AI2Day Daily Brief — 15 Aug 2026

Good morning, it is Saturday the fifteenth of August, and the big story today is a French startup claiming it can make the AI your company already runs ten times faster, using nothing but software. Let us get into it.

Paris-based Kog says the bottleneck slowing down AI inference is not your hardware, it is the software on top of it. In May, the company ran a public demo on AMD M I 300 X and NVIDIA H 200 graphics cards, the kind sitting in many corporate data centres already, and showed a small model producing three thousand tokens per second. That caught enough attention on Hacker News to pull in two hundred business leads. The small test model, Laneformer Two B, is now open-source, but the real target is large language models. CEO Gaël Delalleau says a major model running at ten times speed should be ready to show before September, timed ahead of a Series A raise. Seed funding came from Varsity V C, Bpifrance, and the French Tech 2030 programme. The claim is bold. We will know more when the September demo either lands or does not.

Straight to a leadership change at Google. Koray Kavukcuoglu, a DeepMind veteran since 2012, has taken over as senior vice president of Google DeepMind, replacing co-founder Demis Hassabis, who moves to a chair role. Analysts say the timing reflects pressure: Google has not released a frontier model since Gemini Three Point One Pro in February 2026, while Anthropic and OpenAI both shipped well-received new models this year. The gap, analysts say, came partly from under-investing in coding capabilities. That said, the commercial story is not weak: nearly ninety percent of Fortune 100 companies use Gemini Enterprise. Alphabet shares dipped on the news but are still up roughly seventy-six percent over the previous twelve months. Kavukcuoglu's job, plainly stated, is to get the research side moving again.

Not a bad point to note that even with the model gap, Google's enterprise business held up. But the pressure is real, and that brings us to prices. OpenAI and Anthropic both cut what they charge this week, and the reason is not hard to find. Chinese AI developers including Moonshot and DeepSeek have been winning users in Silicon Valley and Europe simply by costing less. OpenAI reduced prices for G P T Five Point Six Luna, its entry-level option, by eighty percent, describing it, in the company's own words, as its fastest and most affordable model. Anthropic launched Claude Opus Five at half the price of its most powerful tier, Opus Five Fable. These are not marginal adjustments. Businesses were already pulling back on AI spend as bills grew. Cheaper alternatives gave them somewhere to go. The American labs are responding.

Today's sponsor is Zero2Hero: an AI poker coach sits at your table, and you can ask it mid hand whether the player across from you is bluffing. It runs in your browser on any computer. Windows, Mac, Linux, even a Chromebook. Nothing to install. Lifetime access is $99, in US dollars, until 15 September, at zero, the number two, hero, dot A I.

From pricing pressure to a story about harm. A twenty-nine-year-old man in Queensland, Australia lost his savings to a crypto investment scam built around a fake trading app. The app showed a convincing profits dashboard. His numbers climbed, he put more money in, then the funds were gone. This is the fraud type known as pig butchering: feed victims fake paper gains to build confidence, then take everything real. What is new here is the engine behind it. Investigators say scammers used AI tools to handle the administrative work of running the scheme, making it cheaper and faster to target more people at once. The Australian Competition and Consumer Commission recorded more than one point three billion dollars in investment scam losses in 2023 alone. AI is not creating this crime, but it is scaling it.

And finally, the infrastructure behind all of this AI is meeting resistance on the ground. Scottish police submitted a formal letter this month to planning authorities warning that a proposed data centre in Larbert, about thirty miles west of Edinburgh, will face, in their words, a great deal of public opposition. Officers say the site will need serious security to prevent trespass and incursion before a single brick is laid. The project still needs planning approval. This is not an isolated case. Data centre proposals have become flashpoints across Scotland, with residents raising concerns about noise from industrial cooling systems, heavy electricity use, and the visual impact of large warehouse-scale buildings. The facilities are essential to running AI services, but the communities hosting them are not staying quiet about the cost.

For a fuller look at the week's stories, catch AI Today Weekly, our longer Monday show. That is your AI briefing for today. Every story is at A I 2 Day dot live. That is A, I, the number two, D A Y, dot live. We are back tomorrow morning. If this was useful, hit the thumbs up and subscribe, so the next one finds you.