What Is the AI Taxonomy for Technology and Markets?

Key Takeaways

• AI taxonomies divide technology, vendors, markets, regulation, and workloads.
• The market separates chips, cloud, models, tools, services, and end users.
• Space economy links appear in data processing, autonomy, sensing, and compute.

Why the AI Technology Taxonomy Starts with Workloads

On January 15, 2026, Gartner forecast worldwide AI spending of $2.52 trillion for 2026, a figure large enough to make taxonomy more than an academic exercise. An AI technology taxonomy helps separate what customers buy, what vendors sell, what infrastructure supports the work, and what economic activity belongs in the AI market rather than in ordinary software, consulting, semiconductors, or cloud services.

The broadest definition begins with artificial intelligence as software and computing systems that perform tasks associated with learning, prediction, reasoning, recognition, generation, planning, or decision support. That definition remains useful, but it is too broad for market analysis. A search assistant, fraud detector, robotic navigation system, satellite image classifier, enterprise agent, and data center training cluster all fit under the same umbrella, yet they differ sharply in cost structure, vendor base, data demand, power use, regulation, customer behavior, and value chain position.

Workloads as the Demand Unit

The most practical starting point is the workload. A workload is the actual computing job that an AI system performs. Training builds or updates a model. Inference runs a model to produce an output. Fine-tuning adapts a model to narrower use. Retrieval-augmented generation connects a model to external information. Evaluation tests whether outputs meet quality, safety, legal, or business requirements.

This workload view matters because spending does not flow evenly through AI categories. Large-model training consumes specialized chips, advanced networking, data center power, and engineering talent. Small-model inference may run on a laptop, phone, factory gateway, enterprise server, satellite processor, or cloud endpoint. A practical taxonomy must separate a frontier model training run from a document-classification workflow, because they do not create the same market for chips, cloud services, software, data, or professional services.

New Space Economy’s coverage of AI workload types uses the same commercial logic. Some workloads push toward huge centralized clusters. Others move toward edge devices, specialized accelerators, or local enterprise systems. The AI technology taxonomy must capture that split, because the cost of serving a model output can matter as much as the model’s raw capability.

Why One Taxonomy Cannot Serve Every Purpose

A technology team sorts AI by architecture, model family, data type, and deployment pattern. An investor sorts AI by revenue pools, capital intensity, gross margin, growth rate, and customer concentration. A regulator sorts AI by risk, rights impact, transparency duty, security, and accountability. A buyer sorts AI by business function, integration burden, data exposure, vendor dependence, and measurable value.

A useful AI technology taxonomy must connect all of those views without treating them as identical. Technical categories describe how AI works. Industry categories describe who builds, hosts, distributes, audits, and sells AI. Market categories describe where money changes hands. Regulatory categories describe duties and restrictions. Workload categories link the whole structure together because every category becomes more concrete when tied to a specific job.

The Data Year Problem

The AI market changes faster than official statistics. The Bureau of Economic Analysis published a 2025 working paper explaining why AI production can be hard to measure inside ordinary national accounts. Some AI appears as software output, some as cloud service, some as semiconductor input, some as business process improvement, and some as internal production that never appears as a separate sale.

That measurement problem explains why market taxonomies often disagree. A chip sale to a cloud provider may count as semiconductor revenue, AI infrastructure spending, data center investment, or a cost inside a future AI service. A subscription to an enterprise AI assistant may count as software, productivity tooling, workflow automation, or generative AI. The taxonomy choice changes the market size.

New Space Economy’s article on measuring AI in the U.S. economy connects this issue to the broader challenge of finding AI inside production accounts and industry data. The same issue appears in company strategy. Many firms buy AI indirectly through software suites, cloud platforms, managed services, and devices, rather than through a clean line item labeled AI.

How the Technical Stack Divides Into Layers

An AI technology taxonomy needs a technical stack because every commercial category depends on a chain of inputs. Data feeds models. Models run on compute. Compute sits inside data centers, devices, satellites, vehicles, or industrial systems. Applications wrap models into customer workflows. Governance tools monitor, evaluate, secure, and document the result.

This layer view helps explain why some AI vendors look like software companies, some look like semiconductor companies, some look like cloud infrastructure companies, and some look like consulting firms. It also explains why buyers often purchase AI through packages rather than buying a single model. A company adopting AI may buy cloud credits, data services, model access, workflow software, monitoring tools, cybersecurity controls, and staff training during the same project.

