AI Taxonomy
A practical view of AI, from energy and chips to models and real-world applications.
AI is not one thing. It is a stack. From energy and silicon to infrastructure, models, and applications, each layer enables the one above it. This page is my attempt to map the field clearly.
Strategic note · framing the stack
Energy is the hard floor. But the winner is whoever turns electricity into useful intelligence most efficiently.
The claim "whoever has the most energy wins AI" is partly true, but incomplete. A better version: whoever can turn electricity into useful intelligence at the lowest cost, fastest deployment speed and largest scale has the advantage.
AI is becoming an industrial-scale electricity problem, so energy matters. But energy alone does not win. The real edge is intelligence-per-watt, intelligence-per-dollar, intelligence-per-chip, and distribution into real products. Lack of energy can make you lose. Having energy does not guarantee you win.
Three framings of the same question, in order of usefulness.
Five compact reads on what each layer actually decides — colour-coded to the cake-stack above.
Energy
Power decides how much compute can physically run. Cheap, reliable, fast-to-connect electricity becomes a moat once data centres reach hundreds of megawatts or gigawatt scale. But raw energy without chips, infra, models and customers is just unused capacity.
Chips
Better accelerators turn the same electricity into more tokens, lower latency and cheaper training or inference. A chip advantage can beat a pure energy advantage — the question is not megawatts alone, but useful compute per megawatt.
Infrastructure
Poor infrastructure wastes expensive chips and electricity. Scheduling, networking, cooling, serving, batching, caching, observability and distributed training decide whether hardware becomes reliable AI output or idle capital.
Models
Better architectures, better data, better training, distillation, sparsity, MoE, state-space models and reasoning systems can change how much capability you get from the same compute budget.
Applications
Even frontier capability is not automatically a business. Applications turn model output into workflows, revenue, productivity, trust and distribution. The best stack still loses if it does not solve a real problem.
Three scenarios that move the bottleneck.
ConclusionBetter chips weaken the energy bottleneck per task, but may increase total AI deployment.
ConclusionInfra is what turns raw power and chips into actual throughput.
ConclusionA large architecture breakthrough could shift the bottleneck — until compute demand disappears, energy still matters.
The winning layer is usually the current bottleneck.
My read
For the next few years, energy is one of the major bottlenecks in frontier AI because data centres are becoming industrial power projects. But energy is not the whole game. The winner is not simply the country or company with the most electricity. The winner is the one that combines cheap reliable power, advanced chips, high-utilisation infrastructure, efficient models, and applications that turn capability into real economic value.
Energy will not automatically win the AI war. But lack of energy can make you lose it.
Sources
These are evidence that architecture and efficiency research is active — not proof that Transformers are obsolete.
Energy layer · deep dive
Every AI system begins as an electricity problem. Chips turn electricity into computation; data centres turn grid capacity into intelligence. The countries and companies that can secure cheap, reliable, scalable power will have a structural advantage in AI.
Eight short cards. Read them once and you will understand most newspaper energy stories better than the journalist writing them.
The fastest way to think clearly about AI infrastructure is to stop carrying these around.
Latest full-year picture of how much electricity the world generated in 2025, what it came from, and how data centres fit into the global mix.
Global electricity generation
Renewables + nuclear share
Sources: Ember Global Electricity Review 2026; IEA Global Energy Review 2026; IEA Key Questions on Energy and AI 2026.
Twenty-one countries and regions, ordered by relevance to AI infrastructure planning. Type to search by name, region, source mix or bottleneck.
| Country / region | AI power-readiness | Generation, TWh | Peak, GW | Mix (rounded) | Data-centre relevance | AI power-readiness read | Main bottleneck |
|---|
For each: what is strong, what is weak, why it matters for AI, and the one bottleneck to fix first.
A 100 MW data centre is already a major industrial load. A 1 GW AI campus is power-plant scale. Electricity turns into tokens; the price you pay flows through to every API call.
Annual energy = IT load × PUE × utilisation × 8,760 h. Pick a preset or edit the inputs. Numbers are illustrative; not a quote.
