The Contest Map | AI Wars of 2026
The routes being tested across contested AI positions. Source-backed analysis from AI Wars of 2026.
Part II — The Contenders
The Contest Map
Big Idea
The contest spans seven interdependent positions: intelligence, compute, context, authority, execution, interface, and distribution. Winning them only matters if they turn into real economic value.
The map below is a synthesis of the routes the core and provisional contenders are testing and the dependencies among their positions. It is not a measurement of capability, ownership, or likelihood of winning. Product announcements, reported seats, infrastructure revenue, and audience scale use different units, definitions, and periods. They cannot responsibly become a score, ranking, or cross-company strength claim.
The routes being tested
The core cohort begins from different strategic positions. Microsoft puts enterprise workflow, identity, and cloud around agents, while Foundry and multi-model Copilot give it a route into model access and distribution. OpenAI starts with an AI-native work interface. Google connects an existing information and productivity ecosystem, while Gemma gives it a documented open-model and developer-distribution route. Amazon starts with infrastructure and execution environments. Anthropic makes controlled model-led work its test. Meta combines social surfaces with open-weight distribution. Apple, NVIDIA, Baidu, Tencent, and ByteDance are provisional contenders whose distinct routes need more evidence before they belong in the core cohort. These are strategic interpretations grounded in the individual contender records, not demonstrated ownership of any layer. Source 1 Source 2 Source 3 Source 4 Source 5 Source 6 Source 7 Source 8 Source 9
Capability map
The visual maps strategic routes being tested. A connection means that a route is relevant to the contender's stated position; it does not claim that the company owns, leads, or has measured strength in that capability. Microsoft's Foundry and multi-model Copilot entries establish available multi-model access, not frictionless switching, adoption, or model portability. Amazon's workflow and context route is mediated through customer-built Bedrock applications and AgentCore Memory rather than native enterprise workflow data access. Its newly available Agent Registry adds a governed catalog and discovery layer to the agents customers manage themselves. But it does not demonstrate that organizations have adopted a coherent AWS agent ecosystem. AWS's Azure Interconnect preview identifies a scoped private-connectivity route, not application or data portability. Anthropic's orchestration entry reflects MCP tool connectivity and Claude Code's multistep execution, not a documented standalone orchestration platform. NVIDIA's open-model and developer-ecosystem entry now rests on an announced acquisition rather than on reporting, and is shown as pending completion. An announcement does not transfer third-party model ownership or determine developer choice. Two September 2026 developments extend the interoperability route in opposite directions from the same week. GitHub's HydraFusion research preview selects among rival providers' models at runtime. AWS's MCP Server added serverless diagnostics for third-party coding agents it names, including Claude Code. Both describe platforms treating rival models and agents as components; neither establishes portability, adoption, or measured outcomes. Source 10 Source 11 Source 12 Source 13 Source 14 Source 15 Source 16 Source 17 Source 18
Compute roles and access
Cloud access and internal deployment are not the same offer. Google documents rented TPU capacity, and AWS documents Trainium- and Inferentia-based EC2 instances. Source 19 Source 20 Microsoft states that Maia 200 is deployed in US Central and offers an SDK preview. Source 21 Those records support distinct infrastructure routes; they do not establish equal access, reliable migration or comparable operating costs.
Apple's Private Cloud Compute uses custom Apple-silicon servers for inference within Apple Intelligence. Source 22 Meta's MTIA 300 is optimized for training ranking and recommendation models. Source 23 These are internal product-serving routes, not evidence that either company offers a general-purpose accelerator cloud or can replace another supplier across workloads.
