Good morning & happy Monday! Today, we’re exploring the collapse of the Stripe-Advent bid for PayPal to see why Stripe refusing to stretch past $60.50 is the most important data point in the whole saga (the $7-9B gap that was never close, why PayPal's Q2 “beat" didn't change anything for buyers, a third buyer that was part of the original consortium and quietly dropped out before anyone knew, and what PayPal's board now has to prove every quarter to justify the cash it turned down + bonus deep dives into Stripe, the latest PayPal’s financials & How to Turn Claude Fable 5 Into a McKinsey-Level Consulting Engine inside), and going to dive deep into Affirm's latest financials to see how the most profitable operator in BNPL just beat the top of its own guidance on every line (breaking down the most important facts & figures from Affirm's Q4 2026, including the 5.8% cost of funds that built every point of margin over 2 years and has now stopped moving, the fiscal 2027 guide that quietly assumes rates go the other way, & the two agentic-commerce launches that appear in no earnings document, to see whether is still worth your time & money + bonus deep dives into the latest financials of Klarna & Nubank inside). So let’s jump straight into the interesting stuff 🌶️

PayPal rejected Stripe at $60.50. Now it trades at $53 😬📉

Following the money 💸 Two months ago we wrote that the offer price might be wrong, but the thesis behind it wasn’t. A month ago we wrote that PayPal rejected that price but kept negotiating. Now the third act: on August 28, Stripe and Advent officially walked away from the M&A. PayPal dropped 12.7% to $53.66, its worst session since February and roughly 11% below the bid the board called inadequate. Ouch 😓

At the core, the $53 billion deal collapsed because the two sides were never within $7 to $9 billion of each other on price, and Stripe chose to spend its capital elsewhere (and given its recent M&As, it sure knows where to put the money).

So let’s take a closer look at this and unpack the four constraints that killed the deal, a detail about a third buyer that quietly dropped out before the formal bid was ever made, and what PayPal’s board now has to prove every quarter to justify the cash it turned down.

Read more

Join the Future of Finance & AI:

1. How to Turn Claude Fable 5 Into a McKinsey-Level Consulting Engine 🧠 [12 Claude Skills that install the methods McKinsey, Bain & BCG train their consultants on. In head-to-head tests they lifted Claude from 38% to 97% on a partner-level checklist. 10 minutes to set up]

2. GLM-5.3-Flash: Local AI Goes Multimodal 🤖 [Why the first open multimodal AI model that rivals Claude Opus 4.8 and GPT-5.6, runs on a single Mac, and was served without NVIDIA chips resets the math for founders, builders, and investors]

3. The Age of the One-Person Company 🦄 [What the latest data reveals about solo founders building million-dollar companies with AI agents instead of employees, and where the next big opportunity is in 2026]

4. Y Combinator’s QM: The Complete Guide to YC’s Open-Source AI Agent Harness 🤖 [The step-by-step guide for the AI agent harness YC uses to run itself: setup, 15+ copy-paste prompts, the highest-impact workflows for startups & small teams, and the security realities you must know]

5. The Ultimate Guide to Qwen3.8-27B 🤖 [How to run Alibaba’s 27B open AI model locally: a step-by-step guide to the settings, hardware requirements, & prompts that deliver frontier AI results for founders, builders, and investors]

6. Are We Bullish Enough on Revolut? 🦄💸 [deep dive into Bullhound Capital’s 56-page report on Revolut, breaking down whether the report is actually bullish about the right things, the 3 options it leaves completely unpriced, including PRAGMA, the in-house AI foundation model quietly becoming the bank’s operating system, plus the 2 numbers most likely to break the $400B path, & actionable takeaways for founders, operators & investors + bonus deep dives into the AI Monopoly Playbook, the latest financials of Monzo & Starling Bank, and The New Enterprise AI Go-To-Market Playbook inside]

7. The New Enterprise AI Go-To-Market Playbook 📚 [The a16z framework that decides which AI startups win the enterprise, and how founders can use it in 2026]

8. Klarna’s Q2 2026: the 42% growth that’s really 20%, a hidden $2B Google win & why KLAR stock crashed 23% anyway 💳📉 [breaking down the most important facts & figures from Klarna's Q2 2026, including the $2B Google judgment buried in a footnote & the guidance raise that's actually flat on the old accounting basis, to see whether is still worth your time & money + bonus deep dive into the latest financials of Nubank & The Complete Guide to YC’s Open-Source AI Agent Harness inside inside]

9. Anthropic’s $30 trillion IPO pitch is 650 times the business it runs today 🤯💸 [how a private slide from the leaked Coatue deck became a public roadshow number, why the TAM isn't dishonest but has a job to do, & the one tell in the upcoming S-1 to look for + bonus deep dives into The New Enterprise AI Go-To-Market Playbook & The Age of the One-Person Company inside]

10. Nvidia just became the buyer of last resort for its own chips 😳🖥️ [why Nvidia wants its best models to be FREE, their coding agent built to make GPUs worth more after they ship & the failed $2B raise that explains why NVIDIA BANK exists + bonus deep dives into The AI Monopoly Playbook & The Ultimate Guide to Qwen3.8-27B inside]

11. How to Build an AI Monopoly: The New Rules of Startup Strategy 🦄 [Anyone can build an AI product now. Almost no one can build an AI company. Here is how to escape competition, build moats that compound, and capture the value you create]

12. How to Build an Agentic OS with Claude Fable 5 🤖 [The complete guide to building an autonomous AI agent operating system with Claude Fable 5 — architecture, working code, cost control, security, and the economics of agentic labor]

As for today, here are the 3 incredible FinTech stories that are changing the world of financial technology as we know it. This was yet another wild week in the financial technology space, so make sure to check all the above stories.

Revolut Research proves fintech’s real AI moat is proprietary data 🤖📊

The BIG News 🗞️ FinTech giant Revolut just launched Revolut Research, a dedicated AI research unit, anchored by PRAGMA, a proprietary foundation model trained on behavioral data from 80 million customers across 100+ countries. Early results are more than solid: 2.3x better credit-default detection, 65% more fraud cases caught. Nice!

But the real story here isn’t one company’s lab announcement.

→ Nubank already shipped a near-identical architecture into production.

→ Visa, Mastercard, Stripe, and Plaid each independently built their own version.

→ NVIDIA even published a blueprint so anyone else can follow.

Importantly, none of these efforts were coordinated. So let’s take a closer look at PRAGMA’s technical design, how it compares to every major peer effort, why Ramp took a different path, and the regulatory fault lines forming around proprietary AI in finance. We also draw on direct responses from Pavel Nesterov, Revolut’s Head of AI, on architecture decisions, production constraints, and the competitive moat question.

Because when the fastest-growing digital banks & fintechs independently reach the same architectural conclusion, pay attention.

More on this 👉 PRAGMA is a family of encoder-style transformers ranging from roughly 10 million to 1 billion parameters. The models are pre-trained via masked modeling on tens of millions of user records spanning tens of billions of events across 100+ countries. They ingest multi-source sequences (transactions, app activity, trading patterns, communication signals, profile state) to build a single representation of customer financial behavior. Downstream teams freeze the backbone and attach lightweight heads for specific tasks, or apply parameter-efficient fine-tuning through methods like LoRA, updating only a few percent of parameters.

The concept behind PRAGMA is older than most people realize. Nesterov started thinking about the problem 4 to 5 years ago, originally convinced that graph neural networks were the right architecture. When Anton Shapkin joined 2 years ago to work on the project, they had what Nesterov describes as “really difficult discussions” about transformers versus graphs on tabular data. Transformers won.

But the team didn’t rush to announce. The work on PRAGMA started roughly a year ago, experiments were validated on historical data, and Revolut formalized the Research unit only after confirming that PRAGMA works in production. In other words, they shipped proof before they shipped a press release.

Nesterov framed the reasoning plainly: institutions cannot rely on third-party blueprints if they want to lead intelligent banking.

The economics back him up. One shared representation means fewer data pipelines, faster iteration (Revolut claims models built in days instead of months), and compounding returns as the dataset grows. Nearly a billion transactions flow through Revolut monthly, and each one therefore makes PRAGMA marginally better at credit, fraud, recommendations, and engagement prediction simultaneously.

Zoom out 🔎 On the question of whether competitors can replicate the work, Nesterov was blunt:

The barrier, he said, is data. Obtaining multi-source behavioral data, especially historical depth, is the real challenge. When the PRAGMA team started building, all the necessary data was already available inside Revolut’s systems. Nesterov credits Revolut’s data culture for that head start, and it’s the kind of advantage that can’t be installed after the fact.

