Vertiv buys its way out of the power bottleneck

Good morning 👋 Vertiv just paid a public-company price for a young startup, and the reason has nothing to do with chips.

In today's issue:

• 🔭 Vertiv buys its way out of the power bottleneck
• 🧠 Gemini's three-week release cadence, Fable 5.1 lands on Bedrock, agentic video search
• 🔬 What LLM judges actually compute, cheaper agent benchmarks, repo failures traced to a few tokens
• 🛠️ Read the exhibit before the press release
• 🔗 Six fresh filings worth a glance

🔭 THE ONE THING

⚡ Vertiv buys its way out of the power bottleneck

Vertiv is paying $1.45 billion in cash for UtilityInnovation Group, the microgrid and behind-the-meter power outfit Sidney Hinton built to keep data centers running without waiting on the local utility. Up to $1.15 billion more rides on EBITDA targets hit over the next two years, pushing the total to $2.6 billion, roughly 13x UIG's projected 2027 earnings for a young company. That price tag says more than the press release does: Vertiv isn't betting on chips or cooling anymore, it's betting that power delivery is the thing that actually caps how fast an AI data center gets built. CEO Gio Albertazzi called it shrinking the distance from "site selection to first token," which is a cleaner way of saying the grid, not the GPU, is the bottleneck now. The earnout is the honest part of the deal. It means even Vertiv isn't fully pricing in UIG's numbers yet, so the real read comes when this closes in Q4 and the integration has to earn out the other $1.15 billion for real.

🧠 MODELS & RELEASES

• Google shipped Gemini 3.7 Flash three weeks after 3.6, pitched as its sharpest coding-and-agents workhorse yet, at half the token price of the model it replaces. Three weeks between generations is the real story: Google's treating price cuts like a monthly habit now, not an annual one. Google AI updates, August 2026
• AWS brought Claude Fable 5.1 to Bedrock, extending Anthropic's zero-retention Enterprise Frontier Safeguards (your data stays inside your own AWS boundary) to the new model. Anthropic says the gains show up in messy, multi-hour sessions: better planning, and more honest about where it's stuck instead of quietly faking a finish. Claude Fable 5.1 on AWS
• Google DeepMind taught Gemini's video models to search a clip instead of grinding through every frame: up to 88% less token spend, 66% cheaper, 7% more accurate on hour-long footage. Live now on 3.7 Flash, 3.6 Flash and 3.5 Flash-Lite, no extra fee on top of standard token pricing. Introducing agentic video in Gemini
• OpenAI ran case studies on Basis, Clay and Exa Labs pointing agents at onboarding, account management and developer integrations instead of adding headcount. Light on hard numbers, but a decent checklist for where agents are actually load-bearing versus demo-ware. How AI-native companies turn workflows into operating capability

🔬 RESEARCH HIGHLIGHTS

• LLM judges, cracked open. A team reverse-engineered how models like Themis and Prometheus actually land on a score for a summary, using controlled perturbation attacks instead of just reading outputs. The mechanism is two stages: lower layers flag local errors, higher layers integrate them into a verdict, and fine-tuning mostly reshuffles which layers get trusted rather than building new judgment from scratch. If you lean on an LLM judge for eval or a training signal, that number is more mechanical, and more gameable, than it looks. Paper

• Benchmarking SWE agents without running the whole benchmark. PTA-IRT scores coding agents from a small calibration subset by reading how they solved a task (explored files, attempted edits) instead of just pass or fail. Across four SWE benchmarks it beats prior subset-selection baselines at the same low budget. Builder read: if re-running a full agent eval suite after every change is eating your compute, this is a legitimate shortcut. Paper

• Most repo-level code failures trace back to a few tokens. ACToR's finding: when an LLM generates code against a large repo, errors cluster at a handful of decisive "critical tokens." Miss one and the rest of the function follows the wrong semantic path. Retrieving repo context right at those spots, instead of once per task, lifted results meaningfully on both RepoExec and CoderEval. Paper

🛠️ TRY THIS

Read the exhibit before the press release, not after

Vertiv's 8-K this morning ships an EX-2.1. That exhibit type means a plan of acquisition, reorganization, or divestiture, not routine business, and it's the kind of filing that gets summarized by a press release before anyone reads what it actually says. Build the habit of going to the primary document first.

1. Pull the filing history (or RSS feed) for the tickers you actually track, Vertiv included if you're watching AI datacenter infrastructure.

2. When an 8-K lands, check the exhibit index before the item text. EX-2.1 is a deal document. EX-99.1 is usually the press release wearing a filing number.

3. Feed the raw exhibit, not the press release, to an agent and have it pull out parties, deal type, dollar figures, dates, and closing conditions.

4. Compare that extraction against the press release's framing. Where they diverge is the real story.

Prompt: Read this SEC exhibit and extract: (1) the parties involved, (2) the deal type (merger, reorganization, divestiture, other), (3) any dollar amounts or share counts, (4) key dates and closing conditions. Flag anywhere the exhibit's language is narrower or more hedged than a typical press release would suggest.

Worth a look

• Omnigent 9.6k stars. Open-source meta-harness that orchestrates Claude Code, Codex, Cursor, or a custom agent and lets you swap between them without rewriting the job, exactly the kind of backend-agnostic setup you'd want for a standing filing-extraction task. link
• OpenExecutive 3.4k stars. Eight specialist Claude agents behind one coherent persona (FastAPI + Next.js). One config change from being a standing analyst that reads every 8-K for a watchlist. link

🔗 QUICK LINKS

• Artificial Intelligence Technology Solutions (AITX) The security-robot maker behind Robotic Assistance Devices put out a fresh 8-K. The public summary doesn't say what triggered it. Filing
• Churchill Capital Corp XI (CCXI, CCXIU, CCXIW) Filed a Form 425, the paperwork SPACs use to talk publicly about a pending combination. No deal terms in today's document. Filing
• Raphael Pharmaceutical (RAPH) Another bare 8-K current report. Nothing beyond the form type made it into the summary feed. Filing
• Alternus Clean Energy (ACLEW, ALCED) Attached a press release exhibit to its 8-K. Filing
• Yum Brands (YUM) The KFC and Taco Bell parent slipped an EX-99.1 press release into its 8-K. Filing
• bioAffinity Technologies (BIAF, BIAFW) Same move: an EX-99.1 exhibit, contents not pulled through in the feed. Filing

That's the read for today.

Pradeep Perugu