OpenAI's October 2025 agreement with Broadcom describes OpenAI-designed accelerators and systems developed with Broadcom's accelerator and networking capabilities. The announcement targeted deployment beginning in the second half of 2026. That target is not confirmation that deployment has begun. The map records an announced co-design route, not delivered capacity or reduced dependence on another supplier. Source 24
The map is also a dependency map. Model providers depend on compute and deployment environments; infrastructure providers depend on customers choosing their stack; and interface or workflow positions depend on legitimate access to context and authority. A contender can be strong in one position while relying on a rival or partner in another. The resulting structure is a balance of positions, not a single route to a universal host.
Provisional Contenders
Provisional contenders have a source-backed route to a material strategic prize, but their evidence record is not yet sufficient for a core-contender conclusion. They remain visible so the publication can test whether a distinctive route gains repeated use, reliable outcomes, legitimate authority, or durable economics.
Apple and NVIDIA
Apple tests a personal-device-context route, while NVIDIA tests a compute and AI-infrastructure route. Each has a distinct structural position, but neither current record establishes a general-purpose agent host or a durable outcome from that position. Source 25 Source 26
Baidu
Baidu reports revenue from AI Cloud Infra and AI Applications, an enterprise DuMate launch, and rising AI-feature penetration in Wenku and Drive. These company-reported measures identify a route across search, productivity applications, cloud infrastructure, and autonomous mobility in China. They do not identify agent retention, task success, or workflow ownership. Source 27
Tencent
Tencent distributes its Hy4 preview through productivity, assistant, and cloud surfaces, while its corporate overview identifies Weixin and WeChat consumer distribution and Tencent Cloud. This establishes a potential route connecting messaging context, productivity tools, cloud, and open-model distribution. It does not establish repeated agent use, permissioned action, or commercial outcomes. Source 28 Source 29
ByteDance
ByteDance's portfolio reaches creative work, entertainment, commerce, and workplace software through products including TikTok, CapCut, TikTok Shop, and Lark. That distribution makes it relevant to the book's question of whether AI changes industry-specific work and interfaces, not only general-purpose office work. The current record does not identify an AI-agent product, repeated AI use, or a measured outcome. Source 30
Watch List
The watch list is a research queue, not a second cohort or a scorecard. A watch-list company becomes provisional when a source-backed route is material to a distinct strategic prize. A provisional company moves into the core cohort only with dated evidence sufficient to assess its position, dependencies, and outcomes without forcing unlike measures into comparison.
Compute and infrastructure: Oracle and Huawei
Why this route matters: Oracle's public AI material provides a cloud-and-application route worth monitoring. It can shape the cost, availability, and operating environment of AI systems without necessarily owning the user-facing interface or the compute position. Source 31
Evidence limitation: Oracle's public page is not a measure of broad AI adoption, workflow control, or infrastructure economics. Promotion criterion: add a contender only with a source-backed, materially distinct route and dated evidence that can assess position, dependencies, and outcomes.
Huawei's announced route: Huawei's September 17 announcement introduces the Ascend 960 SuperPoD and states that CANN is fully open-source and operating through an open-source community. This connects a compute-infrastructure offer with a software ecosystem, making it relevant to the infrastructure choices available to AI developers. It does not establish that developers have adopted that route. Source 32
Huawei evidence limitation: The announcement does not establish independent workload performance, availability to every buyer, or a durable ecosystem advantage. Huawei remains on the watch list; its product announcement alone does not satisfy the criteria for becoming a contender. Source 32
Enterprise workflow and governance: Salesforce, ServiceNow, and IBM
Why this route matters: Salesforce presents Agentforce as an agent product; ServiceNow's newsroom positions governance and predictable outcomes as enterprise-AI concerns; IBM describes watsonx.ai as a studio joining models, agent tooling, APIs, and runtimes. These are possible routes through the systems where organizations govern and execute work. Source 33 Source 34 Source 35
Evidence limitation: these are primarily vendor product or positioning materials, not common measures of active use, reliable task completion, or return on investment. Promotion criterion: add a contender when an independently inspectable or clearly bounded disclosure shows a materially distinct route with dated position or outcome evidence and explicit controls where relevant.