Looking at the bigger picture, the convergence across the industry is the more telling signal.

• We can remember that Nubank’s NuFormer went into production across Brazil and Mexico before Revolut went public with PRAGMA, reporting real lifts in underwriting and support automation.

• Mastercard trained a large tabular foundation model spanning fraud, authorization, loyalty, and personalization.

• Visa has TransactionGPT.

• Stripe trains contextual transaction models that feed its fraud-blocking numbers.

• Plaid built a transaction foundation model with reported gains in classification.

Then there’s NVIDIA, which formalized the whole pattern with a “Build Your Own Transaction Foundation Model” blueprint and published case studies, which tells you the approach has crossed from experiment to industry expectation.

They all reached the same answer: a single behavioral backbone trained on proprietary event sequences beats a patchwork of siloed models bolted onto generic AI.

Interestingly, Ramp offers a useful contrast here.

The company is heavily AI-native in corporate spend management: its internal Ramp Research tool answers analytical questions from data warehouses, and Ramp Labs ships applied AI products while exploring reinforcement learning for sub-agents, model routing, and fine-tuning.

But Ramp is not training a multi-source behavioral foundation model analogous to PRAGMA or NuFormer. Its focus is clearly agents, retrieval, and operational automation on top of existing frontier models. The difference here clearly maps to data type: Ramp sees corporate expense data; Revolut sees full consumer financial lives across dozens of markets and product lines.

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, what comes next depends on who closes the loop fastest. The winning institutions will feed proprietary data into a foundation model, adapt it rapidly for downstream tasks, and use those tasks to generate more data. Late movers cannot replicate that cycle by renting API access to someone else’s model. That said, for traditional banks sitting on fragmented data lakes, the window to build or partner is narrowing with each quarter of behavioral data they fail to unify. Revolut’s next step is an automatic deep learning platform (AutoDL) that connects directly to data sources and fits models without manual feature engineering. Nesterov describes this as “full automation” of day-to-day data science work, with the human role shifting toward managing AI agents and making the critical decisions they can’t.

The risks, though, are specific and confirmed by Nesterov himself.

• On latency: some production use cases require a 200ms response, and PRAGMA currently runs at around two seconds. The engineering team is building what they call the Token Store to solve the real bottleneck, which is not model inference but delivering and tokenizing raw data from multiple production sources in real time.

• On regulation: PRAGMA has not been deployed in heavily regulated use cases such as European credit scoring. Nesterov says the team needs to publish three more papers, including one on explainability, before even initiating those conversations with regulators. The EU AI Act’s high-risk requirements for credit and fraud models aren’t a future concern; they’re a current constraint on where PRAGMA can actually go.

At the end of the day, every financial institution uses AI now. But the dividing line is between those training on their own behavioral data and those renting someone else’s understanding of their customers. Revolut clearly showed which formula is the winning one.

ICYMI: Revolut built EURR stablecoin to own Europe’s euro stables on-ramp 🪙💶 [the PayPal/Paxos playbook applied to on-chain euros, who Revolut is actually racing here & why it's not Circle, and what’s next + bonus deep dives into Revolut’s AI Research Lab and the Ultimate List of Stables Resources inside]

Stripe just acquired the Legal OS behind $140 billion in startup fundraising 🤑💸

The news 🗞️ Atlas gets you born. Clerky keeps you legally alive. Stripe now owns both.

Let’s unpack this.

More on this 👉 At the core, the acquisition looks like a legal tech tuck-in. It’s basically a switching-cost play.

Atlas already forms roughly a quarter of new Delaware corporations, but the product ends where the hard part begins: stock plans, SAFEs, option grants, board consents, foreign qualification. Founders who start on Atlas routinely have an attorney reconstruct their paper trail before they can use Clerky. So Stripe just turned that friction into an internal product problem.

Clerky’s numbers explain why Stripe paid up.

→ Companies on the platform account for 23% of Silicon Valley seed and pre-seed financings and have raised over $140 billion.

→ Formation volume recently grew 6.5x faster than Clerky’s historical rate.

→ Hundreds of attorneys and paralegals already run client work inside Clerky; Cooley, WSGR, and Fenwick counsel treat its documents as standard.

Zoom out 🔎 The logic fits Stripe’s recent pattern almost perfectly. Bridge, Privy, Metronome, OpenRouter: each was specialized infrastructure Stripe would have spent years rebuilding.

Clerky is legal infrastructure in the same mold. Once formation data, cap-table events, SAFEs, banking, and payments sit in one account, Stripe gets better underwriting for Capital and cleaner tax workflows. It also gets a distribution channel where every client a law firm puts on Clerky becomes a potential Atlas and Treasury customer without Stripe lifting a finger. Win-win!

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead (at least in the near term), expect nothing visible. Darby Wong has said the team, focus, and product stay intact. Over the next 12 to 24 months, the tell will be a clean migration path from Atlas into Clerky’s legal dashboard without the current reconstruction tax. Carta still owns the later cap table, but Clerky already issues the stock and options that seed it. And Stripe doesn’t need to buy Carta. It just needs to make Carta slightly less inevitable. Meanwhile, neobanks that partnered with Clerky for formation-plus-deposit flows now find their legal layer sitting inside a competitor. Thus, company creation is becoming a control point, and Stripe just took it.

ICYMI:

Nobody wants to pay for the smartest AI? 🤖🤔

Following the money 👉 By most benchmarks, Claude Fable 5 is the most capable AI model ever released. But two months after launch, it accounts for roughly 11% of what companies spend on Anthropic’s products. And that share has actually flatlined.

Let’s take a quick look at this, understand why it matters, and what’s next.

More on this 👉 Ramp’s AI Index, drawing on spending data from 70,000 US businesses, shows Anthropic’s flagship AI model being outspent by its own mid-tier Opus 5 within weeks of that model’s late-July launch.

Fable 5 costs about $10 per million input tokens, roughly double OpenAI’s GPT 5.6. The cheaper OpenAI model generated more revenue despite weaker benchmark scores. In other words, enterprises are picking models on price, not benchmarks.

Zoom out 🔎 Of course, we must note that Anthropic isn’t the only lab seeing this. Most production AI workloads don’t need the smartest model available. They need the cheapest one that clears a quality bar. Multi-model routing is cutting inference costs 40–80% for early adopters by trying a cheap model first and escalating only when needed. About 60–70% of enterprise tasks land in the “simple” bucket.

Also, Anthropic is still growing fast. Revenue hit $65B annualized in July, up from $47B in May, nearly 7x since January. The company posted its first adjusted operating profit in Q2.

But that growth is coming from volume on cheaper models, and the revenue mix matters when you’re pricing an IPO at $2T or more.

ICYMI: Anthropic’s historic $2 trillion IPO hands Coatue its 2030 exit four years early 🤯💸 [we can finally price Claude AI maker’s IPO thanks to a leaked Coatue deck, why the multiple isn’t the scary part here & one number that decides whether $2T truly holds + bonus deep dives into The New Enterprise AI Go-To-Market Playbook & The Age of the One-Person Company inside]

THE TAKEAWAY ✈️

What’s next? 🤔 First and foremost, for app-layer builders and fintechs, intelligence commoditizing on a curve measured in months reshapes cost structure. The margin advantage therefore goes to teams that treat model selection as an engineering variable: routing intelligently, caching results, fine-tuning small models for narrow tasks, and reserving frontier calls for the 10% of work that demands them. Proprietary data and product experience are hence where defensibility lives now. Looking ahead, the key question for frontier labs now is whether they can grow the pool of high-value agentic workloads fast enough to justify their training budgets before “good enough” eats the rest of the market. While for everyone building on top of these models, cheap and capable inference is already here, and the winners will be the teams that treat the model layer as an optimizable cost input rather than a fixed dependency.

ICYMI:

🧠 What else I’m watching

• DBS Deploys AI Agents 🤖 DBS is rolling out agentic AI for 1,500 bankers globally to assist with corporate credit assessments, handling over 70 tasks to free up time for client relationships. A pilot showed a 30% reduction in research time, aligning with DBS’s push to become an AI-first bank. It will be very interesting to see if AI agents can transform corporate banking efficiency. ICYMI:

• Young Adults Lead AI Finance 📱 Lloyds Banking Group reports that 76% of 18–24-year-olds have used AI for personal finance, which is the highest adoption rate in the UK. Older generations are also embracing AI, with 47% of 55–64-year-olds and 35% of 65–74-year-olds using it for tasks like insurance comparisons and mortgage advice. Is AI bridging the financial confidence gap across generations? 🤔 ICYMI:

• Banks Adopt Ant’s AI Model 📈 Barclays, Deutsche Bank, Citi, and Standard Chartered are using Ant International’s FalconTST 2.0 AI model to enhance cashflow forecasting and FX liquidity management. The model excels at analyzing time-series data across industries, helping banks optimize liquidity and capital allocation. ICYMI:

💸 Following the Money

• Multiplier, an AI startup making agent harnesses for asset managers, has raised $6M in a seed financing round.