Total facility load
— MW
Annual energy
— TWh
Annual electricity cost
— / yr
Qualitative score (1–10) on four axes: available megawatts now, time-to-power, electricity price, and grid cleanliness. These are reads on the country / corridor, not a guarantee for any specific site.
AI is turning electricity from a background input into a frontline strategic constraint. The real issue is not only that AI uses more electricity. It is concentrated demand, grid-connection delays, cooling, permits, and time-to-power becoming a moat.
The AI energy problem is not only about how much electricity data centres consume. It is about where, when, and how that electricity is drawn from the grid. Globally, data centres still look like a small share of total demand. Locally, they can behave like a new industrial city appearing on the grid overnight. The constraint is no longer just chips. It is megawatts, substations, cooling, permits, and connection timelines.
By 2030, the IEA projects data-centre electricity consumption to roughly double to around 945 TWh per year — close to Japan’s current annual consumption — with AI processing the most significant single driver. Around 20% of planned data-centre projects could face delays if grid constraints are not addressed. Natural gas and renewables are expected to take the lead in powering data centres in the near term; the tech sector may also help bring forward new nuclear, SMR, and geothermal capacity. None of these alone solves the problem.
This is why time-to-power is becoming a strategic moat. The next AI winners will not only be the companies with the best models or the most GPUs. They will be the companies and regions that can secure reliable power, cooling, land, grid access, and long-term energy contracts before everyone else does.
Source: IEA — Energy and AI report; IEA Global Energy Review 2026. Percentages and projections are IEA estimates and will be revised; treat them as direction, not precision.
Solar and wind are cheap and scalable but variable. AI clusters need 24/7 reliable power. Hourly-matched clean energy is much harder than buying annual renewable certificates.
A practical pre-flight checklist. Any single missing item can kill a project for years. Cheap GPUs do not help if the substation does not exist.
Seven stages between a green-field site and the first electron flowing into a GPU rack. Times are ranges, not promises; they vary by jurisdiction, utility and equipment availability.
Sources & methodology
Figures across this section are anchored on the latest reliable year widely available in primary sources (typically 2024 actuals reported in 2025–early 2026 editions). Numbers are rounded; do not treat them as precise. Where exact 2025 values are not yet final, the latest available year is used.
Chips layer · deep dive
Chips are not just GPUs. The AI chip layer is the full compute engine: accelerator architecture, memory, packaging, interconnects, networking, software, manufacturing and supply chain. The countries and companies that secure all of it together will define who scales AI.
AI chips are not just “faster computers”. They are specialised machines for turning electricity into matrix multiplication at trillion-scale. CPUs, GPUs, and TPUs exist because different workloads need different hardware. As AI matures, training and inference are splitting into different chip-design problems.
CPU
A handful of sophisticated cores with deep control logic, large caches, branch prediction, and flexibility. Excellent for operating systems, orchestration, databases, application logic, and serial tasks. Not enough parallel arithmetic units to drive trillion-scale neural-network matrix multiplication.
GPU
Thousands of simpler cores that apply the same operation across many pieces of data at once. Originally built for graphics, where millions of pixels can be processed in parallel. Neural networks rely on the same parallel matrix math, which made GPUs the accidental engine of deep learning.
TPU / AI accelerator
Sacrifices general-purpose flexibility for efficiency on tensor operations. Systolic-array designs push data through grids of simple compute units, doing matrix multiplication with very little wasted control overhead. Higher performance per watt on AI workloads; less useful for arbitrary code.
Graphics and neural networks look unrelated, but they share the same deeper pattern: huge amounts of parallel math. Graphics applies the same transformations across millions of pixels and vertices. Neural networks apply multiply-add operations across enormous matrices. Hardware built for one turned out to be ideal for the other.
Analogy · CPU
Fast, flexible, many routes, many cargo types. The right vehicle when each trip is different and the itinerary matters more than the volume moved.
Analogy · GPU
Less flexible. Slower for any single trip. But moves enormous volume of the same cargo at once. The right vehicle when you have huge amounts of similar work to do in parallel.