Model and sovereignty alternatives: Alibaba/Qwen, xAI, Cohere, Mistral, and DeepSeek
Why this route matters: Qwen released open-weight Qwen3 models; xAI describes Grok 4's reinforcement-learning and tool-use work; Cohere frames sovereign AI as a buyer concern. Mistral Studio catalogs models, agents, prompts, skills, datasets, and workflows; DeepSeek-R1 documents a reasoning model trained with reinforcement learning. Together, they keep open models, model challengers, sovereignty, and model-management surfaces in view. Source 36 Source 37 Source 38 Source 39 Source 40
Evidence limitation: vendor benchmarks, model releases, commissioned research, and technical papers do not establish durable adoption, governance effectiveness, or economic outcomes. Promotion criterion: add a contender when evidence shows a stable route beyond model release with product-defined repeated-use, strategic-position, or outcome evidence.
AI-native interfaces and developer distribution: Cursor and Perplexity
Why this route matters: Cursor's Mixture-of-Kittens post describes infrastructure supporting its Composer coding model, while Perplexity's public hub is a continuing record of its product announcements. They are watched because developer and research interfaces can become starting points for delegated work even without an incumbent workplace suite. Source 41 Source 42
Evidence limitation: Cursor's engineering post and Perplexity's announcement hub do not supply compatible evidence of broad work-host adoption, retention, permission controls, or successful task completion. Promotion criterion: add a contender when a source-defined product route demonstrates a materially distinct position. It must disclose dated usage, governance, or workflow evidence that can be tested against a meaningful counterargument.
What would change the map
The map should change when evidence demonstrates or weakens a route, not when a company merely announces a feature. The next useful records are product-defined measures of repeated use, retention, successful delegated tasks, incident and recovery performance, permission and governance operation, and customer economics. For the NVIDIA-Hugging Face route, the test so far is met; the remaining test is closing. Regulatory clearance and completion would move the route from announced to transferred. After that, developer-path conversion becomes the relevant outcome test. Until compatible measures exist, the right editorial act is to keep routes visible and strength claims bounded.
Sources
- Source 1
- Source 2
- Source 3
- Source 10
- Source 17
- Source 4
- Source 5
- Source 6
- Source 43
- Source 7
- Source 11
- Source 12
- Source 13
- Source 14
- Source 18
- Source 8
- Source 15
- Source 25
- Source 44
- Source 22
- Source 45
- Source 9
- Source 46
- Source 26
- Source 47
- Source 16
- Source 48
- Source 27
- Source 28
- Source 29
- Source 30
- Source 31
- Source 33
- Source 34
- Source 35
- Source 36
- Source 37
- Source 38
- Source 39
- Source 40
- Source 41
- Source 42
- Source 32
- Source 21
- Source 19
- Source 20
- Source 23
- Source 24
Eyebrow: A conceptual map, not a scorecard
Title: Core and provisional contenders across contested AI positions
Description: Each group connects an interdependent contested position to the contenders for whom that route is strategically relevant. Provisional routes need further evidence before they support a core-contender conclusion.
Note
Label: How to read this
Body: Connections show routes being tested, not demonstrated ownership, measured strength, or a forecast. Microsoft's Foundry entries do not prove model portability, and NVIDIA's open-model entry rests on an announced Hugging Face agreement that has not closed.
Center
Name: Contested AI positions
Description: Distinct prizes that can create leverage, economic value, or freedom of action.
Routes
Name: Compute and AI infrastructure
Description: Capacity, networking, software, and operations beneath advanced AI systems.
Contenders
Name: Microsoft
Route: Maia inference infrastructure; SDK preview
Name: OpenAI
Route: Custom accelerators with Broadcom; announced
Name: Google
Route: TPU cloud capacity
Name: Amazon
Route: Trainium and Inferentia cloud capacity
Name: Apple (provisional)
Route: Apple-silicon private inference
Name: NVIDIA (provisional)
Route: AI infrastructure platform
Name: Meta
Route: MTIA recommendation-training infrastructure
Name: Model creation and access
Description: Model creation, access, and adaptation for capable work.