• Rillet, the AI-native ERP (enterprise resource planning) platform for modern finance teams, announced a $100M Series C led by ICONIQ at a $1B valuation.

• US-based start-up Fasset, which is described as an AI-powered stablecoin neobanking platform, has reached unicorn status after its latest fundraising effort. Series C funding raised $68M, led by Tokyo-based SBI Group, and followed a Series B round of $51M earlier this year. ICYMI:

👋 That’s it for today! Thank you for reading, and have a relaxing Sunday! And if you enjoyed this newsletter, invite your friends and colleagues to sign up:

Subscribe now

Subscribe now

For six days, nobody knew who built the most-used AI model on OpenRouter.

On August 20, 2026, a listing called “Ox Alpha” appeared on OpenRouter and OpenCode: free, no owner listed, no model card, a 1M-token context window, and text, image, and video input. 6 days later, it was the most-used model on the platform, having processed about 23 trillion tokens, the biggest launch in OpenRouter’s history at roughly 2.3 times the volume of the next model.

On OpenCode, it ended DeepSeek’s 56-day run at the top. Stripe had agreed to buy OpenRouter the day before the model appeared, and Stripe CEO Patrick Collison even called it very impressive.

On August 26, Bloomberg confirmed it. Ox Alpha was GLM-5.3-Flash from Zhipu AI, the Beijing lab that operates internationally as Z.ai (we will use Zhipu & ZAI for simplicity's sake going forward):

→ 320 billion parameters, 18 billion active per token

→ The first natively multimodal model in the GLM-5 series, and

→ The cheapest model at its score on Artificial Analysis’s Intelligence Index.

Weights landed on Hugging Face under an MIT license that same evening, and Zhipu’s Hong Kong shares closed more than 12% higher the next day, roughly 10x January’s IPO price. Not too shabby!

The numbers that matter: GLM-5.3-Flash pushed the Pareto frontier for intelligence per dollar. It scores 57 on the Artificial Analysis Intelligence Index against 60 for the ten-times-pricier GLM-5.3, tying Claude Opus 4.8 at about a twentieth of Opus’s per-task cost at list ($0.09 versus $1.78) and about a fortieth at the promotional rate ($0.045). Let that sink in! 😳

The architecture is the how here: 3x less attention compute and a 4.4x smaller KV cache than GLM-5.3 at 1M-token context. In other words, intelligence is getting cheaper, brutally fast.

And because Unsloth shipped dynamic GGUF quantizations on Day 1, a 1-bit version now runs on a 100 GB machine, and a 3-bit version fits a 128 GB Mac or NVIDIA DGX Spark.

Then there is the claim that reaches past this particular model. Zhipu says the entire stealth week of Ox Alpha traffic was served on a cluster of roughly 100,000 Chinese-made AI chips, with no NVIDIA hardware involved. If that holds up at production quality, the infrastructure moat that U.S. export controls were built to protect is narrower than most investors assume, a shift readers of our piece on the leaked DeepSeek investor call have seen building from the inside.

This deep dive is the complete playbook on GLM-5.3-Flash: how it stacks up against Claude, GPT, Gemini, and DeepSeek on real benchmarks, what it costs, how to set it up in Claude Code and Cline, how and when to run it locally, five copy-paste prompts that unlock its full potential, and the risks every investor and founder has to underwrite before betting on it.

If you build, fund, or compete in AI, GLM-5.3-Flash just reset the price floor under your market, and you need to understand the most powerful open model available today.

What GLM-5.3-Flash is

Read more

Good morning & happy Thursday! Today, we’re breaking down Revolut's EURR euro stablecoin launch to see why the real story here isn't another token but a white-label distribution play (the PayPal/Paxos playbook applied to on-chain euros, who Revolut is actually racing here & why it's not Circle, and what’s next + bonus deep dives into Revolut’s AI Research Lab and the Ultimate List of Stables Resources inside), Anthropic's $30 trillion IPO pitch that tops SpaceX's record and runs about 650x the business Anthropic actually operates today (how a private slide from the leaked Coatue deck became a public roadshow number, why the TAM isn't dishonest but has a job to do, & the one tell in the upcoming S-1 to look for + bonus deep dives into The New Enterprise AI Go-To-Market Playbook & The Age of the One-Person Company inside), and Stripe's Clerky acquisition (the M&A of the legal OS behind $140B in startup fundraising, why this is a switching-cost play disguised as a tuck-in, the attorney distribution channel Atlas could never build on its own & why Stripe no longer needs to buy Carta to make Carta less inevitable + bonus deep dive into The Complete Guide to YC’s Open-Source AI Agent Harness inside inside). Let’s jump straight into the fascinating stuff 🌶️

Revolut built EURR stablecoin to own Europe’s euro stables on-ramp 🪙💶

Following the money (or should I say euros?) 💶 Revolut’s new euro stablecoin, EURR, went live this week across Denmark, Poland and Portugal with €369 in reserves and 80 million banking customers on the other side of the funnel.

Doesn’t look promising? Don’t be so quick to judge because the float is just a rounding error here. The distribution pipe is the whole product. Revolut outsourced issuance to Stripe’s Bridge and kept both the brand and the customer relationship, a white-label split you’ll see more fintechs copy.

At the core, EURR is a pure distribution play. It lacks DEX depth, exchange pairs, and DeFi integrations, but it sits inside a banking app where 80M+ people hold their money & 16M people already trade crypto.

No euro stablecoin issuer has even anything remotely close.

So let’s unpack this, understand who Revolut is actually racing here (not Circle), and why the next token matters far more than this one.

Read more

Good morning & happy Wednesday! Today, we’re breaking down Revolut Research to see why the real story here isn't one company's AI lab launch but the fact that Nubank, Visa, Mastercard, Stripe, and Plaid all independently built the same architecture (PRAGMA's technical design and early numbers, the industry convergence, why Ramp's AI-native strategy took a different path, & the proprietary data moat that late movers can't rent their way into + bonus deep dives into PRAGMA & Why We Aren’t Bullish Enough on Revolut inside), and Nvidia's $7B Poolside deal to see how the company that just turned compute into an asset class became the buyer of last resort for its own chips (why Nvidia wants its best models to be FREE, their coding agent built to make GPUs worth more after they ship & the failed $2B raise that explains why NVIDIA BANK exists + bonus deep dives into The AI Monopoly Playbook & The Ultimate Guide to Qwen3.8-27B inside). So let’s just jump straight into the finnovative stuff 🌶️

Revolut Research isn’t an AI Lab. It’s the clearest sign yet that fintech’s ultimate moat is proprietary data 🤖📊

The BIG News 🗞️ FinTech giant Revolut just launched Revolut Research, a dedicated AI research unit, anchored by PRAGMA, a proprietary foundation model trained on behavioral data from 80 million customers across 100+ countries. Early results are more than solid: 2.3x better credit-default detection, 65% more fraud cases caught. Nice!

But the real story here isn’t one company’s lab announcement.

→ Nubank already shipped a near-identical architecture into production.

→ Visa, Mastercard, Stripe, and Plaid each independently built their own version.

→ NVIDIA even published a blueprint so anyone else can follow.

Importantly, none of these efforts were coordinated. So let’s take a closer look at PRAGMA’s technical design, how it compares to every major peer effort, why Ramp took a different path, and the regulatory fault lines forming around proprietary AI in finance.

Because when the fastest-growing digital banks & fintechs independently reach the same architectural conclusion, pay attention.

Read more

Subscribe now

Companies paid McKinsey, BCG and Bain about $37 billion last year for advice. Most of the work on any engagement was done by someone who joined the firm less than 2 years ago. McKinsey has hired straight from graduate school since 1953, and it was the first firm to do so on purpose.

That only makes sense if the associate is not what the client is buying.

The client is buying a method: the same questions in the same order, the same test on every claim before it reaches a slide, the same structure on every document so a CEO reads page one and knows the answer. The associate delivers it. The method is the product.

