Daily Briefing

[Daily Briefing] OpenAI chip beats Nvidia 🌶, Gitea deadline 🔒, IG bedtime 🌙

28 August 2026

OpenAI published the first real benchmarks for its own inference chip yesterday, and they are not modest numbers. Across GPT-OSS 120B, DeepSeek R1 670B and Kimi K2.5 1T, Jalapeño claims 1.5x to 1.9x more AI work per watt than the best commercially available systems, and 1.7x to 3.6x lower end to end latency. For interactive workloads, the kind where you are sitting there watching tokens appear, it claims 2.1x to 4.1x.

What makes this worth reading rather than filing under vendor slides: the tests ran on SemiAnalysis's public InferenceX benchmark, and SemiAnalysis verified some of the runs on site. That is a meaningfully higher bar than a press release.

The caveat matters as much as the headline. SemiAnalysis points out the honest comparison is not Blackwell but Vera Rubin, since both use HBM4. Jalapeño still edges ahead on output tokens per megawatt there, and it does so without multi token prediction, an optimisation Nvidia already ships and OpenAI has not adopted yet.

And yet OpenAI has committed to keep buying Nvidia hardware. Only a very small Jalapeño deployment lands by the end of 2026, with the real rollout in 2027.

Why it matters: Inference is the cost centre now, not training. Every watt saved per token is money that does not go to Nvidia. But the gap between a winning benchmark and a shipped fleet is the actual story here. Designing a better chip and manufacturing millions of them are entirely different problems, and OpenAI's own purchase orders say so.

Quick hits

Your Gitea patch deadline is tomorrow. CISA gave federal agencies until 28 August to fix CVE-2026-60004, a CVSS 9.8 code injection in the diffpatch endpoint that turns ordinary repo write access into shell access. Everything from 1.17 up is affected, 1.27.1 fixes it. The three question check
Instagram is getting a midnight curfew. Meta agreed to pay up to 16.68 billion dollars to settle claims from 29 states, and the settlement dictates product defaults: night mode from midnight to 06:00, a two hour daily cap on teen accounts, notifications blocked during school hours. All seven mandated changes
Amazon is shutting Mechanical Turk after 21 years. Gone on 30 September, along with SageMaker Ground Truth and Augmented AI. Bezos once called it "artificial artificial intelligence", humans doing the work software could not. It labelled a great deal of the data that made modern models possible. What replaced it
The Galaxy S26 FE landed today. Samsung came in at 699 dollars rather than the rumoured 799, with seven years of OS and security updates on a mid-tier phone. That last part quietly reframes what a cheap phone costs to own. Price and specs
Nobody can identify what JWST keeps photographing. The little red dots appeared 600 million years after the Big Bang and vanished a billion years later. Harvard led studies published on 5 August argue they are single stars 100,000 times the mass of the Sun, pulsating themselves apart before collapsing into the first supermassive black holes. Two other theories are still standing.

Tool of the day: a Wi-Fi 7 router, but probably not yet

This is the most common bad tech purchase in a home, so here is the honest version. Wi-Fi 7 lands around 3 to 5 Gbps in the real world against 1.2 to 2 Gbps for Wi-Fi 6E, and latency drops from 10 to 20ms down to 3 to 8ms thanks to Multi-Link Operation.

Now run a speed test on your wired connection. Under 1 Gbps and Wi-Fi is not your bottleneck, your internet plan is, and a new router changes nothing. Over 1 Gbps and you game, video call for work or run twenty plus devices, it is worth it. And the trap nobody prints on the box: a Wi-Fi 6E laptop on a Wi-Fi 7 router still runs at Wi-Fi 6E speeds.

If you are buying a router today anyway, buy Wi-Fi 7. The price gap has closed and you keep it for years. Wi-Fi 7 routers on Amazon, or read the full upgrade test first.

Reply and tell me this: if you self-host anything, when did you last actually check who still has write access to it? I suspect the honest answer for most of us is never.

Nath, Future Technology

Forwarded? Get your own free briefing at futuretechnologyhq.com/newsletter

Some links may be affiliate links.