Choose an AI application, trace its documented model, infrastructure, chip, energy, and water relationships, and see exactly where the public evidence runs out.
Pick an AI product and follow its documented suppliers toward chips, campuses, and electricity. The route stops wherever the public evidence stops.
The AMICE stack is Applications, Models, Infrastructure, Chips, and Energy. The essay explains why those layers belong in one mental model. This page is the workbench: start with a product you recognize, compare its possible routes, or enter the system through a company, facility, ownership relationship, or unresolved question.
Everything here represents public evidence, not live telemetry. A selectable model is not necessarily a product's default. A cloud agreement does not reveal the data center that served a particular request. A regional generation mix is context, not a claim about the electrons entering one building. The explorer labels missing disclosures of routing shares, facility assignments, accelerator generations, energy sources, and water use.
Search for an application you use, select it, and inspect one connection at a time. Open the supporting source before following the next line. Then add a second product to see which suppliers they share. The facility map is a separate view of disclosed places; it does not locate your request.
For a useful check, compare the kind of claim you find here with Epoch AI's Chip Users explorer, published September 9, 2026. Epoch estimates developers' compute use with uncertainty ranges; AMICE traces documented relationships. An estimate of a lab's total compute does not fill in a missing application-to-campus connection.
The editorial review on September 17, 2026 adds this reading guidance and further sources. The product and facility counts below still refer to the September 4 research batch.
The application catalog exposes every product with a trace in the current dataset. Search by product, company, or category, then pick up to four applications. Shared suppliers merge and products with several documented routes fan out. Current, historical, announced, optional, and user-selectable relationships remain distinct so an old integration does not quietly become a claim about today's default.
The September 4, 2026 pass traced 35 more products and brought the catalog to 163. Most of the new ones are consumer apps, such as Meta AI inside WhatsApp, Gemini for Home, Fitbit's coach, Klarna's assistant, and the BMW voice assistant, plus a wave of work and coding tools like Google Workspace with Gemini, Devin, Manus, Bolt.new, Otter, and Granola. Those are the routes where the public record usually ends at a provider's name and a privacy page. Some still finish at an undisclosed request-level chip or electricity assignment. The September 18 documentation review adds provider hardware history, training systems, serving deployments, regional availability, and electricity records beside those gaps. Select a product to see relevant hardware evidence, then use the documentation filters below the relationship table.
Start with one product, then add up to three more. Shared suppliers rejoin; routers fan out; gaps stay visible instead of becoming guesses.
Showing all 163 matching mapped applications.
Your comparison
Evidence date
Relationships supported on the review date. Every line is unweighted because request shares are almost never disclosed.
Application node
OpenAI publishes the current ChatGPT model lineup. Its compute portfolio is documented, but a conversation cannot be assigned to a cloud or campus publicly.
Water overlay
The LBNL/DOE estimate for all US data centers in 2023 was 66 billion liters of direct water consumption and nearly 800 billion liters of modeled indirect grid water. It is not AI-only, and the grid estimate omits PPAs and behind-the-meter generation. A low onsite WUE can also trade water for more electricity.
Read the LBNL methodology7 of 7 relationships
204 additional sources reviewed September 18, 2026. Each record names the workload or geography it supports. Training records describe training; availability and development plans have their own labels.
Your location can suggest a region to research. Endpoint configuration, residency rules, capacity, and cross-region routing determine where a request can run. Nearby data centers are not evidence of your application's serving location.
9 of 204 sources · showing 1-9
The connecting lines are deliberately equal-width. Providers almost never publish what share of requests follows each model, cloud, facility, or chip path, so a Sankey-style width would imply precision the evidence does not contain.
Lovable makes the difficulty unusually concrete. Its own engineering record describes an automatic provider router spanning Anthropic, Google Vertex AI, and Amazon Bedrock. Its agent can also split one user prompt into several model calls. The explorer below separates what is documented from what is merely possible, then shows the regional fuel context that becomes available only if a route or location is independently proven.
One prompt, several possible routes
The honest answer today is: public evidence cannot identify it. Lovable documents automatic balancing across three provider surfaces, and one prompt can launch several model calls. This explorer follows the closest defensible chain and labels every inferred hop.
Documented
Anthropic, Vertex AI, and Bedrock sit behind Lovable’s automatic provider routing.
Multi-call
A main agent can delegate to multiple subagents and models, so one prompt is not one server event.
Still private
Per-call provider, model, region, building, chip, power meter, and hourly generator.
Documented current provider candidates · select a branch
Application event
Your Lovable prompt
May create context, main-agent, subagent, retry, and fallback calls.
One possible child call
Anthropic candidate
This provider is in the documented automatic router; public evidence does not say this prompt selected it.
Region and facility
Not disclosed
Multi-cloud, multi-country service boundary; exact cloud, region, availability zone, building, and grid are unknown
Fuel burned
Unknown for this prompt
A provider relationship cannot identify an hourly generator or fuel. Conditional regional proxies appear only in the evidence cards below.
Known service boundary
Lovable-managed editor agent to Anthropic’s commercial model service
Supported claims
Unknowns
Anthropic portfolio, subprocessor, model-availability, and cloud-partnership records are not evidence that this Lovable prompt used any named provider region or physical facility.
Premises
Conclusion
The Anthropic service boundary is a defensible automatic route candidate, but no physical data center or fuel can be assigned to a Lovable prompt from public evidence.
Rationale
The evidence closes the application-to-provider edge while leaving every lower infrastructure-allocation edge open.
Other explanations
Evidence that would prove or disprove it
These are deliberately conditional, historical, or counterfactual—not claims about the prompt above.
Current evidence boundary
Research batch: 2026-08-regional-power-01
Answerable: Lovable uses Claude and Anthropic has operational Rainier capacity. Not answerable: the request’s model step, cloud route, facility, meter, generator, hourly fuel, or physical share.
Public evidence confirms Claude is central to Lovable, but it cannot identify the fuel burned for a particular prompt. Rainier is genuine operational Anthropic capacity and St. Joseph County is one confirmed component; neither fact proves that a Lovable request reached Rainier or Indiana. Only under that unproven condition would PJM 2025 be a broad regional proxy.
Boundary: Approximate marker for the multi-state PJM footprint, not Northern Virginia, Indiana, a cloud availability zone, or an individual generator.
PJM RTO footprint · 2025
Combustion-classified categories make up 60.7% of this annual regional record. This is not a measurement of a facility or prompt.
Boundary: Multi-state annual gross generation. It is broader than the AEP or Dominion zone, does not follow imports or contract attributes, and cannot locate an AWS availability-zone building or identify fuel for a prompt or incident interval.
Evidence state: confidence inferred · allocation known: no · current
Validity: as of 2026-09-04; from unknown (unknown) through unknown (unknown); basis retrieved
Geography and scope: Conditional candidate facilities: aws-rainier-st-joseph-indiana · Conditional application-to-provider-to-facility-or-region evidence; no request, traffic-share, facility, chip, meter, generator, fuel, or water allocation.
Premises
Conclusion
Rainier/St. Joseph is a provider-capacity candidate, not a Lovable request assignment; PJM’s mix is therefore conditional context rather than an answer to “what powered my prompt?”
Rationale
Lovable’s Claude evidence and AWS’s operational Rainier evidence establish a plausible provider-capacity path, but the missing request-to-cluster and cluster-to-site mappings make Rainier a candidate and PJM only a conditional regional proxy.
