💡 The Infrastructure Bet That Could Rewrite AI's Cost Equation

Sponsored by

☕ Morning! 💡 Your 5-Minute Shot of AI + Tech Clarity

Quick Hits 🎯

• 🚢 Sail Research just raised $80M to build infrastructure specifically for AI agents that work for hours or days, not seconds. Sequoia and Kleiner Perkins led the rounds.

• 🤖 Claude Code changed the PM-to-engineer ratio. The traditional 1:8 model is now effectively 1:20. LinkedIn already replaced its APM track to respond.

• 🌏 Asian AI startups are capitalizing on Anthropic's export ban. Tokyo's Sakana AI and Beijing's 360 Security both launched frontier-adjacent models this week, with one marketing pitch: "no export control risk."

• 💰 Amazon just committed another $13B to India. Total India investment now hits $48B through 2030, focused on AWS data centers in Mumbai and Hyderabad.

• 🔐 Prompt injection is now OWASP's #1 LLM vulnerability for the third straight year. GitHub Copilot's recent CVE scored a 9.6, from a RAG pipeline attack, not a user prompt.

• 🔒 SimpleX Chat hit 1,183 new GitHub stars in a week. Privacy-first, no user identifiers. The timing with rising AI data concerns is not a coincidence.

• 🦀 RustDesk continues its surge as distributed teams prioritize self-hosted, privacy-respecting remote access over TeamViewer alternatives.

• 🛡️ The Corgi/Y Combinator open-source dispute put a spotlight on a question more startups need to answer before Series A: do you actually know where your codebase came from?

🎁 + 4 other stories you might find useful

Our Partner 🎉 

The most trustworthy AI admin agent for busy executives.

Busy executive. Packed calendar. An inbox that never empties.

Everyone is talking about AI agents. Every week brings another promise of an assistant that can do it all.

Catch is the real deal.

A smart, proactive AI admin agent focused solely on taking administrative work off your plate.

It schedules meetings, triages your inbox, drafts emails in your voice, resolves conflicts, sends follow-ups, and handles the countless small tasks that consume your day.

Available wherever you work — Gmail, Outlook, Slack, WhatsApp, and even over the phone.

No setup. No training. No learning curve.

Catch learns how you work, takes action when it's confident, and keeps things moving without constant supervision.

From swamped to sorted in seconds.

Get started with Catch and have your assistant ready before your next meeting.

Test Catch for free

The Big Picture 🖼️

💡 The Infrastructure Layer Just Got Its Own Funding Category.

Sail Research raised $80M this week at a $450M valuation, backed by Kleiner Perkins, Sequoia, Redpoint, and angel investors including Alphabet's chairman. The pitch: existing inference infrastructure was built for humans typing at prompts, not agents running autonomously for hours.

The gap is real. A standard chatbot session consumes maybe a few thousand tokens. An agent debugging a codebase, running compliance checks, or doing deep research burns 50 to 500 times that, continuously. At that scale, the economics of inference become the core product problem, not an afterthought.

The takeaway: "agent infrastructure" is now its own category, the same way "cloud infrastructure" split from "servers" a decade ago. The companies that build on commodity inference and the ones that build on purpose-built agent infrastructure will face very different unit economics in 18 months.

💡 Claude Code Didn't Just Change Engineering. It Changed the Org Chart.

Anthropic's own teams now write the majority of their code with Claude Code. Rakuten cut feature delivery time from 24 working days to 5. Stripe deployed it across 1,370 engineers. Ramp cut incident investigation time by 80%.

But the more interesting consequence is upstream. Engineering has effectively tripled in output — product management has not. The traditional 1:8 PM-to-engineer ratio now functions closer to 1:20. LinkedIn's response was to eliminate its associate product manager track entirely and launch a "Product Builder" program instead, training generalists who can do product, design, and engineering.

The takeaway: the bottleneck has moved from "can we build this?" to "should we build this?" Engineers who understand customer problems are suddenly the most scarce resource in tech. That's a career signal worth reading carefully.

💡 The Export Ban Is Writing Asia's AI Origin Story.

