Where the math is defensible.
Long-form research on live enterprise decisions. Publication is selective. Every number traces to a named source. No takes without evidence.
Frontier AI training cost trajectory 2026: the run rate, the deal stack, and the power-bound horizon
Frontier pretraining budgets crossed the half billion mark in 2025 and are heading toward one to three billion dollars per model by 2027, with cluster power, not GPUs, now the binding constraint on the next order of magnitude.
Frontier model training compute scaled at roughly 4x to 5x per year between 2018 and 2024, sat near 10x per year for the leading lab releases, and now confronts a deceleration driven by power, capital, and data, not by silicon. Epoch AI puts GPT-4 near 2e25 FLOPs at a roughly 80 million dollar training cost, Claude 3.5 Sonnet near 3e25, a...
Korea memory in 2026: Samsung versus SK Hynix, NVIDIA qualification, and the HBM share war
SK Hynix turned a two year qualification lead at NVIDIA into roughly half of the global HBM market and most of its profit pool, while Samsung is rebuilding its memory business around a delayed HBM3E ramp, an HBM4 catch up plan, and a foundry separation that signals how serious Suwon now treats the gap.
High bandwidth memory has become the single most concentrated profit pool in the AI compute stack outside NVIDIA itself. TrendForce sized the HBM market at roughly USD 16 billion in 2024 and projects USD 33 billion in 2025, with SK Hynix holding about 53 percent share, Samsung 38 percent, and Micron 9 percent in the fourth quarter of 2024...