What an AI valuation actually means, and what it hides
A $190B valuation is the price of a 2.6% slice extrapolated across everything that did not trade. Run-rate, preferences and compute commitments are all invisible in it.
Read More
Reasoning models explained: what thinking out loud actually buys
A transformer does fixed computation per token, so extra tokens are the only way to spend more effort. That constraint explains the gains and the bill.
Read More
How AI companies make money, and which model is failing
Five business models with wildly different economics. Selling compute works, selling implementation works, and selling tokens is squeezed from both sides.
Read More
Who keeps the money in the AI stack: chips, models or apps
Silicon, infrastructure, models and applications have completely different economics. Margin concentrates where substitution is hardest, which is not where the attention is.
Read More
What it really costs to train a frontier model
The published figure is the final run. It leaves out failed runs, research compute, data licensing, annotation and salaries, which together usually cost more.
Read More
The AI compute shortage moved from chips to power
It started as a chip shortage. It is now power, permitting and financing, which is why a chip vendor ended up guaranteeing the resale value of its own hardware.
Read More
Synthetic training data: where it works and where it fails
Public human text runs out between 2026 and 2032. Generated data fixes that only where a checker exists, which is a much narrower set of tasks than the pitch suggests.
Read More
How RLHF works, and what it costs the model
OpenAI found a 1.3B tuned model was preferred over 175B GPT-3. The three stages, the alignment tax, and whose preferences get encoded.
Read More
Mixture-of-experts explained: why parameter counts stopped mattering
A trillion-parameter model can cost the same to run as a small one, because only a fraction of it fires per token. How sparsity works and what it costs.
Read More
Reasoning vs standard AI models: when the extra cost pays
Reasoning tokens bill as output. A model thinking for 2,000 tokens before a 200-token answer charges you eleven times what you see.
Read More
Open-weight vs closed AI models: what each one really costs
Fixed hardware cost against per-token billing, plus deprecation risk, data residency and the licence traps hiding inside models called open.
Read More
The AI bubble question, and the numbers that would answer it
A 27x revenue multiple, a $500B guarantee against hardware with a four to six year life, and 5x annual compute growth. What breaks first.
Read More