The method is not secret. Minto’s Pyramid Principle, Lafley and Martin’s Playing to Win, Kaplan and Norton’s Balanced Scorecard, Heuer’s declassified CIA handbook on competing hypotheses: all published, all readable by the end of the quarter. Reading was never the constraint. The constraint was the two years it takes to turn those books into reflex, under deadline, with a partner checking your logic at every step. Nobody outside the firms had that, so everyone else bought it back at partner rates.

A frontier model like Claude Fable 5 can now run that method with more discipline than a second-year associate, provided someone writes it down in a form the model can execute: steps in order, a checkpoint at each one, the tests a partner applies, a worked example, and the mistakes that produce plausible nonsense.

That is exactly what this guide gives you. Twelve Claude Skills built on the published frameworks, written to one specification and verified against primary sources. Two were tested head-to-head against the same model with no skill loaded: with the skills, the model passed 95-97% of a partner-level checklist; without them, 38-67%. All the details are in the “Proof” section.

All twelve are free to download at the end of this guide. Install them, and Claude stops answering strategy questions like a well-read generalist and starts working like a firm.

In this guide:

1. Why frameworks beat prompts

2. The engine: how the 12 skills fit together

3. The 12 skills, one by one

4. Running an engagement end to end (with the exact prompts)

5. Proof: how the skills were built, verified, and tested

6. How this connects to the earlier guides

7. Download and install (all 12 skills, three ways)

8. Getting the most out of it, and where it falls short

9. FAQ

1. Why frameworks beat prompts

Read more

Good morning & happy Monday! Today, we’re diving deep into Klarna's latest financials to see how the strongest merchant-distribution machine in BNPL just beat on every headline number and still crashed 23%, & why its 42% margin growth was really 20% with a $69M assist from a line item that didn't exist a year ago (breaking down the most important facts & figures from Klarna's Q2 2026, including the $2B Google judgment buried in a footnote & the guidance raise that's actually flat on the old accounting basis, to see whether is still worth your time & money + bonus deep dive into the latest financials of Nubank & The Complete Guide to YC’s Open-Source AI Agent Harness inside inside), and how Anthropic built the smartest AI model on the planet and then watched its own customers buy the cheaper one instead (Fable 5's spending plateau, why Opus 5 is already outselling it, & what this means for app-layer builders and fintechs + The Ultimate Guide to Qwen3.8-27B & The Age of the One-Person Company inside). So let’s just jump straight into the interesting stuff 🌶️

Klarna’s Q2 2026: the 42% growth that’s really 20%, a hidden $2B Google win & why KLAR stock crashed 23% anyway 💳📉

Earnings time 🤑 BNPL giant Klarna just reported the beat everyone wanted - revenue up 27% to $1.04B vs $996M expected, EPS of $0.01 against a forecast $(0.06) loss, transaction margin dollars up 42%, adjusted operating income more than tripled to $91M, net income positive - and the stock crashed 22.81% by the close, then another 4.8% by Friday. Not nice 🥲

At $14.33, KLAR now sits 74.9% below its 52-week high of $57.20 and more than 66% under its $40.00 September 2025 IPO price 🤷‍♂️

The crash itself is no mystery though: Klarna cut its full-year outlook - GMV down to $149-151B from “>$155B”, revenue to $4.08-4.16B from “>$4.34B” against a $4.415B consensus - and Germany, its biggest market, is soft. That’s the story every headline ran with. But it’s not the interesting one.

The most important number of the quarter wasn’t the beat, and it wasn’t the crash either. It was $69 million - the “gain on sale of consumer receivables” line. It was $0 in every quarter through Q3 2025, and this quarter it accounts for 53% of the entire year-over-year increase in transaction margin dollars - the metric management says it optimises for, and a phrase the company uses 25 times in a 20-page release. Strip the line from both periods (a clean comparison, since the prior year is zero) and 42% TMD growth becomes 19.7%. Revenue growth of 26.6% becomes 18.2% - behind GMV growth of 17.5%. And the “+4.5 points” of margin expansion becomes +0.5 🥲

Which leaves Klarna as a strange one: the strongest merchant-distribution machine in BNPL just printed a record quarter, and half the record rests on a line that didn’t exist a year ago.

So let’s dive deeper into Klarna’s Q2 2026, unpack the figures that actually matter, and see whether is worth your time & money in the years to come.

Read more

1. The Age of the One-Person Company 🦄 [What the latest data reveals about solo founders building million-dollar companies with AI agents instead of employees, and where the next big opportunity is in 2026]

2. The Power Shift: Why Electricity Is AI’s Real Bottleneck ⚡️ [AI’s bottleneck is shifting from chips to gigawatts. Who secures power, and on what terms, will decide who captures value from the next phase of AI infrastructure]

3. Y Combinator’s QM: The Complete Guide to YC’s Open-Source AI Agent Harness 🤖 [The step-by-step guide for the AI agent harness YC uses to run itself: setup, 15+ copy-paste prompts, the highest-impact workflows for startups & small teams, and the security realities you must know]

4. The Ultimate Guide to Qwen3.8-27B 🤖 [How to run Alibaba’s 27B open AI model locally: a step-by-step guide to the settings, hardware requirements, & prompts that deliver frontier AI results for founders, builders, and investors]

5. Are We Bullish Enough on Revolut? 🦄💸 [deep dive into Bullhound Capital’s 56-page report on Revolut, breaking down whether the report is actually bullish about the right things, the 3 options it leaves completely unpriced, including PRAGMA, the in-house AI foundation model quietly becoming the bank’s operating system, plus the 2 numbers most likely to break the $400B path, & actionable takeaways for founders, operators & investors + bonus deep dives into the AI Monopoly Playbook, the latest financials of Monzo & Starling Bank, and The New Enterprise AI Go-To-Market Playbook inside]

6. The New Enterprise AI Go-To-Market Playbook 📚 [The a16z framework that decides which AI startups win the enterprise, and how founders can use it in 2026]

7. Nubank’s Q2 2026: the best-run bank on Earth, and the $1 billion quarter that wasn’t 🏦👀 [breaking down the most important facts & figures from Nubank's Q2 2026, including the $500M buyback nobody mentioned & the AI flywheel incumbents can't copy, to see whether is still worth your time & money + bonus guide on how to Turn Claude Opus 5 Into a Financial Analyst That Never Sleeps inside]

8. Ramp is building the CFO for AI Agents 📊🤖 [the Stripe vs Ramp routing war, how USDC agent wallets turn AI agents into controlled corporate spenders, who wins between developers and CFOs + bonus deep dives into How to Build an Agentic OS with Claude Fable 5 & The AI Monopoly Playbook inside]

9. The Everything Router: Stripe to acquire OpenRouter for more than $7 billion 🔀💸 [why the $3B haircut from July's price is the real tell here, what falling token prices reveal about OpenRouter's fee base, & whether this is Stripe's Instagram Moment + How to Build an Agentic OS with Claude Fable 5, and AI Monopoly Playbook inside]

10. PayPal rejected a $53B buyout bid. Now it’s negotiating 👀🤝 [why the Q2 "beat" was mostly arithmetic underneath and gives Lores less leverage than it looks, what a forced Braintree carve-out means for Advent's growing payments empire, & what’s next + bonus deep dive into PayPal’ Q2 2026 earnings & The New Enterprise AI Go-To-Market Playbook inside]

11. How to Build an AI Monopoly: The New Rules of Startup Strategy 🦄 [Anyone can build an AI product now. Almost no one can build an AI company. Here is how to escape competition, build moats that compound, and capture the value you create]

12. How to Build an Agentic OS with Claude Fable 5 🤖 [The complete guide to building an autonomous AI agent operating system with Claude Fable 5 — architecture, working code, cost control, security, and the economics of agentic labor]

As for today, here are the 2 incredible FinTech stories that are changing the world of financial technology as we know it. This was yet another wild week in the financial technology space, so make sure to check all the above stories.

Stripe’s leaked investor letter says the Singularity started January 1st 🤖😳

The news 🗞️ Patrick Collison just told Stripe’s investors that the singularity began on January 1, 2026, and that their company now manages two currencies, money and intelligence.

Let’s unpack this.

More on this 👉 Stripe’s August 19 investor update, written to announce the OpenRouter acquisition, goes past the deal. It identifies capital and intelligence as the two digital flows that run every business, then argues that managing inference spend requires the same discipline as managing financial capital: cost-of-return reasoning on every unit and granular billing against time-value of delay.

Stripe built the pipeline for the first flow. OpenRouter is the pipeline for the second.