Layered AI Categories

The base layer is data. Data includes public text, licensed content, enterprise documents, code, images, video, audio, sensor readings, geospatial data, transaction records, customer interactions, and synthetic data. Data quality and permissions shape model capability, compliance risk, and defensibility. In sectors such as healthcare, finance, defense and security, and space services, data access may matter more than the model itself.

The model layer includes machine learning models, foundation models, large language models, multimodal models, computer vision systems, recommender systems, robotic control models, anomaly detection systems, and planning agents. Foundation models are trained on broad datasets and can support many downstream uses. Smaller task-specific models may deliver better economics for routine work because they cost less to run and can be easier to govern.

The compute layer includes graphics processing units, tensor processing units, central processing units, neural processing units, field-programmable gate arrays, application-specific integrated circuits, memory systems, networking, storage, power, cooling, and software libraries. NVIDIA reported $81.6 billion in revenue for the quarter ended April 26, 2026, showing how concentrated AI infrastructure spending has become around accelerated computing.

Applications and Governance

The application layer turns AI capability into a usable product. It includes chat assistants, coding assistants, research tools, marketing tools, customer service systems, legal review tools, medical imaging aids, robotics software, manufacturing inspection systems, satellite analytics, cybersecurity monitoring, and business intelligence. Applications capture value by fitting AI into a workflow rather than leaving users with a raw model endpoint.

The governance layer includes evaluation, monitoring, access control, audit logging, red-teaming, safety testing, privacy review, model documentation, data lineage, model cards, policy enforcement, incident response, and vendor risk management. This layer has gained commercial weight because large buyers need evidence that systems behave consistently, meet internal policy, and comply with external rules.

The following taxonomy separates the stack by commercial function rather than by academic field. It is designed for market analysis, product planning, procurement, and policy review.

Why Stack Position Shapes Profit

Value does not settle evenly across the stack. Infrastructure suppliers earn from scarcity, scale, and technical barriers. Model providers earn from capability, brand, developer adoption, and distribution. Application vendors earn from workflow fit and customer relationships. Governance vendors earn from legal pressure, security need, and procurement discipline.

A business selling only a thin application wrapper may face pricing pressure if model providers add the same feature. A chip supplier may face cycle risk if cloud customers overbuild capacity. A model provider may spend heavily on compute and still struggle to capture margin if customers switch models easily. A systems integrator may earn steady revenue from implementation even if model prices fall.

That is why the AI technology taxonomy must include stack position. The same user outcome, such as faster document review, can create revenue for cloud providers, model companies, database vendors, application software firms, consultants, and compliance tools.

How Models Divide by Method and Behavior

Models are the visible center of AI, yet model taxonomy can mislead when it focuses only on size or brand. A clearer model taxonomy separates method, input type, output type, deployment pattern, autonomy, and economic profile. A 7-billion-parameter open-weight language model, a proprietary multimodal model, and a small anomaly detector can all generate value, but each belongs in a different market category.

The ISO/IEC 22989 standard establishes terminology and concepts for artificial intelligence. Standards like that matter because buyers, regulators, insurers, and vendors need shared language. Without shared categories, the same product may be described as AI, machine learning, analytics, automation, agent, or decision support depending on the seller’s incentives.

Model Families

Machine learning models learn patterns from data. Deep learning models use neural networks with many layers. Large language models process and generate text. Multimodal models can work with text plus images, audio, video, charts, or structured data. Computer vision models classify, detect, segment, or interpret visual information. Speech models convert speech to text, text to speech, or audio to structured outputs.

Generative AI creates text, images, video, audio, software code, molecular structures, designs, or simulations. Predictive AI estimates a future value, probability, classification, or risk. Prescriptive AI recommends actions. Agentic systems connect models to tools, memory, workflow steps, and external systems so software can perform tasks under defined control.

Model taxonomy should avoid treating generative AI as a substitute for the full AI category. Fraud detection, search ranking, recommender systems, industrial inspection, aircraft maintenance prediction, satellite image analysis, and logistics optimization predate the generative boom and remain commercially significant. Generative AI has raised attention and spending, but predictive and optimization systems still support many high-value operations.