Other explanations
Evidence that would prove or disprove it
The route above follows documented application, model, cloud, chip, and energy relationships. The map below adds a different entry point: disclosed AI campuses with approximate locations, project status, scoped workload evidence, capacity claims, electric utility and grid context, and water or cooling disclosures. A marker identifies a researched place; it does not claim that a particular prompt ran there or that the named grid supplied a specific hour.
The same September pass added 29 campuses and brought the atlas to 71 sites in 18 countries. Microsoft's Fairwater site outside Atlanta, the OpenAI lease at the old Portsmouth uranium plant in Ohio, Google's Texas Panhandle sites, the Nscale and Microsoft buildout at Sines in Portugal, Google's Visakhapatnam hub in India, and SoftBank's Sakai plant conversion in Japan are all in. So are the sites where the record stops early. A 191 megawatt lease to an unnamed frontier lab stays unnamed here, and a Saudi cluster with no published city stays out of the canonical list. Every site now has an electricity context. A campus linked to a named grid or utility in public records carries that operator's dated annual mix as a proxy. Sites without that evidence, including the Mississippi, El Paso, and Texas Panhandle sites, have no assigned mix. A regional mix remains a proxy for the wires a building sits on, rather than a measurement of the power it drew in any hour.
The stack touches the ground
Select a marker or named site. These are approximate, evidence-backed campus locations—not a claim that any particular prompt ran there. Workload, utility, grid, capacity, and water fields retain their public evidence boundaries.
Approximate coordinates protect evidence boundaries; utility or grid identity does not establish the generators serving a particular hour.
Open full facility dossiersThe product trace answers one question well: what documented relationships could sit underneath an application? The evidence atlas provides other entry points. It separates mobile-usage rankings from App Store charts, audits private-company valuation cohorts instead of filling quotas with guesses, compares control across the largest AI players, distinguishes operating facilities from construction and announcements, and keeps parent companies, investors, joint ventures, suppliers, and utilities from collapsing into one vague idea of ownership.
Start with the app in your hand, a startup headline, a platform company, a place on the map, a corporate parent, or an unanswered question. Every view points back to dated evidence and keeps control, contracts, capacity, and physical supply distinct.
Products people use: Start with the ranked mobile cohort or image and video tools, then open any product in the AMICE trace above.
Mapped application census
Browse all 163 applications with a mapped evidence trail, isolate image and video tools, or inspect the 30-product consumer cohort. Apple does not publish an “AI” App Store category, so that cohort freezes a published global mobile ranking by monthly active users and separately overlays Apple’s U.S. overall Top Free chart. Those ranks are not interchangeable.
163
all mapped
30/30
mobile cohort
6
Apple Top 100
30
creative traces
Inderjeet Singh / devhook; AIBY identity disputed · Creative
Current + earlier routes
Apple currently identifies Inderjeet Singh and devhook, while still-linked 2023 legal text names AIBY Inc. The conflict is unresolved; neither identity evidence nor Amazon contact-data storage identifies the app’s image-generation provider.
1
documented current links
5
unresolved current links
Adobe · Creative
Current route
Firefly publishes a surface-specific Adobe and partner-model catalog. Ordinary partner use is visibly selected, while the first video generation can begin on Veo 3.1 instead of the ordinary Firefly Video default.
13
documented current links
3
unresolved current links
Adobe · Creative
Current route
Lightroom now publishes feature-specific Firefly, Veo, and Topaz mappings while leaving other GPU-assisted tools architecturally incomplete. Send to Firefly is an explicit web-product handoff, not an invisible Lightroom backend.
5
documented current links
7
unresolved current links
Aider · Coding
Current route
Aider is a local client with explicit model/provider selection. Cloud providers, gateways, compatible endpoints, and same-device runtimes remain distinct physical paths.
5
documented current links
3
unresolved current links
Airtable · Work
Current route
Airtable-hosted orchestration exposes different provider sets for Omni, field agents, and generated interfaces. Live documentation conflicts in provider wording, and data residency is not model-serving geography.
7
documented current links
6
unresolved current links
Amazon · Consumer
Current + earlier routes
Rufus was renamed Alexa for Shopping in May 2026. Its current Bedrock router can choose Claude Sonnet, Amazon Nova, or a custom shopping model; the earlier 80,000-chip Prime Day disclosure is aggregate and historical.
6
documented current links
2
unresolved current links
Amazon · Consumer
Current + earlier routes
Alexa+ routes tasks across Amazon Nova and Anthropic Claude on Bedrock. General availability, early access, model routing, and lower-stack allocation remain separate claims.
4
documented current links
2
unresolved current links
Amazon Web Services · Coding
Current route
Kiro exposes Auto and selectable model families through Amazon Bedrock cross-region inference. Account geography constrains an eligible region set, but storage geography, logical inference region, physical facility, accelerator, and request allocation remain separate.
4
documented current links
4
unresolved current links
Amazon Web Services · Coding
Current route
Amazon Q Developer exposes supported Claude choices in IDE chat and uses Bedrock cross-region inference within a documented geography for eligible existing users. New signups stopped May 15, 2026, and support ends April 30, 2027. Eligible and service regions are not observations of a request region.
4
documented current links
4
unresolved current links
Apple · Consumer
Current + planned routes
Three paths coexist: production features use Apple’s 2025 on-device/PCC models; the future Siri AI beta uses Apple-custom AFM 3 models built with Google from Gemini technology; and ChatGPT remains a separate optional extension for Siri, Writing Tools, and visual intelligence.
6
documented current links
4
unresolved current links
Asana · Work
Current route
Asana AI Studio has a documented current default, a selectable model catalog, and an admin-dependent fallback when providers are disabled. Smart Chat and workflow steps can have feature-specific routes, while connected data sources are context—not model hosting.
6
documented current links
6
unresolved current links
Atlassian · Work
Current route
Rovo uses a portfolio of Atlassian-hosted open models and third-party OpenAI, Anthropic, and Google models. Public documentation establishes possible provider branches but not feature defaults, per-request selection, cloud region, facility, chip, or energy.
4
documented current links
4
unresolved current links
Augment Code · Coding
Current route
Augment Code offers direct model selection, organization defaults, automatic Prism routes, and feature-specific choices such as Code Review. Its Context Engine is a separate retrieval stage and must not be used to infer the downstream generative provider or serving hardware.
8
documented current links
8
unresolved current links
BMW Group · Consumer
Current route
BMW’s assistant is built on Amazon’s Alexa+ architecture with an LLM and reached existing iX3 vehicles via software updates from the end of May 2026 in Germany and the United States. BMW does not disclose the model routing, region, or hardware behind Alexa+.
2
documented current links
2
unresolved current links
Baidu · Search
Current route
Baidu’s China-store listing explicitly names ERNIE for multimodal search and assistant features. The specific checkpoint and serving stack remain private.
1
documented current links
3
unresolved current links
Base44 (Wix) · Coding
Current route
Base44 defaults to Auto model selection with a Builder-plan picker across Base 1, Sonnet 5, Opus 5, GPT-5.6 Terra and Sol, and Fable 5.1, names Anthropic, Google, and OpenAI as providers, and runs Superagents on GLM 5.2 by default. Base44 Payments is powered by Wix. Servers are in the United States with optional EU or UK residency.
