In partnership with
☀️ TRENDING AI NEWS 🏢 OpenAI Valuation: OpenAI reportedly raised another $11 billion at a $1.25 trillion valuation, with the round expected to close in coming weeks. 🚨 Market Rout: The Nasdaq fell 5.5% on Thursday, its worst day in 19 months, as Asian and European markets followed on Friday. 🛠️ Meta Muse: Meta's new Spaces feature lets Muse build shopping hubs, school dashboards and trivia games from a plain-English description. 🤖 Open Source: Buoyant released Kinference, an Apache 2.0 control plane for vLLM on Kubernetes that cut median TTFT by 95% in its own benchmark. |
Nearly $2 trillion vanished from the Nasdaq in a single session, and the cause wasn't a bad earnings report or a failed model launch. It was a ballot box.
Voters weighed in on 41 data center ballot measures this week and backed more restrictions in all but one. Today's issue is about what happens when the AI boom runs into democracy, plus a few releases worth your attention.
| 🤓 AI Trivia | |
A 594B-parameter mixture-of-experts model only activates a fraction of its weights per token. How many parameters are active in Thinking Machines' Inkling 2 Small?
The answer is hiding near the bottom of today's newsletter... keep scrolling. 👇
| 🚨 The AI Sell-Off Goes Global | |
From Wall Street to Tokyo to London
The Nasdaq composite slumped 5.5% on Thursday, tumbling into correction territory, as investors bet that the Democrats' midterm sweep means tougher regulation and curbs on AI data centers. By Friday, Asian and European markets were sliding too, and analysts quoted by The Guardian expect more waves of selling.
Context matters, though. The Nasdaq is still up 4.6% for the year, and a late-Friday recovery was modest but real. Apple, Tesla and Nvidia shed roughly $1 trillion in one session, so the pain was real even if the year-to-date picture isn't a disaster.
Meanwhile, newly elected Democrats and Senator Bernie Sanders are already renewing calls for data center moratoriums, and Meta has warned that the backlash could make AI expansion harder. This is the political bill for the boom coming due, as we covered earlier in the week in our breakdown of how data centers lost the midterms.
The bottom line
The sell-off reprices political risk, not technology. Watch data center permitting and state-level moratoriums over the next few months, because that is where the AI build-out will actually slow or not.
The voice-to-text app that built its own speech model
Most dictation apps are a wrapper around someone else's speech model. Wispr Flow trained its own.
It's called Canto, and it was built for the way people actually talk to AI: thinking out loud, changing direction mid-sentence, half-whispering in an open office.
Wispr tested it on 10 hours of real dictations from more than 2,300 speakers. Canto had the lowest word error rate of every model tested, including models from Google, OpenAI, AssemblyAI and Deepgram. That means your prompts to ChatGPT, Claude and Cursor come through the way you said them.
Talk naturally in any app and Wispr Flow types clean, formatted text, 4x faster than typing.
| 🏢 OpenAI Hits $1.25 Trillion Anyway | |
Investors keep writing bigger checks
While public markets wobbled, private money kept flowing. OpenAI reportedly raised an additional $11 billion at a $1.25 trillion valuation, according to Bloomberg, in a round that hasn't formally closed but is expected to in the coming weeks. That follows a $122 billion raise in March at an $852 billion valuation.
The IPO chatter is getting louder too. President Greg Brockman said this week the company could be IPO-ready in 2027, while rival Anthropic is reportedly nearing its own listing. There is a wrinkle: Ramp data shows AI spending among big spenders has stalled even as usage climbs, with price cuts and smarter cost controls eating into the hypergrowth narrative.
The bottom line
Private valuations and public sentiment are pulling in opposite directions. If usage keeps rising while spend flattens, the IPO class of 2027 will have to justify margins, not just momentum.
| 🛠️ Meta's Muse Can Now Build Your Apps | |
Describe a space, get an interface
Meta introduced Spaces, a feature that lets its Muse AI assistant build customized, app-like experiences inside the Muse app. You describe what you want, and Muse generates something like a shopping hub, a school dashboard, a music tracker or a trivia game that you can interact with directly.