Open, Closed, and Hybrid Models

Commercial model markets divide by openness. Closed models are accessed through hosted services or licensed products. Open-weight models provide downloadable weights under license terms, although open-weight does not always mean open-source in the strict software sense. Fully open projects may publish code, weights, data recipes, and training details, but many popular releases limit commercial use, restrict certain applications, or omit training data.

Meta’s Llama and Hugging Face have helped expand the open-weight and model-sharing side of the market. OpenAI, Anthropic, Google, and other providers compete through hosted models, developer platforms, cloud partnerships, and enterprise controls. Some buyers use closed frontier models for high-value reasoning tasks and smaller open-weight models for local, lower-cost, or privacy-sensitive workloads.

New Space Economy’s comparison of open-source and commercial AI software reflects a buyer-side reality. Open-weight systems can support control and portability. Commercial platforms can reduce operating burden, provide support, and simplify governance. Many enterprises use a mixed approach.

Capability Is Not the Same as Fit

A model’s benchmark score does not automatically determine market success. Fit depends on latency, accuracy, cost per output, context size, privacy, auditability, integration, language support, multimodal handling, and failure behavior. MLCommons provides benchmark suites for performance and measurement, but buyers still need task-specific evaluation.

For example, a legal review system may value citation accuracy, source traceability, access control, and audit records. A customer support system may value low latency, tone control, policy compliance, and cost. A satellite analytics system may value on-board performance, bandwidth reduction, and resilience under constrained power. A coding assistant may value repository context, debugging skill, and secure handling of proprietary code.

This is why the model layer belongs inside the larger AI technology taxonomy rather than standing alone. A model becomes commercially meaningful only when paired with data, compute, product design, governance, and customer demand.

How Infrastructure and Data Define Vendor Power

AI infrastructure is the most capital-intensive part of the AI technology taxonomy. It includes semiconductors, server design, high-bandwidth memory, networking, data centers, power supply, cooling, storage, orchestration software, cloud capacity, and operations talent. Infrastructure categories matter because they explain why AI growth has produced large changes in electricity demand, capital spending, supply chains, and vendor bargaining power.

The International Energy Agency projects that data center electricity consumption could reach about 945 terawatt-hours by 2030 in its base case. That forecast connects AI taxonomies to energy markets, land use, grid planning, water use, construction, and public policy. Compute-heavy AI does not exist as pure software. It relies on physical infrastructure.

Compute Categories

Graphics processing units dominate large training and many inference workloads because they handle parallel matrix operations efficiently and have mature software support. Tensor processing units, including Google’s TPU, are specialized accelerators for machine learning workloads. Neural processing units support device-level AI in phones, laptops, vehicles, and edge systems. Central processing units still matter for orchestration, data preprocessing, smaller models, and general-purpose workloads.

Memory and networking now sit close to the center of infrastructure taxonomy. Large models require fast access to model weights and intermediate data. High-bandwidth memory, optical interconnects, Ethernet, InfiniBand, storage systems, and rack-level architecture determine practical throughput. The commercial unit is shifting from a single chip to a rack, cluster, region, or data center campus.

Open hardware work also matters. The Open Compute Project supports open systems for AI, reflecting an industry push toward shared designs, rack-scale architectures, and supply-chain flexibility. That does not eliminate proprietary advantage, but it gives customers and component suppliers more reference designs.

Data Categories

Data can be sorted by origin, permission, structure, sensitivity, and update frequency. Public web data differs from licensed publisher data. Enterprise documents differ from transactional records. Satellite imagery differs from social media. Medical records differ from synthetic test data. A taxonomy that treats data as a single commodity misses the market.

Permission is becoming a sharper dividing line. Some data can be used freely. Some data is licensed. Some data requires consent, anonymization, data localization, or strict retention controls. Some data cannot be used for model training at all. The market for data rights, data cleaning, annotation, and data governance grows when buyers need evidence of lawful and reliable AI.

Data freshness also shapes product design. A model trained months earlier may lack current facts. Retrieval systems and tool-connected agents address that limitation by linking the model to databases, documents, search indexes, or software systems. That creates markets for vector databases, knowledge graphs, search systems, connectors, and access-control layers.