8
documented current links
8
unresolved current links
Be My Eyes · Consumer
Current route
Be My AI launched on the multimodal features of OpenAI’s GPT-4, and Be My Eyes’ 2026 posts still reference that work without naming the model currently in service. No region, facility, or accelerator is disclosed.
2
documented current links
2
unresolved current links
StackBlitz · Coding
Current + earlier routes
Bolt's release notes document Claude Agent as the default since September 2025, Sonnet 4.6 as default from April 2026, Opus 4.7 availability, and a May 2026 switch to Standard and Max agents with Bolt choosing the model behind each. The legacy v1 Agent was retired on August 3, 2026. The browser runs the WebContainers development environment; inference is remote.
4
documented current links
2
unresolved current links
Booking.com · Consumer
Current route
The AI Trip Planner launched as a Booking.com ML feature partially powered by OpenAI’s ChatGPT API, and in 2026 Booking.com also ships an app inside ChatGPT. Neither source names the current checkpoint, region, or accelerator.
4
documented current links
4
unresolved current links
Box · Work
Current route
Box AI first enforces permissions and retrieves relevant content, then invokes a selected model path such as Azure OpenAI or Gemini. Retrieval/storage geography is not the model-serving geography, and example providers do not expose request allocation.
6
documented current links
4
unresolved current links
Brave · Assistants
Current route
Leo exposes Claude and many open model families with manual or automatic selection. Brave says it directly hosts Anthropic models used in Leo.
2
documented current links
4
unresolved current links
Brave · Search
Current route
Brave renamed Answer with AI to AI Answers and added the deeper Ask Brave surface. Qwen3 is named for Ask Brave, not necessarily AI Answers; current hosting, region, accelerator, and request shares remain private.
1
documented current links
7
unresolved current links
Canva · Creative
Current + earlier routes
Canva’s proprietary Design Model, Leonardo-powered Dream Lab, dated Claude and Veo feature routes, and opaque Canva AI 2.0 orchestration are distinct. Ownership and a subprocessor listing are not runtime allocation.
2
documented current links
6
unresolved current links
ByteDance · Creative
Current + earlier routes
CapCut’s June 29 guide still described Seedance 2.5 as not launched; its mutable AI-video page was revised August 10 to describe global web, desktop, and phone availability. Dreamina remains a separate product profile.
4
documented current links
3
unresolved current links
Character.AI · Assistants
Current + earlier routes
Character.AI moved from older proprietary models toward fine-tuned open-source models and later named PipSqueak2/DeepSqueak work. The open bases, current per-chat selection, provider, region, facility, chip, and shares remain undisclosed.
1
documented current links
3
unresolved current links
OpenAI · Assistants
Current + earlier + planned routes
OpenAI publishes the current ChatGPT model lineup. Its compute portfolio is documented, but a conversation cannot be assigned to a cloud or campus publicly.
6
documented current links
1
unresolved current links
Anthropic · Assistants
Current + earlier + planned routes
Claude’s current model is public; Anthropic also discloses a diversified compute portfolio led by AWS, with Google TPU and Azure/NVIDIA capacity.
5
documented current links
2
unresolved current links
Anthropic · Coding
Current route
Claude Code has account- and provider-specific model defaults plus selectable and fallback behavior. Direct Anthropic, Bedrock, Vertex AI, and Foundry are documented paths; an unobserved session cannot be assigned to one cloud, region, cluster, chip, or fuel.
9
documented current links
7
unresolved current links
ClickUp · Work
Current route
ClickUp Brain defaults chat to Max, can expose ChatGPT, Claude, or Gemini choices, and lets Super Agents carry configured models and tools. ClickUp core-service AWS hosting and its provider portfolio do not allocate third-party model inference.
6
documented current links
6
unresolved current links
Cline · Coding
Current route
Cline routes through a user- or administrator-selected provider, gateway, subscription, compatible endpoint, or local runtime. The configured endpoint and account determine the downstream deployment.
9
documented current links
7
unresolved current links
Superhuman · Work
Current route
Coda AI continues within the Superhuman Docs rename and exposes faster versus higher-quality modes plus automatic AI-column behavior. Current primary sources confirm third-party AI providers but do not publicly name the model or provider behind either quality mode.
6
documented current links
6
unresolved current links
CodeRabbit · Coding
Current route
CodeRabbit says OpenAI and Anthropic frontier models produce most review comments, Nemotron 3 Super and Ultra handle context, summarization, and non-complex reviews for self-hosted customers, and Plan uses Opus, Sonnet, and Haiku by stage. Self-hosted containers send prompts only to the customer-configured provider.
8
documented current links
4
unresolved current links
Continue · Coding
Current route
Continue assigns models independently by role and provider configuration. Cloud and local-compatible endpoints are user-controlled, and a local protocol does not prove same-device execution.
7
documented current links
5
unresolved current links
Anysphere · Coding
Current route
Cursor documents an unusually explicit router that classifies tasks and selects among its own and partner models. Serving shares remain private.
18
documented current links
4
unresolved current links
DeepSeek · Assistants
Current route
DeepSeek’s official iPhone listing ties the assistant to its flagship family. The exact current checkpoint and lower serving stack can change and remain unallocated here.
1
documented current links
3
unresolved current links
Descript · Creative
Current route
Underlord exposes Auto and a current picker spanning Claude, GPT, and Gemini labels. Product privacy documentation does not identify the serving provider infrastructure, region, facility, or accelerator.
4
documented current links
4
unresolved current links
Cognition · Coding
Current route
Cognition documents Devin as a cloud agent, a CLI with local sandboxed tools, and Devin Desktop with a model catalog spanning Anthropic, OpenAI, Google, Cognition SWE, and open-weight models plus the Adaptive router. The model-serving brain always runs in Cognition's cloud, third-party API keys are unsupported, and Cognition's pages variously cite an Azure tenant and an AWS environment.
10
documented current links
8
unresolved current links
SPRING (SG) PTE. LTD. · Assistants
Current route
The com.larus.wolf package and legacy Cici URLs now present Dola, operated under SPRING (SG) terms. Standard Chatbots may use integrated tools such as Google Gemini, but no exclusive model, checkpoint, inference cloud, site, chip, or ownership link is disclosed.
2
documented current links
8
unresolved current links
ByteDance / Beijing Chuntian Zhiyun Technology · Assistants
Current route
Doubao discloses mode-specific filed model families plus selectable Doubao 2.0 Expert and SeedEdit 3.0 P图 checkpoints. Its default router, other permitted services, request shares, and physical serving stack remain undisclosed.
3
documented current links
9
unresolved current links
ByteDance · Creative
Current route
Dreamina is a separately governed ByteDance creative platform that exposes Seedance and partner video models. Its current page advertises Seedance 2.5 while initially displaying Seedance 2.0 Mini; CapCut credentials do not collapse the products.
5
documented current links
4
unresolved current links
Dropbox · Work
Current + earlier routes
Dash exposes Auto and curated model selection. A 2025 Groq-hosting statement is historical and cannot allocate current picker traffic; current security language has surface-specific partner-sharing tension.
4
documented current links
4
unresolved current links
DuckDuckGo · Assistants
Current route
Duck.ai publishes a plan-specific model list and identifies Tinfoil hosting for named open models. Other serving details remain provider-specific.
