More data, more problems. And problems are what hyperscalers are having when it comes to powering their data center and chip production ambitions. Now they are having to borrow large amounts to build these data centers and to attempt to build their own chips. Bankers are salivating over these deals, since they may end up owning the infrastructure should the profits never come, and Nvidia, although still enjoying large demands, may be soon be shaking as all its clients are searching for ways to become self-sufficient.
If we follow the social media trend, AI will be in greater demand, from us humans and from the bots that have entered the chat and are populating our daily activities very rapidly, behind the curtains and without us realizing it. Also, as we advance, all the devices that surround us may become AI devices, including the table and chair I’m sitting on. Smaller models may fit in our palms, but bigger and more sophisticated one are power hungry, and it seems this is just getting started. I’m not sure the world and its natural resources can sustain this, or that there is enough money to go around to build all these ambitions. But Let's go a bit deeper on what this means for you.
In today’s issue:
📰 AI News and Trends
AI's Tab Is Now Too Big for Big Tech's Wallet
🧰 AI Tools of The Day - Latest Image Creation Model and Agents
Texas Just Hit Pause on Data Centers. Here’s What It Means for You
Why the Pause?
Will my power bill go up?
Is the water being polluted?
Why so many Data Centers, and will we need them?
What happens next?
Your new Voice Planner
🐈⬛ GitHub repos of the week - 50 Marketing Skills and AutoCad plugins
The Chat Window Is Dead, GPT-6 Answers With Interactive Tools Instead
ANd obviously, Claude cannot be left begind Claude Haiku 5.5 is optimized for high-volume, cost-sensitive tasks, offering lower operating costs and faster performance for applications like summaries and customer support
Meanwhile Grok Bot will use Anthropic’s Claude, Midjourney and Suno models to deliver sound, images, and video and top of the line inference
California has declared a human-vs-robot fight illegal
OpenAI’s math breakthrough points beyond math
Samsung forecasts record third-quarter profit of $80 billion as AI boom fuels chip demand
Share Yaro on AI and Tech Trends | Your Top AI Newsletter
Building AI is getting so expensive that even the biggest tech companies can’t pay for it out of their own pockets anymore. Broadcom, Oracle, and SpaceX are each lining up huge loans, worth tens of billions of dollars apiece, to buy AI chips, and investment firms like Apollo, Blackstone, and Goldman Sachs are in talks to provide the money.
Broadcom is working on more than $50 billion in financing for OpenAI’s custom chip, SpaceX has talked to lenders about $40 billion for Nvidia chips, and Oracle is negotiating its own deal, which would likely be set up through a separate company that buys the chips and leases them to Oracle so the debt stays off Oracle’s books. This shift is happening because the usual options are running out. Big cloud companies have already sold hundreds of billions of dollars in bonds, and AI labs like OpenAI and Anthropic want to own their own computing power but don’t have the cash to buy it. The takeaway is that demand for AI hardware is clearly strong, but more of it is now being paid for with borrowed money, which could become a problem if AI revenue doesn’t keep up. These deals are still early, but should the get approved and the revenues and profit do not come, we will have a very rude awakening. Additionally, we should consider that Chinese models are becoming more competitive and ubiquitous.
Share
1. HyperFrames Studio (HeyGen) - An open-source framework that turns HTML into MP4 video, with a timeline editor you and your AI agent share. Use it to template and batch-render branded videos like stat cards and client reports for free, with no per-render fees.
2. Mistral Large 4 - Use it for long-document work like full-site audits or RFP summaries, at a lower API price than most near-frontier models, plus a self-host path for clients with data-residency needs.
3. Spira Maxima - A script-to-video tool that generates the presenter, B-roll, captions, and music, then posts to TikTok, YouTube, and Instagram. Use it to turn a paragraph or newsletter into short-form social video in about 10 minutes. It starts free (50 credits, watermarked) and paid plans start at $9/mo.
4. EmbeddingGemma 2 - Google’s free, open model that puts text, images, video, audio, and code into one searchable space, small enough to run on a laptop. Use it for cross-media search and RAG (retrieval-augmented generation) without an API bill. It’s builder-only, so you’ll need to set it up yourself.
*Mistral’s “open weights” is still a promise, and Spira’s pricing page is inconsistent, so verify before quoting it.
Texas is freezing new data center permits. Behind the headline is a question that will affect electric bills, local water supplies, and the AI tools we use every day and also do we actually need this many data centers?