Nobody else describes OpenRouter that way. The market calls it a router. Stripe’s own CEO calls it an intelligence pipeline, and the letter treats tokens as a parallel currency to dollars: expensive, heterogeneous, and constantly changing. Accept the framing and the $7 billion price stops looking like a premium for routing. It becomes the entry cost for a second Stripe-shaped business running on a second currency.

The Collisons are careful to strip the word of its apocalyptic baggage. They mean an observed inflection, especially in the rate of new firm creation, and they decided to take it seriously. The numbers back the posture.

• H1 2026 net revenue grew 41% year over year, free cash flow grew 43%, and Stripe Billing grew 71%.

• Atlas now accounts for more than a quarter of all Delaware incorporations, and Link passed 300 million users.

• Token consumption on OpenRouter has been compounding at 9% per week year-to-date, which annualizes to roughly 90x.

Zoom out 🔎 The product stack the letter maps out is the Khala Research diagram we published earlier this week, made official.

• Discovery and onboarding for agents through Stripe Projects, Directory, and a Provisioning API.

• Usage management through Metronome.

• Payments through Bridge, MPP and Tempo.

• Fund storage through Privy and a new stablecoin.

• OpenRouter slots in as the routing and metering layer that feeds every stage downstream.

The letter confirms that Radar, originally built for payment fraud, is already being used to detect token fraud at major AI companies, which is yet another product that deepens lock-in without a new sale.

The letter also answers the dilution question we raised two weeks ago. Stripe’s share count is lower today than three years ago, despite Bridge, Metronome, Privy and now OpenRouter. In other words, the core payments engine prints enough cash to absorb all four without net dilution. Share price has compounded at 31% annually since the Series D a decade ago, against 14% for the S&P 500.

THE TAKEAWAY ✈️

What’s next? 🤔 The deal should close in the coming weeks, and the framing debate closes with it. OpenRouter is not a model router that Stripe overpaid for. It is, in Stripe’s own words, the other half of what every developer needs. And whether the world agrees depends entirely on whether tokens actually behave like a currency. If they do, Stripe just doubled its addressable surface. If they don’t, it bought an API aggregator for 50 times revenue. Our AI Monopoly Playbook readers know which side of that bet the infrastructure layer tends to win.

ICYMI:

Razorpay launches Vulcan, India’s first AI payments foundation model 🤖🇮🇳

The news 🗞️ The most valuable AI models of 2026 may not speak a word, and Razorpay just built one that only reads money.

Let’s take a closer look at this.

More on this 👉 On August 18, Indian FinTech giant Razorpay launched Vulcan, a transformer foundation model trained on roughly 3 trillion data points from 4 billion payments, weighing about 3,000 signals per transaction. It is not an LLM. It treats each payment as an unordered set of fields, tokenizing even missing ones. That design fits India’s mess of UPI apps, cards, netbanking, and cash-on-delivery better than the sequence models Nubank (nuFormer) and Revolut (PRAGMA) built for user histories. One self-supervised backbone now feeds routing, cross-merchant fraud detection, risk scoring, and checkout personalization, replacing siloed single-task models.

The beta numbers, across 51,000+ businesses including Blinkit and redBus, are the story: payment success rates up 8-10%, international card fraud caught 8x more often, and 100,000-200,000 extra completed purchases per month just from showing shoppers their preferred UPI app.

In high-volume Indian e-commerce, where a failed payment sends cash-first shoppers straight back to cash, an 8-10% success lift is worth more than most product launches. And because the model learns from every transaction on the network, patterns invisible to any single merchant, like one stolen card hitting unrelated sellers, become visible. In other words, it’s a moat that compounds with volume- exactly what an IPO-bound company wants investors staring at.

Zoom out 🔎 Vulcan also confirms a pattern: Stripe, Adyen, Mastercard, Nubank, and Revolut are all building domain foundation models on proprietary transaction data, because general LLMs cannot do millisecond, high-stakes tabular reasoning under regulatory constraints. Trained and hosted in India for RBI localization rules, Vulcan turns compliance into a competitive filter foreign models cannot easily pass.

ICYMI: Inside Revolut’s PRAGMA: The Foundation Model Trained on 40 Billion Banking Events 🧠 [Architecture, performance benchmarks vs. Stripe, Mastercard, and Visa, regulatory risks, and why PRAGMA may be the most consequential AI bet in consumer finance]

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, here’s what to watch: whether the gains hold outside a curated beta, and whether Razorpay extends the backbone into lending and credit risk, where the real margin lives. The second-order effect most will miss is agentic commerce. When AI agents start initiating payments, they will route through whichever rail decides fastest and fails least, and a self-improving payments brain is the natural counterparty. Zooming out, the losers here are pure-play gateways and banks without a real-time payment graph; their classical ML now competes against a network effect. Of course, we must note that the honest caveat is that all reported lifts are Razorpay’s own numbers, measured against its own prior models. Directionally, though, the era of fintechs renting intelligence is ending. The data-rich ones are minting their own. AI continues to eat fintech 🤖

ICYMI:

🧠 What else I’m watching

• Binance Launches AI Trading 🤖 Binance has launched Agent OS, a platform enabling AI agents to connect to its systems and execute trades on behalf of users. Users can authorize agents via tools like ChatGPT to access market data and perform trading activities within set permissions, supporting both custom and ready-made integrations across crypto and traditional markets. It seems that AI-driven trading is becoming the new norm on exchanges. ICYMI:

• Z.ai Unveils GLM-5.3 🤖 Z.ai is upgrading its GLM-5.3 model with improved coding to rival Anthropic and OpenAI, releasing its weights within two weeks and outperforming its predecessor. Despite a 9% stock dip, the move underscores China’s push for cost-effective, high-performance AI, with Z.ai’s market cap still at $75 billion and $1 billion in annual recurring revenue. ICYMI:

• Revolut Seeks Finnish License 🇫🇮 Revolut has applied for a banking licence in Finland to establish a branch, enabling its 250,000 Finnish customers to access IBANs and local banking services. The move follows its recent full European banking licence in France and is part of a €1 billion investment to expand across Europe, including a new HQ in Paris by 2027. ICYMI:

💸 Following the Money

• British alternative credit scoring technology startup Noggin HQ has raised £2.3M in an oversubscribed seed round.

• Sterling, the New Zealand AI startup building an autopilot for finance teams, has raised $3.8M (NZD) in a round led by trans-Tasman venture capital firm Blackbird.

• Indian digital financial services firm Navi is set to receive a $100M funding boost from Dutch asset manager Prosus.

👋 That’s it for today! Thank you for reading, and have a relaxing Sunday! And if you enjoyed this newsletter, invite your friends and colleagues to sign up:

Subscribe now

Good morning & happy Friday! Today, we’re diving into Ramp's launch of Router.com to see why the product they shipped alongside it, not the router LLM itself, is the bigger move here (the Stripe vs Ramp routing war, how USDC agent wallets turn AI agents into controlled corporate spenders, who wins between developers and CFOs + bonus deep dives into How to Build an Agentic OS with Claude Fable 5 & The AI Monopoly Playbook inside), and Stripe's leaked investor letter to see why Patrick Collison told his investors the singularity started on January 1st and what that means for the $7B OpenRouter deal (Stripe's H1 2026 financials, the "two currencies" thesis that reframes the entire acquisition, plus the product stack that turns tokens into a payment rail + bonus deep dive into The Ultimate Guide to Qwen3.8-27B inside). So let’s jump straight into the finnovative stuff 🌶️

Ramp is building the CFO for AI Agents 📊🤖

The news 🗞️ When we first covered Ramp Router on July 22, we clearly said it is not a product but rather a wedge into AI spend management. A month later, the picture is even bigger than we thought.

On Wednesday, Ramp launched Router.com, an AI model routing service built on 3 years of internal production use. Hours earlier, Stripe had officially confirmed its $7.5B acquisition of OpenRouter, the developer-favorite model marketplace, at nearly 6x the valuation it held just three months prior.

These two companies are not strangers. Stripe co-led Ramp’s 2021 Series B at a $1.6B valuation and still provides card-issuing and stablecoin infrastructure that Ramp relies on today.

Now they want the same layer: the routing decision that determines which AI model gets paid per request. And that’s before you account for what Ramp shipped alongside the router. That second product may actually matter more than the router itself.

Read more

Subscribe now

Qwen3.8-27B is the “DeepSeek Moment” for open source AI. A 27-billion-parameter file you can easily download tonight now does the kind of coding and agent work that needed a closed, hosted frontier AI model just a year ago. Let that sink in! 😳

Some AI developers even called this open weights release the most civilization-changing 20 GB of data ever published. While clearly exaggerated, the story of Qwen3.8-27B now definitely matters more than any single benchmark number attached to it.