A GPU is not a faster CPU. It trades flexibility for parallel throughput by stacking thousands of simple arithmetic units, feeding them data through very wide memory pipes.
The same silicon family runs both, but the optimisations diverge. Training builds the model. Inference runs it. Two different bottlenecks, two different chip-design problems.
Training
Compute-bound and interconnect-heavy. Needs raw floating-point throughput, large batch processing, high-bandwidth memory, and very fast chip-to-chip communication so gradients can be shared across many accelerators each step. Rewards scale and cluster networking.
Inference
Often memory-bound and latency-sensitive. The system must load weights, manage the KV cache for context, generate tokens quickly, and minimise cost per token. Agentic AI raises the stakes — one user task can trigger many sequential model and tool calls, so inference efficiency drives the unit economics.
AI hardware is moving from one-chip-fits-all to workload-specific design. The frontier training cluster wants maximum throughput and interconnect. The production inference fleet wants low latency, memory efficiency, reliability, and cost control. This is why the AI chip market is splitting into training accelerators, inference accelerators, hyperscaler custom silicon, edge NPUs, and specialised AI processors.
AI compute is no longer a single market. Nvidia GPUs became the default engine because they are flexible, programmable, and excellent at large-scale parallel math — and that flexibility still matters, especially for frontier training. But production AI is increasingly an inference problem: serving tokens quickly, reliably, and cheaply across millions of requests. As that side of the workload grows, specialised silicon becomes attractive — a buyer who knows the workload can trade general-purpose flexibility for better latency, utilisation, power efficiency, and cost per token.
General-purpose AI GPU
Best ecosystem, broad programmability, deep software moat — CUDA, kernels, libraries, model code — and useful across many AI and non-AI workloads. Examples: Nvidia data-centre GPUs. The default platform when the workload is varied or not yet stable.
Hyperscaler ASIC
Useful when a cloud provider controls the workload end-to-end and wants lower cost, better performance per watt, and less dependence on external chip supply. Examples: Google TPU, AWS Trainium and Inferentia, Microsoft Maia, Meta MTIA.
Wafer-scale AI chip
Pushes more compute and memory onto one very large piece of silicon to cut chip-to-chip communication overhead. Targets training and inference where interconnect is the binding constraint. Examples: Cerebras-style wafer-scale engines.
Inference accelerator
Optimised for low-latency inference, high throughput, and cost-efficient serving. Trades training generality for serving speed and tokens-per-dollar. Examples: Groq-style LPUs, SambaNova, d-Matrix, and other inference-focused ASICs.
Edge NPU
Moves smaller AI workloads onto phones, laptops, cameras, and embedded devices — lower latency, better privacy, offline use, and reduced cloud cost. Examples: Apple Neural Engine, Qualcomm Hexagon, Intel / AMD / Arm NPUs.
The durable lesson is not “Nvidia is losing” or that any single challenger will win. The durable lesson is that AI compute is fragmenting by workload. Frontier training clusters, cloud inference fleets, enterprise agents, consumer-device AI, and specialised scientific workloads will not all use the same hardware forever. Even Nvidia has shown interest in inference-specific technology through major licensing, partnership, and talent-related moves — the incumbent itself treats inference as its own category.
CPUs, GPUs, TPUs, custom hyperscaler silicon, AMD Instinct, Intel Gaudi, wafer-scale, LPUs and edge NPUs — nine cards, each with its best use, its weakness, and a concrete example.
A practical first cut, not a benchmark. Always validate on your workload, your software stack and your latency target.
The fastest way to think clearly about AI chips is to stop carrying these around.
From transistors to global supply chain. Each layer enables the one above it, and any single layer can become the binding constraint.
Vendor-published specs for the major AI accelerators. Use these to compare at a glance — not as a benchmark substitute.
Read this before reading the cards
Peak FLOPS are marketing numbers unless paired with memory pressure, interconnect, software stack, utilization and the actual workload. Two chips with the same peak BF16 number can have very different tokens-per-second on the same model.
Same silicon family, different optimisations. Below: the two workloads as visual flows, then a side-by-side read across seven axes that decide which chip you actually want.
