Perplexity is back, after the launch of a new agent tool Perplexity Computer, and a shift to usage-based pricing, Perplexity's estimated ARR rose to over $450M in March, jumping 50% in a month.

Perplexity Computer is an AI-driven agentic workstation designed to act as a digital "operator." Unlike standard search engines or chatbots that only provide information, it is built to execute complex, multi-step workflows.

It functions as a centralized command center where the AI uses tools, browses the web, and manages files to complete specific projects.

From Personal Agents to Team Agents: AI Is Moving Beyond Tasks — Toward a “Team Brain”

AI has been evolving along a very clear trajectory: boosting individual productivity. The interface has shifted from simple chatboxes to agents, and now to systems like OpenClaw that can independently execute parts of real-world work. The ability to complete single tasks keeps getting stronger.

We’re already seeing a transition from task-level AI to workflow-level AI. But in real work environments, that’s still not enough. Finishing one task doesn’t automatically trigger the next step. Once workflows span multiple tools—or rely on human coordination—the process breaks. Tasks stall, context gets lost, and execution fragments.

This is exactly where Kollab comes in, cofounded by my friend Zhaofei Wang.

The Missing Layer: From Tasks to Flow

Kollab isn’t trying to make AI better at doing isolated tasks. It’s trying to solve something deeper: the fragmentation of collaboration.

Instead of helping individuals “get things done,” Kollab is designed to help teams keep things moving—end to end. The goal is simple but ambitious: once something starts, it should flow through the organization automatically, without needing constant human handoffs.

This is what you might call team-level AI.

A System Built for Teams, Not Individuals

Kollab approaches this problem through four core layers:

1. AI-Native Workspace

At the foundation is a shared workspace that brings everything—projects, tasks, documents, context, and AI outputs—into a single environment.

This isn’t just another chat app. It’s an AI-native workspace, where the project becomes the central unit. Around it, discussions, documents, code, design assets, and agent activity all live in the same context.

The idea is simple: instead of scattering work across tools, bring execution into one cohesive layer.

2. Bots That Live Inside Your Existing Tools

Kollab doesn’t ask teams to abandon their current stack. Instead, it embeds itself directly into tools teams already use—Slack, Lark, Discord, Telegram, and more.

The logic is pragmatic: no team is going to migrate away from tools like Slack, Notion, or GitHub just because of a new AI product. So instead of replacing them, Kollab integrates with them.

Using it feels natural. You can trigger actions by simply tagging Kollab—just like tagging a teammate. Behind the scenes, it executes cross-system workflows.

In this sense, Kollab isn’t a replacement layer—it’s an execution layer on top of existing tools.

3. Skills as Team Assets

This is where things get really interesting.

Kollab turns repeatable workflows into reusable Skills—things like:

• Pulling completed items from a task system

• Extracting key commits from a codebase

• Identifying recurring issues from feedback channels

• Automatically generating release notes

Once defined, these workflows become reusable building blocks for the entire team.

One person optimizes it once—everyone benefits.

This shifts best practices from being trapped in someone’s head (a great PM or analyst) into something systematized and scalable. As Kollab’s co-founder Wang puts it, the goal isn’t just automation—it’s organizational compounding: turning individual experience into team capability.

4. Memory That Compounds Over Time

The final layer is memory—and it’s what makes the system truly long-term.

Kollab doesn’t just remember what was said. It tries to learn:

• How the team evaluates problems

• How priorities are defined

• Preferred structures and formats

• Implicit standards and decision patterns

Over time, what accumulates isn’t chat history—it’s organizational experience.

This creates a very different kind of lock-in. In the SaaS era, switching costs came from data—documents, files, workflows. Here, the cost is deeper:

Are you willing to lose a system that has internalized how your team thinks and operates?

That’s a fundamentally different moat.

From “Answering” to “Delivering”

Wang believes this direction is now viable because AI has crossed an important threshold.

It’s no longer just about generating answers—it’s about delivering outcomes.

Once AI enters real workflows, organizational compounding becomes critical. Every team operates differently, and those differences—once captured—become durable advantages.

Individual experience is already being productized through Skills. Now, team-level experience is being productized as well.

Kollab’s interface reflects this philosophy:

• Left panel: teams, bots, skills, connectors, projects

• Center: task execution and output generation

• Right panel: progress, artifacts, and context

Instead of a prompt-in / answer-out model, it visualizes AI work as a trackable, auditable, and deliverable system.

The Bigger Shift

If early ChatGPT felt like a smarter search engine, today’s agent systems—like OpenClaw—represent something fundamentally different.

They can:

• Operate browsers

• Call tools

• Execute across systems

• Move workflows forward autonomously

This marks a shift from AI as assistant to AI as operator.

And once execution is unlocked, the way teams function starts to change.

Wang argues that the rise of one-person companies will actually increase the demand for high-efficiency collaboration systems—because even small teams will rely heavily on AI to scale output.

Inside Kollab Itself

Interestingly, Kollab’s own team already runs heavily on its product.

Across product, engineering, design, and QA, workflows like:

• Requirement management

• Bug tracking

• Automated feedback analysis

• Scheduled error log analysis

• Content operations

…are all handled through Kollab bots inside their team chats.

The result? A workflow that feels fundamentally different from traditional companies.

My Final Thoughts

Kollab is still in public beta, but it points to a clear direction:

The future of AI isn’t just about making individuals faster.

It’s about making organizations flow.

👉 You can try it here: https://kollab.im/product

From a revenue perspective, the top spot in large AI models has changed hands.

Anthropic announced its annualized revenue run rate hit $30B, up from $9B at the end of 2025—a $21B increase in just four months.

In February, 500+ enterprise customers were each spending over $1M annually. Now, that number has doubled to 1,000+ in under two months.

Meanwhile, OpenAI sits at around $25B ARR as of March, with growth appearing slower than expected and internal tensions emerging around its aggressive IPO plans and high costs.

Anthropic’s key growth playbook

In a podcast with Lenny, Anthropic’s head of growth Amol Avasare reveals Anthropic’s key growth playbook:

1. Big swings over micro-optimization: 70% of effort goes into high-impact bets, not small tweaks. In a world where AI value could grow 100–1000x, 1% gains don’t matter.

2. AI-native growth (CASH): Using Claude to automate growth: find opportunities, run experiments, and optimize—already performing like a junior PM.

3. Engineering leverage > PM/design: 5 engineers can now do the work of 15–20. Result: PMs are stretched → engineers act as “mini PMs” on small projects.

4. Extreme focus: Coding + B2B

• Coding improves both product + internal research loops

• B2B focus helped them survive against Meta, Google, OpenAI

5. Smart onboarding (“right friction”)

• Import ChatGPT memory → reduces switching cost

• Add friction to better understand users → improves conversion

6. Built for hypergrowth chaos: ~70% of time spent handling “success disasters” (systems breaking under growth)

AI Insurance hit $40M ARR in 6 months

Read more

Collaboration Without the Cloud

During the period when note-taking apps were expanding toward larger teams and enterprises with collaborative features, Obsidian avoided the usual Notion or Google Docs model. It doesn’t force you to store data in its cloud; instead, collaboration happens via “sync services” or “shared folders.”

Even after the AI boom, when almost every product was rebuilding around AI or adding it as a core feature, Obsidian has been extremely restrained in incorporating AI.

The team is fully remote, with only one in-person meetup per year, and virtually no recurring meetings. Everyone focuses deeply on the highest-priority work.

Yet, Obsidian has been profitable from day one and has earned a large, loyal user base. CEO Steph Ango recently estimated over 4 million users, including 10,000+ organizations.

A Tiny Team, Big Impact

Obsidian’s ARR is estimated at $25M (third-party estimate). Despite no funding, the market reportedly values it at $350M.

The team currently has 9 people, including a cat named Sandy, with three engineers. They are hiring a fourth engineer—bringing the team to ten—and plan to keep the full-time team around 10–12 people.

Obsidian didn’t succeed by “turning notes into a collaborative suite” or by “embedding AI everywhere.”

It succeeded by sticking to product principles, transforming a niche audience into a highly sticky, self-propagating ecosystem willing to pay.

An “IDE for Thought”

Co-founder Erica Xu defines Obsidian as an “IDE for thought”—not a tool that thinks for you, but an environment you can fully control and continuously shape.