Contenders
Name: Microsoft
Route: Foundry multi-model access
Name: OpenAI
Route: AI-native work platform
Name: Google
Route: Integrated ecosystem
Name: Anthropic
Route: Trusted model core
Name: Meta
Route: Open-weight distribution
Name: Tencent (provisional)
Route: Hy4 product and cloud distribution
Name: Model access and developer ecosystem
Description: Where developers discover, select, collaborate around, and deploy AI artifacts.
Contenders
Name: Microsoft
Route: Foundry model distribution
Name: Google
Route: Gemma open-model distribution
Name: NVIDIA (provisional)
Route: Hugging Face agreement; deal not yet completed
Name: Workflow and context
Description: Permitted information and applications around the task.
Contenders
Name: Microsoft
Route: Enterprise intelligence platform
Name: OpenAI
Route: AI-native work platform
Name: Google
Route: Integrated ecosystem
Name: Amazon
Route: Infrastructure-first host
Name: Apple (provisional)
Route: Personal-device context
Name: Baidu (provisional)
Route: Search and productivity applications
Name: Tencent (provisional)
Route: Messaging and productivity route
Name: Identity, permissions, and governance
Description: Who may access data, use tools, and approve actions.
Contenders
Name: Microsoft
Route: Foundry Agent Service and Copilot Studio
Name: Google
Route: Integrated ecosystem
Name: Amazon
Route: Infrastructure-first host
Name: Anthropic
Route: Trusted model core
Name: Apple
Route: Provisional personal-device route
Name: Orchestration and tools
Description: Coordinating models, tools, and multistep work.
Contenders
Name: Microsoft
Route: Enterprise intelligence platform
Name: OpenAI
Route: AI-native work platform
Name: Google
Route: Integrated ecosystem
Name: Amazon
Route: Infrastructure-first host
Name: Anthropic
Route: MCP and multistep execution
Name: Distribution and interface
Description: Where people begin, supervise, and receive work.
Contenders
Name: Microsoft
Route: Enterprise intelligence platform
Name: OpenAI
Route: AI-native work platform
Name: Google
Route: Integrated ecosystem
Name: Amazon
Route: Infrastructure-first host
Name: Apple (provisional)
Route: Personal-device context
Name: Meta
Route: Open-weight distribution
Name: ByteDance (provisional)
Route: Creative, commerce, and work route
Takeaway
Label: Evidence boundary
Body: The map names interdependent strategic positions that the publication will test. Microsoft's multi-model access does not establish switching behavior, and NVIDIA's potential Hugging Face pathway is not a completed transaction or evidence of developer dependence. No contender is shown to have won a position, and unlike product and business metrics cannot be compared.
Source references
Source 1: Agents for Microsoft 365 Copilot
Microsoft Learn
Published: 2026
Rolling documentation covering enterprise data, actions, multi-step workflows, orchestration, models, and Microsoft 365 integration. The document can change after its verification date.
Source 2: Microsoft Foundry
Microsoft Azure
Microsoft's product page describes Foundry Models, Foundry Agent Service, Foundry Tools, and support for third-party models, tools, and frameworks. It establishes product positioning and available integration surfaces, not frictionless model switching, actual model portability, customer adoption, or workflow outcomes.
Source 3: Powering Frontier Transformation with Copilot and agents
Microsoft 365 Blog
Published: 2026-03-09
Microsoft's 2026 discussion of multi-model intelligence and model choice in Copilot.
Source 4: Introducing ChatGPT agent: bridging research and action
OpenAI
Published: 2025-07-17
Describes web interaction, deep research, conversational intelligence, and a virtual computer.