Two weeks after the US government banned Anthropic from distributing its Mythos and Fable 5 models globally, Tokyo-based Sakana AI launched Fugu, an orchestration model that routes tasks across multiple AI systems. Beijing's 360 Security launched Tulongfeng, a vulnerability-discovery tool it positions as a direct Mythos competitor.

Both companies launched with essentially the same message to enterprise buyers: we can't be export-controlled. Asia represents an $847B AI market by IDC projections, and every week the ban holds is a week for these alternatives to embed in enterprise workflows.

The takeaway: export controls intended to protect American AI dominance may be doing the opposite. Once enterprises standardize on a foundation model, the switching costs are enormous, rebuilt integrations, retrained systems, rewired pipelines. This market moment may not reverse.

💡 Amazon Is Treating India Like a Second Home Market.

Amazon announced another $13B for India's AI and cloud infrastructure this week, its third major commitment in three years, pushing total planned spending to $48B through 2030. Andy Jassy flew to New Delhi and met with PM Modi personally.

The investment funds AWS expansion in Mumbai and Hyderabad, bringing custom AI chips and managed services closer to one of the world's fastest-growing enterprise markets. The broader picture is a hyperscaler land grab: Google committed $15B to India in late 2025, Microsoft pledged $17.5B in December. The battle for AI infrastructure leadership is being fought in Mumbai and Hyderabad as much as it is in Northern Virginia.

The takeaway: geography is becoming a competitive moat. Regional infrastructure, local data residency, and proximity to customers are no longer compliance concerns, they're strategic differentiation for cloud vendors and their enterprise clients.

💡 Prompt Injection Isn't a Research Problem Anymore.

OWASP has ranked prompt injection as the #1 LLM vulnerability for the third consecutive year. This isn't theoretical. GitHub Copilot's CVE-2025-53773 scored a 9.6, remote code execution triggered by malicious instructions embedded in externally fetched content, not a user prompt.

The attack surface has expanded. RAG pipelines, MCP tool descriptions, browser-fetched pages, customer tickets, any content an agent reads can now carry embedded instructions. The kill chain documented in recent research shows five stages: initial access via poisoned document, privilege escalation, persistence, reconnaissance, and lateral movement to other agents. No executable code required. The payload is natural language.

The takeaway: the teams winning on AI security in 2026 are the ones treating their LLM like any other untrusted execution environment, minimal permissions, validated outputs, full observability. Security built in from the start, not bolted on after the breach.

💡 The Agent Infrastructure Market Is Moving at VC Speed.

Sail isn't alone. Baseten raised $1.5B at a valuation as high as $13B. Nebius paid $643M for 20-person Eigen AI. Scaled Cognition raised $100M to automate high-stakes customer interactions.

The pattern Kleiner Perkins is betting on is that inference will fragment into workload-specific platforms the same way compute fragmented into GPU, TPU, and specialized silicon. The highest-growth workload is agents. The companies that own the inference layer for long-horizon agents own the financial control plane for autonomous labor.

The takeaway: infra is the new software. The picks-and-shovels play for the agent era isn't the models, it's the plumbing that makes running those models sustainable at scale.

💡 Privacy Tooling Is Having Its Moment.

SimpleX Chat gained 1,183 GitHub stars last week, built in Haskell, zero user identifiers, end-to-end encrypted by design. This isn't just a developer curiosity. The uptick tracks closely with rising concern about what AI systems do with data after interactions end.

The broader signal: privacy-first infrastructure is crossing from security niche to mainstream product consideration. The teams building AI features on platforms with murky data policies are going to face the same trust problem that social platforms faced five years ago, but faster, because the stakes are higher.

The takeaway: if your product touches user data and your privacy story isn't airtight, that's not a legal problem, it's a growth ceiling.

💡 Open-Source Provenance Is Now a Due Diligence Question.

The Corgi/Y Combinator dispute, where a buzzy insurtech startup faced accusations of shipping a rebranded open-source product without attribution, landed this week. Regardless of how it resolves, the story surfaced a question most fast-moving startups skip: do you know where your codebase actually came from?

In the age of Claude Code and AI-assisted development, this gets harder, not easier. Agents writing code don't always surface the libraries, patterns, or training data they're drawing from. Legal and compliance reviews are going to catch up to AI-generated code, and the teams with clean provenance trails will have a significant advantage.