From the start, it wasn’t about creating a “large, all-in-one collaborative suite,” but a “holy grail note-taking app” for individuals. This fostered near-religious loyalty among geeks, researchers, writers, and knowledge workers.

Founders Shida Li and Erica Xu first built Dynalist, an outlining tool. It taught them the value of structured information and heavy-user workflows—but also its limitations: users want more than hierarchy—they want links, backlinks, graph views, and local file ownership.

This led to the transition from Dynalist to Obsidian:

“From organizing items → to connecting thoughts.”

Dynalist solved “how to list things clearly.”

Obsidian solves “how to make knowledge continuously readable, linkable, and evolvable over time.”

The Five Principles

Obsidian’s manifesto highlights five core principles: Yours / Durable / Private / Malleable / Independent.

• Yours: Everyone should have tools to think clearly and organize ideas effectively. Core tools are free.

• Durable: Data should be future-proof and accessible anywhere. Simple, open file formats prevent lock-in.

• Private: Thoughts belong to the user. Data is stored locally and encrypted end-to-end.

• Malleable: Tools should adapt to your thinking, not the other way around. Obsidian is highly customizable.

• Independent: Fully user-funded, no investors.

CEO Steph emphasizes “File over app”: apps become obsolete, files endure. Notes are stored as Markdown files locally, not in proprietary cloud databases.

“If you want your work to be readable on computers in the 2060s or 2160s, ensure it’s readable on 1960s computers.”

This approach sacrifices some short-term growth-friendly features (web-first collaboration, centralized data, strong lock-in), but gains trust, migration freedom, and long-term mental model value.

Adding Power Without Losing Control

Obsidian layers features like backlinks, graph view, Canvas, Properties, Web Clipper, Bases, CLI, all grounded in user-controlled files.

Bases (2025) brought “database views” while still building on local Markdown + properties. Obsidian adds Notion-style structured views without betraying its file-first philosophy.

Community-Driven Growth

Unlike many SaaS companies whose “community” is just a Discord, Obsidian’s community is an extension of the product itself.

Steph explained:

“The plugin architecture is genius not because of tech, but because it keeps the core app simple while letting the community satisfy long-tail needs.

Each user can have a unique Obsidian, so the company doesn’t need to build ‘10,000 features’ but can support ‘10,000 workflows.’”

This creates a craft-driven PLG model: users make tutorials, plugins, templates, and themes to improve their system, attracting new users organically.

Monetization Without Compromise

Obsidian is free for individuals, commercial, nonprofit, educational, and government use. Paid options are optional support and services, not core features. Revenue mainly comes from: Sync, Publish, Catalyst, and optional commercial support.

• Less than 1% of users use Sync, but it generates 80% of revenue.

• Sync Standard: $4/user/month (annual)

• Sync Plus: $8/user/month (annual)

• Emphasis: end-to-end encryption, version history, offline sync, cross-platform, shared vaults

Core users are highly sensitive to data ownership, customizability, and long-term readability: researchers, writers, developers, content creators, knowledge workers.

AI as an Auxiliary Layer

Obsidian remains restrained with AI. AI is never the main actor replacing thought—it’s an auxiliary layer, an ecosystem layer, a team lever.

Steph:

“Any AI integration must align with the Manifesto, especially privacy. User data will not go to OpenAI servers without explicit consent.”

Early AI features were community-built (e.g., Text Generator). Official roadmap features—Web Clipper Interpreter, Keychain, CLI—are peripheral AI infrastructure, connecting without violating file-first philosophy.

Selling Control, Not Notes

Obsidian embeds values into product constraints. Every decision—local-first, Markdown, encryption, plugins, no investors—is interconnected.

Obsidian sells “principled convenience,” not principle betrayal.

Sync and Publish enhance convenience without breaking “files belong to you, data local by default, portable.”

It’s not aimed at beginners, but this “thresholded love” fosters deep loyalty. Even the AI trend hasn’t swayed its core.

Steph:

“Obsidian doesn’t sell notes—it sells control.”

Impact

When Obsidian posted a recent engineering job on X, it sparked discussion and reached 8 million views.

Similar philosophy-driven products include Ghost, which I previously covered. Like Obsidian, Ghost started from the founder's needs and now has ARR over $10M.

Kimi, the flagship AI product from Moonshot AI, has rapidly emerged as a leading player in the global large language model and agentic AI space.

Positioned against major competitors like ByteDance and Alibaba, Kimi differentiates itself through a strong performance-to-cost ratio, particularly via its API offerings.

The platform enables developers and enterprises to build AI-native applications, automate workflows, and deploy intelligent agents capable of reasoning, planning, and executing tasks.

With the rise of the “agent-driven” paradigm—accelerated by trends like OpenClaw—Kimi has become a preferred infrastructure layer for teams seeking scalable, cost-efficient AI capabilities.

The core of Kimi’s product evolution lies in its rapid model iteration and focus on agentic workflows. Starting with Kimi K2 Thinking, an early reasoning-focused model, the company quickly advanced to Kimi K2.5, which introduced enhanced multimodal capabilities and significantly improved performance across complex tasks.

Beyond models, Moonshot AI is actively building an agent ecosystem, including Kimi Agent (single-agent execution) and Agent Swarm (multi-agent collaboration), enabling parallelized and autonomous task execution at scale. This approach directly addresses a key bottleneck in AI adoption: moving from passive generation to active execution.

On the infrastructure side, innovations like Kimi Delta Attention and the upcoming Kernel-Attention Dual Architecture for K3 signal deep investment in foundational model efficiency and scalability.

This product momentum has translated into explosive business growth. Moonshot AI reportedly reached an ARR of $100 million within just one month of launching Kimi K2.5, highlighting extraordinary demand for its APIs.

The company’s valuation has surged to approximately $18 billion, quadrupling within three months, alongside an ongoing $1 billion funding round and potential plans for a Hong Kong IPO.

Legora ARR hit $100M in under 18 months from $1M

$100 million in ARR. 1,000 customers. In 18 months. This milestone puts Legora among a very small group of enterprise software companies to have scaled at this speed.

Customer base: Over 1,000 customers, including major law firms such as White & Case, Linklaters, and Cleary Gottlieb.

Team scale: Expanded to over 400 employees.

Go-to-market: Direct sales motion.

Valuation context: $5.55 billion valuation from a $550 million Series D. Here is an ARR comparison between Harvey and Legora（made by arr.club）:

The Rise of the One-Person AI Company：Solo Founder+AI Built a Business Making Nearly $300M in Profit

Two years ago, Altman predicted that in the AI era, one-person unicorns would become possible.

We’ve seen early examples(Here and here)—small, fast-growing solo businesses hitting meaningful ARR—but nothing close to that scale.

Until recently. A solo founder with AI already built a business Making Nearly $300M in Profit. Not Revenue or ARR, it is Net Profit.

Read more

As a platform providing the infrastructure to build, deploy, and host web apps and AI agents, Vercel is the quintessential “shovel-seller” story of our time.

According to CEO Guillermo Rauch, Vercel has become a primary beneficiary of the Claude Code explosion. Its popularity within the developer ecosystem has naturally positioned Vercel as the default deployment tool recommended by Claude. We are only seeing the tip of the iceberg:

While Vercel customers using Claude represent just over 1% of the total user base, they already account for nearly 15% of all deployments.

Simultaneously, deployments generated by AI Agents are skyrocketing—climbing from roughly 5% in June 2025 to over 21% by February 2026. Within this “agentic” segment, nearly 70% of deployments originate from Claude Code.

This massive surge in Vibe Coding has propelled Vercel’s ARR past the $340M mark. After crossing the $100M milestone, the company’s growth rate has actually accelerated year-over-year.

Last September, the company closed a $300 million funding round led by Accel and GIC, boosting Vercel’s valuation to $9.3 billion—a nearly three-fold jump from $32.5 billion just a year prior. Rauch’s ultimate vision? To make the “one-person unicorn” a tangible reality.

Other AI infrastructures also exploded, such as Supabase, which has become a foundational piece of the AI coding stack.

In 2025, its developer base grew from 1M to 2M in four months, then to 3M just three months later. Nearly every “vibe coding” product leverages it as part of their GTM motion.