Source 5: Turn access to Google Workspace with Gemini Beta on or off
Google Workspace
Published: 2026
Rolling Google Workspace administration documentation that includes Gemini Beta actions and integrations across Workspace and third-party services. The document can change after its verification date.
Source 6: Inside the Gemmaverse: Celebrating one billion Gemma downloads
Published: 2026-08-20
Google reported that Gemma had passed one billion downloads and that developers had published more than 100,000 variants, alongside the launch of an official Awesome Gemma repository. These are vendor-reported distribution and ecosystem figures, not measures of active use, production deployment, revenue, or comparative model quality.
Source 7: Amazon Bedrock Agents
Amazon Web Services
Rolling AWS product documentation describing Bedrock Agents, including use of foundation models, APIs, data, memory, guardrails, and multi-agent collaboration. It does not establish customer adoption, production reliability, or business outcomes, and the page can change after its verification date.
Source 8: Claude for Enterprise
Anthropic
Published: 2024-09-04
Anthropic's Enterprise-plan announcement describes a 500K-token context window, SSO, role-based permissions, and administrative controls. It establishes product availability and intended controls, not deployment scale, reliability, retention, or ROI.
Source 9: The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation
Meta AI
Published: 2025-04-05
Meta's April 5, 2025 announcement makes Llama 4 Scout and Maverick available for download and describes them as its first mixture-of-experts Llama models with native text-and-vision capability. It also describes Meta AI integrations in WhatsApp, Messenger, Instagram Direct, and the Meta AI website. The announcement does not establish adoption, retention, monetization, or reliable agent outcomes. It also does not establish legal open-source classification, independent quality, or downstream safety.
Source 10: Select a primary AI model for your agent
Microsoft Learn
Published: 2026-05-28
Microsoft documentation on selecting models for agent orchestration, including workload tradeoffs and external-model administration. This rolling document was updated August 3, 2026 and its model lineup can change.
Source 11: Amazon Bedrock AgentCore Memory now supports fine-grained access control
Amazon Web Services
Published: 2026-08-28
AWS announced that AgentCore Gateway can apply OAuth JWT authentication and Cedar policies to AgentCore Memory, exposing 12 Memory operations as Cedar actions. This product announcement does not provide an independent audit, usage measure, or control-effectiveness result.
Source 12: Amazon Bedrock AgentCore Memory now supports flexible namespace variables
Amazon Web Services
Published: 2026-08-28
AWS announced that customers can define up to five namespace keys per AgentCore Memory resource to organize long-term memory by application-defined dimensions. This product claim does not establish isolation effectiveness or production uptake.
Source 13: AWS Agent Registry for centralized agent discovery and governance is now generally available
Amazon Web Services
Published: 2026-08-31
AWS announced general availability of Agent Registry in five regions as a private, governed catalog and discovery layer for agents, tools, skills, MCP servers, and custom resources. AWS describes approval workflows, CloudTrail audit trails, infrastructure-as-code, cross-account sharing, and AgentCore auto-discovery. This vendor announcement does not establish customer adoption, governance effectiveness, task reliability, or economic outcomes.
Source 14: AWS announces the public preview of AWS Interconnect - multicloud with Microsoft Azure
Amazon Web Services
Published: 2026-08-31
AWS announced a public preview of managed private AWS-to-Azure connections in four AWS regions and stated that Azure adopted the AWS Interconnect specification. This vendor announcement does not establish customer migration, application or data portability, adoption, pricing, performance, availability, resilience, or reduced switching costs.
Source 15: Introducing the Model Context Protocol
Anthropic
Published: 2024-11-25
Anthropic's announcement that it was open-sourcing MCP for connections between AI applications and external systems. Adoption of a common protocol does not itself establish secure or portable workflows.