Infrastructure and Data in One Market Map

Infrastructure vendors often depend on data-rich customers. Model vendors need data to train, tune, evaluate, and ground systems. Enterprise software vendors hold workflow data. Cloud vendors combine infrastructure, security, data platforms, and model services. This creates a market where stack layers overlap and bargaining power shifts as customers decide where to store data and where to run models.

New Space Economy’s article on NVIDIA space computing shows that infrastructure categories are also entering space markets. On-orbit processing, satellite autonomy, and geospatial analytics require compute near sensors in some cases and ground-scale clusters in others.

The following taxonomy separates infrastructure and data by the market question each category answers. It is designed to make procurement, valuation, and policy analysis more precise.

How the Industry Network Splits Into Vendor Groups

The AI industry network is best understood as a set of vendor groups that overlap, compete, and resell one another’s capabilities. Hyperscale cloud providers sell compute and platform services. Semiconductor firms sell chips and software stacks. Model developers sell access to capability. Enterprise software companies embed AI into existing products. Data companies sell content, labeling, cleaning, and retrieval. Consulting and systems integration firms make projects operational.

This is not a clean one-company-per-category market. Microsoft Foundry, Amazon Bedrock, and Google Cloud AI infrastructure combine cloud, models, agent tools, security, deployment, monitoring, and partner model access. OpenAI, Anthropic, Mistral AI, Cohere, Meta, and other model providers distribute through direct application products, application programming interfaces, cloud marketplaces, enterprise agreements, and partnerships.

Hyperscalers and Platform Firms

Hyperscalers sit at the center because AI workloads need compute, storage, networking, identity management, security, and deployment tools. A cloud provider can host third-party models, offer its own models, sell training capacity, support data pipelines, and integrate AI into office software, developer tools, databases, analytics platforms, and customer applications.

This position gives hyperscalers commercial advantages. They already hold enterprise contracts. They can bundle AI into existing commitments. They control cloud marketplaces. They can provide compliance tooling and regional infrastructure. They can absorb large capital spending over many business lines.

The downside is cost and concentration. AI clusters need chips, power, construction, engineering labor, and long-term customer commitments. A taxonomy for cloud AI must separate training clouds, inference services, model marketplaces, data platforms, agent frameworks, and enterprise AI suites. Treating all of that as cloud AI hides both value and risk.

Model Developers and Application Vendors

Model developers compete through capability, cost, speed, context size, multimodal input, tool use, safety controls, distribution, and developer experience. Some sell direct subscriptions. Some sell token-based application programming interface access. Some use cloud partnerships. Some release open-weight models to build adoption, influence standards, or support sovereign AI goals.

Application vendors turn model capability into domain products. A law firm does not usually want a bare model. It wants legal research, document review, privilege screening, contract analysis, and workflow records. A hospital does not usually want a bare model. It wants radiology support, scheduling automation, patient documentation, coding assistance, or risk stratification under strict governance.

AI application markets often favor vendors with distribution and domain knowledge. Salesforce, ServiceNow, Adobe, Intuit, Palantir, GitHub, SAP, Oracle, and many smaller software firms can embed AI where users already work. Their market power comes from workflow ownership and customer data, not merely from model invention.

Services, Integration, and Governance Vendors

AI projects create services demand because companies need data preparation, architecture design, process change, security review, legal review, evaluation, migration, training, and support. Consulting firms, systems integrators, managed service providers, cybersecurity firms, and specialized AI implementation firms sit in this category.

Governance vendors form a distinct layer. Their tools test model outputs, monitor drift, manage prompts, track data lineage, control access, document compliance, and support audits. That market grows when regulated sectors move from experiments to production. It also grows when boards and public agencies ask for evidence that AI systems can be explained, monitored, and stopped when necessary.

New Space Economy’s article on AI vendor lock-in explains how vendor choice can shape future cost, portability, security, and operating freedom. That risk belongs inside the industry taxonomy because AI buyers often depend on model providers, cloud platforms, data connectors, and proprietary orchestration tools at the same time.

How Market Taxonomies Separate Spending from Value

AI market taxonomy must separate spending categories from value categories. Spending measures what buyers pay for. Value measures what the technology changes in productivity, revenue, cost, risk, quality, or speed. These are related but not identical. A company can spend heavily on AI infrastructure and create limited operating value. Another company can use a modest model inside a high-volume workflow and create measurable savings.