Fireworks, positioned as a GenAI platform-as-a-service, crossed $130M ARR in May 2025, hit $200M by July, and surpassed $280M by Oct, 2025.

Fal, a generative media platform, grew from $1M ARR to $40M earlier in 2025, then quickly hit $100M ARR in Aug 2025, and now it has hit $400M in ARR.

The 5 Moats That Matter in the AI Era

In a recent piece, Michael Bloch, partner at Quiet Capital, argued that in the age of AI, only five moats truly hold water:

Compounding proprietary data, Network effects, Regulatory permission, Capital at scale, and Physical infrastructure.

Michael notes that many skeptics doubt “intelligence” will become 100x faster, stronger, and cheaper. The reality? It will, and it’s happening fast.

Moats have always been built on two pillars: things that are hard to do and things that are hard to get. AI is rapidly democratizing the former. Building software, maintaining complex integrations, or embedding a product so deeply into a customer’s workflow that it takes a year to rip out—these used to be “hard to do.” Now, they are becoming trivial.

Conversely, owning 10 million users, a government license, a chip fab, or $1 billion in deployable capital—these remain “hard to get.” AI compresses the “time to do,” but it cannot compress the “time for things to happen.” This distinction has become his most critical filter for investment.

Regarding compounding proprietary data, the real moat is “living data”—proprietary information generated through operations that already possess their own moats. While network effects and regulatory licenses are well-understood, the scale of capital is a factor almost everyone underestimates.

The endgame is the “Physical World.” A chip fab costs $20 billion; a nuclear power plant costs $10 billion; satellite constellations cost even more. This is why Elon Musk, while musing that money might not matter in 15 years, is simultaneously raising $75 billion and pushing SpaceX toward an IPO.

When the bottleneck shifts from software to “atoms,” the ability to raise and deploy capital at a massive scale becomes a terminal advantage. Capital here isn’t just cash—it’s institutional trust, a proven track record, and relationships built over decades.

Designing a system with AI might take a week, but you cannot manufacture, install, and connect thousands of physical devices in a week. The laws of physics set a floor on time that intelligence cannot bypass. Those who start building first establish a lead that widens over time. This is why physical infrastructure is one of the most vital moats of the AI era.

The assets that grow stronger as AI evolves share a common trait: they require “real-world time” to accumulate—a dimension no amount of intelligence can compress. Network density requires years of user adoption; regulatory approval requires years of political maneuvering; infrastructure requires years of construction; data requires years of compounding; and capital relationships require decades of trust.

Time that cannot be parallelized is the “meta-moat” behind these five categories.

I find Michael’s framework incredibly compelling. The core judgment is simple: Is it hard to do, or hard to get? If your moat is limited by “intelligence,” its collapse is only a matter of time. But if your moat is limited by “time” itself—by human behavior, physical laws, political will, or capital—then you are likely building something built to last.

Over the past year, OpenAI’s product lineup has become increasingly fragmented. From the video generation tool Sora, to various experimental hardware efforts, to the split between web and desktop experiences—this dispersion has created two critical issues:

1. Internal inefficiency (“side quests”):

Teams have been spread thin across too many parallel initiatives. In an internal memo, Chief Applications Officer Fidji Simo noted that this fragmentation has lowered the overall quality bar.

2. Broken user experience:

Developers write code in Codex, knowledge workers draft content in ChatGPT, and research still requires jumping back into the browser. In enterprise environments where efficiency is everything, this context-switching cost is unacceptable.

OpenAI Superapp: All-in-One

The core idea behind the super app is simple: All-in-One. OpenAI wants to unify the entry point so its research and product teams can focus on optimizing a single core experience—rather than juggling a dozen disconnected apps.

The real breakthrough, however, lies in its agentic capabilities—a shift from “chat interface” to “autonomous agent.” Traditional AI works like this: you ask, it answers. The super app aims for something more powerful: you give instructions, and it executes.

• Autonomy: The app will be able to take actions directly on a user’s computer.

• Closed-loop execution: It won’t just write code—it will run it, debug it, and even perform financial analysis or document processing using real-time data gathered from the browser.

• Reinventing Codex: OpenAI plans to first embed agent capabilities into Codex, expanding it beyond programming into general office workflows—before fully merging it with ChatGPT and the browser.

The underlying logic is clear: whoever controls the desktop layer becomes the operating system of the AI era.

This strategic shift also carries a strong defensive undertone.

First, there’s the aggressive push from Anthropic. With highly focused, productivity-driven tools like Claude Code and Cowork, Anthropic has built strong momentum among developers and enterprise users.

Second, there’s a race for execution speed. While OpenAI has been investing in projects like Sora—which some internally view as costly experiments—Anthropic has been embedding itself directly into core enterprise workflows. According to Simo, OpenAI has realized that spreading effort across too many products and stacks has slowed them down and made it harder to meet quality expectations.

In recent weeks, top executives—including CEO Sam Altman, Chief Research Officer Mark Chen, and Simo—have been reviewing OpenAI’s product portfolio and reassessing priorities. At a recent all-hands meeting, Simo reportedly described the company as operating in a “Code Red” state, underscoring the urgency of the situation.

Still, building a super app is far from easy. OpenAI faces three major challenges:

1. Organizational complexity:

After a year of rapid expansion and multiple parallel product lines, consolidating everything under a more unified structure—led by Simo and Greg Brockman—will require strong execution.

2. Commercial and IPO pressure:

Both OpenAI and Anthropic are rumored to be preparing for potential IPOs as early as this year. The super app must quickly prove its revenue-generating power to satisfy investor expectations.

3. Security and privacy:

An AI agent capable of operating your computer, browsing the web, and reading/writing files raises fundamental concerns. Defining the security boundaries of such a system will be critical for both users and regulators.

According to OpenAI’s latest data, Codex has already surpassed 2 million weekly active users. Since the launch of GPT-5.3-Codex in early February, its user base has more than tripled. Desktop app downloads have exceeded 1 million, and token usage has grown roughly fivefold this year.

M&A for the Superapp strategy

To support this strategy, OpenAI has also announced the acquisition of Astral, aiming to deepen its integration into the developer toolchain and achieve a true end-to-end engineering loop—while directly competing with Anthropic in AI-powered coding.

Astral isn’t an AI company—it’s a high-performance developer tools powerhouse. OpenAI is particularly interested in its three dominant open-source products:

• uv: An ultra-fast Python package and project manager that solves the long-standing pain of environment management.

• Ruff: A Rust-based Python linter that runs 10–100x faster than alternatives and has become a modern standard.

• ty: A tool focused on type safety, improving maintainability for large codebases.

The goal is clear: embed AI directly into compiler-level workflows. With Astral, Codex can evolve from simply writing code to autonomously managing environments, running and debugging programs, and validating code quality. Only by mastering these underlying tools can AI agents behave like real programmers—executing end-to-end tasks on local machines.

Beyond Astral, OpenAI has also acquired Promptfoo and Torch.

Promptfoo specializes in AI evaluation and red teaming, with a strong enterprise customer base—including many Fortune 500 companies. Its core asset is a mature suite of evaluation tools and a widely adopted open-source CLI designed to uncover vulnerabilities in AI systems.

This will strengthen OpenAI’s Frontier platform—its environment for building AI agents—by embedding security directly into the development lifecycle.

As agents gain more permissions (such as operating user machines), managing risks like prompt injection, data leakage, and tool misuse becomes critical. Promptfoo also provides auditable reports, addressing enterprise concerns around governance and compliance.

Torch, on the other hand, focuses on data retrieval and indexing optimization. Its technology enables large language models to extract information from massive unstructured datasets more accurately and efficiently.

For a super app that needs to understand local files, emails, and enterprise databases, Torch provides the necessary indexing layer—ensuring low-latency responses and reducing hallucinations.

If Astral provides the “hands” (execution), then Torch provides the “eyes”—helping AI understand the user’s entire working environment. Meanwhile, Promptfoo acts as the “safety system,” ensuring that increasingly powerful agents remain controllable.

Together, these pieces form the foundation of OpenAI’s next move: not just a better chatbot, but a true AI operating layer for the desktop.

What’s remarkable now is the speed of its breakout: Fal has already surpassed $400M ARR, reaching that milestone just six months after crossing $100M. Few infrastructure companies scale this fast, even by AI standards.