Source 16: NVIDIA to Acquire Hugging Face
NVIDIA
Published: 2026-09-03
NVIDIA's first-party statement, authored by its founder and CEO, that it has agreed to acquire Hugging Face for $12,930,300,000. It reports platform scale of more than 18 million users, 3 million models, 500,000 datasets, 1 million applications, and 200,000 companies, and states Hugging Face will remain open with no NVIDIA-compute requirement and continued multi-cloud and multi-accelerator support. It establishes an announced agreement and the acquirer's stated intentions; it does not establish closing, regulatory clearance, consideration structure, or enforceable governance terms, and the scale figures are acquirer-reported.
Source 17: Project HydraFusion: Frontier quality via multi-model orchestration
GitHub
Published: 2026-09-04
GitHub's announcement of a research preview that performs runtime orchestration across models from multiple providers using Single, Cascade, and Critique execution patterns, available through an experimental flag in GitHub Copilot CLI. GitHub reports that on TerminalBench 2.1 it improved verified task quality by 4.9 percentage points at 67% lower estimated cost than Claude Opus 5. The results are vendor-run, offline, and cost-estimated rather than measured; a research preview behind an experimental flag is not evidence of production use or adoption.
Source 18: AWS MCP Server adds a serverless capability for AWS Lambda functions
Amazon Web Services
Published: 2026-09-04
AWS states that its MCP Server added serverless diagnostics allowing coding agents, naming Claude Code and Kiro, to inspect Lambda functions and connected resources across seven AWS services and correlate error signals against a seven-day baseline. The server runs in two regions and reaches services in all commercial regions at no additional cost. It establishes that AWS supports third-party agents as clients of its own protocol server; it does not establish portability, reliability, measured token savings, or customer adoption.
Source 19: Tensor Processing Units
Google Cloud
Google's Cloud TPU page describes purpose-built accelerators, the associated software stack and rented compute capacity. It distinguishes generally available generations from products described as coming soon. This rolling product page does not establish access to every generation or region, independent workload performance, migration effort or customer returns.
Source 20: Amazon EC2 instance types
Amazon Web Services
AWS's EC2 documentation describes Trainium-based training instances, Inferentia-based inference instances and software access through its Neuron ecosystem. Instance generations, regions and configurations have distinct conditions. The documentation establishes an offered compute route, not a current capacity allocation, effortless model migration or a matched cost comparison with another platform.
Source 21: Maia 200: The AI accelerator built for inference
Microsoft
Published: 2026-01-26
Microsoft's January 26, 2026 announcement states that Maia 200 is deployed in its US Central datacenter region and describes an SDK preview with PyTorch and Triton support. It establishes Microsoft's reported inference-infrastructure deployment and developer offer, not unrestricted accelerator rental, independent performance validation or comparable operating costs.
Source 22: Private Cloud Compute: A new frontier for AI privacy in the cloud
Apple Security Research
Published: 2024-06-10
Apple's architecture disclosure describes stateless personal-data processing, no privileged runtime access, and verifiable transparency for Private Cloud Compute. It is product-design evidence, not evidence of user trust, adoption, or an absence of implementation vulnerabilities.
Source 23: MTIA 300: Meta's First Training Chip with Built-in NICs and Communication-Offloading Engines
Meta Engineering
Published: 2026-08-24
Meta's August 24, 2026 engineering article describes MTIA 300, an in-house accelerator optimized for training ranking and recommendation models, with integrated networking and the HCCL communication library. Its production example is company-reported and workload-specific. The article does not establish a general accelerator-cloud offer, frontier-LLM performance or independently measured superiority over GPU platforms.
Source 24: OpenAI and Broadcom announce strategic collaboration to deploy 10 gigawatts of OpenAI-designed AI accelerators
OpenAI
Published: 2025-10-13
OpenAI's October 13, 2025 announcement describes custom accelerators and systems designed by OpenAI and developed with Broadcom, including networking. Deployment was targeted to begin in the second half of 2026. The agreement establishes an announced development and supply relationship, not confirmation that deployments have begun, capacity has been delivered or another supplier has been displaced.