Training
Dataset + chips + time → model weights
Inference
User query + weights + KV cache → tokens
Strategic read—
No single country can ship a frontier AI chip end-to-end. The map below traces the value chain; the cards below it read each major jurisdiction.
For each: what is strong, what is weak, why it matters for AI, and the bottleneck to fix first.
Each major jurisdiction with its role in the chain, where its leverage sits, and where its risk sits. Most leverage points are concentrated in fewer than ten countries.
A high-level read of the post-2022 export-control cycle. Use as a directional summary, not as legal guidance.
—
Twelve line items decide the actual dollars-per-token of an AI workload.
Training builds the model. Inference runs the business. The chip you want for one is rarely the chip you want for the other.
Why it matters
Pick a preset or edit the inputs. Numbers are illustrative; not a quote.
Total IT power
— MW
Annual energy
— TWh
Annual energy cost
— / yr
Total run-rate (energy + allocated)
— / yr
Strategic summary · what chips mean for AI power
The Chips layer is the conversion layer of AI: it turns electricity into computation, and computation into model capability.
Sources & methodology
Hardware specs are taken from official vendor product pages, datasheets and architecture briefs (NVIDIA, AMD, Google Cloud TPU docs, AWS Neuron docs, Intel Gaudi, Cerebras, Groq, Apple). Geopolitical and supply-chain reads draw on TSMC + ASML disclosures, SIA / CSIS reports, Reuters coverage, and the official Singapore EDB / Malaysia MIDA sector pages. "Latest available" labels apply throughout.
Infrastructure layer · deep dive
Infrastructure is where chips become usable systems. It is the layer that turns raw accelerators into reliable AI products through distributed computing, networking, storage, scheduling, serving, MLOps, observability, security and data-centre operations.
Ten foundational concepts. Read these once and the rest of the section makes more sense.
From a laptop to a hyperscale AI factory. Each rung adds an order of magnitude of complexity.
Where is your team today? Six levels from "demo" to "AI factory", with the giveaway tells for each.
Pick the row that matches your context. Each card shows the recommended pattern, the tools likely involved, the bottleneck to plan for, and the obvious things to avoid.
The fastest way to think clearly about AI infrastructure is to stop carrying these around.
From the building to the audit log. Any one layer can become the binding constraint on a real workload.
What actually happens when a distributed training job kicks off. Each stage is a place where something can fail; each is a place where good infra earns its keep.
Eleven dimensions, two columns. The most common architecture mistake is assuming the training stack is also the serving stack.
Seventeen stages between the typed prompt and the streamed answer. The model is one step. Everything else is the production system around it — and it decides speed, reliability, safety and cost.
Four symptom families, each with the most common causes. Use as a triage checklist when latency or cost regress.
The metrics an on-call engineer actually watches at 3am. Logs, metrics and traces are the three pillars; OpenTelemetry standardises all three.
Read this first
Do not confuse workflow orchestration (Airflow / Dagster / Prefect), container orchestration (Kubernetes / Nomad), and GPU job scheduling (Slurm / Volcano / KubeRay / Ray). They solve different problems. Mature AI orgs use more than one — and confusing them is one of the most common sources of "Kubernetes can't run my training" pain.
Concrete starting points. Pick the row that matches your context, then steal the building blocks. None are universally optimal; they are honest defaults.
Real cost is not just GPUs. It is silicon + servers + network + storage + power + cooling + space + ops + utilisation losses. The calculator below is an intuition tool — pick a preset, edit any field.
Pick a preset or edit the inputs. Energy = IT load × PUE × utilisation × 8,760 h. Numbers are illustrative; not a quote.
IT power
— MW
Facility power (× PUE)
— MW
Annual energy
— TWh
Annual electricity
— / yr
Annual depreciation
— / yr
Cost / GPU-hour
$—
Cost / million tokens
$—
These are intuition tools, not quotes. Real costs depend on workload, contracts, utilisation, hardware, cooling and region.
A four-column on-call diagnostic. Symptom → likely layer → possible cause → first thing to check. Read it left-to-right when something is on fire.