Fal has quietly become a critical backend layer for developers building AI-native products. Its core offering is high-performance inference infrastructure—allowing teams to run and scale generative models (across image, video, and multimodal use cases) without touching the complexity of GPU orchestration.

By abstracting away deployment and optimization, Fal addresses one of the biggest bottlenecks in the AI stack: reliable, low-latency, cost-efficient model execution at scale. This is exactly what teams need as AI moves from demos to production.

Investors have taken notice. Fal has reportedly raised between $300M and $350M across two rounds, led by firms like Sequoia Capital and GIC, at a blended valuation of $8B—a strong signal that the AI infrastructure layer is becoming one of the most valuable positions in the stack.

Fal, alongside Fireworks AI, represents a broader explosion in AI infrastructure. But infrastructure is only one side of the story.

From Infrastructure to Agents: The Next Layer

This wave of AI is quickly shifting from copilots to autonomous systems. AI agents are no longer just chat interfaces—they’re evolving into “AI employees” capable of executing real-world tasks.

Tools like OpenClaw are accelerating this transition. Meanwhile, Sequoia Capital has recently argued that 2026 marks the beginning of an “Autopilot” era, where services become software—unlocking a massive new category.

We’re already seeing a new generation of AI-native products emerge across the stack—from search(Exa) and browsers(Browserbase) to databases purpose-built for agents.

But as agents become mainstream, a fundamental problem emerges:

How do AI agents identify themselves and communicate with the existing internet?

OpenClaw Ignites “Gmail for AI Agents” to Build the Identity Layer for the AI Era

Read more

In an era where most startups trade profitability for growth, Deel has quietly done both.

Founded in 2019 by Alex Bouaziz, the global payroll and HR infrastructure company has scaled to $1.4B in ARR in under seven years, reached a $17.3B valuation, and—perhaps most unusually—remained profitable for three consecutive years.

Today, Deel serves over 40,000 customers, ranging from early-stage startups to companies like Shopify and Coinbase, as well as enterprises in industries like aviation, oil & gas, and luxury.

So how did Deel pull this off?

1. Capital efficiency as a strategy

From the very beginning, Deel operated with extreme financial discipline. At one point after raising its Series A, the company had $4M in the bank—but had only spent a few hundred thousand dollars.

This restraint created leverage. Instead of raising capital out of necessity, Deel raised when it didn’t need to—allowing it to choose long-term partners like Ribbit Capital, Coatue, and Andreessen Horowitz on its own terms.

The philosophy is simple: fundraising is selling part of your company—timing matters.

2. M&A as a growth engine

While many companies struggle with post-acquisition integration, Deel has turned M&A into a repeatable system.

Instead of rebuilding acquired products before selling them, Deel does the opposite:

• Ship a V0 version within the first month

• Train the sales team immediately

• Rebuild the backend over the following year

This “front-end first” approach compresses time-to-market by nearly a year and turns acquisitions into instant revenue contributors—not long-term projects.

More importantly, Deel doesn’t acquire for revenue. It acquires capabilities—like local payroll engines or new HR modules—and plugs them into a growing product ecosystem that now spans over a dozen tools.

3. A global-first operating system

Deel’s moat isn’t just software—it’s infrastructure.

The company operates entities in 120+ countries, enabling it to offer a full-stack solution: hiring, payroll, compliance, and even IT management. This regulatory depth is hard to replicate and becomes stronger with scale.

Internally, Deel runs a fully remote team of 7,000 employees across 120+ countries—and uses its own product to pay them. This “Deel on Deel” approach creates a powerful feedback loop, where the company continuously stress-tests and improves its own system.

4. Betting on the future of work (and AI)

Rather than seeing AI as a threat, Deel sees it as a tailwind.

Alex Bouaziz believes AI will reduce repetitive jobs—but dramatically increase the number of companies. In that world, Deel’s role expands:

Not just paying humans, but potentially managing compensation and performance for autonomous agents.

It’s a subtle but important shift—from HR software to global labor infrastructure.

Deel’s rise isn’t just a story about remote work tailwinds. It’s a case study in operational rigor:

• Stay lean to maintain control

• Use M&A as a capability accelerator

• Build deep, hard-to-replicate infrastructure

• Align with long-term shifts like AI and global work

From an “outsider” startup to a category leader, Deel proves that in complex, regulated markets, execution—not hype—is the ultimate advantage.

AI SEO Bootstrapped to $5M ARR in a year

With the explosion of AI, AI SEO(aka GEO) is one of the fastest-growing categories. I tracked some AI SEO startups that are growing fast and backed by VC.

Such as Brand Radar from Ahrefs is reportedly adding $1M ARR every two weeks. Profound and others were heavily attracted to VC funding.

Now, a new AI SEO startup just crossed $5M ARR, fully bootstrapped in roughly a year. But the number wasn’t the story. The story was the transformation: from a niche Reddit marketing tool into a platform redefining how brands show up inside AI answers.

Read more

The Vibe Coding war is still on. While Base44 hit $100M ARR in a year of launch, Cursor reportedly hit $2B ARR.

Now Lovable’s business growth has been equally explosive. The company recently announced that its annual recurring revenue (ARR) has reached $400M, jumping from $300M just one month earlier and effectively doubling in only a few months.

That is saying Lovable is one of the fast-growing AI coding platforms that enables users to build full-stack applications and websites simply by describing what they want in natural language.

Founded in Stockholm by Anton Osika and Fabian Hedin, the company popularized the concept of “vibe coding”—a workflow where AI generates and iterates on code based on conversational prompts.

The platform’s adoption has surged as AI-assisted development becomes mainstream. Lovable now reports around 15 million daily active users and sees over 200,000 new projects created every day, indicating massive demand for AI-powered development tools across both individuals and businesses.

It has raised major venture funding, including a $330 million Series B round led by CapitalG and Menlo Ventures, bringing its valuation to about $6.6 billion. With this momentum, Lovable is projecting that it could surpass $1 billion ARR in the near future, positioning itself as one of the defining companies in the emerging AI coding infrastructure market.

Services As Software

Sequoia’s Julien Bek published an insightful article Services: The New Software, " which argues that the next generation of massive tech companies will not sell software tools, but rather the outcomes traditionally provided by human service firms.

The Core Shift is From Tools to Work

Julien posits that the “Copilot” era (selling tools to professionals) is transitioning into the “Autopilot” era (selling the finished work).

• The Risk: If you sell a tool, you are in a race against AI model improvements that might turn your product into a “feature.”

• The Opportunity: If you sell the work, every AI improvement makes your service faster and more profitable.

Intelligence vs. Judgement

He distinguishes between two types of labor to identify where AI will win first:

• Intelligence: Tasks involving rules, translation, and processing (e.g., writing code, medical coding, filling forms). AI has already crossed the threshold to handle these autonomously.

• Judgement: Tasks requiring experience, instinct, and “taste” (e.g., deciding which feature to build, cultural fit in hiring).

• The Trend: Today’s “judgement” will become tomorrow’s “intelligence” as models accumulate proprietary data.

The “Outsourcing” Wedge

The most effective way for “Autopilot” startups to enter the market is to target tasks that are already outsourced. This is effective because the company already accepts that the work is done externally.

There is an established budget line for the work. And the buyer is already purchasing a specific outcome rather than managing headcount.

High-Opportunity Verticals

Julien maps out several multi-billion dollar markets ripe for AI-native services:

• Accounting/Audit: Facing a massive talent shortage, AI can “close the books” directly.

• Insurance Brokerage: Highly standardized work that is currently fragmented across thousands of human brokers.

• Healthcare Revenue Cycle: Medical coding is a “pure intelligence” task with mature outsourcing.

• Legal/Transactional: Standardized work like NDAs and contract drafting where quality is easily verifiable.

For founders, the “Innovator’s Dilemma” is now: will you continue selling tools to your customers, or will you replace your customers’ manual labor entirely? Sequoia suggests that the most legendary companies of the next decade will choose the latter.

2 Cases of the Trend: AI Recruiting Platform/AI Career Agent grows at 20x+ in 6 months

AI Career Agent is one of the cases of Services As Software trend. One example is Juicebox, an AI recruiting startup backed by Sequoia Capital in the U.S. With only a small team of a few people, it has already surpassed $10 million in ARR, growing revenue 10× within a year.

