Elena' s AI Blog

The AI Landscape: Model Vendors, Hardware and Key Concepts

Elena Daehnhardt

 Midjourney AI-generated art
Image credit: Illustration created with Midjourney, prompt by the author.
Image prompt

“An illustration representing cloud computing”

The AI Landscape

AI news moves quickly, but the cast of players changes slowly. This map shows who builds the models, who makes the chips they run on, and the concepts that tie it together. Select any node to see a short explanation and the posts on this blog that cover it.

Select any node to see what it is and which posts on this blog explore it.

Model vendors 10 topics

The labs that train and ship foundation models, from closed frontier systems to open-weight families you can run yourself.

OpenAI Frontier lab

Maker of the GPT models and ChatGPT, plus Codex for coding and Sora for video. Its models set the pace for consumer chat assistants.

Anthropic Frontier lab

Maker of the Claude family and Claude Code. Known for its focus on safety research, long context and agentic coding.

Google DeepMind Frontier lab

Builds the Gemini models and the open Gemma family, and designs its own TPU chips, so models and hardware are developed together.

Meta Open-weight lab

Releases the Llama models with open weights, which made Llama a common starting point for local and fine-tuned models.

xAI Frontier lab

Builds the Grok models, tightly integrated with the X platform and trained on very large GPU clusters.

Mistral AI Open-weight lab

European lab shipping both open-weight and commercial models, popular for efficient models that run on modest hardware.

DeepSeek Open-weight lab

Chinese lab best known for efficient open-weight reasoning models that showed strong results at a fraction of the usual training cost.

Alibaba Qwen Open-weight lab

Alibaba's Qwen family spans many sizes, from phone-class to very large models, mostly released with open weights.

Moonshot (Kimi) Open-weight lab

Chinese lab behind the Kimi models, focused on long context and agentic tool use, with open-weight releases.

Zhipu (GLM) Open-weight lab

Chinese lab behind the GLM models, which compete closely with Western models on coding benchmarks.

AI hardware 10 topics

The chips and foundries that decide how fast models train, how cheaply they run, and where they can run at all.

NVIDIA GPU maker

Dominant supplier of AI training and inference GPUs. Its CUDA software stack is as important to its position as the chips.

AMD GPU maker

The main merchant-GPU alternative to NVIDIA, with Instinct accelerators and the open ROCm software stack.

Google TPU In-house accelerator

Google's custom tensor processors, built for training and serving its own models and offered through Google Cloud.

AWS Trainium In-house accelerator

Amazon's custom training and inference chips, used to lower the cost of running models on AWS.

No blog posts on this yet.

Apple silicon On-device chips

M-series chips combine CPU, GPU and neural engine around unified memory, which makes laptops and desktops surprisingly capable for local models.

Qualcomm On-device chips

Supplies the neural processing units in many phones and laptops, and is pushing into data-centre inference.

Intel CPU and accelerator maker

Still ships most of the world's server CPUs and builds Gaudi accelerators and NPUs for PCs.

No blog posts on this yet.

Cerebras Specialist chip maker

Builds wafer-scale processors aimed at very fast model training and low-latency inference.

No blog posts on this yet.

TSMC Foundry

The foundry that manufactures most leading AI chips, so its capacity and advanced packaging are a bottleneck for the whole industry.

Huawei Ascend In-house accelerator

China's most prominent domestic AI accelerator line, central to efforts to reduce reliance on imported GPUs.

Key concepts 17 topics

The ideas that explain how models are built, adapted, connected to tools and kept under control.

Neural networks Foundation

Layers of simple units whose connection strengths are learned from data. Every model on this map is built from them.

Transformers Architecture

The attention-based architecture behind modern language models. Attention lets each token look at all the others in the context.

No blog posts on this yet.

Large language models Model type

Transformers trained on huge amounts of text to predict the next token, then adapted to follow instructions.

Training and fine-tuning Process

Pre-training teaches general patterns; fine-tuning, including parameter-efficient methods such as LoRA, adapts a model to a task or style.

Reasoning models Capability

Models trained to spend extra compute thinking step by step before answering, which helps on maths, code and planning.

Context and tokens Concept

Models read text as tokens and can only attend to a limited context window. Tokenisation and window size affect cost and quality.

RAG Technique

Retrieval-augmented generation fetches relevant documents at question time and gives them to the model, grounding answers in your own data.

AI agents Pattern

Models that plan, call tools and act over several steps. Useful, but they widen the security surface.

MCP Protocol

The Model Context Protocol is an open standard for connecting models and agents to tools and data through small servers.

Multimodal AI Capability

Models that handle text, images, audio and video in one system, including image generators built on diffusion.

Open weights Release model

Models whose trained weights can be downloaded and run locally, in contrast to API-only systems. Licences differ, so read them.

Quantisation Optimisation

Storing weights with fewer bits (for example 4-bit) so large models fit on smaller GPUs and laptops, at a small cost in quality.

No blog posts on this yet.

Local AI Deployment

Running models on your own machine with tools such as Ollama, for privacy, cost control and offline use.

Compute and GPUs Infrastructure

Training and serving models is limited by accelerator supply, memory bandwidth and power, which shapes costs and who can compete.

Hallucinations Risk

Fluent but false output. Grounding with retrieval, citations and verification reduces it but does not remove it.

Prompt injection Risk

Untrusted text that tricks a model or agent into following the attacker's instructions, especially dangerous when tools are attached.

Safety and regulation Governance

Evaluations, guardrails and laws such as the EU AI Act that set expectations for how powerful models are built and deployed.

Last updated: 2026-10-02. Nodes without posts are topics I have not written about yet.

How to read the map

  • Model vendors are the labs that train foundation models, from closed frontier systems to open-weight families you can run yourself.
  • AI hardware covers the chips and foundries that decide how fast models train and where they can run.
  • Key concepts explain how models are built, adapted, connected to tools and kept safe.
  • Dashed lines show connections, for example between a vendor and the chips it depends on.

For step-by-step reading, see the blog series.

All Posts
Copyright © Elena Daehnhardt
Privacy and Terms