precision-medicine platform with its own multimodal genomic/clinical models plus LLMs for clinician copilots.
Internet · On-Prem · Native (own / self-hosted weights)
Tencent’s production stack is hybrid in practice, combining large-scale self-hosted infrastructure (notably OpenStack/TStack and Tencent Cloud-native/TCE/TCS) with major open-source infrastructure components such as Apache Kafka. Its database layer is centered on Tencent-built systems like TDSQL/TDSQL Boundless, with enterprise controls spanning regional deployment, access control, encryption, and auditing.
Automotive · Hybrid · Native (own / self-hosted weights)
Tesla runs a vertically integrated, edge-first architecture where custom FSD hardware in vehicles and robots performs real-time inference, while a hybrid cloud (including AWS and in-house GPU clusters) ingests telemetry into Kafka/Flink streams, S3/LakeFS lakes, and Snowflake/BigQuery warehouses for continuous model training and OTA updates.[1][7][8][10] Production for cars, energy products, and robots is tightly coupled to this AI and data infrastructure, effectively making Gigafactories and AI clusters part of a single closed-loop manufacturing and autonomy system.[2][6][7]
AdTech · AWS · Native (own / self-hosted weights)
Programmatic ad DSP; Kokai is its own large-scale bidding/forecasting ML over trillions of ad-auction events.
Analytics · Multi-Cloud · Native (own / self-hosted weights)
Search/AI analytics; Spotter agent blends its own engine with OpenAI over the in-memory data store.
AI Infra · Hybrid · Native (own / self-hosted weights)
Together AI operates its own GPU cloud and managed Kubernetes clusters as an AI-native infrastructure layer, while also integrating with external cloud storage and streaming services for data and analytics workloads.[2][11][12] The production stack is centered on open-weight models deployed as containerized workloads on Together-managed GPU clusters, exposed via serverless inference APIs and voice/agent pipelines.[3][4][9][11][16]
Mapping · Azure · Native (own / self-hosted weights)
Mapping/location; own routing/map ML plus a Microsoft-partnered conversational in-car assistant on Azure OpenAI.
Gaming/3D · Multi-Cloud · Native (own / self-hosted weights)
Real-time 3D engine; Muse AI blends partner LLMs with its own generative tooling for creators.
Telecom · Hybrid · Native (own / self-hosted weights)
Telecom running its own network ML; customer/agent assistants on Google Cloud Gemini.
E-Commerce · GCP · Native (own / self-hosted weights)
Home-goods retailer on Google Cloud; in-house recommendation/visual ML plus Gemini (Vertex) for shopping assistants.
Autonomous · GCP · Native (own / self-hosted weights)
Waymo’s production driver uses a modular six-layer stack (sensor processing, perception, world modeling, prediction, planning, control) integrated with HD maps, simulation, and a foundation-model-based end-to-end driving architecture that fuses camera, lidar, and radar into a unified world representation.[1][2][8][9][10] Training, evaluation, and large-scale data handling run on Google/Alphabet’s internal and Google Cloud infrastructure, with closed-source models and tooling.[1][4][5][8][10]
Autonomous · Azure · Native (own / self-hosted weights)
Wayve runs a proprietary end-to-end embodied driving stack on Microsoft Azure, using large vision-world models (AI Driver, GAIA-3) trained on petabyte-scale fleet and simulated data via Azure Kubernetes Service, Azure Machine Learning, Databricks, and GPU clusters, then deploys a compressed model on in-vehicle compute for production driving.[7][9][12] The architecture replaces modular perception/mapping/planning with a single neural policy and related foundation models (driving, simulation, scenario classification, language explanation), all controlled via internal MLOps pipelines on Azure.[1][4][9][12]
HR · Hybrid · Native (own / self-hosted weights)
Runs Illuminate, its own ~800B-parameter LLM trained on HR/finance data from 10,500+ customers (800B transactions/yr) for privacy + enterprise fit; acquired Sana (2025) for custom AI agents.
Enterprise AI · AWS · Native (own / self-hosted weights)
Writer operates as a multi-tenant SaaS platform primarily hosted on AWS, with options for customers to deploy Writer’s self-hosted LLM on-premises or in their own cloud/VPC for stricter data residency and control.[9][16] The core stack combines enterprise-grade proprietary LLMs, a knowledge graph over customer data sources, and agent orchestration tooling (AI HQ / Agent Builder) exposed via web UI and APIs.[10][16][19]
AI/LLM · On-Prem · Native (own / self-hosted weights)
The provided sources do not expose xAI’s actual production infrastructure, and most search results are generic XAI architecture guidance rather than xAI company disclosures. The only xAI-specific material in the results points to an internal product architecture (Home Mixer, Thunder, Phoenix, Grox) but does not establish cloud, database, or compliance stack details for xAI’s core production environment.[9][5]
Local/Internet · AWS · Native (own / self-hosted weights)
Local-discovery platform; own ranking/review ML plus OpenAI for its AI assistant.
Proptech · AWS · Native (own / self-hosted weights)
Real-estate marketplace; the Zestimate is large in-house valuation ML; natural-language search on OpenAI.
SaaS (India) · On-Prem · Native (own / self-hosted weights)
Bootstrapped SaaS suite on its own Indian data centers; builds its own Zia AI models for privacy + cost control.
FoodTech (India) · AWS · Native (own / self-hosted weights)
Indian food-delivery; own logistics/recommendation ML plus OpenAI assistants.
Communications · Hybrid · Native (own / self-hosted weights)
Runs its own global data centers for real-time media, bursting to AWS/Oracle Cloud at peak. AI Companion uses a 'federated' approach blending its own models with Anthropic and OpenAI.