Today it announced an $80M Series B at an $850M valuation, led by DST Global, with participation from Sequoia Capital, Coatue, and Y Combinator.

Although recruiting is currently its core business, the company is already expanding the model into other areas — such as sales lead generation and market research.

At its core, the model depends on one thing: how effectively AI can gather large volumes of accurate information and use it for precise matching.

Traditionally, most recruiting platforms relied on pulling data from sources like LinkedIn and other databases across the web.

Two AI Agents Replacing the Traditional Headhunter

Another example entering the recruiting space has taken a much simpler and more direct approach. Instead of relying primarily on external data scraping, it gathers information through conversation — letting users interact directly with an AI agent.

This design significantly improves both efficiency and accuracy, grows its revenue 20x+ in 6 months.

Read more

After hitting $200M ARR last year in Dec, Glean is now at $250M ARR, with a 40% DAU/MAU.

Glean is an enterprise “Work AI” platform designed to help organizations unlock knowledge and automate workflows using artificial intelligence. Founded by former Google engineer Arvind Jain, the platform connects to tools like Google Workspace, Slack, Microsoft 365, Salesforce, and dozens of other enterprise systems to create a unified knowledge layer across a company.

Employees can ask questions in natural language, generate documents, summarize internal information, or automate tasks using AI agents that understand organizational context. By solving the problem of fragmented corporate data and inefficient information retrieval, Glean enables teams to find knowledge instantly and turn it into actionable insights directly within their existing workflows.

The company has experienced extraordinary growth as enterprises adopt AI for daily operations. Glean surpassed $200 million in annual recurring revenue (ARR) just nine months after hitting the $100 million milestone, making it one of the fastest-growing enterprise AI software companies of the decade.

More recently, the platform has exceeded $250 million ARR with more than 150% year-over-year growth, reflecting strong demand for AI-powered productivity tools across global enterprises. The platform also sees heavy engagement, with employees running multiple AI queries daily and businesses integrating Glean into mission-critical workflows such as product development, incident response, audits, and customer support.

Glean has raised hundreds of millions of dollars from top venture investors including Sequoia Capital, Lightspeed Venture Partners, and Kleiner Perkins. In 2025 the company raised a $150 million Series F round at a $7.2 billion valuation, up from $4.6 billion just months earlier.

Copying Plaud’s Playbook? How this AI hardware Hit $27M ARR in 5 Months

In the emerging world of AI hardware, several early winners have already been absorbed by tech giants. Bee and Limitless were acquired by Amazon and Meta respectively, while Plaud’s AI recording device reportedly generated nearly $200 million in revenue, proving the massive opportunity in AI-powered note-taking hardware.

Now a new contender has appeared.

A product strikingly similar to Plaud has quietly reached $27M in ARR in just five months, selling 30,000 devices while maintaining 50% month-over-month growth. At first glance, many observers assumed the company was simply copying Plaud’s strategy.

But a closer look reveals something more interesting: not imitation, but a carefully executed product reframing and strategic compression of the market.

Read more

Base44 is an AI-powered “vibe coding” platform that allows anyone to build full-stack applications using natural language prompts. Instead of writing code, users simply describe what they want to build, and Base44 automatically generates the frontend, backend, database, authentication, and deployment infrastructure.

The platform targets founders, product teams, developers, and non-technical creators who want to rapidly build production-ready web applications without managing complex engineering stacks.

The company was founded by solo founder Maor Shlomo, a former CTO of Explorium, and quickly became one of the fastest-growing AI developer tools.

Base44 reached $1M ARR just three weeks after launch, grew to 300,000+ users within six months, and hit $3.5M ARR before being acquired by website-building giant Wix for $80 million in cash in June 2025.

At the time of the acquisition, the company had fewer than ten employees and was largely bootstrapped, making it one of the most notable solo-founder exits during the AI startup boom.

Following the acquisition, Base44 continued scaling rapidly inside the Wix ecosystem. The platform expanded to over 2 million users, enabling millions of apps to be generated through its AI development workflow. And claimed adding $1M ARR every 2.5 days

Less than a year after the acquisition, Base44 announced it had reached $100 million in annual recurring revenue (ARR), highlighting explosive demand for AI-driven software creation tools and signaling a broader shift toward “AI-native” development platforms where building software becomes as simple as describing it.

ElevenLabs’ competitor hit $10M ARR in 12 months from a niche open-source project

In the past year, the race in generative AI voice has looked a lot like every other AI race: massive models, massive funding, and massive GPU bills.

The leading AI Voice platform ElevenLabs, has already crossed $330M ARR recently, while enterprise customers now account for nearly half of total revenue.

Yet one small team quietly rewrote the script. In early 2025, they were doing roughly $400K in ARR. Just three months later, it crossed $5M ARR, and now it’s already surpassed $10M ARR within its first year of commercial operations.

Read more

Meanwhile, Anthropic’s growth continues to shatter records. Its run rate was $9B at the end of last year, jumped to $14B a few weeks ago, and now has crossed $19B—almost $20B—in just a few weeks.

As Morning Brew founder Alex Lieberman put it, the crazier thing than a $5B increase in weeks is that AI is gobbling up trillions of dollars in white-collar work.

And now, the AI diet app Cal AI has officially been acquired by nutrition-tracking giant MyFitnessPal.

Founded less than two years ago by a handful of high school students, Cal AI uses AI to calculate the calories and nutrition of food from photos instantly. It quickly rose to the top of the app download charts, competing with MyFitnessPal itself, reaching around 15M downloads and hitting $30M ARR in 2025.

I first covered this app in July 2024, when its monthly revenue was just $300K. According to the latest update from co-founder Zach, Cal AI’s ARR has now surpassed $50M.

Zach revealed that the acquisition deal was actually finalized in December 2025, when he was only 18, but the news is just being announced now. The exact purchase price hasn’t been disclosed, but MyFitnessPal CEO Mike Fisher said the Cal AI team didn’t have to sell—they were simply happy with the offer.

Given Cal AI’s $30M ARR at the time, I think the deal likely ranges in the tens of millions, if not over $100M. Post-acquisition, Cal AI will continue to operate independently, with the seven-person team—including 19-year-old CEO Zach—joining MyFitnessPal and integrating into its massive food database of over 20 million items.

This acquisition is seen as one of the most strategic and forward-thinking moves in digital health in recent years. With over 20 years of history and 200M registered users, MyFitnessPal once dominated mobile health through its vast food database. But the rise of generative AI and multimodal visual recognition has made traditional manual entry apps vulnerable to replacement.

AI-Native Experience Meets Established Expertise

MyFitnessPal views the acquisition as a step toward expanding its AI-powered health product suite and attracting younger users. Cal AI offers a fully AI-native, effortless calorie-tracking experience, complementing MyFitnessPal’s traditional focus on detailed macro tracking.

Essentially, this reflects a broader strategy: instead of building from scratch, the legacy giant is absorbing external innovation to meet shifting user expectations.

Convenience Creates Value

Cal AI’s core feature is simple: snap a photo of your meal, and AI instantly estimates calories and nutritional content. No tedious searches, weighing, or manual input. This resonates strongly with young users who prioritize speed over precision. Within its first month, Cal AI earned nearly $30K in revenue and quickly amassed millions of downloads, surpassing 15M overall.

The team claims the AI’s error margin is just ~10%, even exceeding FDA standards. Users describe the photo-to-AI workflow as “effortless tracking that makes sticking to a diet simple.” Cal AI’s target audience—young fitness enthusiasts focused on fat loss—prefers quick, AI-assisted logging, while MyFitnessPal’s traditional user base favors granular, customizable tracking.

Critics call it a “wrapper” app, noting it doesn’t build its own large-scale multimodal model, instead relying on APIs like OpenAI’s GPT series and Anthropic’s Claude. But this misses the point: commercial success in the AI era hinges on user experience and distribution efficiency.

Zach developed the app from personal frustration: as a skinny teen trying to gain muscle, he hated manual entry in existing apps. By removing that friction, Cal AI made calorie tracking as easy as snapping a photo—and proved a key lesson: in AI-driven verticals, seamless integration of AI into a specific user scenario creates outsized value.