12
documented current links
6
unresolved current links
Duolingo · Learning
Current + earlier routes
Duolingo’s latest explicit public model relationship is GPT-4. It has not refreshed the exact current checkpoint or serving path.
0
documented current links
4
unresolved current links
Duolingo · Learning
Current route
Video Call with Lily is explicitly LLM-powered, but Duolingo does not name its production model or provider. The older GPT-4 relationship belongs to Role Play and Explain My Answer and is not inherited by this feature.
0
documented current links
4
unresolved current links
Instagram / Meta · Creative
Current route
Edits is Instagram’s mobile creation app. No reviewed primary source supports carrying Meta AI’s Muse relationship over to Edits, so its model path remains separate and unknown.
0
documented current links
4
unresolved current links
ElevenLabs · Creative
Current route
ElevenLabs publishes its proprietary speech, transcription, and music models. Its cloud, chips, and electricity are not disclosed in product sources.
1
documented current links
3
unresolved current links
Factory · Coding
Current route
Droid offers Factory Router (Auto Model) with cross-provider failover, an explicit managed catalog across Anthropic, OpenAI, Google, xAI, and open-weight models, BYOK to OpenAI, Anthropic, OpenRouter, Fireworks, Together, Ollama, vLLM, or localhost runtimes, and enterprise traffic routed directly to configured providers without a Factory proxy.
9
documented current links
7
unresolved current links
Fathom Video Inc. · Work
Current route
Fathom names Anthropic, OpenAI, and Google as AI sub-processors under no-training terms, hosts its service in the United States, and uses de-identified data for its own models. The feature-to-provider mapping is private.
3
documented current links
2
unresolved current links
Figma · Creative
Current + earlier routes
Figma Make launched with Claude 3.7 and now lists GPT, Claude, and Gemini choices. Figma confirms an unnamed default that may change over time.
4
documented current links
10
unresolved current links
Google · Coding
Current route
Firebase Studio remains available to existing workspaces after new signup and workspace creation closed on June 22, 2026, with a March 22, 2027 sunset announced. Built-in Gemini, experimental customer-key/custom-Gemini, and App Prototyping routes remain distinct from workspace execution.
7
documented current links
6
unresolved current links
Fireflies.ai · Work
Current route
Fireflies says Global AskFred can run on Anthropic Claude models, claims 0-day retention with vendors, and exposes MCP connectors where ChatGPT and Claude process data under their own terms. The default summarization and AskFred model is not named in the reviewed sources.
6
documented current links
6
unresolved current links
Google · Consumer
Planned route only
Fitbit’s personal health coach is a public preview built with Gemini for eligible United States Fitbit Premium users, using a three-agent architecture described by Google Research. No Gemini version, region, TPU generation, or energy source is disclosed.
0
documented planned links
0
unresolved planned links
Garmin · Consumer
Current route
Active Intelligence delivers AI-powered insights inside the Garmin Connect+ subscription, initially in beta, without any documented provider, model, region, or accelerator.
2
documented current links
2
unresolved current links
Gauth · Learning
Current route
Gauth’s listing calls the study companion multi-model, while keeping provider names, allocation, hosting, chips, and energy private.
0
documented current links
4
unresolved current links
Google · Assistants
Current route
Google publishes available Gemini models and TPU infrastructure, while the exact consumer-app default, region, and chip generation can change without a request-level trace.
3
documented current links
1
unresolved current links
Google · Coding
Current route
Current Gemini 3 behavior is surface- and preview-dependent. Standard and Enterprise prompts are globally load-balanced across eligible Google locations, with no customer region choice or public prompt-to-region allocation.
2
documented current links
2
unresolved current links
Google · Learning
Current route
Google renamed NotebookLM to Gemini Notebook in July 2026 and documents Gemini 3.5 plus Antigravity; serving location and TPU generation remain private.
3
documented current links
1
unresolved current links
Google · Consumer
Current + earlier routes
Gemini for Home replaces Google Assistant on speakers and displays with free and Google Home Premium tiers, while camera descriptions and Home Brief require the Advanced plan. Google names Gemini but no version, region, TPU generation, or energy source.
4
documented current links
4
unresolved current links
GitHub / Microsoft · Coding
Current route
GitHub publishes a broad, changing model catalog and retirement ledger. Selection is visible; infrastructure behind a chosen third-party model is not.
18
documented current links
4
unresolved current links
GitHub / Microsoft · Coding
Historical route only
GitHub Spark stopped accepting new users and apps on August 4, 2026, kept export access until August 31, 2026, and lost GitHub Models, the inference service behind its llm() function, on July 30, 2026. Preview-era documentation named Claude Sonnet 3.5, GPT-4o, o1-preview, and o1-mini as selectable models.
4
documented earlier links
4
unresolved earlier links
Glean · Work
Current route
Glean-hosted SaaS runs on GCP, while customer-hosted deployments can use GCP or AWS. Model Hub and Chat selection can route across OpenAI/Azure, Vertex AI, Anthropic via Vertex or Bedrock, Amazon via Bedrock, and open models; tenant hosting is not model-serving geography.
6
documented current links
7
unresolved current links
Google · Creative
Current route
Google Flow currently exposes Veo 3.1 and Gemini Omni Flash. The Google fleet and TPU family are public, but a generation’s model choice, campus, and TPU generation are not.
6
documented current links
2
unresolved current links
Google · Coding
Current route
Jules uses a current Gemini model while executing repository tasks in a separate short-lived cloud Ubuntu VM. The service does not disclose a model-serving region, facility, accelerator, or energy source.
3
documented current links
2
unresolved current links
Google · Search
Current route
Google Photos routes supported simple queries through classic search and complex/background requests through Gemini-backed Ask Photos. Model family does not establish TPU, region, facility, or energy.
4
documented current links
4
unresolved current links
Google · Consumer
Current route
Pixel AI is not one route. Call Notes and other named Pixel 10 features can run Gemini Nano on Tensor G5, while Magic Cue and expanded Recorder summaries can use Private AI Compute in Google’s cloud.
6
documented current links
2
unresolved current links
Google · Search
Current route
Google names Gemini families for AI Mode and AI Overviews, but not a complete per-query router.
3
documented current links
1
unresolved current links
Google · Work
Current route
Gemini features in Gmail, Docs, Sheets, Slides, Meet, and Vids are gated by Workspace or Google AI plans, admin controls, age, and language. Google's help pages name "Gemini" without a version for these surfaces and name Veo 3 only for Vids clips, so the route stops at Google's fleet.
4
documented current links
4
unresolved current links
Superhuman · Work
Current route
Grammarly’s documented generative route uses Azure OpenAI, while the broader Superhuman suite names additional AI subprocessors without product-level allocation. Grammarly BYOK means AWS KMS encryption, not a model API key.
4
documented current links
4
unresolved current links
Granola · Work
Current route
Granola captures audio on the device, streams it to a transcription provider, stores notes in a US AWS VPC, and summarizes with OpenAI and Anthropic models. Chat uses Auto selection by default, with Business and Enterprise able to choose standard or thinking models including Fable 5.
8
documented current links
8
unresolved current links
xAI · Assistants
Current route
xAI identifies the Grok family and Memphis compute, but does not publish the default for every chat or whether a response was served in Memphis.
4
documented current links
0
unresolved current links
Groq · Assistants
Current route
GroqChat and Playground expose models supported by GroqCloud, while Compound can automatically choose models and built-in tools. The application-to-GroqCloud relationship does not by itself establish the accelerator used for a request.
