📅 August 26, 2026 · ⏱️ Read time: 5 min · 🔗 Issue No. 23
You're in the loop — two Chinese humanoid robots just ran the 100 meters faster than Usain Bolt's world record, one clocking 9.39 seconds — a machine that could barely manage 21 seconds a year ago. And in a related twist, all that robot-and-AI ambition is quietly landing on your receipts: Amazon just raised prices on some Echo, Kindle, and Fire TV gear by as much as 60% as the AI boom devours the world's memory supply.
Today: Nvidia's $6B move to build America's open-weight answer to China, Alibaba's document-to-video model, a how-to for turning any file into a 30-second clip, five fresh tools, and a competitor-teardown prompt.
🔁 The Loop
Nvidia goes upstream, and robots outrun Bolt
Humanoid sprinters stole the week's headlines — but the bigger money moved quietly upstream.
Nvidia is spending about $6 billion to build an American open-weight rival to China's models. The chipmaker struck a sweeping deal with startup Poolside — roughly $1 billion in direct investment plus technology access — and is moving 100-plus engineers into its own Nemotron model effort, per the Wall Street Journal. It's Nvidia's clearest step yet from selling GPUs toward building the models that run on them — a hedge against DeepSeek, Kimi, and Qwen owning downloadable weights. Watch whether Nemotron ships a genuine frontier open model this year: that's when this stops being a talent-and-cash play and starts squeezing rivals. See the deal.
Alibaba shipped Wan3.0, a video model that turns your documents into 30-second clips. The Cloud unit's new generator makes 1080p video up to 30 seconds long — double Wan2.7's 15-second cap — straight from PDFs, spreadsheets, slides, or web pages, the company said. It landed a day after Alibaba raised roughly $10 billion in a share placement to fund its AI buildout. Take a look.
Humanoid robots beat Usain Bolt's 100-meter record at Beijing's robot games. Tiangong Ultra crossed in 9.39 seconds and Honor's Lightning hit 9.47 — both under Bolt's 9.58 — with one Lightning test run clocking 9.32, according to reports. The same Tiangong model needed 21.50 seconds a year ago. The catch: the machines still slammed into mats trying to stop. Watch the sprint.
Unlocked reads: the WSJ on Nvidia becoming a model-maker, and The Next Web on why doc-to-video changes who makes ads.
🌊 DEEP CURRENT
The AI buildout is escaping the data center — into your utility grid and your shopping cart.
The receipt tells the story. Amazon just raised prices on parts of its Echo, Kindle, and Fire TV lineup by as much as 60%, The Verge reported, and the culprit isn't tariffs — it's memory. AI data centers are consuming so much advanced DRAM and storage that the shortage is now rippling out to phones, PCs, and household gadgets.
Why it matters. The buildout no longer lives only in server halls; it's a claim on the world's electricity and components. Two deals on the same day made that concrete: nVent agreed to buy Maverick Power for $1.75 billion (plus up to $550 million in earnouts) for AI-data-center switchgear, and Germany's Infineon snapped up Bengaluru's C2i to manage power inside dense racks.
The AI race is quietly becoming an electricity race — and the unglamorous companies that move and manage power are suddenly worth billions.
The other side. For most readers, AI has meant free chatbots and clever demos. Now it's a memory crunch that makes a Kindle pricier and turns a switchgear maker into an acquisition target. GPUs get the headlines, but substations, transformers, cooling, and DRAM are the real bottleneck — and they don't scale on a software timeline.
The bottom line. Expect more quiet price creep on everyday electronics and a steady wave of power-and-infrastructure M&A well outside the chip aisle. If you're budgeting a device upgrade or a compute-heavy project, watch memory spot prices and utility rates — they're becoming the truest gauge of how hot the AI economy really is.
🛠️ The Workbench
Turn any document into a 30-second video
Wan3.0's headline trick — turning a file into a finished clip — is available in Alibaba Cloud's Model Studio, but the workflow generalizes to most document-to-video tools. Here's a clean first pass.
Pick one source file — a one-page brief, a single slide, or a short product PDF. Trim it to the one message you want on screen.
Open a doc-to-video model (Wan3.0 in Alibaba Cloud Model Studio, or your tool of choice) and upload the file as the input.
Add a one-line intent prompt describing tone, pacing, and who it's for.
Set length to 15–30 seconds and resolution to 1080p — shorter clips keep motion and characters consistent.
Generate two or three variations, then keep the one with the cleanest mapping from your text to the on-screen action.
Export, then add captions yourself — AI video still fumbles on-screen words.
Sample Prompt: "Turn this one-page product brief into a 20-second explainer for busy founders: calm pacing, one key benefit on screen at a time, upbeat but not salesy."
🗣️ Overheard
What the timeline's buzzing about
💰 Nvidia eyes Perplexity: Nvidia is reportedly weighing an investment that would value Perplexity above $30 billion — up from ~$20B last year — as its annualized revenue tops $750M.
🔩 Xiaomi's own silicon: Xiaomi unveiled the Xring O3, a TSMC-made 3nm phone chip, part of a ~$3B, 3,000-person push to cut its reliance on Qualcomm.
⚖️ Chips on trial: Taiwan indicted nine people for allegedly smuggling 130 Super Micro servers with Nvidia chips to China; customs intercepted 56.
🖥️ Arm meets the mainframe: IBM revealed a 2nm chip that runs both Arm and IBM Z code on the same cores, with 11 cores clocking above 5.7GHz.
🤖 Robots get funded: XPeng's robotics unit raised $900M+ at a $6.3B valuation from IDG, Tencent, and Alibaba to mass-produce its IRON humanoid.
🔎 Fresh Finds
Five tools worth a look
🎭 Wizstar: Digital avatars that move and act like professional actors.
🔌 Supernova: Pipes all your data into Claude and Codex so agents work with real context.
🕸️ Mindcase: Extracts structured data from anywhere on the web in minutes.
🛠️ fx (by Vercel): Vercel's tiny, open-source coding agent.
🧮 Router by Ramp: Routes each prompt to the cheapest capable model — tokens are money.
★ = sponsored placement, if any.
🧪 Prompt Lab
The Competitor Teardown
Paste this in when you need a fast, honest read on a rival — it forces specifics and flags its own guesses.
You are a sharp competitive-strategy analyst. I'll name a company and its top competitor. Produce a one-page teardown: 1) Positioning: how each frames itself, in one sentence. 2) Wedge: the single feature or price move where the challenger is winning. 3) Moat: what actually protects the incumbent. 4) Blind spot: one thing the leader is ignoring that I could exploit. 5) 30-day play: three concrete moves I can make this month. Keep it specific, state the assumption behind each claim, and flag anything you're unsure about. Company: [YOUR COMPANY] · Competitor: [THEIR NAME]
Want a header image for the teardown? Try this named-style image prompt:
Modern gouache illustration: two friendly robots facing off across a chessboard on a sunlit desk, warm indigo-and-amber palette, visible brush texture, soft paper grain, generous negative space, no text, no logos.
⏪ Rewind
Yesterday’s most-opened link
Readers couldn't stop clicking the Wall Street Journal's breakdown of Nvidia's $6B Poolside deal — proof that "the chipmaker becomes a model-maker" is the plotline everyone's watching.
Stay in the loop — the InTheLoop team