Distribution Is King: Influencers and the “App Mafia”

If AI is Cal AI’s engine, its marketing is the turbocharger. Zach and his co-founders—including the so-called “App Mafia” members Blake Anderson (25) and Jake Castillo (28)—rented a $10M Miami mansion as a traffic lab, building an intricate influencer distribution system.

I’ve detailed their marketing approach. They collaborated with 150+ influencers, paying anywhere from $100 to $100K per post, and ran 12+ official TikTok accounts to test content. Their flashy lifestyle—luxury cars, mansion life, and social media spectacle—served as marketing itself, appealing to Gen Z’s belief that the right tools can lead anyone to rapid success.

In recent interviews, Zach says that after leveraging fitness influencer UGC to reach $2M in monthly revenue, the team scaled further with highly intuitive, pain-point-focused performance ads. Influencer content builds brand awareness, but performance ads deliver direct conversions.

Interestingly, even with this success, Zach was rejected by 15 top universities, including Harvard, Stanford, and Princeton, eventually choosing the University of Miami to “be with peers his own age” and treat college as a social experience rather than an academic one.

From $10 in the Bank to $100M+

Cal AI was founded in May 2024 in Zach’s New York home with just 3–4 people, funded entirely through early project income without outside investment. By the time of acquisition, the team had grown to ~30, with monthly revenue around $3M and net profit exceeding $1M. By January this year, monthly revenue hit $5.7M.

In a post-acquisition reflection, co-founder Jake Castillo shared a screenshot of his personal bank account from the night before Cal AI’s breakout—less than $10.

Cal AI’s journey and its young founders have sparked wide discussion, providing a motivational blueprint for other young entrepreneurs—especially impressive given it all happened without any outside funding.

Unsurprisingly, a wave of Cal AI clones has since emerged, also generating revenue.

Suno ARR Hit $300M

Suno, the AI-powered music generation platform that lets users create original songs from simple natural language prompts, has quickly become one of the fastest-growing SaaS products in the creative tech sector. Launched just two years ago, Suno makes it possible for anyone—from hobbyists to professional musicians—to produce full-length, high-quality tracks with vocals and instrumentals without any musical training.

The intuitive text-to-music interface and advanced generative models have driven massive adoption, with more than 100 million users worldwide having interacted with the platform since its inception.

Suno’s rapid growth has translated directly into commercial success. The company recently announced it has reached 2 million paid subscribers and an annual recurring revenue (ARR) of approximately $300 million—a substantial increase from roughly $200 million reported just months earlier.

The platform operates on a freemium model, with tiered subscription plans including Pro and Premier options that unlock commercial rights and higher usage limits. This strong conversion to paid plans, paired with exceptionally fast revenue acceleration, underscores Suno’s robust product-market fit in both consumer and creator segments.

Behind the scenes, Suno has also drawn significant investor attention and navigated industry challenges. In late 2025, the company closed a $250 million funding round at a $2.45 billion valuation, reflecting confidence in its long-term growth potential.

Pure software is rapidly becoming un-investable

This Spring Festival, the pace of AI progress has felt so overwhelming that writing about it almost seems pointless.

Between the valuation shockwaves OpenClaw sent through SaaS, and the geopolitical drama surrounding Anthropic’s Claude, the ground is shifting faster than most founders—or investors—can process.

Earlier this week, after tensions between Anthropic and the Pentagon surfaced, President Donald Trump announced a federal ban on Claude. Yet during the recent conflict, reports emerged that the Pentagon was still heavily using Claude. The result? Claude shot to #1 on the App Store, overtaking OpenAI’s ChatGPT in downloads.

Some joked that Anthropic no longer needs a marketing department. The numbers back it up: record daily sign-ups this week, free users up over 60% since January, and paid subscriptions more than doubling this year.

At the same time, sentiment toward traditional SaaS keeps deteriorating.

Today, Naval Ravikant posted: “Pure software is rapidly becoming un-investable.”

The message is clear. Software-only companies—code, apps, SaaS, no hardware moat—are becoming harder to fund at attractive returns. AI has made building and cloning software radically cheaper and faster. The old moats are dissolving.

A developer named Zoomer replied under Naval’s post: AI is eroding SaaS defensibility by collapsing development costs. The value of “pure software” trends toward zero when anyone can spin up a functional clone in 30 minutes. He cited Perplexity AI Computer tool as an example—he vibe-coded a Notion-like product with core functionality in half an hour.

The implications are bigger than hobby projects.

AI agents and dev tools are flooding the market with lookalikes. Horizontal SaaS—project management, basic CRM—loses pricing power. Why would a 10-person startup pay $20,000 a month for a CRM when they can build a tailored clone for a fraction of the cost?

Negotiation leverage shifts. In renewal calls, customers now say: “I can build this myself in 30 minutes.” Vendors respond with 50% discounts. ARPU shrinks.

Capital markets reprice accordingly. Revenue growth for horizontal SaaS has slipped from double digits to low single digits. Combined with discount pressure and AI-driven cost compression, price-to-sales multiples have fallen from ~20x to around 3x. And the reset may not be over.

Not all software dies, of course. Zoomer points out two resilient categories:

• Vertical SaaS with regulatory/data moats (e.g., HIPAA-compliant healthcare workflows).

• AI-tailwind infrastructure like data warehousing and observability.

Even coding tools are under pressure. On a recent episode of 20VC, Insight co-founder Jerry Murdock argued that Cursor is already “outdated.” In AI, you either embrace autonomous agents—or get left behind.

And now, something even more radical has emerged.

AI That Runs Your Company While You Sleep hit $1M+ ARR at 10x growth in 2 weeks

From Vibe Coding to AI-Run Companies

Until recently, vibe coding meant AI helping you build a product.

Now there’s an AI platform that builds—and runs—the entire company.

You give it an idea.

It launches the business.

It runs ads.

It acquires customers.

It fixes bugs.

It decides what to work on daily.

It makes money.

You split the profits.

Instead of hiring a developer, designer, marketer, operator, and support rep—AI does it all.

The platform itself has zero employees. Now over 1,300 companies are already running on it. Its revenue grew 10x in the past two weeks to hit $1M+ARR, and still growing faster. Today alone, it added almost $250K in ARR.

Read more

After its overnight rise to prominence, its creator, Peter Steinberger, recently shared a bold prediction in a conversation with Y Combinator:

80% of apps will disappear.

That’s not clickbait. It’s a thesis.

The Real Difference: Local, Not Cloud

According to Peter, what sets OpenClaw apart from most AI assistants isn’t intelligence — it’s architecture.

Most AI tools run in the cloud. That limits what they can actually do. But if an agent runs locally on your machine, it can control anything your computer can access:

Connect to your Tesla, adjust your smart bed, search every file on your hard drive. It inherits your tools, your permissions, your data. It becomes an extension of you.

And that’s where the magic starts.

Peter describes his “aha” moment back in November, when he was building an early prototype that could chat with him via WhatsApp. At a party with terrible Wi-Fi, he sent the bot a voice message—even though he hadn’t implemented voice processing yet.

A few seconds later, it replied. Curious, he asked what happened.

The system explained:

You sent a file without an extension. I inspected the header and detected it was audio. I used ffmpeg to convert it to WAV. Whisper wasn’t installed locally, so I found your OpenAI API key and sent the file via curl for transcription. Then I returned the text.

It improvised. It found tools. It solved the problem.

That’s when Peter realized something: this wasn’t just automation. It was creative problem-solving.

And if agents can autonomously execute tasks across your entire computing environment, why would we still need dozens of standalone apps?

If Agents Know Everything, Why Open Apps?

Take fitness tracking apps like MyFitnessPal. If your agent already knows what you bought, what you ate, and what’s in your fridge, why manually log meals?

Or to-do apps. If you say, “Remind me tomorrow,” and it just happens — do you care which database it stores?

Peter’s argument is simple:

Apps that primarily exist to manage structured data will be replaced by natural-language agents.

Only apps tied to specialized hardware or sensors may survive. Everything else becomes a capability — not a destination.

It’s reminiscent of the broader SaaS debate: if the interface dissolves into conversation, the value shifts from UI to infrastructure — from workflow to memory.

Memory Is the Moat

Peter believes model providers currently have moats, but models are rapidly commoditizing.

The real defensibility lies in memory.

