Unpacked Japan

Preferred Networks says it builds every layer of AI in Japan: its own chips, computers, language models and applications. As countries decide how much of their AI to build at home, PFN is a test case for building all of it. We check the bet against PFN's own papers, prices and releases, and government records.

My argument: every layer is real. PFN designed a chip that came first on the Green500 energy-efficiency list three times, it trains Japanese language models from scratch, and it sells a materials-simulation tool that, it says, more than 150 organizations use. But owning every layer is not the same as profiting from connecting them. The models we could check were trained on NVIDIA GPUs, the only named outside installation of PFN's chip we found is free to use, and the inference chip has moved from 2026 to 2027. Some connections work in trials and on an emulator, but the public evidence doesn't yet show that owning all four layers pays for what it costs. One important test: the chip planned for 2027.

Chapters

0:00 The framework PFN gave up

1:15 Three quotes, the thesis and the question

2:58 From Toyota's partner to "every layer"

5:01 The chip that set records

8:23 The chip that would matter

10:36 The model, and the stack that isn't fully stacked

13:58 What customers are actually buying

16:09 The business with the clearest adoption evidence

18:47 The money

20:18 Back to the question

23:23 Next time: SAKURA internet

Sakana AI is one of Japan's most valuable AI startups, and its homepage says it is "Building Frontier AI in Japan." But it doesn't train the giant models. We check its biggest claims against the papers, data and procurement records.

My argument: Sakana is a frontier integrator, not a frontier builder. It gets more out of other people's models than they do, a real and sometimes world-class skill. Its research is peer-reviewed, but its headline claims shrink when checked: "small beats big" holds on one task, the "AI Scientist" paper passed a workshop with a 70% acceptance rate after humans picked it from 24 drafts, and the "100× faster" CUDA code was exploiting a hole in the test. The business is real: defense and banking customers trust a Japanese partner. The open question: how long does that last once the model makers sell to Japan directly?

Chapters

0:00 The argument up front

1:13 Three quotes and the question

2:33 The real constraint is cost

4:29 Model merging is real. The headline is oversold.

7:23 A good machine, a misleading label

9:53 The search found the grader

12:29 Is the real strength the harness?

15:11 Who controls supply, who controls demand

18:08 The better question

19:56 Next time: Preferred Networks

Clarifications

12:00 "every speedup fell" should be "the average speedup fell" (3.13× to 1.49×).

18:40 "in its own contest" should be "in its other contests" (21st to 395th in eight others).

Funding, valuation and product benchmarks are Sakana's own figures.

Sources

Opening

AtCoder AHC058 results: https://atcoder.jp/contests/ahc058/results

Sakana on AHC058: https://sakana.ai/ahc058/

Lucas Beyer on X: https://x.com/giffmana/status/1892510741242036468

Three quotes

David Ha interview (Disrupting Japan): https://www.disruptingjapan.com/the-future-of-ai-looks-very-different-in-japan/

Nature Peer Review File, Referee 3: https://media.springernature.com/original/springer-static/esm/art%3A10.1038%2Fs41586-026-10265-5/MediaObjects/41586_2026_10265_MOESM2_ESM.pdf

CUDA Engineer post-mortem (JP): https://sakana.ai/ai-cuda-engineer-post-mortem/

Sakana homepage: https://sakana.ai/

Cost

Attention Is All You Need: https://arxiv.org/abs/1706.03762

Seed round: https://sakana.ai/seed-round/

Founding date (NTA registry): https://www.houjin-bangou.nta.go.jp/henkorireki-johoto.html?selHouzinNo=6010401175698

David Ha leaves Stability AI (Bloomberg Tax, 2023): https://news.bloombergtax.com/ip-law/stability-ai-head-of-research-resigns-from-startup

Namazu alpha, on costs: https://sakana.ai/namazu-alpha/

Series B: https://sakana.ai/series-b/

Dealroom, Japan's unicorns: https://dealroom.co/unicorns/japan/

Model merging

Akiba et al., Nature Machine Intelligence (2025): https://www.nature.com/articles/s42256-024-00975-8

Merge tuning set: https://huggingface.co/datasets/SakanaAI/gsm8k-ja-test_250-1319

MGSM test set: https://github.com/google-research/url-nlp/tree/main/mgsm

Independent rebuild: https://github.com/arcee-ai/mergekit/blob/main/docs/evolve.md

The AI Scientist

Original paper: https://arxiv.org/abs/2408.06292

Sakana's announcement: https://sakana.ai/ai-scientist-first-publication/

Workshop experiment: https://arxiv.org/abs/2504.08066

Lu et al., Nature (2026): https://www.nature.com/articles/s41586-026-10265-5

TechCrunch on consent: https://techcrunch.com/2025/03/19/academics-accuse-ai-startups-of-co-opting-peer-review-for-publicity/

AI CUDA Engineer

Launch page (archived): https://web.archive.org/web/20250220014413/https://sakana.ai/ai-cuda-engineer/

Released dataset: https://huggingface.co/datasets/SakanaAI/AI-CUDA-Engineer-Archive

Stricter benchmark: https://arxiv.org/abs/2509.14279

Harness and post-training

AB-MCTS: https://arxiv.org/abs/2503.04412

ShinkaEvolve: https://arxiv.org/abs/2509.19349

TAID distillation: https://arxiv.org/abs/2501.16937

TinySwallow: https://huggingface.co/SakanaAI/TinySwallow-1.5B-Instruct

Namazu: https://sakana.ai/namazu/

Brawn GP: https://en.wikipedia.org/wiki/Brawn_GP

OpenAI second at the AtCoder World Tour Finals: https://tech.yahoo.com/ai/articles/human-programmer-beats-openais-custom-182009030.html

Supply and demand

Marlin: https://sakana.ai/marlin/

Fugu technical report: https://arxiv.org/abs/2606.21228

ATLA contracts FY2025 (JP PDF): https://www.mod.go.jp/atla/data/info/ny_disti/pdf_ichiran/r07/07-ekimu-zuikei-disti-02.pdf

ATLA contract Jul 2026 (JP PDF): https://www.mod.go.jp/atla/soubiseisaku/newentry/startup_contr_r080806.pdf

MUFG partnership (JP PDF): https://www.bk.mufg.jp/news/news2025/pdf/news0519.pdf

MUFG loan-document AI: https://sakana.ai/mufg-ai-lending/

SoftBank–OpenAI venture: https://group.softbank/en/news/press/20250203_0

The better question

ALE-Agent contest history: https://atcoder.jp/users/fishylene/history?contestType=heuristic

Next time

Preferred Networks: https://www.preferred.jp/en/

Unpacked Japan: how Japan's companies and markets actually work.

Leave a comment and share your thoughts: https://open.firstory.fm/user/cmvdpz2iu0oco01xge0ybg08w/comments