4
documented current links
4
unresolved current links
HeyGen · Creative
Current route
HeyGen describes Avatar V as its current foundation while Video Agent can compose Avatar V with Seedance 2.0 and current documentation still exposes Avatar IV features and an API. The apparent V/IV tension is scoped by feature rather than resolved into an unsupported universal retirement claim.
8
documented current links
6
unresolved current links
Higgsfield · Creative
Current + earlier routes
Higgsfield separates proprietary Soul models and creative tools from a live partner picker and the Supercomputer orchestration layer. A dated OpenAI-to-Sora workflow is historical evidence, not a universal current route; infrastructure remains unallocated.
13
documented current links
6
unresolved current links
HubSpot · Work
Current route
Breeze model cards map some features to OpenAI or Anthropic models, while other agents remain automatically routed. Infrastructure subprocessors and storage regions do not identify request allocation.
5
documented current links
4
unresolved current links
Hugging Face · Assistants
Current + earlier routes
Hosted HuggingChat can use direct model choice or Omni routing through Hugging Face Inference Providers. Generic Chat UI self-hosting options must not be silently attributed to the hosted service.
7
documented current links
5
unresolved current links
ByteDance · Creative
Current route
Hypic is ByteDance’s AI photo editor. Its current iPhone listing does not identify the foundation model, serving cloud, chip, or grid.
0
documented current links
4
unresolved current links
IBM · Work
Current route
IBM documents watsonx Assistant migration into watsonx Orchestrate, per-agent model choice among Granite, Llama, and GPT-OSS, and virtual models registered from fourteen providers including OpenAI, Anthropic, Groq, Azure OpenAI, AWS Bedrock, and Ollama. Registration and catalogs are eligibility, not the deployment serving an agent.
6
documented current links
6
unresolved current links
Ideogram · Creative
Current + earlier routes
Ideogram 4.0 is current across the hosted app/API and an open-weight customer-controlled option. Older living model-picker documentation is bounded as pre-4.0 history; open weights and published minimum hardware do not reveal the hosted service’s accelerator.
4
documented current links
4
unresolved current links
Intercom · Work
Current route
Fin uses Intercom’s proprietary AI Engine and workflow orchestration, but its current downstream LLM is unnamed. Subprocessor and regional-processing lists cannot allocate a Fin request.
4
documented current links
4
unresolved current links
Jasper · Work
Current route
Jasper automatically routes and falls back across eligible model providers by use case. Provider membership and US data storage do not identify the model or serving location for a request.
3
documented current links
2
unresolved current links
JetBrains · Coding
Current route
JetBrains AI Assistant can use JetBrains-managed cloud routing, customer-provided third-party credentials, or Ollama/LM Studio models on the developer machine. Only the local branch moves the immediate inference-energy boundary to the user device.
7
documented current links
5
unresolved current links
KAYAK · Consumer
Current route
AI Mode combines KAYAK’s data with ChatGPT after graduating from the KAYAK.ai test site. KAYAK names no model checkpoint, region, or accelerator.
2
documented current links
2
unresolved current links
Kagi · Search
Current route
Kagi Assistant exposes explicit model choices and custom assistants, while Kagi Research automatically decomposes work across multiple calls and roles. Catalog and benchmark pages establish eligibility, not the model or provider used by an individual request.
6
documented current links
6
unresolved current links
Khan Academy · Learning
Current + earlier routes
Khan Academy’s latest explicit model disclosure is GPT-4 from launch-era material. The current checkpoint and deployment are unknown.
0
documented current links
4
unresolved current links
Moonshot AI · Assistants
Current + earlier routes
Kimi Chat currently exposes K3-backed chat and agent behavior, while K2-family releases provide chronology only. Open checkpoints and app tools do not disclose the hosted provider, region, facility, or accelerator.
2
documented current links
2
unresolved current links
Klarna · Consumer
Current route
Klarna’s in-app assistant is powered by OpenAI for customer service and, since September 2024, shopping. Klarna names no model checkpoint, region, or accelerator.
2
documented current links
2
unresolved current links
Kuaishou / KLING AI PTE. LTD. / LOHAS GAMES PTE. LTD. · Creative
Current + earlier routes
Kling’s current 3.0 family spans four video/image variants plus later Motion Control and native 4K modes. DeepSeek-R1 was a dated prompt-generation and optimization adjunct, while consumer inference hosting remains undisclosed and Singapore is a storage-only boundary.
5
documented current links
4
unresolved current links
Learna · Learning
Current route
Learna is a conversational language tutor in the measured mobile cohort. Its model and deployment providers are not disclosed.
0
documented current links
4
unresolved current links
Linear · Work
Current route
Linear separates its first-party Agent, coding sessions with Auto/default and explicit model choices, and installed external agents. Linear MCP and agent APIs supply data and tools; they do not make third-party agent inference part of Linear’s own model infrastructure.
8
documented current links
8
unresolved current links
LinkedIn (Microsoft) · Consumer
Current route
LinkedIn’s GenAI proxy routes application requests either to Azure OpenAI service or to fine-tuned Llama models served on LinkedIn infrastructure with vLLM. Request shares, regions, GPU counts, and energy are not disclosed.
5
documented current links
4
unresolved current links
Lovable · Coding
Current route
Lovable documents a main agent, context work, and automatic subagents behind a dynamically weighted provider load balancer spanning Anthropic, Google Vertex AI, and Amazon Bedrock. A prompt can produce several calls, while the model used at each step, provider order, serving region, facility, accelerator, and electricity remain private.
5
documented current links
4
unresolved current links
Luma AI · Creative
Current route
Luma’s current product information distinguishes first-party Ray/UNI models from integrated Veo, Kling, Seedance, Nano Banana, and ElevenLabs models. AWS confirms a broad infrastructure relationship, not the route of every Luma request.
7
documented current links
7
unresolved current links
MagicSchool · Learning
Current route
MagicSchool says it pairs each tool with a model from a multi-model architecture including OpenAI GPT-4o, Anthropic Claude, and Google Gemini, holds zero-data-retention attestations from Anthropic and OpenAI, and processes data in US-based data centers run by unnamed infrastructure providers. The tool-to-model table is private.
3
documented current links
2
unresolved current links
Manus (Butterfly Effect); Meta acquisition unwound · Coding
Current + earlier routes
Manus documents ten selectable OpenAI, Anthropic, and Google models for features in apps it builds and a default plus Veo 3 option for video, while the models behind its core agent are not named. CNBC reported Meta's December 2025 acquisition and, in August 2026, that Chinese regulators forced it to be unwound with Manus returning to independence.
6
documented current links
8
unresolved current links
Meitu · Creative
Current route
Meitu is a large mobile AI editor; its current listing does not disclose a feature-level foundation model or infrastructure path.
0
documented current links
4
unresolved current links
Meituan · Consumer
Current + earlier routes
Xiaotuan is the all-user assistant embedded in Meituan App and uses LongCat capabilities plus unnamed mainstream or open-source models. Xiaomei is a separate app, and Meituan’s domestic-chip LongCat 2.0 Preview disclosure is not allocated to Xiaotuan.
3
documented current links
4
unresolved current links
Mercedes-Benz · Consumer
Current route
MBUX uses a context- and vehicle-dependent router across native, Microsoft/OpenAI, and Google automotive-agent paths. Announcements apply only to supported model years, head units, markets, and features.
