As of 30 November 2025, the prediction window (≈Dec 2024–Jun 2026) is still open, so it is too early to say definitively whether it will end up correct. That said, available evidence mostly points against a 100× cost drop so far.

1. Frontier training costs have not fallen 100×

Public estimates for training top models show rising (or at best slightly reduced) dollar costs for frontier-scale training runs, not a collapse from ~$100M to ~$1M:

• The 2025 AI Index (summarized by Visual Capitalist) puts GPT‑4 (2023) at about $79M, Gemini 1.0 Ultra (2024) at ~$192M, Llama 3.1‑405B (2024) at ~$170M, Mistral Large (2024) at ~$41M, and Grok‑2 (2024) at ~$107M for a single training run. These are all still in the tens to hundreds of millions of dollars range, not near $1M. (visualcapitalist.com)
• A detailed cost model from Epoch AI finds that the amortized cost to train the most compute‑intensive frontier models has grown ~2.4× per year since 2016, with GPT‑4 and Gemini dominated by tens of millions in accelerator and staff costs. It projects > $1B training runs by around 2027 if trends continue. (arxiv.org)
• Other 2025 summaries similarly estimate GPT‑4‑class training in 2025 at roughly $50M–$200M per run, even after efficiency gains—far from a 100× cost collapse. (localaimaster.com)

Given that GPT‑4‑type runs were already >$100M in 2023, current public estimates for similar‑capability training runs remain well above $1M; in some cases they are higher than 2023 figures.

2. Some individual models claim big savings, but not clearly 100× and not uncontested

• DeepSeek‑V3 / R1 (China): DeepSeek claims it trained V3 for about $5.6M in GPU‑rental‑equivalent cost using ~2,048 H800s over ~2.8M GPU‑hours, with performance comparable to GPT‑4‑class systems. (en.wikipedia.org) That’s roughly a 10–20× reduction vs GPT‑4‑style estimates (> $100M), but:
  • The figure explicitly excludes prior R&D, ablations, and infrastructure, and has been criticized as misleading; some analyses argue that total capital outlay may be orders of magnitude higher. (reddit.com)
  • Even if you take the $5–6M figure at face value, it is still short of the 100× drop (to ~$1M) the prediction specified.
• MiniMax M1 (Shanghai): MiniMax has claimed that its M1 model—reported as competitive with top frontier models—was trained for about $534,700, nearly 200× cheaper than some >$100M GPT‑4/4o cost estimates. (ainvest.com) However, these numbers are self‑reported, not independently audited, and it’s unclear how directly comparable the task, scale, and quality are to the 2023 $100M reference runs.

These examples show pockets of large efficiency gains and at least claimed 20–200× reductions for specific architectures and vendors. But they are:

• Not yet clearly established as industry‑wide norms.
• Often measured on different architectures and training setups rather than “the exact same run that cost $100M in 2023 now costs $1M.”

3. Direct same‑model comparisons suggest slower cost decline

Where we do have like‑for‑like comparisons, the cost reductions are large but not on the 100× / 18–36‑month schedule Friedberg predicted:

• Andrej Karpathy’s 2024–25 reproduction of GPT‑2 on a single 8×H100 node brought the training cost down to about $672, versus earlier estimates of tens of thousands of dollars for the original 2019 training, an ~90× reduction over roughly five years, not 100× in 1.5–3 years. (tomshardware.com)

This suggests that per‑FLOP or per‑model cost can fall substantially over time, but the observed pace so far is much slower than “100× cheaper in 18–36 months” for a fixed large model.

Why this is labeled inconclusive rather than wrong

• Friedberg’s window runs to June 2026, about 7 months after the current date (Nov 2025), so the prediction period is not over yet.
• So far, public data on major frontier models shows no broad 100× drop in training cost for a run that would have cost ~$100M in mid‑2023; if anything, headline frontier runs are as expensive or more expensive. (visualcapitalist.com)
• There are emerging claims (MiniMax, DeepSeek) that individual models with near‑frontier capability were trained at 10–200× lower marginal compute cost, but these are either below the 100× threshold (DeepSeek at ~$5–6M) or not yet well‑verified (MiniMax’s ~$0.5M claim).

Given the remaining time in the window and the lack of clear, widely accepted evidence that a $100M‑class 2023 run is now reproducible for ~$1M, the fairest judgment today is “inconclusive (too early)”, with current trends pointing against the prediction ultimately being met without a dramatic further breakthrough in the next several months.