At the end of 2024, Texas’s grid operator (ERCOT) had 63 gigawatts of large-customer requests waiting in line. By June 2026, that line had grown to 474 gigawatts, more than five times the state’s record peak demand. About 90% of it was data centers.
Almost no one believes all of that will get built. Developers often file the same project at several sites to see which one lands. Some have no customer lined up, and some have never built a data center. Grid planners can’t tell the real projects from the placeholders, and they don’t want ratepayers paying for upgrades that serve a project that never shows up. Add community frustration and an upcoming election, and the state hit pause.
It depends on who pays for the upgrades, and right now that isn’t settled.
The worry is real. In an August poll, 56% of Texas voters said more data centers would hurt local energy bills. In the mid-Atlantic and Midwest, data centers account for 38% of the latest capacity bill, about $6.3 billion, according to the grid’s market monitor.
A Berkeley Lab analysis found that from 2019 to 2025, states with the most growth in electricity demand generally saw average prices fall after inflation. When a big customer pays its own way, the grid’s fixed costs are spread across more sales, so everyone else pays a smaller share. When it doesn’t, the rest of us cover the gap.
Some early deals are trying to get this right. An Indiana utility proposed using data center revenue to cut household bills by about $100 a year. Others have been murkier. In Arkansas, Google paid $443 million toward a solar plant, but the utility counted it as a prepayment and can still seek the plant’s full cost from all customers.
Water use is poorly disclosed. As of June, only 28 of 377 data center companies had answered a Texas state survey on their water and energy use. Without the numbers, nobody can verify the claims on either side. Some companies say new designs use almost no water for cooling. But elections are around the corner and right now all hyperscalers are fighting for their reputation
Air quality is where the documented problems are. xAI ran dozens of gas turbines at its Memphis site without air permits, and a Mississippi order now requires its turbines there to be retired by mid-2027.
Secrecy fuels distrust. In Hood County, Texas, commissioners were asked to approve a tax break for “Project Patriot” without being told who the company was.
If you live near a proposed site, ask for water, noise, and emissions numbers in writing before any tax break is approved.
Two things are true at once.
Demand is real. AI chips are so expensive that an idle one costs far more than the electricity to run it, so companies will pay almost anything to get power fast. That urgency is why developers are rushing to build their own power plants on site.
The pipeline is inflated. Much of the 474 GW will never be built. The real danger isn’t that data centers become useless. It’s that heavy borrowing and speculative projects leave the weakest ones stranded if AI spending slows, which is the debt risk we flagged in our above piece on Oracle, Broadcom, and SpaceX.
Could smaller AI models running on your own devices make giant data centers unnecessary? Probably not. Cheaper AI has historically meant more AI use, not less. Local models will likely take over everyday tasks, while training and the hardest reasoning stay in big facilities. That’s our analysis, not settled fact.
The grid itself is now the bottleneck. Connecting a large data center can take 5 to 10 years, and the equipment is scarce. Large power transformers take more than two years to arrive. The article’s author argues that new rules should require developers to pay for what they reserve and hit milestones, with faster connections for those that agree to cut power when the grid is tight.
Dates to watch: new Texas large-load rules take effect today, Oct. 8, with a $100,000 study fee plus a $50,000-per-megawatt deposit. The state’s environmental regulator reports to the governor on Oct. 19, and an audit of the queue is due in December.
Voice Planner connects Alexa to your Notion account so you can create, query, complete, and delete your tasks completely hands-free. You can create filtered task lists by adding dates, times, categories, and priorities, then simply ask Alexa what you need to do without having to open Notion every time.
Take the stress out of task management, with Alexa, you can manage everything by voice.
Start Guided Setup
"Voice Planner now also listed on Notion Marketplace and live on Product Hunt.”
Matt Pocock's skills - github.com/mattpocock/skills · Personal library of small, composable agent skills (SKILL.md files) that enforce engineering discipline on Claude Code, Codex and other agents. It's the anti-framework: small, composable skills that make agents interview you before they code, break work into tickets, and test before they ship
earthtojake/text-to-cad - github.com/earthtojake/text-to-cad · A plugin/skills library that lets coding agents generate real parametric 3D CAD files (STEP, GLB, STL, 3MF), run manufacturability checks, produce engineering drawings, and connect to fabrication services
coreyhaines31/marketingskills - github.com/coreyhaines31/marketingskills · Marketing skills for Claude Code and AI agents. CRO, copywriting, SEO, analytics, and growth engineering.” A library of 50 Agent Skills (SKILL.md files) that turn Claude Code, Codex, Cursor and similar agents into marketing operators
