# flash costs half of claude, and the data center promises to shut off > the whole day is cost: tokens at $0.75 per million, 100 mw that shut off in virginia, ascend with 2 petaflops and a brazilian ledger that fits in 8 bytes. - edition: Saturday, September 19, 2026 (2026-09-19) - notebook: dev & ai - topics: llm · market · infra · brazil - items: 11 from 9 sources - original: https://tonho.wtf/en/daily/2026-09-19/ - portuguese edition: https://tonho.wtf/diario/2026-09-19/ - authorship: written by an llm pipeline, reviewed and translated by antonio leandro (tonho.wtf) --- Five of today's items say the same thing in different units. Google cut the price of the token, Huawei promises cheap flops in sanctioned territory, an energy alliance wants to cut the peak watt, a PyTorch commit wants to cut the passes over the logits buffer, and a TabNews post wants to cut the ledger record down to eight bytes. It's the same problem — what does it cost to run this — seen from five layers of the stack, and none of the five solves the other's bill. What matters is where the arithmetic doesn't add up. t3n registers the catch with Flash: the price per token stayed the same as the previous version, but the new model applies more compute to complex tasks, so the price of the token didn't move and the price of the work may have gone up. On the power grid the gap is the same in kind, only bigger: Emerald AI talks about freeing up a hundred gigawatts from the existing American grid, and the first real test is a 100 MW site in Virginia. Three orders of magnitude between the promise and the pilot. No paper came through today, so there's no research section. ## labs **[Gemini 3.8 Flash](https://t3n.de/news/google-gemini-3-8-flash-halb-so-teuer-claude-1761336/?utm_source=rss&utm_medium=newsFeed&utm_campaign=newsFeed)** — Third Flash release in a few weeks (3.7 shipped three ago), and the price didn't change from its predecessor: $0.75 per million input tokens and $3.75 on output, against $5 and $25 for Claude Opus 5 and $4 and $20 for GPT-5.6 Sol. Google says it's its best reasoning and coding model: it leads HLE-Verified and sits just behind Opus 5 on DeepSWE v1.1. t3n's caveat is the one that matters for an agent budget: more compute per hard task means more tokens consumed, and the cost of the job may not follow the cost of the unit. **[Gemini 3.8 Flash Cyber](https://t3n.de/news/google-gemini-3-8-flash-halb-so-teuer-claude-1761336/?utm_source=rss&utm_medium=newsFeed&utm_campaign=newsFeed)** — Same base model, trained to find and fix vulnerabilities, with deliberately less restrictive guardrails. It solves 86.2% of Cybergym in one attempt (GPT-5.5-Cyber: 85.6%) and 47.2% on CWE-Bench, behind Fable 5. It isn't public: it ships only through the Fairwind Program, to more than 650 partner companies, among them Crowdstrike and Snowflake. It's the third lab to hold a specific capability behind a guest list. **[linear_cross_entropy routed in PyTorch](https://github.com/pytorch/pytorch/releases/tag/trunk%2F717d7a288d82e60e92d7936cf41aaf3ec4b23c2d)** — The commit installs the `torch._native` override for the two chunked versions of the loss and nothing else: the implementation still delegates to eager, and the text says explicitly that no performance change is expected or claimed. The reason it exists is in the description of the target: `linear_cross_entropy` puts behind a single symbol one matmul (N, F) x (F, C), a row-shifted softmax, the reduction and three gradient matmuls, and the eager loop walks the logits buffer (B, C) about seven times per chunk before the gradients read it again. Separating the routing from the kernel is so that a future regression is attributable to one or the other. A footnote that has become routine by now: the commit was written by Claude Code and reviewed and approved by pearu. ## brazil **[JEC Enterprise 64-bit](https://www.tabnews.com.br/mlopes75/otimizando-sistemas-de-pagamento-de-alta-escala-com-bit-packing-conheca-o-jec-enterprise-64-bit)** — A reply to a guide on payment systems peaking at 25 thousand transactions/s and 500 GB/day of ledger: pack a microsecond timestamp plus 7 bits of user space (microservice or region id) into a native 64-bit word, instead of a 16-byte UUID with a mapping table. The part that makes the post worth it is the edge case: Dart treats 64-bit ints as signed, EVM as unsigned, and in Postgres's `BIGINT` half the header space sorts inverted because of the sign bit. The fix is a mask in the order clause, `ORDER BY (jec_id & x'7FFFFFFFFFFFFFFF'::int8)`, which keeps the ordering for free. **[Data contract for crypto market data](https://www.tabnews.com.br/musheghmanukyan/como-projetar-dados-de-mercado-cripto-que-continuam-confiaveis-quando-a-api-falha)** — The problem isn't the API that goes down, it's the one that answers 200 OK with a stale price, volume in a different quote currency or the last known value repeated without warning. The proposal is to treat freshness as a field, not as an assumption: separate `observed_at` and `received_at`, explicit states (`fresh`, `delayed`, `stale`, `unavailable`), a bounded `stale-if-error` window, provenance that survives normalization, and zero that never stands in for absence. It holds for any integration with an external source, not just crypto. The author runs ARMCP.net, and that's declared in the post. **[Women pay the bill for the betting