Stanford HAI’s 2026 AI Index reported that U.S. private AI investment reached $285.9 billion in 2025. That investment figure describes financing activity, not customer spending or business value. Gartner’s 2026 AI spending forecast describes spending by category. Adoption surveys describe use. Productivity studies describe economic impact. A sound taxonomy keeps those measures separate.

Spending Pools

AI spending pools include infrastructure, software, model access, data, services, devices, and embedded AI inside larger products. Infrastructure can include chips, servers, networking, power systems, cooling, construction, and cloud capacity. Software can include model platforms, application suites, development tools, databases, analytics, cybersecurity, and governance. Services can include consulting, implementation, maintenance, training, and managed operations.

Some market estimates include only AI-centric systems. Others include AI embedded in consumer devices, enterprise software, advertising, cloud, and services. The wider the definition, the larger the market. A taxonomy must state whether it counts AI as a product category, an enabling input, a feature inside other products, or an economic influence.

New Space Economy’s article on SpaceX’s AI market framing makes this distinction with total addressable market, serviceable available market, and serviceable obtainable market. That distinction is useful beyond space. Broad AI demand does not mean any one vendor can capture a large share of the full economic category.

Value Pools

Value pools sit closer to the customer. They include faster software development, lower customer service cost, better fraud detection, improved sales conversion, lower inventory waste, faster drug discovery, stronger cybersecurity triage, improved document handling, more accurate satellite image analysis, and better equipment maintenance. Value depends on adoption quality, workflow redesign, data readiness, and governance.

A market taxonomy built only on vendor revenue misses value capture by customers. If an industrial firm uses AI to reduce downtime, the AI vendor may earn modest software revenue, but the buyer may protect millions of dollars in output. If a retailer uses AI to improve demand forecasting, the market revenue appears in software and cloud spending, but the value appears in inventory economics.

This distinction matters for policy. Public agencies may care less about vendor revenue and more about productivity, jobs, competition, safety, energy demand, and national capability. Investors may care about which revenue pool has margins and pricing power. Customers may care about return on deployment, switching cost, and auditability.

A Practical Market Taxonomy

The following market taxonomy distinguishes what is purchased, what creates value, and where risks appear. It can be used for a vendor map, procurement plan, investor screen, or public policy review.

How Regulation and Standards Reshape AI Categories

Regulation and standards increasingly define AI market categories. They affect product design, customer documentation, liability, procurement rules, export controls, data handling, cybersecurity, accessibility, and sector-specific approvals. A taxonomy that ignores regulation may misclassify products that look similar technically but face different legal duties.

The European Union Artificial Intelligence Act creates a risk-based structure for AI regulation. It includes prohibited practices, high-risk systems, transparency duties, and obligations for general-purpose AI models. That legal taxonomy does not match the commercial taxonomy one-to-one. A model may be commercially general-purpose but legally subject to duties based on capability, use, distribution, or system integration.

Risk-Based Categories

Risk-based taxonomies sort systems by potential harm, level of autonomy, domain, affected population, and ability to contest decisions. A hiring tool, credit scoring system, medical triage system, border control tool, classroom proctoring system, and defense-related system may face closer scrutiny than a simple writing assistant or internal writing tool.

The NIST AI Risk Management Framework provides a voluntary structure for managing AI risk, and NIST’s generative AI profile addresses risks associated with generative systems. These frameworks create practical categories for governance tools, documentation, testing, monitoring, and procurement.

Risk categories also affect insurance. Insurers need to know whether an AI product creates professional liability, cyber risk, product liability, intellectual property exposure, privacy risk, discrimination risk, or operational risk. That creates a market for audit records, incident reporting, validation services, and independent assessment.

Policy and Classification

The OECD AI Principles and the OECD classification framework give policy makers a way to characterize systems by people and planet, economic context, data and input, AI model, and task and output. This policy taxonomy helps bridge technical design and public accountability.

Government classification also affects national statistics. Existing industry classification systems, such as NAICS, were not built around AI as a standalone sector. A company can be an AI company by product, but its official industry code may sit under software, semiconductor manufacturing, cloud services, research, consulting, advertising, healthcare technology, robotics, or defense contracting.