Qwen 3.8 27b running locally vs. Opus 4.8. Source: Alex Finn via X

Alibaba’s Qwen team shipped Qwen3.8-27B on August 14, 2026: a dense, Apache 2.0-licensed, vision-capable model with native image and video understanding and a 262,144-token context window. Unlike the mixture-of-experts giants dominating recent frontier releases, every one of its 27 billion parameters activates on every token.

The full weights compress to a 17-18 GB file at 4-bit quantization, small enough to load, alongside its KV cache, on a single 24 GB consumer GPU such as an RTX 3090 or 4090, or a mid-range Apple Silicon Mac. Parameter count, file size on disk, and the RAM or VRAM you need to run it are three different numbers, and this guide keeps them that way throughout.

Qwen 3.8 27B Q4 running on an RTX 4060 with just 8GB VRAM and a 64,000-token context window using Unsloth's new IQ4_XS quant at 14.6GB on disk.

This AI release is unlike anything we’ve seen before, and the data proves it. Independent third-party benchmarking from Artificial Analysis puts Qwen3.8-27B at 52 on its Intelligence Index, a composite across reasoning, knowledge, math, and coding. That score ties GPT-5.6 Luna (max) and sits one point behind GLM-5.2 (max) and DeepSeek V4 Pro (max), despite both being vastly larger mixture-of-experts systems.

On Qwen’s own reported agentic-coding benchmarks, it also jumps sharply over its dense predecessor, Qwen3.6-27B: from 63.4 to 73.0 on Terminal-Bench 2.1, for instance. Of course, we should treat the vendor-reported numbers as directional and the Artificial Analysis figure as the more independent read here, but both point in the same direction.

It’s important to note that none of that translates automatically into a good result on your machine. Local performance depends on which quantization you pick, how much memory and context headroom you actually give it, which runtime you use, and how hard you make it think.

By far the biggest reason “download and run” fails the majority of people trying to run this locally is the model’s own defaults. Qwen3.8-27B thinks by default, and its default reasoning setting is the most expensive one available. Left alone, it will burn tens of thousands of reasoning tokens on requests a human would answer in one sentence. In long agent sessions, that same verbosity can quietly exhaust the context window and return empty output that looks like an unrelated error. Therefore, getting frontier-adjacent results out of a 27B model you own outright means fixing that mismatch deliberately, not hoping the defaults are fine.

The same pattern actually runs through the whole open-weight AI surge of the past year: Kimi K3 pushes the frontier at enormous scale, while Qwen3.8-27B makes the more personally useful point that frontier-adjacent capability can now fit on hardware a single builder already owns. Our earlier deep dive covered that same local-AI shift through GLM-5.2 (GLM-5.2: The ChatGPT Moment for Local AI).

That’s why, in this guide, we go deep on the part that actually determines your results: how you run the model, not just whether you can. Here’s what you’ll unlock:

1. The exact reasoning_effort and context-window settings that stop the model from overthinking and burning your token budget on trivial tasks.

2. A hardware-and-quantization decision table matched to real VRAM tiers, 8 GB laptops through 48 GB workstations, including which quant is currently the best default at each size.

3. The launch flags (MTP speculative decoding, KV cache precision, chat template handling) that separate a sluggish setup from one running 2-3x faster on identical hardware.

4. A copy-paste prompt library built around the model’s documented failure modes, so you inherit the fixes instead of rediscovering them the hard way.

5. Full local deployment walkthroughs for llama.cpp, Ollama, LM Studio, vLLM, and SGLang, plus the security and privacy calls worth making before you put this in front of real users or real data.

A misconfigured setup pays full compute cost for a model quietly working against you. A well-configured one runs something close to a frontier coding and agent system, on hardware you already own, for the cost of electricity.

Let’s make sure you run the latter.

What Qwen3.8-27B actually is

Read more

Good morning & happy Wednesday! Today, we're diving deep into Nubank's latest financials to see how the best-run bank on Earth just printed its first-ever $1 billion quarter - and why it was really an $896M quarter with a $165M tax break (breaking down the most important facts & figures from Nubank's Q2 2026, including the $500M buyback nobody mentioned & the AI flywheel incumbents can't copy, to see whether is still worth your time & money + bonus guide on how to Turn Claude Opus 5 Into a Financial Analyst That Never Sleeps inside), and Razorpay Vulcan, India's first AI foundation model built purely for payments (what a "shared brain" trained on 4 billion payments actually does, how Vulcan stacks up against Stripe, Nubank's nuFormer & Revolut's PRAGMA, and why the real endgame is agentic commerce + bonus deep dives into How to Build an Agentic OS with Claude Fable 5 & The Complete Guide to YC’s Open-Source AI Agent Harness inside). So let’s just jump straight into the finnovative stuff 🌶️

Nubank’s Q2 2026: the best-run bank on Earth, and the $1 billion quarter that wasn’t 🏦👀

Earnings time 🤑 Latin America’s digital banking giant Nubank just reported the monster quarter everyone wanted - net income above $1 billion for the first time in its 13-year history ($1.06B, +49% FX-neutral), revenue up 39% to $5.9B, and return on equity at a record 33% - and the stock jumped 9.3% the next session. Nice!

This yet again proves that Nubank might be the best-run company in banking, and one of the best-run anywhere in the world 🏦

In a capital-intensive, heavily regulated industry, across emerging markets, it just posted a 19.5% efficiency ratio (Brazilian incumbents run in the 40s and 50s), a 33% ROE, and a cost to serve stuck near $1.00 per customer per month against $17.1 of ARPAC - with 10.4k employees covering 139M customers (~13,400 each, an order of magnitude above incumbents), ~80% of them historically acquired organically, and $45.3B of deposits priced below the risk-free rate.

Behind it sits a real flywheel: a decade of transaction data feeding NuFormer, the in-house AI model management says cut predicted risk ~70% versus prior generations, while headcount fell last quarter and revenue grew 50%. And the playbook clearly travels: Mexico hit break-even in 6 years versus Brazil’s 8, and is monetising at more than double the ARPAC Brazil had at the same stage. Nubank definitely is a compounding machine - the closest thing banking has to a SaaS business with a balance sheet.

But the most important number of the quarter wasn’t the billion, and it wasn’t the rally either. It was 14.17% - Nubank’s effective tax rate, down from 27.56% a year ago. Re-tax Q2 at last year’s rate and the first billion-dollar quarter becomes an $896M quarter, while the tax line alone accounts for 39% of the entire year-over-year profit increase. And the words “tax rate” appear zero times in the release, the 39-slide deck, and the hour-long earnings call 🥲

That combination - arguably the best-run bank on Earth, and a record headline flattered by a line nobody discussed - is exactly what makes Nubank one of the most interesting stocks in banking & fintech right now.

So let’s dive deeper into Nubank’s Q2 2026, unpack the figures that actually matter, and see whether is worth your time & money in the years to come.

Read more

Subscribe now

For two years, the AI industry’s supply anxiety centred on silicon. Nvidia allocation lists, TSMC wafer starts, and memory shortages were visible, well covered, and easy to track. Electricity looked like somebody else’s problem, much as bandwidth did before the internet became mass infrastructure.

That assumption no longer holds. In 2026, the practical limit on AI infrastructure is shifting from how many GPUs a company can buy to how many gigawatts it can secure, connect, and afford. Global data centre electricity demand grew 17% in 2025, and demand from AI-focused facilities grew 50%, according to the IEA’s April 2026 report on energy and AI. The agency now expects data centre consumption to roughly double from 485 TWh in 2025 to 950 TWh by 2030.

Capital alone cannot deliver that power. A new AI campus needs turbines, transformers, grid connections, permits, and a community willing to host it, and several of those inputs take longer to secure than the chips they support. The result is therefore a collision between software-speed demand and infrastructure-speed supply.

It is also a more complicated investment thesis than “AI needs electricity, so buy power.” Some players sell scarce inputs with pricing power. Others commit capital for decades while their largest customers retain the option to slow spending within a quarter.

This deep dive maps who sits on which side of that spread: why turbine makers repriced while power producers did not, how capacity auctions and behind-the-meter deals actually price power, and why the strongest bear case is duration rather than demand. It then works through the turbine and fuel bottlenecks, what physically gets built before 2028, the storage and flexibility market forming in the gap, the political backlash reshaping site selection after Texas, and the decisions founders, operators, and investors need to make over the next three years.

Let’s dive in.