If you connect outside apps like Gmail or Google Calendar, a space can pull in your real data. Meta's Muse agents have been picking up features and integrations at a steady clip, and earlier this week US users also got the ability to video chat with a customizable Muse avatar.
Building a quick web presence is getting this easy elsewhere too. Our sponsor 60sec.site, an AI website builder, takes a similar describe-it-and-get-it approach for full sites.
The bottom line
Spaces turns an assistant into a lightweight app platform. The real test is whether the connected-app integrations are reliable enough that people stop opening the originals.
| 🔬 Free Chinese Models Are Getting Scary Good at Hacking | |
Open weights close in on gated cyber models
Wired reports on an experiment where AI models were asked to break into simulated networks. The hacking skills of DeepSeek and GLM, both free to download, came out not far behind OpenAI's most advanced cyber model, which isn't available to the public.
That matters because labs have leaned on access restrictions as a safety lever for their strongest security tools. Separately, security researchers found hundreds of servers running the Artex AI penetration testing tool exposed to the open internet, letting anyone use it.
The bottom line
If free models are close to gated ones, restricting access to one lab's cyber model buys less safety than it used to. Defenders should assume attackers have this capability already.
| ⚡ Open-Source Infrastructure Worth Bookmarking | |
Routing and kernels get faster
Buoyant released Kinference, an Apache 2.0 inference control plane for vLLM on Kubernetes. It adds declarative routing policies, KV-cache-aware load balancing and per-tenant metrics, and cut median time-to-first-token by 95% in Buoyant's own benchmark (so treat that as a vendor number until others reproduce it).
On the hardware side, Zyphra open-sourced ZYFA, FP8 and BF16 prefill attention kernels for AMD's Instinct MI355X. On real model activations, its FP8 kernels run 1.16 to 1.76x faster than AMD's AITER, with 62-78% lower error at head dimension 128.
The bottom line
If you self-host models, Kinference is worth a weekend test. Cheaper, faster serving is one of the few levers that works regardless of what happens to data center permits.
| ⚠️ Activists Storm a Microsoft Site in Amsterdam | |
The backlash goes physical in Europe
Protesters from the Dutch direct action group Geef Tegengas broke into a construction site for a new Microsoft data center in Amsterdam on Friday, clashing with police. The project involves three 85-meter data center towers on the city's outskirts, and similar actions took place across Europe.
Elsewhere, the pressure is power and water rather than protest. DataVita admitted its Lanarkshire AI development near Glasgow will be served entirely by the National Grid despite earlier renewable-energy promises, and Southeast Asia's data center rush has hit pause over power and water constraints.
The bottom line
Opposition to data centers is now a global, multi-front story: ballots in the US, grid limits in the UK, and direct action in the Netherlands. Expect more projects to be delayed or reshaped.
| 🛠️ Tool of the day | |
Warp - a modern terminal with built-in AI agents that helps developers run commands, write scripts and debug faster. It has a free plan.
| ⚡ Quick hits | |
Airbnb says it will launch its AI travel agent to consumers in the first half of 2027, with an interface built for each request. Read more →
Google made Gemini Omni 1.2 Flash generally available, with video clips up to 15 seconds and 30% lower prices than Omni 1.1. Read more →
METR estimates Anthropic is improving its models on well-specified tasks at about twice its earlier pace, but most key evidence was redacted. Read more →
| 🌎 Trivia Reveal | |
The answer is 17 billion! Thinking Machines' Inkling 2 Small is a 594B-parameter mixture-of-experts model that activates only 17B parameters per token, which is how it keeps inference costs far below what its total size would suggest.
Want more AI news like this every day? Visit dailyinference.com for our daily AI newsletter.
| 💬 Quick Question | |
Do you think the data center backlash will actually slow AI progress, or will companies just build somewhere else? Hit reply and tell me - I read every response!
That's it for today - see you tomorrow with more!
| 🎁 Share Daily Inference | |
Know someone who would like this? Share your link and earn free rewards, starting with the AI Tools Starter Kit at 1 referral.