That mismatch affects grants, procurement, labor statistics, productivity analysis, trade data, antitrust review, and regional economic development. A city trying to attract AI investment may count data centers, software startups, research labs, chip design firms, and automation companies. A national statistician may count the same activity under multiple existing sectors.

Compliance as a Product Category

Compliance creates product markets. Vendors sell model monitoring, data lineage, risk scoring, access control, audit trails, privacy review, bias testing, secure deployment, and documentation management. These are no longer optional add-ons for many enterprise and public-sector buyers. They can determine whether a model moves from experiment to production.

This changes product packaging. A model provider that can supply safety documentation, enterprise controls, data residency, evaluation results, and security certifications may win customers even if another model scores higher on a public benchmark. A smaller open-weight model running under stronger local control may beat a stronger hosted model for sensitive data.

AI technology taxonomy should include governance as a layer and regulation as a cross-cutting category. Without those categories, the taxonomy overstates technical capability and understates adoption constraints.

How Space Economy Links Fit Into AI Taxonomies

Space-related AI belongs inside the broader AI technology taxonomy, but it adds distinct constraints. Satellites generate data, operate under limited power, face bandwidth limits, and often need autonomy because communication links can be intermittent or delayed. Ground segments handle large data flows, mission planning, cybersecurity, and customer analytics. Defense and security users may require low latency, trusted provenance, and controlled access.

New Space Economy’s coverage of satellite data analytics shows how AI can reduce bandwidth, detect objects, classify imagery, and support faster decisions. Those use cases involve edge inference, computer vision, sensor fusion, ground analytics, and mission software rather than generic chat.

AI in Space Systems

Space AI categories include on-board autonomy, on-board data reduction, image classification, anomaly detection, collision avoidance support, spacecraft health monitoring, mission planning, satellite tasking, geospatial analytics, and customer-facing data products. These categories align with general AI layers, but the operating conditions are different. Radiation, power budgets, thermal limits, launch mass, update cycles, and verification requirements shape hardware and software choices.

A satellite operator may prefer a smaller, well-tested model that reduces downlink demand over a large frontier model. A defense and security customer may prefer systems that preserve data provenance, run in controlled environments, and support audit trails. A commercial Earth observation firm may use AI to turn imagery into alerts, indices, counts, classifications, and insurance or agriculture products.

AI also connects to space infrastructure markets. On-orbit computing, optical links, ground cloud integration, and satellite edge processing create new categories. New Space Economy’s articles on orbital data centers and orbital data center failure modes place those ideas inside market and engineering constraints rather than treating space-based compute as a simple extension of terrestrial cloud.

AI as a Space Economy Input

AI can also be an input into launch operations, manufacturing, logistics, quality inspection, supply-chain forecasting, mission assurance, customer support, and regulatory documentation. In this sense, AI resembles cloud computing or cybersecurity. It supports many parts of the space economy without becoming a space product itself.

The space economy value chain divides space activity into upstream and downstream segments. AI appears in both. Upstream uses include design, manufacturing, testing, mission planning, and operations. Downstream uses include imagery analytics, navigation services, communications optimization, weather products, and defense and security applications.

The AI technology taxonomy helps space companies avoid exaggerated claims. A satellite company using computer vision is not automatically an AI platform company. A launch provider using predictive maintenance is not automatically an AI infrastructure provider. An orbital compute concept must still answer workload, latency, power, radiation, repair, cost, customer, and regulatory questions.

Where Space AI May Create Distinct Markets

Distinct markets may appear where space conditions create needs that ordinary cloud AI does not address. On-board inference can reduce downlink costs. Satellite autonomy can reduce operator burden. Geospatial foundation models can support faster analysis of Earth observation data. Secure processing close to sensors may attract defense and security demand. Space-based compute may interest customers only when the workload can tolerate latency, data movement, maintenance limits, and launch economics.

New Space Economy’s article asking which AI workloads stress-test orbital data centers points to the right taxonomy question. The issue is not whether AI is large. The issue is which workloads can justify a space-based architecture after cost, latency, regulation, repair, and competition from ground data centers are considered.

How Buyers Can Use AI Taxonomies

Buyers need AI taxonomy because vendor language often blends models, applications, automation, analytics, agents, and platform services. A disciplined taxonomy reduces confusion. It helps procurement teams ask what is being purchased, what data is used, where the system runs, who controls outputs, what risks apply, and how success will be measured.