6
documented current links
4
unresolved current links
Meta · Assistants
Current + earlier routes
Meta AI moved from Llama 4 to the Muse family, but model revision depends on mode, medium, surface, market, and rollout. No reviewed source assigns Muse serving to a campus or chip.
4
documented current links
12
unresolved current links
Meta · Consumer
Current route
Meta AI inside the messaging and social apps is a gradual Muse Spark and Muse Image rollout by surface and market, while WhatsApp summaries and writing help run in Private Processing confidential VMs. Neither branch discloses a region, facility, accelerator, or electricity source.
6
documented current links
6
unresolved current links
Microsoft · Work
Current route
Copilot Chat preselects an opaque Auto router. GPT-5.6 is preferred with two disclosed serving channels, Claude is tenant-controlled, and one EU/EFTA/UK app setting creates a narrow Anthropic-default exception.
6
documented current links
12
unresolved current links
Microsoft · Assistants
Current route
The standalone consumer Copilot app is distinct from Microsoft 365 Copilot. Its listing does not establish the default model or live fallback policy for each request.
0
documented current links
4
unresolved current links
Microsoft · Search
Current route
Edge’s iPhone listing confirms built-in Copilot, but not the live model, fallback policy, cloud allocation, or site.
0
documented current links
4
unresolved current links
Microsoft · Consumer
Current route
Recall snapshot storage, OCR, analysis, and natural-language search run locally on an opted-in Copilot+ PC. User-invoked Click to Do exits, metadata retrieval, diagnostics, feedback, and EEA export are distinct paths and do not turn the core inference into a cloud route.
8
documented current links
4
unresolved current links
Microsoft · Work
Current route
Teams separates license-free AI features, Copilot in Teams with Facilitator and Channel Agent, and Teams Premium intelligent recap. Microsoft says Copilot prompts stay inside the Microsoft 365 service boundary and that recap notes and tasks use an OpenAI GPT model through a dedicated Azure OpenAI service; a compliance boundary is not a facility or accelerator.
6
documented current links
6
unresolved current links
Midjourney · Creative
Current route
Midjourney publishes its proprietary model version history but not an outside model provider, cloud, accelerator, or grid path.
1
documented current links
3
unresolved current links
Mistral AI · Assistants
Current + earlier routes
Le Chat was renamed Vibe on June 5, 2026. Vibe routes among Chat, Work, and Code modes with agent, model, and tool conditions; Code can execute tools locally or remotely while model inference remains a separate managed-service boundary.
5
documented current links
4
unresolved current links
Notion · Work
Current route
Notion Agent and Custom Agents use an opaque Auto default or explicit model selection. Enterprise Search and the external Claude agent have separate provider, storage, and retention boundaries.
9
documented current links
16
unresolved current links
OpenAI · Coding
Current route
Codex separates local CLI or IDE tool execution from delegated cloud task environments. A local agent still uses OpenAI-hosted model inference unless contrary configuration is documented.
6
documented current links
4
unresolved current links
OpenRouter · Assistants
Current route
OpenRouter Chat supports explicit model/provider choice and automatic provider balancing, filtering, BYOK priority, and model fallback. Response metadata—not the catalog—is needed for observed allocation.
6
documented current links
6
unresolved current links
Opera · Assistants
Current + earlier routes
Opera AI replaced Aria and automatically uses multiple LLMs including OpenAI and Google. Aria-era GPT, Gemini, and image-model mappings are retained only as historical chronology.
2
documented current links
2
unresolved current links
Oracle · Work
Current route
Oracle's Fusion Applications agents run on LLMs Oracle curates and hosts in OCI, with Llama and Cohere named as optimized options, announced support for OpenAI, Anthropic, Cohere, Google, Meta, and xAI models, customer-added LLM credentials, and an earlier statement that custom-AI GPT-4 would be hosted in Azure. Provider hosting statements are not facility, region, or per-agent allocations.
6
documented current links
6
unresolved current links
Otter.ai · Work
Current route
Otter trains proprietary transcription and chat models on de-identified data and lists Anthropic as a backend AI subprocessor, OpenAI for LLM evaluation, and AWS, Google Cloud, and Crusoe as US cloud providers. A subprocessor list does not say which provider or cloud serves a particular feature.
4
documented current links
4
unresolved current links
Oura · Consumer
Current route
Oura Advisor is an LLM-powered assistant whose general provider is undisclosed, with a separately documented proprietary women’s-health model hosted on Oura-controlled infrastructure. Neither branch discloses a site, region, accelerator, or energy source.
4
documented current links
4
unresolved current links
Perplexity AI · Search
Current route
Perplexity exposes several model providers. Model Council is a rare documented fan-out: three parallel models rejoin through an unnamed synthesizer.
15
documented current links
7
unresolved current links
Google LLC · Learning
Current route
Photomath documents a staged cloud image-recognition and symbolic expert-system solver rather than a named foundation-model call. Its separate Plus tutorial AI remains opaque; Google ownership does not establish Gemini or Google Cloud routing.
3
documented current links
7
unresolved current links
Photoroom · Creative
Current route
Photoroom AI Backgrounds exposes platform- and plan-conditioned defaults and model selection, while API documentation presents a distinct model parameter and potentially stale default. In-house model strategy does not justify assigning every feature or request to a proprietary model or physical fleet.
6
documented current links
6
unresolved current links
Picsart · Creative
Current route
Picsart is a widely used AI photo/video editor, but its mobile feature-to-model router and infrastructure are not public.
0
documented current links
4
unresolved current links
Pika · Creative
Current route
Pika’s principal creator uses its first-party video family, while Agent/MCP and API Club are separate multi-provider surfaces. PikaStream’s one-H100 disclosure is restricted to AI Self real-time video and cannot be carried into ordinary creator jobs.
17
documented current links
10
unresolved current links
Quora · Assistants
Current route
Poe is explicitly multi-provider. Users select bots and models, while its Assistant can route automatically without exposing every decision.
15
documented current links
7
unresolved current links
Proton · Assistants
Current route
Current Lumo 2.0 exposes Max and Lite labels plus automatic quota fallback on Proton’s fully European infrastructure. Current base checkpoints, country, facility, cluster, hardware ownership, accelerator, and electricity remain undisclosed; the 2025 open-source wording is historical context only.
7
documented current links
6
unresolved current links
Tencent · Search
Current + earlier routes
QQ Browser’s current build names both Hunyuan and DeepSeek and integrates Yuanbao across several modes. A selectable 2025 DeepSeek-R1 launch and the distinct Hy3 Preview integration are bounded checkpoints; formal Hy3 cutover remains unresolved.
2
documented current links
4
unresolved current links
Qodo · Coding
Current route
Qodo unified Merge, Gen, Command, and Aware into one platform. The IDE picker lists GPT 5.1, GPT 5.1 Instant, GPT 5, Sonnet 4.5, Haiku 4.5, Opus 4.1, Gemini 2.5 Pro, Grok, o3, and o4-mini, while on-premise deployments run on customer Kubernetes with undisclosed LLM endpoints.
4
documented current links
4
unresolved current links
Quizlet · Learning
Historical route only
Q-Chat is a closed historical path: it launched with OpenAI’s ChatGPT API and was discontinued in June 2025.
