Almost all AI coverage is about models. A model was released, it scored this on a benchmark, here is what it can do, here is a thread of examples. Very little of it is about the layer underneath, which is fabs and packaging and high bandwidth memory and power contracts and networking topology and the question of how many accelerators actually exist. SemiAnalysis covers that layer and it is close to alone in doing so at this depth.
Dylan Patel and his team publish long analytical pieces on semiconductor manufacturing, AI accelerator supply chains, datacentre buildouts, the economics of training and serving large models, memory, networking and energy. Some of it is free and the archive of free content is substantial. Full access, including the deeper analysis and the data products, is paid, and the pricing is aimed at institutions rather than individuals. The value is that it explains the constraints.
When a lab announces a model, the interesting questions are how much compute it took, what that compute cost, whether the hardware to serve it at scale exists, what the inference economics look like per token, and whether the pricing announced alongside it can possibly be sustainable. SemiAnalysis works on those questions with actual numbers, and the answers routinely contradict the narrative in the general coverage. The cost modelling is the most useful part for anyone building on these systems. Understanding what a training run costs, what serving costs per million tokens, how utilisation and batch size and context length change that, and where the margin sits, tells you which parts of the current pricing landscape are stable and which are subsidised land grabs that will be repriced.
If you are making architectural decisions that assume today's API prices persist, this is directly relevant to you even though it is written for people deciding where to put a billion dollars. The supply chain reporting has a decent track record. Analysis on packaging capacity, high bandwidth memory constraints and datacentre power availability has repeatedly been ahead of the mainstream coverage, and several pieces have shifted how the industry discussed a topic. That is a real record and it is why the publication has the readership it does.
Now the honest limitations, and the first is accessibility. This is written for people who already speak the language. Fab processes, packaging technologies, memory bandwidth, interconnect topologies, and a fair amount of finance vocabulary appear without explanation. A software engineer with no hardware background will find the first few pieces heavy going.
It is learnable, and the publication makes no accommodation for you learning it. The second is framing. This is investment research in form and function. The question being answered is generally what this means for capital, which is a legitimate and useful lens and is not the same as what this means for the technology or for the people using it.
Read enough of it and you absorb a view of the field that is entirely about spend, capacity and margin. That view is real and it is partial. The third is verifiability. A meaningful portion of the analysis rests on industry sources, channel checks and supply chain conversations that a reader cannot independently confirm.
The track record is good, which is the reason to extend credit, and it remains a different epistemic situation from reading a paper with a methodology section. Treat confident claims about what a specific company is doing as well informed reporting rather than established fact. The fourth is price. The paid tier is priced for firms with research budgets.
For an individual engineer curious about the industry, it is not a defensible personal purchase, and the free content is generous enough that this matters less than it might. Read the free archive, and if your employer has a subscription, use it. And the obvious point for a site about learning. Nothing here teaches you a skill.
You will not learn to build anything, train anything or deploy anything. What you get is context, and context is genuinely valuable in a field where most people's model of the industry comes from launch announcements and benchmark charts. Understanding why inference is priced as it is, why certain model sizes exist, why context windows cost what they cost, and why capacity constraints shape what gets released, makes you better at judging the claims that reach you. My four out of five is for consistently excellent analysis of a layer that is badly under reported, marked down for accessibility that excludes most of the people who would benefit, for an investor framing that colours everything, and for a price that puts the best of it out of reach of individuals.
Read the free pieces. They will change how you read everything else.