epidemic](https://www.canalmeio.com.br/2026/09/19/mulher-paga-conta-da-epidemia-de-bets/)** — A piece by Mariliz Pereira Jorge at Meio, with the text behind a subscription: 42% of Brazilian women already bet and 67% say the platforms are blowing up families. The best-optimized digital product in the country keeps being measured by a unit that shows up in no other item today. ## market **[Tin](https://planetscale.com/blog/introducing-tin)** — PlanetScale shipped to GA a full-text search extension for Postgres, with boolean expressions, phrase, fuzzy, wildcard, regex, `COUNT(*)` and top-k by BM25 — the list exists because, according to the post, none of the three text indexes that already exist for Postgres covers all of it. The benchmark numbers are theirs: an 85 GB corpus with 150 million Stack Exchange documents, an 8 vCPU and 32 GB container, a 50.7 GB index built in 8m10s, against 19m20s for ParadeDB and 2h09m for GIN — and the other three only finished building with 64 to 128 GB of RAM. On a mixed top-10 query it does 25× ParadeDB's QPS with p99 26× lower; with concurrent writes at 1,000 UPDATEs/s, it completes 270,279 updates in ten minutes against 735 for pg_textsearch, which stalls because continuous reads never release the lock. A vendor benchmark is a vendor benchmark, but the decision it goes after is concrete: don't stand up an Elasticsearch next to the database. **[Ascend 960](https://korben.info/huawei-avance-ses-puces-dia-ascend-960-pour-doubler-nvidia-en-chine.html?utm_source=rss&utm_medium=feed&utm_campaign=flux-complet)** — The 960DT arrives in the first quarter of 2027, nine months ahead of schedule, with 2 petaflops, 288 GB and 9.6 TB/s of bandwidth — against the roughly 17 petaflops Nvidia promises for Rubin in the same form factor. The bet isn't chip against chip: it's UnifiedBus and the SuperPoD Atlas 960, a cluster of more than 4,000 accelerators over optical fiber, even though the 15,000-chip machine talked about last year has shrunk. The chip came early and the cluster got smaller. For the other moat, Huawei open-sourced CANN, its equivalent to CUDA, already compatible with PyTorch. Nobody knows who etches these wafers without TSMC, and the company itself admits it doesn't produce enough for Chinese demand. **[AI Energy Management Alliance](https://korben.info/data-centers-dia-google-et-nvidia-promettent-de-moduler-la-conso-pour-eviter-le-black-out.html?utm_source=rss&utm_medium=feed&utm_campaign=flux-complet)** — Google, Nvidia and Emerald AI pulled together a couple dozen players, including Anthropic, Constellation and National Grid, to standardize the flexible data center: pause, slow down or shift the least urgent compute during hours of grid stress. The trade is explicit — a faster connection, given that in the US a new data center can wait up to ten years to be switched on. Emerald AI sells the software that orchestrates the flexibility and just raised $150m above a $1bn valuation, which takes it out of the referee role. First test by the end of the year: about 100 MW in Virginia, with Nvidia, Emerald and Digital Realty. ## world **[Trump announces an "AI Force" and an AI czar](https://www.bbc.co.uk/news/articles/cqlykr2vrv04o?at_medium=RSS&at_campaign=rss)** — The announcement lands in a week when the pressure was coming from inside the industry: Dario Amodei asking for a slowdown and independent monitoring, with Sam Altman and Elon Musk saying they agree; Jack Clark telling the BBC that a third-party-verifiable kill switch may need to be mandatory; and Altman saying Washington can't keep up with the pace and that the world should trust companies to do the right thing. In the UK, the Joint Committee on Human Rights published a report calling for a dedicated AI law. Opinion: Google's Fairwind, in the labs section, is exactly the shape that comes out of this deadlock — the dangerous capability stays with 650 partners chosen by the company, and public oversight gets a task force and a czar. ## who wrote **[What Zig felt like, coming from Rust](https://besok.github.io/posts/what-zig-felt-like-coming-from-rust/)** — Seven years of Rust, and a reimplementation of jsonpath-rust as zig-jsonpath (RFC 9535) to compare on familiar ground. What's most surprising in the account isn't performance: it's the near-nonexistent IDE support, which pushed the project toward the CLI and `build.zig`, and the flattened structure — five files in Zig against a tree of almost twenty in Rust, because the threshold for needing hierarchy turned out to be much higher. The real loss is the functional paradigm: where Rust uses combinators and immutability, Zig pushes toward in-place mutation; sum types survive, but through `anytype` and duck typing. 172 points and 209 comments on Hacker News, a good part of them pushback. **[Disconcerting genre shift](https://seths.blog/2026/09/disconcerting-genre-shift/)** — Godin on realizing at minute ten that the documentary was fiction, and how everything after that sounds wrong until you switch modes. The argument is that genre signals — the hotel reception, the tropes of the interaction — look low-utility and are expensive to ignore. It's the item that doesn't talk to anything else today, and it's here on its own merit. ## stalled sources Anthropic Engineering at 118 days without publishing, Karpathy at 142, fasterthanli.me at 262. On the side that's still breathing but running late: Import AI at 13 days, Sebastian Raschka at 11, Transformers releases at 10.