1
documented earlier links
3
unresolved earlier links
Alibaba Cloud · Assistants
Current route
Qwen Chat is now presented as Qwen Studio. Current first-party releases establish selectable or feature-conditioned Qwen text, thinking, and image families, but neither open-weight availability nor a Singapore contracting entity identifies a hosted request region or physical deployment.
6
documented current links
6
unresolved current links
rabbit · Consumer
Current route
r1 runs headless in the cloud on rabbit OS, whose LAM works with Perplexity, Anthropic, and OpenAI for intent understanding. rabbit does not publish which vendor serves a request, nor any region or hardware.
3
documented current links
2
unresolved current links
Meta · Consumer
Current route
Ray-Ban Meta combines cloud Meta AI/Muse Spark assistant paths with narrow on-device ExecuTorch features. Qualcomm AR1 device hardware is not evidence of cloud model hardware.
5
documented current links
3
unresolved current links
Bending Spoons · Creative
Current + earlier routes
Remini’s current feature catalog leaves its model and serving stack undisclosed. Google sources support bounded historical G2/L4 and Imagen 3 filter routes, while Bending Spoons’ group-wide provider portfolio cannot be carried into current Remini.
0
documented current links
4
unresolved current links
Replit · Coding
Current route
Replit publishes a broad model selector and says Auto can escalate harder tasks. Model availability does not disclose the serving site.
15
documented current links
7
unresolved current links
Roo Code · Coding
Historical route only
Roo Code used named API Configuration Profiles to select a provider, gateway, compatible endpoint, Bedrock account, Ollama, or LM Studio route before the extension shut down on May 15, 2026. Only a verified same-device endpoint supported an on-device inference conclusion.
5
documented earlier links
3
unresolved earlier links
Runway · Creative
Current + earlier routes
Runway’s subscriber web app, developer API, API Model Router, enterprise entitlements, and MCP have different catalogs and controls. Sora ended on Runway April 3, 2026; current inference infrastructure is not allocated.
13
documented current links
6
unresolved current links
Salesforce · Work
Current route
Agentforce model routing depends on tenant, agent, feature, and selected Salesforce or external model provider. Provider availability is modeled without converting it into an observed invocation, cloud region, facility, chip, or fuel.
10
documented current links
10
unresolved current links
Samsung · Consumer
Current route
Galaxy S26 pairs an upgraded Bixby device agent with a user choice of Gemini and Perplexity agents. Samsung names no Bixby model, cloud, or hardware, and agent choice is eligibility rather than request allocation.
6
documented current links
6
unresolved current links
Samsung · Consumer
Current route
Galaxy AI is not one model route. Some compatible feature/device combinations use local NPU processing, while other features require network services. A feature must be resolved before assigning the immediate device or cloud energy boundary.
8
documented current links
4
unresolved current links
Undisclosed in reviewed source · Search
Current route
Seekee is rank 25 in the published Sensor Tower-derived chart. The a16z link target appears mismatched, so this release makes no company, iPhone-listing, or backend claim beyond the measured name and rank.
0
documented current links
4
unresolved current links
ServiceNow · Work
Current route
Now Assist can use ServiceNow models or per-instance, skill, and agent configurations for Azure OpenAI, Gemini, Claude on AWS, and Bedrock. Historical centralized-hub wording is not universal current routing.
7
documented current links
6
unresolved current links
Salesforce · Work
Current + earlier routes
General native Slack AI now routes unnamed models across Bedrock and Vertex. Native Slackbot specifically names Claude; the Anthropic-operated Marketplace app is a separate processor and product.
7
documented current links
8
unresolved current links
Snap · Consumer
Current + earlier routes
Snap says My AI uses models “like” OpenAI GPT and Google Gemini. The precise current versions and live router are not disclosed.
9
documented current links
2
unresolved current links
OpenAI · Creative
Historical route only
The standalone Sora product is a closed historical route: Sora 2 launched September 30, 2025, and OpenAI states the product became unavailable April 26, 2026.
1
documented earlier links
3
unresolved earlier links
Sourcegraph · Coding
Current route
Cody uses feature defaults or user selection through Cody Gateway, while enterprise customers can configure direct/BYOK providers. Self-hosted Sourcegraph does not imply self-hosted model inference.
6
documented current links
6
unresolved current links
Speak · Learning
Current route
Speak's dated 2023 posts say its AI Tutor was built on GPT-4 and that it launched with OpenAI's Whisper-based speech-recognition API; its current privacy notice says audio is shared with OpenAI and Microsoft for speech recognition and transcription. The current tutor model generation is not disclosed.
6
documented current links
4
unresolved current links
Spotify · Consumer
Current + earlier routes
Current Spotify DJ behavior is documented, but its current model and provider are not. OpenAI editorial tooling and Sonantic voice are retained only as historical launch-era components.
2
documented current links
2
unresolved current links
Strava · Consumer
Current route
Athlete Intelligence generates activity summaries with generative AI for subscribers on mobile, without a documented provider, model, region, or accelerator.
2
documented current links
2
unresolved current links
Suno · Creative
Current route
Suno names its proprietary flagship model. No reviewed first-party source identifies an outside foundation model or hosting route.
1
documented current links
3
unresolved current links
Synthesia · Creative
Current route
Synthesia’s Express-2 avatar route uses its proprietary stack on AWS, with a separate feature-specific Veo 3 branch. Storage and processing terms do not reveal a serving facility or accelerator.
4
documented current links
4
unresolved current links
Tabnine · Coding
Current route
Tabnine supports model selection in SaaS and private deployments, including VPC, on-premises, air-gapped, and NVIDIA NIM configurations. Product examples and capacity do not allocate an individual request to hardware.
6
documented current links
6
unresolved current links
Topaz Labs · Creative
Current route
Current Topaz Photo succeeds the discontinued catalog name Topaz Photo AI. Wonder has version- and hardware-specific local processing as well as cloud rendering; some platform/model combinations require cloud, so device and data-center electricity remain separate route boundaries.
7
documented current links
5
unresolved current links
Tripadvisor · Consumer
Current route
Tripadvisor’s itinerary generator launched on OpenAI’s generative AI technology, and its Viator brand later became a connected app inside Google Gemini. No reviewed source names a current checkpoint, region, or accelerator.
4
documented current links
4
unresolved current links
Uber · Consumer
Planned route only
Uber announced a driver-app AI Assistant powered by OpenAI’s GPT-4o for EV questions as coming soon in October 2024. No reviewed first-party source confirms launch, current model, region, or hardware.
2
documented planned links
2
unresolved planned links
Ubiquiti Labs · Creative
Current route
VN is ranked as a major AI-related mobile editor; its listing does not disclose the models or providers behind its AI features.
0
documented current links
4
unresolved current links
Venice AI · Assistants
Current + earlier routes
Venice exposes distinct Anonymous, Private, TEE, and E2EE trust modes. TEE and E2EE remain remote compute; historical decentralized-GPU and older default-model claims do not establish the current provider or site.
8
documented current links
8
unresolved current links
Volkswagen · Consumer
Current + earlier routes
For supported vehicles and markets, Volkswagen IDA/Plus Speech can invoke Cerence Chat Pro integrated LLMs for open-domain conversation. Vehicle eligibility is not universal, and the selected model, cloud, region, facility, accelerator, and energy are not disclosed.