The most useful buyer taxonomy begins with the business function. A customer service deployment, software development tool, fraud detection model, satellite imagery workflow, and hospital documentation assistant all require different evaluation criteria. After that, buyers should identify workload type, model dependency, data exposure, integration burden, governance requirements, vendor concentration, and cost model.

Procurement Questions

A buyer should identify whether the system uses a hosted model, open-weight model, local model, embedded vendor model, or mixed model architecture. That choice affects cost, privacy, auditability, data residency, performance, and switching options. A buyer should also ask whether the system stores prompts, trains on user data, connects to internal databases, executes actions, or makes recommendations that affect people.

Cost taxonomy matters. Some systems charge by seat. Others charge by token, usage, compute hour, application programming interface call, workflow, document, endpoint, or enterprise license. A low subscription price may hide high inference cost. A low token price may become expensive if workflows generate long context windows, repeated tool calls, or agent loops.

Governance taxonomy should be part of procurement rather than an afterthought. Buyers need monitoring, evaluation, audit records, access controls, incident response, change management, and human review rules. These controls become more important when AI affects customers, employees, citizens, patients, students, passengers, investors, or security operations.

Build, Buy, or Partner

The AI technology taxonomy helps decide whether to build, buy, or partner. Building makes sense when the organization has unique data, high-value workflows, sufficient talent, strong security requirements, or a need for differentiation. Buying makes sense when the need is common, the vendor product is mature, and integration burden is manageable. Partnering makes sense when the buyer needs a combination of domain expertise, vendor tooling, and internal data.

Enterprises often use multiple approaches at once. A bank may buy an AI assistant for office work, build fraud models internally, partner on contact-center automation, and use cloud-hosted models for software development. A satellite operator may build mission-specific analytics, buy ground cloud services, and partner with a geospatial AI firm.

The taxonomy also supports exit planning. If a vendor changes price, model access, terms, location, or functionality, the buyer needs to know which workloads can migrate and which are locked in. This is a business risk as much as a technical risk.

Summary

AI taxonomies matter because artificial intelligence is no longer a single product category. It is a stack of data, models, compute, platforms, applications, services, governance, and end-user workflows. The AI technology taxonomy provides the technical map, the industry taxonomy identifies the vendor groups, and the market taxonomy separates spending from value.

The strongest taxonomy begins with workloads. Training, inference, retrieval, fine-tuning, evaluation, agent execution, edge processing, and sensor analysis create different cost structures and vendor needs. From that base, technical layers show how data, models, infrastructure, platforms, applications, and governance fit together. Market layers then show where customers spend money and where they capture value.

Regulation and standards now shape the taxonomy as strongly as technology does. The EU AI Act, NIST AI risk work, OECD classifications, ISO terminology, and national measurement efforts all show that AI categories are becoming legal, statistical, and operational categories. Buyers, investors, policy makers, and vendors need shared language because the same system can be a model, a cloud service, an application feature, a regulated tool, and an infrastructure workload at the same time.

Space economy applications make the taxonomy more concrete. Satellite AI, on-board processing, geospatial analytics, orbital compute concepts, and defense and security use cases show that AI value depends on location, latency, power, data movement, and mission context. The same lesson applies in healthcare, finance, manufacturing, energy, government, and enterprise software. AI markets become clearer when they are mapped by workload, stack position, customer value, and governance duty.

Appendix: Useful Books Available on Amazon

• Artificial Intelligence: A Modern Approach
• Prediction Machines
• Power and Prediction
• Competing in the Age of AI
• Deep Learning

Appendix: Top Questions Answered in This Article

What Is an AI Technology Taxonomy?

An AI technology taxonomy is a structured way to classify AI by workload, data, model type, infrastructure, platform, application, and governance. It helps buyers and analysts separate a model from the cloud service, data source, chip cluster, software product, or compliance layer that supports it. Without taxonomy, AI markets can look larger and less precise than they are.

How Is an AI Market Taxonomy Different?

An AI market taxonomy classifies where money flows and where value appears. It separates chips, cloud capacity, model access, data products, software, services, governance tools, and customer benefits. This matters because spending on infrastructure may grow quickly even when end-user productivity gains appear later or appear in a different company’s financial results.