Thirteen anti-patterns. Most outages, blown budgets and bad postmortems trace back to one of these. Knowing them is most of the work.
Strategic summary · what infrastructure means for AI power
Without it, model capability stays as a research result. With it, model capability becomes product performance.
Sources & methodology
Tool descriptions are functional summaries of official documentation (Kubernetes, Slurm, Ray, PyTorch Distributed, JAX, NCCL, vLLM, NVIDIA TensorRT-LLM, SGLang, HuggingFace TGI, NVIDIA Triton, MLflow, W&B, Kubeflow, OpenTelemetry, Prometheus, Grafana). Vendor-claimed performance is flagged as such; benchmarks should be cross-checked against MLPerf and your own workload.
Models layer · deep dive
Energy gives AI power. Chips convert power into computation. Infrastructure coordinates computation. Models turn computation into language, reasoning, vision, code, planning, memory and action.
Ten foundational terms. Read these once and the rest of the section makes more sense.
Five mix-ups that block clear thinking about models. Learn them, drop them.
Where is your team today? Six levels from "demo" to "model operating system" with the giveaway tells for each.
The longer myth-truth list. Most bad model decisions trace back to one of these.
From data to deployment. Capability is downstream of every layer above; weakness in any layer caps the model that emerges.
Architecture uncertainty
The Transformer unlocked modern AI by replacing recurrent and convolution-heavy sequence systems with attention-based architectures. But it should not be treated as the final form. New architectures and efficiency patterns matter because they can change the compute curve. If an architecture gives a 2–5× efficiency gain, energy still matters. If it gives a 100× gain in useful capability per watt, the stack reshuffles.
Different shapes of model for different jobs. The serious products use more than one.
A modern frontier model is not one training run. It is a pipeline. Each stage is a place where capability or safety is shaped.
Three deployment shapes for the model layer. Each has real strengths and real costs — pick the one that matches your context, not the one that matches your taste.
When to choose each path. Most serious products end up hybrid.
Ten roles the frontier landscape actually breaks into. Use this as a living framework when picking a model — not a leaderboard.
For reference. The ecosystem read for each major closed lab, open-weight ecosystem, and specialist category.
Fourteen stages between a user goal and the improvement loop. The model is one stage. The surrounding loop decides reliability, cost, safety and usefulness.
Seven canonical shapes of AI product. Most real products are one of these or a composition.
Fifteen common workloads, each with a model-type recommendation, key capability, sensitivity ratings, evaluation method and what to avoid. The best AI product rarely uses one model — it uses a model system.
Five ways the model can reach beyond its weights. They are complementary, not competitors.
Seven concrete blueprint cards from chatbot to autonomous workflow agent. Steal the building blocks for your context.
Six real product flows as horizontal pipelines. Steal the lane that looks like your product.
Seventeen anti-patterns. Most blown budgets, bad postmortems and sad demos trace back to one of these.
Symptom → likely cause → first thing to check → fix pattern. Fifteen rows. The on-call triage list when a deployed model misbehaves.
Model cost depends on input + output + reasoning tokens, context length, model size, batching, cache hit rate, tool calls, retrieval, reranking, modality processing, self-host vs API, utilisation, and the evaluation + monitoring overhead. The calculator below is an intuition tool — pick a preset, edit any field.
Pick a preset or edit the inputs. Numbers are illustrative; not a quote.
Daily input tokens
—
Daily output tokens
—
Monthly cost
—
$ per active user (/1k DAU)
—
$ per task
—
These are intuition tools, not quotes. Real cost depends on provider pricing, caching, routing, latency, model mix, retries and evaluation overhead.
Strategic summary · what models mean for AI power
The winning products will not simply use the biggest model. They will use the right model system: routing, retrieval, memory, tools, evaluation, safety and cost control.
Sources & methodology
Drawn from official model docs and system cards (OpenAI, Anthropic, Google DeepMind, Meta Llama, Mistral, Qwen, DeepSeek, xAI), foundational papers (transformer / attention, diffusion, RLHF, RAG, DPO), MLCommons / MLPerf, Stanford AI Index, METR and the LMSYS Chatbot Arena (used carefully). Vendor-claimed performance is flagged. Public benchmarks change quickly; this section is a framework, not a leaderboard.