The core thesis: power has repriced as compute infrastructure

Read more

Good morning & happy Monday! Today, we’re diving into Stripe's now-finalized deal to acquire OpenRouter for more than $7 billion (why the $3B haircut from July's price is the real tell here, what falling token prices reveal about OpenRouter's fee base, & whether this is Stripe's Instagram Moment + How to Build an Agentic OS with Claude Fable 5, and AI Monopoly Playbook inside), and the biggest shift in the PayPal takeover saga: PayPal rejected the $60.50 bid, called it inadequate... then never stopped negotiating (why the Q2 "beat" was mostly arithmetic underneath and gives Lores less leverage than it looks, what a forced Braintree carve-out means for Advent's growing payments empire, & what’s next + bonus deep dive into PayPal’ Q2 2026 earnings & The New Enterprise AI Go-To-Market Playbook inside)So let’s just jump straight into the awesome stuff 🌶️

The Everything Router: Stripe to acquire OpenRouter for more than $7 billion 🔀💸

Following the money 💸 Just 3 weeks ago, we covered that the talks were near $10 billion territory. On Saturday, Bloomberg reported the terms are now set: FinTech giant Stripe has agreed to buy OpenRouter for more than $7 billion, merely three months after the routing startup raised at a $1.3 billion valuation. Not too shabby! 😳

Of course, neither company has confirmed the M&A yet, and the price could still move, but this would be the largest acquisition in Stripe’s history, and its clearest statement yet on what AI agents will do to commerce.

Let’s dive deeper into this and unpack why $3 billion came off the price in 3 weeks and what that reveals about OpenRouter’s fee base, the full Stripe logic from LLM routing to Google’s Universal Commerce Protocol, whether this is Stripe’s Instagram Moment, and the signals that will tell you if $7 billion was cheap.

Read more

If you’re not a subscriber, here’s what you missed this week:

1. The Age of the One-Person Company 🦄 [What the latest data reveals about solo founders building million-dollar companies with AI agents instead of employees, and where the next big opportunity is in 2026]

2. Y Combinator’s QM: The Complete Guide to YC’s Open-Source AI Agent Harness 🤖 [The step-by-step guide for the AI agent harness YC uses to run itself: setup, 15+ copy-paste prompts, the highest-impact workflows for startups & small teams, and the security realities you must know]

3. Are We Bullish Enough on Revolut? 🦄💸 [deep dive into Bullhound Capital’s 56-page report on Revolut, breaking down whether the report is actually bullish about the right things, the 3 options it leaves completely unpriced, including PRAGMA, the in-house AI foundation model quietly becoming the bank’s operating system, plus the 2 numbers most likely to break the $400B path, & actionable takeaways for founders, operators & investors + bonus deep dives into the AI Monopoly Playbook, the latest financials of Monzo & Starling Bank, and The New Enterprise AI Go-To-Market Playbook inside]

4. The New Enterprise AI Go-To-Market Playbook 📚 [The a16z framework that decides which AI startups win the enterprise, and how founders can use it in 2026]

5. Anthropic’s historic $2 trillion IPO hands Coatue its 2030 exit four years early 🤯💸 [we can finally price Claude AI maker’s IPO thanks to a leaked Coatue deck, why the multiple isn't the scary part here & one number that decides whether $2T truly holds + bonus deep dives into The New Enterprise AI Go-To-Market Playbook & The Age of the One-Person Company inside]

6. Shopify’s Q2 2026: 30%+ growth everywhere, 1.48 cents going nowhere & why SHOP stock soared 23% anyway 🛍️🚀 [breaking down the most important facts & figures from Shopify's Q2 2026, including the 1.48 cents of gross profit per GMV dollar that stayed frozen through a year of AI launches, payments wins and enterprise logos, to see whether is still worth your time & money + bonus deep dives into the latest financials of Coinbase, Robinhood, and How to Build and Agentic OS with Claude Fable 5 inside]

7. Grok Bot: SpaceXAI bets on agent operating system for AI teammates 🤖💻 [what Grok Bot actually is, how it stacks up against OpenAI's ChatGPT Work, Anthropic's Claude Cowork & Google's Gemini Spark, and what this means for the future of AI + bonus deep dives into Y Combinators Open-Source AI Agent Harness, and How to Build an Agentic OS with Claude Fable 5 inside]

8. How to Build an AI Monopoly: The New Rules of Startup Strategy 🦄 [Anyone can build an AI product now. Almost no one can build an AI company. Here is how to escape competition, build moats that compound, and capture the value you create]

9. The Ultimate Guide to Kimi K3 🤖 [Moonshot’s 2.8T open model is now #1 in the world for frontend coding and AI agents at a third of Claude’s price. What it changes for founders, builders, and investors, and exactly how to operate it]

10. Turn Claude Opus 5 Into a Financial Analyst That Never Sleeps 📊 [Claude Opus 5 just took #1 on the leading independent finance benchmark. Here’s the framework that turns it into an AI analyst that works for you 24/7]

11. How to Build an Agentic OS with Claude Fable 5 🤖 [The complete guide to building an autonomous AI agent operating system with Claude Fable 5 — architecture, working code, cost control, security, and the economics of agentic labor]

12. Loops at Scale: The Governance Layer Nobody Built Yet 🔁 [Why founders running AI agent loops in 2026 need harnesses, outer-loop ownership, and swarm governance, not just better prompts]

As for today, here are the 3 incredible FinTech stories that are transforming the world of financial technology as we know it. This was yet another insane week in the financial technology space, so make sure to check all the above stories.

Compute is an Asset Class now, and NVIDIA is the Bank of AI 🤖🏦

The BIG News 🗞️ $5.26 trillion tech giant Nvidia’s $500 billion new financing agreement with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR turns GPUs into loan collateral for the AI build-out, and the price of Wall Street’s participation is that Nvidia must guarantee what its own chips will be worth secondhand.

Jensen Huang just agreed to those terms, which makes the world’s most valuable chipmaker something closer to a bank than a semiconductor company 👀

Let’s take a closer look at this to see why the stock fell on the news, who actually gets the cheap money, the number to hunt for in the first prospectus, the contingent liability nobody has added up yet, and why the whole deal is a page from our AI Monopoly Playbook.

More on this 👉 On August 10, Nvidia signed memorandums of understanding with the six firms to mobilise more than $500 billion of third-party capital into “compute financing platforms” that will lend to Nvidia customers at rates they can’t get alone.

It’s important to note that nothing is binding here, no vehicles exist and no money has been raised yet, but the architecture is specific & clear: debt sold to pensions, insurers and sovereign wealth funds, secured by the chips being bought or leased, with Nvidia backstopping up to 25% of a project’s value through a residual-value support mechanism.

Nvidia’s stock closed down 2.9%, which was not too surprising, because nobody builds a half-trillion-dollar lending machine for customers who can pay cash 🤷‍♂️

Zoom out 🔎 The good news? The hyperscalers can. Morgan Stanley sees them spending $3.5 trillion through 2028, funded by an investment-grade bond flood that reached $344B of AI-linked issuance by early August, up more than $200B on all of 2025. That flood crowds out everyone beneath them.

We must also remember that CoreWeave paid over 9% on a $2.6B loan this month, well wide of where its bankers marketed it; Galaxy Digital paid nearly 10% on junk bonds for a Texas data centre leased to CoreWeave. Apollo’s Jim Zelter puts the full build-out above $8T. At double-digit yields, a large slice of that never gets built, and Nvidia’s revenue stalls with it.

BlackRock’s Larry Fink compares the moment to inventing mortgage-backed securities in the 1970s, and the comparison actually makes a ton of sense. Mortgages had houses under them; chip values have historically rotted, and the thing that rots them fastest is Nvidia’s own product cadence. The residual-value floor converts that objection into a contingent liability, so the faster Huang ships better chips, the bigger the hole he may owe on the old ones.

Legendary Michael Burry, who shorted MBS the first time around, calls it “structuring unnatural credits to prolong momentum.” Meanwhile, Broadcom already ran the pilot, taking up to $29B of exposure on June’s $35B Anthropic chip-lease financing.