Why Do AI Workloads Matter So Much?

Workloads define what an AI system actually does. Training, inference, retrieval, evaluation, fine-tuning, and agent execution use different infrastructure and create different cost patterns. A workload-based taxonomy prevents buyers from treating a low-cost document workflow and a frontier-model training cluster as the same kind of market demand.

Where Do Models Fit in the Taxonomy?

Models sit between data and applications. They convert inputs into predictions, text, images, code, recommendations, classifications, or actions. Model categories include foundation models, large language models, multimodal models, computer vision systems, speech systems, anomaly detectors, and task-specific machine learning models.

Why Is Compute a Separate AI Category?

Compute is separate because AI workloads require physical infrastructure, not just software. Chips, memory, networking, data centers, power, cooling, and cloud orchestration all affect cost and performance. Compute-heavy AI demand has become large enough to influence energy planning, capital spending, semiconductor supply, and data center construction.

How Do Regulations Change AI Taxonomies?

Regulations classify AI systems by risk, use, transparency duty, and affected population. A system used for hiring, credit, healthcare, education, border control, or public services may face stricter duties than a simple writing assistant. Regulation can turn governance, monitoring, documentation, and audit tooling into important product categories.

Why Are Open-Weight Models Commercially Important?

Open-weight models allow organizations to run or adapt models with more control than fully hosted closed systems often allow. They can support local deployment, cost management, data control, and portability. Their value depends on license terms, operating skill, infrastructure access, and whether the model fits the workload.

Where Do AI Applications Capture Value?

AI applications capture value by embedding models into work that customers already need to perform. Examples include software development, customer support, document review, satellite image analysis, fraud detection, medical documentation, and factory inspection. Workflow fit often matters more than model size because adoption determines whether the product changes operating results.

How Does AI Connect to the Space Economy?

AI connects to the space economy through satellite autonomy, on-board processing, mission planning, geospatial analytics, defense and security workflows, and orbital compute concepts. Space applications face constraints involving power, radiation, latency, bandwidth, and launch cost. Those constraints make workload taxonomy more important than broad claims about AI demand.

Why Do AI Market Estimates Differ?

AI market estimates differ because analysts count different things. Some include only AI-centric software, hardware, and services. Others include AI embedded in devices, enterprise software, cloud platforms, advertising, or productivity gains. A clear taxonomy states whether it measures vendor revenue, infrastructure spending, customer value, financing, or economic impact.

Appendix: Glossary of Key Terms

AI Technology Taxonomy

An AI technology taxonomy is a structured classification of AI by technical layer, workload, deployment pattern, and governance need. It helps separate data, models, compute, platforms, applications, and operating controls into categories that can be compared.

Agentic AI

Agentic AI refers to systems that use models with tools, memory, planning steps, and software actions. These systems can perform multi-step tasks under defined controls, although their reliability depends on evaluation, permission design, monitoring, and human oversight.

Application Programming Interface

An application programming interface is a defined way for one software system to access another system’s functions or data. In AI markets, APIs often let developers send prompts, retrieve outputs, call tools, and integrate models into products.

Foundation Model

A foundation model is trained on broad data and can support many downstream tasks. It may be adapted through prompting, retrieval, fine-tuning, tool connection, or integration into an application workflow.

Inference

Inference is the process of running a trained model to produce an output. It may generate text, classify an image, recommend an action, detect an anomaly, or support a decision in a live application.

Large Language Model

A large language model is a model designed to process and generate text at scale. Many can also support reasoning, coding, tool use, document handling, and multimodal workflows when connected to additional systems.

Multimodal Model

A multimodal model can work with more than one input or output type. Common modes include text, image, audio, video, charts, code, and structured data used in business or scientific workflows.

Retrieval-Augmented Generation

Retrieval-augmented generation connects a model to external information before generating an output. It can improve freshness and source grounding by retrieving relevant documents, database entries, or knowledge records.

Training

Training is the process of building or updating a model from data. Large training runs can require specialized chips, fast networking, high-bandwidth memory, large data centers, and substantial engineering work.

Workload

A workload is the actual computing task performed by an AI system. Examples include training, inference, fine-tuning, retrieval, evaluation, image classification, agent execution, and anomaly detection.