Applications layer · deep dive
Energy gives AI power. Chips convert power into computation. Infrastructure coordinates computation. Models turn computation into capability. Applications turn capability into products, workflows, agents, decisions, revenue and real-world outcomes.
Ten foundational ideas. Read these once and the rest of the section makes more sense.
Six mix-ups that block clear thinking about AI products. Replace them with the right framing.
Where is your team today? Six levels from "cool demo" to "AI-native operating system" with the giveaway tells for each.
The longer myth-truth list. Most failed AI applications trace back to one of these.
From user surface to feedback loop. Most failed AI products are weak in one specific layer; finding it is the work.
The same idea looks identical in a demo and a product. The difference is everything about how it behaves over time.
Most real AI applications are one of these shapes or a composition. Understanding the shape comes before picking the model.
AI products win when they pull on at least one of these levers strongly. Most pull on two or three.
For each industry: best AI use cases, the workflow wedge, the data advantage, the trust / risk envelope, monetisation lens and what to avoid. A briefing, not a directory.
A persistent strategic split. Horizontal serves everyone shallowly. Vertical owns one workflow deeply. Both win — for different reasons.
See also · Biological Intelligence Atlas — ants, bees, slime moulds, immune systems and embodied cognition as references for multi-agent design.
The recurring shapes that ship. Most production AI systems are one of these or a composition. Use them as starting points, not finished designs.
Start narrow. Win a workflow. Then expand. The order matters more than the speed.
The recurring traps. Each one looks reasonable up close and corrosive in aggregate. Print this. Re-read before every roadmap.
The dashboard for an AI product. User, quality, operational, business and risk. If a category is missing, you cannot tell whether the system is healthy.
For every risk, a corresponding control. Trustworthy AI applications make this pairing explicit, not implicit.
Symptom → likely cause → first thing to check → fix pattern. Sixteen rows. The on-call triage list when a deployed AI app misbehaves.
AI app cost is not just model cost. It is model + retrieval + tools + human review + ops + errors + change management. Value is hours saved, revenue uplift, errors avoided, decisions improved. Pick a preset, edit any field — these are intuition tools, not quotes.
Pick a preset or edit the inputs. Numbers are illustrative; not a quote.
Time saved / mo
—
Gross value / mo
—
Operating cost / mo
—
Net value / mo
—
Break-even price / user
—
Cost per task
—
These are intuition tools, not pricing. Real economics depend on workflow fit, model mix, retrieval / tool costs, retries, review rate, integration effort and adoption curve.
Strategic summary · what applications mean for AI power
The winners will not simply ship a chatbot. They will own a workflow — with the right context, the right interface, the right evaluation and the right economics.
Sources & methodology
Drawn from operator-grade product research and labelled vendor documentation: Stanford AI Index, Microsoft Work Trend Index and copilot research, Anthropic / OpenAI / Google product docs, NBER and academic productivity studies (where available), Y Combinator and a16z operator essays, and credible product playbooks. Vendor-claimed performance is flagged. Industry cards are illustrative briefings — they map territory, not specific deployments.
Most AI discussion stays at the top of the stack. People debate which model is best, which agent framework to use, which startup just launched. That layer matters, but it is only one-fifth of the picture.
Real leverage comes from understanding how the layers connect. A breakthrough in chip interconnects changes the economics of distributed training. A new compiler optimization shifts what model sizes are practical to serve. Cheaper energy in a specific geography changes where AI factories get built, which changes who has access to frontier compute.
The people building the most consequential AI systems are not just prompting models. They are reasoning across the full stack: spotting bottlenecks, understanding where marginal progress in one layer unlocks disproportionate gains in the layers above it. That is the kind of thinking this page is designed to support.
For a deeper look at the foundational papers behind many of these ideas, see The AI Atlas, an interactive knowledge graph of the 50 most important AI/ML papers.
Inspired by Jensen Huang's framing of AI as a five-layer industrial stack.