2
documented current links
2
unresolved current links
Warp · Coding
Current route
Warp exposes explicit model choices, Auto variants, profile defaults, custom routers, fallbacks, customer API keys, and enterprise Bedrock accounts. Local terminal and cloud-agent execution environments are distinct from the remote model-serving path.
7
documented current links
6
unresolved current links
Windsurf · Coding
Current route
Windsurf exposes explicit model choices, supported BYOK branches, an automatic Adaptive router, and feature-specific models such as Tab. Local agent execution is distinct from model serving.
6
documented current links
6
unresolved current links
Meitu · Creative
Current route
Wink offers AI video and photo enhancement, while its current model, serving cloud, accelerator, and energy path remain private.
0
documented current links
4
unresolved current links
Writer · Work
Current route
Writer can use Palmyra or external models deployed through AWS Bedrock, Microsoft Azure, or NVIDIA NIM. Model-provider configuration and customer-managed encryption keys are distinct controls.
4
documented current links
4
unresolved current links
Apple · Coding
Current + earlier routes
Xcode is several stacks: Apple’s predictive completion runs locally, while generative chat and agents use developer-selected ChatGPT/Codex, Claude, Gemini, compatible providers, local models, or ACP agents. Xcode 27 is a separate developer-beta expansion.
10
documented current links
13
unresolved current links
Yandex · Search
Current route
Yandex’s iPhone listing explicitly embeds Alice AI, while the exact model checkpoint, serving region, accelerator, and grid remain private.
1
documented current links
3
unresolved current links
Google · Consumer
Current route
Ask YouTube on the watch page is documented as built with Gemini LLMs across 40+ languages, while the newer conversational search experiment is limited to United States users without a named model. Neither discloses a region, facility, or accelerator.
2
documented current links
2
unresolved current links
Zed Industries · Coding
Current route
Zed supports hosted models, direct provider credentials, existing subscriptions, gateways, compatible endpoints, and same-device runtimes. Endpoint configuration—not the word local—determines whether inference is actually on the user device.
7
documented current links
5
unresolved current links
Zendesk · Work
Current route
Zendesk separates proprietary classifiers from feature-specific third-party generation. Provider portfolio includes several services, but only call transcription currently supports a precise Deepgram-to-OpenAI feature chain.
7
documented current links
6
unresolved current links
Zoom · Work
Current route
Zoom describes a federated layer across Zoom-owned, OpenAI, and Anthropic models, with feature-specific Perplexity and ElevenLabs services.
12
documented current links
6
unresolved current links
monday.com · Work
Current route
monday AI routes feature work across a documented portfolio that can include Claude through Vertex, Bedrock, or direct service; GPT through Azure or direct service; Gemini through Vertex; and self-hosted sentence-transformer features. Data residency and provider eligibility are not request-level serving allocation.
6
documented current links
6
unresolved current links
Vercel · Coding
Current route
v0 combines a managed chat/agent surface, an explicit v0-1.5 Model API family, and per-chat code sandboxes that can deploy generated applications to Vercel. Sandbox and deployment locations do not establish where the v0 model is served.
7
documented current links
6
unresolved current links
No view locates an individual request or traces specific electrons. “Unknown” means the reviewed public record does not support a stronger claim.
Image + video routes includedThe gaps matter as much as the confirmed connections. Every unresolved product route, unmapped application, facility disclosure, valuation ambiguity, and layer-specific reporting question belongs in the research backlog. Use it to see what still needs a primary source, a public-record request, a company response, or on-the-ground reporting.
Search the unanswered questions, queued applications, missing allocation data, valuation boundaries, facility fields, and layer-specific reporting leads. Open questions derived from the AMICE public-evidence boundary. Inclusion is not an allegation, a prediction, or proof that a missing fact exists in a discoverable document.
Research snapshot reviewed Sep 4, 2026.
2172 of 2172 records match. Showing 72 below; exports always include the complete matching set.
Evidence boundary
Adobe Express AI is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Adobe Podcast is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
After Effects AI is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Airbnb AI Support is listed for future research in the Consumer catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
AIVA is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
ALEKS is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Amira Learning is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Andi is listed for future research in the Assistants catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Arc Max is listed for future research in the Assistants catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Arc Search is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Ask Instacart is listed for future research in the Consumer catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
AskCodi is listed for future research in the Coding catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Avoma is listed for future research in the Work catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Babbel AI is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Beatoven.ai is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Beautiful.ai is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Bing Copilot Search is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Bing Image Creator is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Bito is listed for future research in the Coding catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Blackbox AI is listed for future research in the Coding catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Boomy is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Brainly AI is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Brisk Teaching is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Busuu Conversations is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Canva Magic Studio is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
CapCut AI is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Captions is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Carnegie Learning MATHia is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Century Tech is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Chatsonic is listed for future research in the Assistants catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
CheggMate is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Chorus is listed for future research in the Work catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Clari Copilot is listed for future research in the Work catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Clipdrop is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Clockwise AI is listed for future research in the Work catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
CodeGeeX is listed for future research in the Coding catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Codeium is listed for future research in the Coding catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Cohere Coral is listed for future research in the Assistants catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Confluence AI is listed for future research in the Work catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Connected Papers is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Consensus is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Copilot Workspace is listed for future research in the Coding catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Copy.ai is listed for future research in the Work catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Cosine Genie is listed for future research in the Coding catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Coursera Coach is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Craiyon is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Curipod is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
DeepSeek Chat is listed for future research in the Assistants catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Devv AI Search is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Dia Browser AI is listed for future research in the Assistants catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Dimensions Research GPT is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Discord Clyde is listed for future research in the Consumer catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
DoorDash AI Concierge is listed for future research in the Consumer catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
DreamBox Math is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
DreamStudio is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Dropbox AI is listed for future research in the Work catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
DuckDuckGo AI Answers is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
eBay AI Shopping is listed for future research in the Consumer catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Eduaide.Ai is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
ElevenLabs Studio is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Elicit is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
ELSA Speak is listed for future research in the Learning catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
ERNIE Bot is listed for future research in the Assistants catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Etsy Gift Mode is listed for future research in the Consumer catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Exa Search is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Expedia AI Assistant is listed for future research in the Consumer catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Felo Search is listed for future research in the Search catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
FLUX Playground is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Forefront Chat is listed for future research in the Assistants catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Freshworks Freddy is listed for future research in the Work catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Gamma is listed for future research in the Creative catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
Evidence boundary
Genspark is listed for future research in the Assistants catalog. No backend relationship is asserted until dated evidence supports it.
Record
No source attached yet
This item names work to begin or a cross-source question. It does not pretend that a supporting source has already been found.
2100 more matching records are ready to inspect.
This is a maintained evidence map, not a claim that the AI supply chain can be made perfectly observable. A good update either adds a sourced relationship, narrows a claim, or turns an apparent fact back into an open question.
These reading notes were checked on September 17, 2026. The interactive source ledgers retain their separately dated records.
Epoch AI: Chip Users explorer methodology. Independent estimates of compute use, with explicit uncertainty and a distinction between renting and owning.
EIA: electricity data and tools. A starting point for investigating a documented U.S. grid relationship. Regional data cannot establish an undisclosed route.
Berkeley Lab: Queued Up 2026. A check on supply-side project claims. Queue entries are proposed capacity, not operating power plants.