Large platforms want to lock your chat history into proprietary silos. OpenClaw does the opposite. Your memory is simply a folder of Markdown files on your machine. You own them. You control them.

Even his development philosophy reflects this ethos of simplicity. Instead of complex Git workflows, he clones multiple local repositories to work in parallel. Instead of adopting heavy model context protocols, he routes everything through the CLI.

Why? Because AI is exceptionally good at using Unix tools.

Rather than inventing new abstractions “for robots,” Peter believes we should give AI the same tools humans love — and let it figure the rest out.

The Slack Fork That Claimed $1M ARR in 3 Hours

As developers rushed into the OpenClaw ecosystem, commercial spin-offs started appearing almost immediately.

Read more

Last year, VCs’ favorite AI note-taker Granola raised a $43 million in Series B at a $250 million valuation, to build Team’s second brain.

Now, the AI note-taking startup is back in the market—reportedly raising at least $100 million led by Index Ventures, at a valuation approaching $1 billion.

Not bad for a product that, on the surface, “just takes notes.” But Granola isn’t really about notes.

It’s about replacing the cognitive infrastructure of modern knowledge work.

Spend time in VC offices or startup war rooms, and you’ll notice something is changing: fewer pens, fewer notebooks, even fewer open laptops during meetings. Investors increasingly rely on Granola to automatically capture fundraising discussions, company metrics, partner meetings—and generate structured, intelligent summaries.

In September, Granola shipped a call feature. Connect it once, and it effectively becomes your phone—recording, transcribing, and organizing conversations in real time. What began as a “second brain” for individuals quietly evolved into a second brain for teams.

When announcing a recent update, the founder shared something unexpected:

Frequent users begin to develop a real relationship with Granola.

Its core users—ambitious, opinionated, overbooked—live inside back-to-back meetings. In that chaos, Granola becomes grounded. Like a trusted notebook you carry everywhere. It supports you.

And the insight is simple but powerful: AI works best when it has deep context. The richest context in knowledge work isn’t your to-do list. It’s your conversations.

Meetings are where decisions happen, ideas are debated, and commitments are made. Even in an automated future, that won’t change.

Granola isn’t just building a tool that records meetings. It’s building AI that understands what’s happening—and helps move things forward.

The Fastest-Growing AI Category You’re Noticing Everywhere

“AI that takes notes for you” has quietly become one of the fastest-growing categories in AI.

Startups and enterprises alike are racing to leverage large language models for speech recognition and summarization. Otter became the first AI note-taking company to surpass $100M ARR.

And my last piece covered a solo founder who built $7M ARR AI note-taker in two years.

Granola’s early traction is no accident. It focused on a specific tribe from day one: the VC and startup ecosystem. It didn’t join meetings as an awkward “AI participant” like early competitors. Instead, it ran quietly in the background on your iPhone or Mac, transcribing calls without friction.

Distribution + positioning + product subtlety. That combination matters.

A 7-Person Team Built a $10M+ ARR AI Language Speaking Coach

On the language learning side, we’ve already seen explosive growth. Speak crossed $100M ARR. The AI Speaking App Fluently grew 100x ARR in a Year.

Now another company is emerging—this one with just 7 people.

Read more

Sierra, the enterprise AI agent platform co-founded by former Salesforce co-CEO & OpenAI Chairman Bret Taylor and Google veteran Clay Bavor, has emerged as a fast-growing force in AI-powered customer experience automation, enabling brands to deploy autonomous chat and voice agents across complex support and engagement workflows.

Launched publicly in February 2024, Sierra’s product suite — built on its Agent OS and Agent Data Platform — enables enterprises to route customer interactions across channels such as voice, chat, email, and WhatsApp, automating resolution and backend actions (such as returns or subscription updates) without human intervention.

Its AI agents are designed to handle nuanced, real-world use cases across regulated industries such as finance, healthcare, and media, with notable deployments at companies including ADT, SiriusXM, and Rocket Mortgage.

The company’s growth trajectory has been remarkable: after reaching $100M in ARR in just seven quarters since launch, Sierra reported its first $50M quarter, propelling its annualized recurring revenue above $150M as it entered its third year of operations.

This momentum is driven by strong enterprise adoption — one in four customers generates more than $10B in revenue and half exceed $1B — and outcomes-based pricing tied to usage and successful resolutions.

Sierra’s platform now touches an estimated 95% of U.S. shoppers during peak retail periods, 50% of healthcare families, and 70% of fintech value chain interactions, underscoring broad usage across major markets.

On the company financials and backing, Sierra has secured substantial venture funding, raising $285M+ through rounds that culminated in a $10B valuation in 2025, led by Greenoaks Capital with participation from Sequoia, Benchmark, ICONIQ, and Thrive Capital. Earlier rounds included $110 M in initial financing and a $175M Series led by Greenoaks at a $4.5B valuation.

1-Person Built A $7M ARR AI Note-Taker with no funding

At the same time, everyone is talking about “one-person companies,” something remarkable has already happened in one of the most crowded AI categories of all: AI note-taking.

In a space packed with competitors, a single founder—working completely alone—built a product to $7 million in ARR in just 2+ years.

Read more

After hitting $100M ARR(Annual Run Rate) in 9 months, Genspark is now hitting $155M ARR in 10 months.

Functioning as an all-in-one autonomous work platform, it integrates specialized AI models to handle end-to-end workflows such as research, content creation, and data management.

By providing users with a personalized “AI assistant army,” the platform aims to liberate knowledge workers from “busywork,” allowing them to focus on high-value creative and strategic decisions.

Genspark recently closed a $300 million Series B extension, bringing its total funding to significant levels to support its global expansion, including a recent strategic launch in Japan.

With over 1,000 companies onboarded to its business tier in only eight weeks and a lean team of approximately 50 employees, Genspark is maintaining a high-efficiency ratio while scaling its infrastructure to millions of users worldwide.

A 20-Person Team Built a $70M ARR AI Video Platform — and It’s Not What You Think

AI video is moving so fast that the numbers are starting to feel unreal.

After Higgsfield reportedly crossed $200M ARR in just nine months, another story quietly unfolded under the radar: a 20-person team built an AI video platform that has now surpassed $70M ARR, growing revenue 7× in a single year.

But the most interesting part isn’t the speed.

It’s what kind of company this turned into.

What began as a tool that could easily have been replaced or commoditized has evolved into something far more durable: infrastructure for visual storytelling.

This isn’t another AI image or video generator

Read more

Redis, the popular in-memory data platform trusted by developers worldwide, has officially surpassed $300M in ARR, marking a major milestone in its evolution from a high-performance cache into a multi-purpose real-time data infrastructure.

Originally known for lightning-fast key-value storage, Redis has expanded into a broader platform supporting caching, session storage, streaming, search, vector databases, and now AI-agent workloads, making it a foundational layer for modern cloud-native and AI-driven applications.

At its core, Redis solves the problem of speed and real-time data access at scale. By keeping data in memory and offering flexible data structures, Redis enables developers to dramatically reduce latency for applications such as recommendation systems, fraud detection, gaming leaderboards, messaging, and increasingly, LLM and AI agent state management.

How the overnight success of Clawdbot was Built

Over the past few weeks, Clawdbot — recently renamed Moltbot following a complaint from Anthropic — has exploded in popularity.

It is being described as one of the first real personal AI agents with “Jarvis-level” capabilities.

And for once, that description doesn’t feel like marketing hype.

Why Clawdbot Feels Different from Most AI Assistants

Most AI products today stop at conversation. Clawdbot doesn’t.

Its core value isn’t chatting — it’s execution.

Clawdbot is designed to act as a true digital butler, capable of automating tasks across applications and directly operating your device and accounts. It can:

• Monitor flight statuses and complete online check-ins

• Manage calendars and send messages via WhatsApp or other messaging apps

• Execute local CLI commands: rename files, organize folders, write and run code

• Control smart home devices like Sonos speakers

Because it can operate locally, Clawdbot blurs the line between software and human action.

The hype got so intense that:

• Mac Mini sales reportedly surged, as developers began dedicating local machines to running Clawdbot

• Cloudflare’s stock jumped ~14%, as investors noticed developers using its infrastructure to support local AI agents

Suddenly, “local-first AI agents” became investable again.