The takeaway: governance isn't a growth constraint. It's what separates repeatable product velocity from expensive cleanup later.

Trending Tools 📈

chopratejas/headroom (+2,473 ⭐ per week, 🐍 Python) Link

• Compresses LLM inputs by 60-95% without sacrificing output quality, works across any model or provider

• Helps teams: Directly attacks the token cost problem as agent workloads scale; the teams running thousands of concurrent agent calls are already watching this one closely

CopilotKit/CopilotKit (+366 ⭐ per week, 🟦 TypeScript) Link

• Frontend stack for building AI agents and generative UI natively into React and Angular apps

• Helps teams: Bridges the gap between backend agent logic and actual user-facing interfaces, lets product and eng ship agent-powered features without rebuilding the UI layer from scratch

simplex-chat/simplex-chat (+1,183 ⭐ per week, 🔷 Haskell) Link

• Messaging network with zero user identifiers, 100% privacy by design, across iOS, Android, and desktop

• Helps teams: Privacy-first communication and secure prototyping for applications where data residency and user anonymity are non-negotiable

ripienaar/free-for-dev (+472 ⭐ per week, 🌐 HTML) Link

• Comprehensive curated list of SaaS, PaaS, and IaaS offerings with free tiers, maintained for devops and infradev

• Helps teams: Discover cost-free tools for testing, hosting, monitoring, and deployment without revenue commitments

IceWhaleTech/CasaOS (+370 ⭐ per week, 🚀 Go) Link

• Open-source personal cloud system, simple, self-hosted, focused on data sovereignty

• Helps teams: Quick self-hosted cloud storage for teams that want control over their data without enterprise pricing

microsoft/PowerToys (+228 ⭐ per week, 🧰 C#) Link

• Windows productivity utilities, keyboard-driven launchers, window management, and customization tools

• Helps teams: Eliminate the low-level friction that adds up across a 40-hour week; engineers who keyboard-optimize save real time

vinta/awesome-python (+176 ⭐ per week, 🐍 Python) Link

• Curated, opinionated list of Python frameworks, libraries, tools, and resources

• Helps teams: Skip the discovery phase; find mature tooling and best practices for rapid prototyping

veracrypt/VeraCrypt (+98 ⭐ per week, 🔒 C) Link

• Strong, cross-platform disk encryption, the maintained successor to TrueCrypt

• Helps teams: Adds encryption for sensitive data and backups without an enterprise contract

Tech Trend of The Week 📊 

🔍 "Long-horizon agents" search volume surged 340%+ in the US over the past 7 days

Sail Research's $80M raise landed on June 25 and pulled the term from infrastructure insider jargon into broader tech discourse. Searches are clustering around questions of cost, use cases, and how these systems differ from standard chatbot APIs, which means enterprise buyers and developers are starting to evaluate, not just observe.

The signal: we're at the "what is this?" stage for agentic infrastructure, which means we're roughly 6-12 months from "how do I buy this?" That's the window. The teams building intuition now, understanding what long-horizon agents can and can't do, where the costs compound, what governance looks like, will be the informed decision-makers when budget conversations start.

Our Partner 🎉 

Scale your customer experience across all channels

ElevenAgents handles high volumes of customer interactions across voice, chat, and WhatsApp, resolving customer requests faster with human-sounding agents. Train agents on your SOPs, improve CSAT in real time, and integrate with your existing tools. Higher resolution rates, no added headcount. Flat pricing at $0.08 per minute.

Get started free

Hit Reply And Tell Me 💬 

What's the most impressive tech you've seen recently that actually works?

I read every single reply

Go where your customers actually are 📍

Most of the internet copies Reddit.

Blogs, SEO pages, AI models, product reviews. They all pull from the same well.

Reddit is the 3rd biggest site on earth. And it trains the bots your users ask for help.

Odd Angles Media runs Reddit campaigns for over 45 brands each month.

Luckily, we’ve convinced them to give away all of their strategies in a (free) ebook ⬇️⬇️

Grab The Free Ebook (Free) →

Reach the People Who Sign the Checks 💰️ 

40,000+ CTOs and engineering leaders. 96% US-based. 80%+ corporate emails.

They evaluate vendors. They shortlist solutions. They buy.

Advertise with us on Paved → (discount applied)