Source 25: Introducing Apple Intelligence for iPhone, iPad, and Mac
Apple
Published: 2024-06-10
Apple's product announcement describing Apple Intelligence, personal context, cross-app actions, on-device processing, and Private Cloud Compute. It does not establish broad availability, adoption, or workflow reliability.
Source 26: NVIDIA Announces Financial Results for First Quarter Fiscal 2027
NVIDIA
Published: 2026-05-27
NVIDIA reported $81.6 billion in total revenue for the quarter ended April 26, 2026, including $75.2 billion in Data Center revenue, $60.4 billion in Data Center compute revenue, and $14.8 billion in Data Center networking revenue. This primary financial release establishes reported economic scale and momentum, not agent-related demand, customer outcomes, or a company's position as an intelligence host.
Source 27: Baidu Announces Second Quarter 2026 Results
Baidu
Published: 2026-08-18
Baidu reports RMB 12.5 billion in Q2 2026 Core AI-powered Business revenue, including RMB 7.3 billion from AI Cloud Infra, RMB 2.5 billion from AI Applications, and RMB 2.6 billion from AI-native marketing services. It also reports an enterprise DuMate launch, increased AI daily-active-user penetration across Baidu Wenku and Baidu Drive, and Apollo Go activity in 28 cities with more than 350 million autonomous kilometers. Its AI business and operating data come from internal management accounts and are unaudited; its safety characterization is also company-reported. The release does not establish product-specific agent retention, task success, workflow ownership, independent safety, or customer outcomes.
Source 28: Tencent Releases and Open-Sources Tencent Hy4 Preview
Tencent
Published: 2026-08-28
Tencent announces Hy4 preview availability through WorkBuddy, CodeBuddy, Yuanbao, ima, Tencent Cloud TokenHub, and OpenRouter. The announcement also reports an internal evaluation. It establishes distribution routes and stated product availability, not independent reliability, repeated use, or commercial outcomes.
Source 29: Tencent Corporate Overview
Tencent
Tencent's corporate overview identifies Weixin and WeChat as consumer products and Tencent Cloud as a business surface. It describes WeChat as having more than one billion monthly active users. This corporate material establishes company-reported distribution context, not an AI-agent adoption or outcome measure.
Source 30: ByteDance Products
ByteDance
ByteDance's corporate overview identifies TikTok, CapCut, TikTok Shop, Lark, and China-market products as parts of its product portfolio. It establishes distribution across creative, entertainment, commerce, and workplace surfaces, not an AI-agent product, repeated AI use, or a measured workflow outcome.
Source 31: Artificial Intelligence
Oracle
Oracle's rolling AI page presents Oracle AI and an OCI Data Science customer example. It is product and customer-story material, not a measure of broad adoption, workflow control, or agent outcomes, and can change after its verification date.
Source 32: 华为发布全球首个采用NPO的超节点——昇腾960超节点
Huawei
Published: 2026-09-17
Huawei's Chinese-language September 17 announcement introduces the Ascend 960 SuperPoD and states that CANN is fully open-source and in regular open-source community operation. This is Huawei's description of its infrastructure and software-ecosystem direction, not independent validation of workload performance, availability to every buyer, developer adoption or ecosystem advantage. The announcement does not by itself establish a contender classification.
Source 33: Agentforce
Salesforce
Salesforce's rolling product page presents Agentforce as an agent offering. It does not provide a compatible measure of active use, task success, retention, or customer return on investment, and the page can change after its verification date.
Source 34: ServiceNow Newsroom
ServiceNow
ServiceNow's rolling newsroom overview presents the company's current enterprise-AI positioning, including governance and predictable-outcome claims. It is vendor communication, not independently comparable evidence of adoption or operational outcomes, and can change after its verification date.