Not for Everyone — Yet

a16z partner Olivia Moore shared her thoughts after testing Clawdbot:

For developers and technical users, Clawdbot is an experimental playground with endless possibilities.

But for mainstream users, it’s probably not ready yet.

She pointed out three major frictions:

1. Security risks are hard for non-technical users to evaluate

2. Setup complexity is high

3. Most people don’t actually know what they want an agent like this to do

For most consumers, products like Claude Cowork may still be a better fit — at least for now.

An “Overnight Success” Built on 40+ Side Projects

Clawdbot’s sudden popularity also made its creator, Peter Steinberger, an overnight name in the AI world.

But if you look at his GitHub profile, the story feels very different.

Before Clawdbot, Peter had already built more than 40 products — mostly small tools. When people started reviewing those projects closely, a pattern emerged:

Clawdbot is essentially a wrapper that connects all of Peter’s previous work.

Rather than a single product, Peter is assembling an entire AI agent ecosystem.

Clawdbot as the “Brain”

At the center of this ecosystem is Clawdbot (now Moltbot) itself — not a chatbot, but an agentic AI.

All of Peter’s CLI tools and MCP servers ultimately connect to this “brain,” enabling the agent to perform real-world actions instead of just generating text.

Think of it this way:

• Clawdbot = the brain

• Everything else = hands, eyes, and nervous system

The “Hands”: CLI Tools for Action

Peter has built an extensive set of command-line tools that give Clawdbot concrete skills:

Social & Communication: bird for Twitter/X actions, imsg for iMessage, wacli for WhatsApp synchronization.

Life & Work: go-cli for Google Calendar, Docs, and Gmail; remindctl for Apple Reminders; ordercli for food delivery management.

Hardware Control: sonoscli for speakers; blucli for audio systems; camsnap for camera snapshots.

These tools allow Clawdbot to do things, not just suggest them.

The “Eyes”: Seeing and Understanding the UI

To operate a computer like a human, an AI must understand what’s on the screen.

Peter built that too. Peekaboo: Gives AI access to macOS screenshots and GUI automation, enabling visual Q&A over the current screen.

While AXorcist: Uses Swift-based UI control to precisely interact with apps that have no APIs — clicking buttons, navigating menus, controlling interfaces.

This is what turns a language model into an actual computer user.

The Infrastructure Layer

To make the brain and tools work seamlessly together, Peter also built middleware:

VibeTunnel: Enables remote control — you can issue commands to your home Clawdbot from a browser on the street.

mcporter: A core connector that lets TypeScript-based AI plugins behave like APIs or be packaged as CLIs.

SweetCookieKit: A deceptively powerful tool that allows AI to extract browser cookies automatically — solving authentication for protected sites like food delivery or banking portals.

Rewriting the Human–Computer Interface

It’s unlikely Peter started with a master plan like this. But in hindsight, his philosophy is clear:

a LEGO-style approach that turns every corner of macOS into an API an AI agent can call.

He’s not just building a product — he’s rebuilding the interface between humans and machines.

And Yes, He’d Already Made $100M+ Before

This isn’t Peter’s first success.

Before all of this, he sold a significant stake in PSPDFKit to Insight Partners for over €100 million.

After the exit, he openly wrote about feeling empty — traveling, doing therapy, searching for meaning. Eventually, he rediscovered that building things was what gave him the most joy.

That spark led to this entire ecosystem.

Most “overnight successes” look like this when you zoom out.

Meanwhile: The AI-OnlyFans That Quietly Hit $100M+ ARR

We all know how profitable OnlyFans is.

But AI created a new opening: what if OnlyFans itself became AI-native?

Read more

FloQast is a leading Accounting Transformation Platform designed to automate and modernize finance operations for accounting teams worldwide.

Built “by accountants for accountants,” FloQast integrates AI-powered workflows into core accounting processes such as close management, reconciliations, audit readiness, and compliance to reduce manual work and improve operational efficiency.

Trusted by over 3,500 global accounting teams – including enterprise names like Lululemon, Chipotle, and Shopify – the platform’s relevance has surged as finance teams seek better accuracy, visibility, and audit-ready reporting in increasingly complex environments.

The company’s flagship offering, FloQast Transform, enables teams to build, manage, and deploy intelligent, auditable AI Agents that automate repetitive, high-volume tasks – freeing accountants to focus on strategic analysis rather than manual execution.

These auditable AI capabilities help organizations standardize and scale workflows across multi-entity operations and compliance functions, driving strong ROI for users.

Synthesia is now valued at $4B, ARR at~$120M

After hitting $100M ARR in April 2025, Synthesia, the AI video platform transforming how organizations create and share knowledge, announced that it has raised a $200 million Series E funding round at a $4 billion valuation.

What impresses me is Synthesia’s strategic pivot: From Hollywood to the Enterprise.

The most crucial strategic decision in Synthesia’s history was the pivot from a service-based, Hollywood-focused AI dubbing tool to a scalable, self-service platform for enterprise video communications.

The initial market, while glamorous, proved to be too niche and required a high-touch, service-heavy model that was antithetical to venture-scale growth. The founders recognized that the true, massive pain point was not in high-end film production, but in the mundane, repetitive, and costly world of corporate video—training, onboarding, internal communications, and product explainers.

This pivot was the "right thing they did" that unlocked hyper-growth. By focusing on the enterprise, Synthesia solved a clear, quantifiable problem: reducing the time and cost of video creation by up to 80%. This shift transformed the company from a cool technology provider into an essential business utility.

The company’s customer base has swelled to over 60,000 customers, including more than 60% of the Fortune 100. This enterprise traction is the primary driver of its valuation.

Building “AI Employees” for Local SMBs hit $100M ARR in Under 2 Years

However, one of the most instructive AI products I’ve studied recently is a platform designed to provide “AI employees” to small businesses.

In less than 2 years, it scaled from zero to $100M in ARR. More importantly, it has become a textbook example of how AI can help a traditional SaaS company successfully reinvent itself—perfectly illustrating what Sequoia once described as “the biggest opportunity in AI: replacing services with software.”

Read more

Emergent is an AI-driven software creation platform that combines autonomous AI agents with a user-friendly “vibe-coding” interface. The result is a system where users— even those with no prior coding experience— can design, build, test, and deploy full-stack web and mobile applications in a remarkably short amount of time. From idea to production, everything happens inside a single, tightly integrated workflow.

Founded in 2025, Emergent is headquartered in San Francisco, with a large engineering hub in Bengaluru. Its target audience is clear: entrepreneurs, small businesses, and creators who want to turn ideas into real products without assembling engineering teams or navigating the traditional complexity and cost of software development.

Emergent’s ambition is broader. Design, development, testing, deployment, and even monetization are woven together through AI, making software creation feel less like an engineering project and more like a creative process.

The company’s growth metrics suggest strong market pull. In just seven months since launch, Emergent has surpassed 5 million users across more than 190 countries. Over the same period, its annual recurring revenue (ARR) climbed from roughly $100,000 in the early days to $50 million, with projections pointing to more than $100 million in ARR by April 2026.

Software development is becoming increasingly “de-specialized,” moving from a skill reserved for engineers to a capability accessible to a much wider population through AI-assisted platforms.

Investors have taken notice. Emergent recently raised a $70 million Series B round led by SoftBank Vision Fund 2 and Khosla Ventures, valuing the company at approximately $300 million post-money. Including earlier rounds, total funding has reached $100 million within just seven months of the company’s founding.

Additional backers include Prosus, Lightspeed Venture Partners, Together Fund, and Y Combinator.

Sequoia Just Backed an AI-Native Calendly

A few days ago, I wrote about an AI-native version of Calendly that reached $1M in ARR in just 3 months. I didn’t expect another product in the same space to surface so quickly—this time with Sequoia’s backing.

Unlike the previous product, which targeted a specific vertical, this one is built as a general-purpose solution. It feels much closer to what Calendly would look like if it were designed from scratch in the age of AI.

It’s already showing early network effects: through pure organic adoption, it has onboarded more than 200 enterprise customers, with zero human involvement in operations.

Why Scheduling Is Harder Than It Looks

What makes it interesting is that it isn’t just “Calendly with AI added on.” The team rebuilt the entire scheduling stack around a dedicated intelligent agent system.

Read more