Follow the processors, power, and factory systems behind AI. Read the sources, compare the claims, and see what remains uncertain.
A sourced reference collection on autonomous AI agents in industrial operations — what vendors have actually shipped, what results they claim, and why the published success and failure rates cannot all be true at once.
A sourced timeline of the September 2026 Navier–Stokes announcement and the attribution dispute around it — built entirely from what the parties published themselves, with the prize's own rules for what counts as solved, and nothing adjudicated.
A sourced collection on what AI has demonstrably delivered to science and what the institutions of science have decided to require of it — built from awarding bodies and publishers rather than from the organisations that build the models.
A sourced reference collection on the buildings the accelerators go into — a North American market at 1.4% vacancy with four-fifths of its construction pipeline already pre-sold, what a watt of capacity costs to build across markets, and why the binding constraints have become electricity, consent and skilled labour rather than capital.
A sourced reference collection on gallium and germanium — two metals almost nobody writes about, sitting under compound semiconductors, fibre optics and satellite solar cells, made almost entirely in one country, and produced as byproducts of something else so that supply cannot respond to price.
A sourced reference collection on co-packaged optics — why AI networking is pulling optical conversion inside the chip package, what the power savings actually measure, and which problems still stand between sampling and deployment.
A sourced reference collection on the least glamorous component in an AI rack — why a rack-scale accelerator contains miles of copper cable, why that choice saved twenty kilowatts, why copper is now being replaced by light, and why the power busbar is a materials problem measured in kilograms.
What runs outside the model, why a slow agent is not always a busy CPU, and how to read agent-performance claims.
A sourced starter collection on the central processor — the two scaling laws that governed it, why one of them broke in 2005 and clock speeds stalled, what multiple cores bought and what Amdahl's law capped, and how the CPU became a component of somebody else's rack.
A sourced reference collection on the AI accelerator — what a GPU actually is and why it is fast, how the reticle limit forced it onto two dies, the four industries that have to deliver before one ships, what the money and the customer concentration look like in the filings, and how export control turned chip performance into a regulated quantity.
A sourced reference collection on advanced semiconductor packaging — CoWoS, chiplets, and the UCIe interconnect standard — the capacity constraint that limited AI chip production more than logic fabrication did.
A curated, evidence-first reference collection on High Bandwidth Memory (HBM): what it is, how it differs from conventional memory, who builds and standardizes it, and the major generational, production, and packaging milestones behind today's AI accelerators.
A sourced reference collection on humanoid robots in manufacturing and logistics — the documented pilots and their measured results, how many units really ship, and the distance between production capacity and work actually being done.
A sourced reference collection on how AI data centres stopped being cooled by air — the physics that made the switch inevitable, the coolant distribution unit that separates a building's water from a processor's, seven years of fleet reliability data from an operator that has actually run it, a chemicals regulation that quietly killed one cooling architecture, and the attempt to move the coolant inside the silicon.
A sourced reference collection on automated visual inspection in manufacturing — what the systems are asked to do, what they physically require, how their results should be judged, and why the measured error rates look nothing like the accuracy claims.
A sourced reference collection on DRAM — how the cell actually works and is built, how three companies came to make almost all of it, why the industry moves in cycles, and why AI repriced the memory in every device you own.
A sourced reference collection on the AI interconnect — why a training step finishes only when its slowest link does, the twenty-year doubling of InfiniBand, what RDMA over Ethernet actually costs to run, and why the biggest buyers of AI infrastructure are funding a consortium to commoditise it.
A sourced reference collection on the electricity AI actually consumes — what the measured figures are, how far the published estimates disagree, and why grid connection rather than capital has become the binding constraint on building.
A sourced reference collection on predictive maintenance in industry — the measured adoption figures, why two credible surveys disagree about whether adoption is rising or falling, and what is actually stopping deployment.
A sourced reference collection on the small chips that exist because copper traces are too short — why the usable reach of a PCIe 6.0 trace is measured in inches, what a retimer actually does to a signal, and why the least glamorous silicon in an AI rack earns a gross margin above 70%.
A sourced reference collection on the bottom of the memory hierarchy — why hard drives needed a laser to keep growing, what separates a data-centre SSD from the one in a laptop, why the storage interface had to be redesigned for a medium with no moving head, and why the accelerator now talks to the disk directly.
A sourced reference collection on the manufacturing step almost nobody writes about — why an AI accelerator occupies a tester for twenty minutes where a phone chip takes under a minute, why a memory stack must be proven good before it is assembled, and why the arithmetic of test time turned a line item into a queue.
A sourced reference collection on the semiconductor fabrication plants being built for the AI era — what companies and governments have committed, which milestones have actually been reached, and where the announced numbers and the built capacity diverge.
A sourced reference collection on custom AI accelerators — why every major cloud operator now designs its own chip, why two companies design most of them and one foundry builds all of it, and why the numbers everyone quotes about this market come from earnings calls and analyst estimates rather than from anything filed.
A sourced reference collection on the pluggable optics that carry AI traffic today — why the shortage sits in the indium phosphide laser chip rather than in module assembly, why a laser maker guides to a forty per cent operating margin while module building stays a volume business, and why nearly every shipment figure quoted about this sector is unverifiable.
A sourced reference collection on the companies that buy accelerators so other people can rent them — why they report large operating margins and large losses at the same time, why half-year capital spending runs to several times revenue, and why the numbers this sector is argued about are mostly not in its filings.
A sourced reference collection on the semiconductors that carry electricity rather than information — why AI data centres are rebuilding their power architecture around 800 volts of direct current, why silicon carbide and gallium nitride are what make that possible, and why the problem that decides whether it works is servicing a live rack rather than saving a percentage point.
A sourced reference collection on the laminate board every AI accelerator sits on — why substrate suppliers are being handed their customers' capital to build factories, why organic laminate is running out of area, and why the glass core that was supposed to replace it has consumed a billion dollars and three years without reaching a single product.
A sourced reference collection on the change that reframed high-bandwidth memory — the die at the bottom of the stack moving off a memory process onto a foundry logic process, which turned memory makers into foundry customers, made the stack customisable per buyer, and let vendors ship parts far faster than the standard they are built to.
A sourced reference collection on the supply side of the AI power problem — a closed nuclear reactor reopened on the strength of one buyer's twenty-year contract, hyperscalers ordering reactors that have never been built, gas turbine slots sold out to 2030, and the transformer queue that is the least glamorous item on the critical path.
Free browser-only utilities. Your input is not uploaded or saved.
Every claim links to the sources it rests on. Unavailable sources and disagreements stay visible; interpretation is labeled.
Methodology and limits →Suggest a topic, flag a correction, or ask us a question.
Leave a message →Join the weekly newsletter →