Source 35: watsonx.ai
IBM
IBM's rolling product page describes watsonx.ai as an integrated studio for foundation models, agent tooling, machine learning, APIs, and runtimes. This product description does not establish broad adoption, reliable task execution, or return on investment, and can change after its verification date.
Source 36: Qwen3: Think Deeper, Act Faster
Qwen
Published: 2025-04-29
Qwen announced Qwen3 models, including open-weight dense and mixture-of-experts releases under Apache 2.0. Vendor benchmark claims and model availability do not establish production adoption, governance, or commercial value.
Source 37: Grok 4
xAI
Published: 2025-07-09
xAI describes Grok 4's reinforcement-learning training, tool use, and access to X search. Its benchmark and product claims are vendor-authored and do not establish independent reliability, enterprise adoption, or durable workflow value.
Source 38: The State of Sovereign AI Adoption 2026
Cohere
Cohere discusses commissioned IDC research on enterprise and government interest in sovereign AI. The commissioned, vendor-published material does not establish market-wide adoption, product performance, or a competitive outcome, and can change after its verification date.
Source 39: Mistral Studio
Mistral AI
Mistral's rolling product page describes Studio as a catalog for models, agents, prompts, skills, datasets, and workflows with lineage and access controls. The product description does not establish usage, governance effectiveness, or commercial outcomes, and can change after its verification date.
Source 40: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
DeepSeek-AI
Published: 2025-01-22
DeepSeek-AI's technical paper describes the DeepSeek-R1 reasoning-model approach. A technical paper does not establish product distribution, enterprise governance, recurring use, or commercial outcomes.
Source 41: Mixture-of-Kittens
Cursor
Cursor describes open-sourcing its Mixture-of-Kittens mixture-of-experts training kernel and its use in Composer training. Engineering performance claims do not establish coding-agent adoption, task reliability, or commercial outcomes.
Source 42: Perplexity Hub
Perplexity
Perplexity's rolling public product-news hub records company announcements. It is not a stable measure of active use, retention, permission controls, or successful delegated work, and can change after its verification date.
Source 43: Gemini passes one billion monthly users
Published: 2026-08-11
Google reported that the Gemini app had surpassed one billion monthly users. This is a company-reported usage milestone, not an independently audited audience figure, comparative market share, or proof of task success.
Source 44: How to get Apple Intelligence
Apple Support
Published: 2026-07-07
Apple's support guidance lists supported devices and features, including intelligent Shortcuts actions and Apple Intelligence models in Shortcuts, while stating that more personal Siri features remain in development. Eligibility and availability do not measure active use, task success, retention, or leadership.
Source 45: Too Private to Tell: Practical Token Theft Attacks on Apple Intelligence
arXiv / Ohio State University-affiliated researchers
Published: 2026-04-17
Independent security-research preprint using traffic analysis and reverse engineering to demonstrate a credential-replay attack in a macOS test environment. The authors state that they disclosed the condition to Apple; current remediation and reproduction status have not been established. It is not a population-level measure of exploitation, privacy outcomes, or user adoption.
Source 46: Meta Reports Second Quarter 2026 Results
Meta Platforms
Published: 2026-07-29
Meta reported 3.60B Family daily active people and $31.08B in Q2 2026 capital expenditures. These are company-level distribution and investment context, not Meta AI active users, paid adoption, agent use, or standalone AI economics.
Source 47: How XPUs Meet a World-Class AI Factory
NVIDIA
Published: 2026-08-24
NVIDIA describes NVLink Fusion as a way to connect custom XPUs with its networking, rack, software, and operations layers. The product material makes performance and operational claims but does not establish broad deployment, customer economics, switching costs, or agent outcomes.
Source 48: Hugging Face Hub documentation
Hugging Face
Hugging Face documentation describes the Hub as a platform hosting models, datasets, and applications, with collaboration and developer tooling. It does not establish a transaction with NVIDIA, a transfer of third-party ownership or licensing rights, or any change in developer infrastructure choice.