THE TAKEAWAY ✈️

What’s next? 🤔 Looking ahead, the first real test comes August 26, when Nvidia reports earnings. Once public, skip the revenue beat and look for the number nobody has consolidated: aggregate exposure across Nvidia’s stakes in the labs that buy its chips, its vendor financing and now these residual-value floors. The second test is the first prospectus, whose depreciation schedule will tell debt investors what Nvidia actually believes about its roadmap before equity holders hear it. Then count the copycats: Broadcom is in, AMD has every reason to follow, and once every chipmaker guarantees its own collateral, compute-as-collateral stops being an experiment and becomes market structure. Finally, watch who gets the money. The six partners will finance the labs and neoclouds they already know, so access to frontier compute becomes a credit decision made inside Nvidia’s orbit while everyone else pays CoreWeave prices or goes without. Readers of our AI Monopoly Playbook will recognize the play: dominate the wedge, then build a moat that compounds as the market grows, and no moat compounds faster than being your ecosystem’s cheapest source of capital. Nvidia is clearly running the framework at a scale its startup counterparts can only sketch. Oh, and the $500B pipeline concentrates the build-out instead of democratizing it 👀 All in all, Nvidia sold shovels to the gold rush, and now it owns the bank that decides who digs. Well played Jensen, well played sir 👏

ICYMI:

Cloudflare just shipped the full stack for AI Agents that spend money 🤖💸

The news 🗞️ Cloudflare skipped the AI model wars last week and shipped something harder to replicate. Its Agents Week 2026 produced a vertically integrated infrastructure for agents that browse and transact under enterprise governance.

Two flagship launches, the Kitesurf headless browser and the open-source Cloudflare OS workspace, join the Wallets and Monetization Gateway announced days earlier to close the loop from agent compute to agent spend.

The lock-in economics underneath are exactly what matter for agentic commerce and fintech. Let’s take a closer look at this.

More on this 👉 Kitesurf is a headless browser built for agents instead of people. It skips the full Chromium process entirely. Rust components compiled to WebAssembly run on Cloudflare Workers. The benchmarks claim 3-4x less CPU and 5-7x less memory than warm Chromium, at the cost of roughly 1.8x slower wall-clock time.

For agent fleets doing price research or form filling at volume, the economics are clear. It speaks Chrome DevTools Protocol, so Puppeteer and Playwright scripts port with minimal changes. It’s free in beta.

Cloudflare OS is the enterprise governance layer. Cloudflare deployed it internally first, and the sales team reportedly saved over 10,000 hours in a single month.

The Gatekeeper model is the piece worth watching. Agents start with zero access, every data source they touch is recorded, and when someone shares an agent’s output, the platform re-checks the recipient’s permissions against the underlying data. Users can turn conversations into full-stack apps running on Workers and Durable Objects.

You deploy into your own Cloudflare account, but the runtime is Cloudflare’s.

Zoom out 🔎 Now add the Wallets and Monetization Gateway announced earlier in the week, and the whole picture becomes complete.

→ Kitesurf gives agents cheap eyes on the commercial web.

→ OS governs what they do with what they find.

→ Wallets provide programmable stablecoin rails with hard spend limits.

An agent researches a market via Kitesurf, makes a decision under OS policy, and pays through a wallet. All of it auditable, all of it on Cloudflare.

ICYMI: Cloudflare gave AI Agents a stablecoin wallet. Now it needs them to spend 🪙👛 [what Cloudflare Wallets actually do & why the 14-month two-sided infrastructure stack they complete matters, how Google already won the consumer protocol race but Cloudflare is playing a completely different game with x402 + bonus deep dives into Building an Agentic OS with Claude Fable 5 & the ultimate list of stables resources inside]

THE TAKEAWAY ✈️

What’s next? 🤔 First and foremost, we must note that the lock-in is clean. The full stack runs on Cloudflare compute. Gatekeepers on Workers, user apps on Workers, Kitesurf sessions on Browser Run. Open-sourcing the OS drives adoption while the meter stays on Cloudflare infrastructure. Thus, for fintech, agent-driven price discovery and market research just got dramatically cheaper, putting procurement bots and compliance scrapers in reach at volumes that weren’t viable before. Yet, the subtler effect is that the Gatekeeper pattern hands regulated industries an actual governance framework for autonomous agent spend. That’s still the missing piece between “we prototyped an agent” and “our CFO lets it buy things.” All in all, fintechs that make their rails agent-native gain distribution through platforms like this. Everyone else gets scraped by Kitesurf.

ICYMI: Block cut 40% of its workforce and then open-sourced Buzz, the AI Agent Workspace that makes it work 🤖🐝 [how it actually works, why calling it a Slack killer misses the ultimate goal, what Block’s internal numbers say about running a company on AI agents + the Full AI OS to Run a Startup with Claude & How to Build and Agentic OS with Claude Fable 5 inside]

Tempo Earn hands Stripe a stablecoin yield business the GENIUS Act forgot to ban 🤑🪙

The news 🗞️ A year after Congress banned stablecoin issuers from paying yield, Stripe’s chain is paying up to 4% anyway, and on paper nobody is.

Let’s unpack this.

More on this 👉 Tempo, the payments chain incubated by Stripe and Paradigm, launched Tempo Earn on Aug. 12. Platforms can now pay rewards on idle stablecoin balances, sourced through tokenized money market funds and onchain lending, and keep whatever share of the return they choose.

Deel is the first deployment: contractors paid in DLUSD, a white-label dollar issued through Stripe’s Bridge, earn a promotional target of up to 4% APY via Morpho vaults on Tempo, with a spend card due in Q3.

Deel runs payroll for 40,000-plus businesses and 1.5 million workers, many of them watching dollar earnings lose value the moment they hit a local account. Easy win!

Zoom out 🔎 Once you look at the bigger picture, it’s clear that the legal architecture is the product here. Section 4(a)(11) of the GENIUS Act bars issuers from paying holders interest, so Earn keeps the issuer out of the money flow entirely; lending protocols pay, the platform picks the assets, and Deel’s help center calls DLUSD a “digital dollar voucher” rather than a cryptocurrency.

For Stripe, the launch finishes the vertical stack it has been assembling since buying Bridge.

Those are basically a bank’s deposit economics without a charter, and Stripe never touches the yield itself. Well played, well played 👏

Of course, we must note that none of this is big yet. Tempo carries $29.9 million in stablecoin supply and $14.4 million in TVL; Morpho’s Tempo deployment holds $6.9 million against $7.96 billion protocol-wide. Nobody is monetizing. Stripe is therefore buying position before the rules are set.

THE TAKEAWAY ✈️

What’s next? 🤔 First and foremost, we must note that the rules will set soon. The OCC’s March proposal would presume an issuer is paying interest whenever yield reaches holders through a related third party, language that fits white-label issuance like DLUSD, and the American Bankers Association wants it wider still, because bankers recognize deposit flight when they see it. The statute bites Jan. 18, 2027, Congress never defined who counts as a holder, and Deel’s wallet already skips the US, UK, EU and Australia, a map of where the lawyers see risk. More platforms will definitely follow Deel, and BlackRock’s BRSRV is slated for the asset menu. Ultimately, the structure survives the statute as written, but whether Earn ever comes onshore is an open question, and Circle’s Arc, welded to USDC, has no answer to a stablecoin-agnostic chain that pays. So watch this space, as it will get even more interesting very soon.

ICYMI:

🧠 What else I’m watching

• Revolut Wins French Banking License 🇫🇷 FinTech giant Revolut has just secured a full banking license in France from the ACPR and ECB, accelerating its European expansion and enabling it to offer loans, mortgages, and savings products to its 7 million French customers. The move follows a €1 billion investment pledge and comes despite past regulatory scrutiny over risk controls. This is yet another move that will solidify Revolut’s place as Europe’s fintech leader. ICYMI:

• Scotiabank Deploys AI Agents 🧠 Scotiabank has introduced AI-powered knowledge agents through Scotia Intelligence to help employees access internal expertise, navigate processes faster, and focus on high-value client work. With over 71,000 employees using the tools and 14 million actions generated, the bank is shifting from general AI assistance to workflow orchestration, supported by training for 80% of its global workforce. ICYMI:

• Banks Face AI Concentration Risk 🤖 Moody’s warns that banks’ reliance on a narrow set of AI vendors risks creating systemic dependency, with potential widespread outages and future price hikes from loss-making providers like Anthropic and OpenAI. Regulators are expected to increase scrutiny on operational resilience and third-party concentration in AI stacks. Irrespective of whether or not AI dependency becomes the next systemic risk for banks, it’s yet another reason to start paying more attention to open source. ICYMI:

💸 Following the Money

• Mastercard completes BVNK acquisition to expand stablecoin payments infrastructure.

• Erebor, the new bank backed by tech billionaires that plans to fill the gap left by Silicon Valley Bank’s collapse, is on the verge of raising $1.5B in funding, according to the Financial Times.

• Anthropic is in talks to purchase Decart AI for around $6B, sources reported to Bloomberg. Nvidia-backed Decart develops its own AI models, which can support Anthropic’s plans for growth. Decart’s Lucy model can edit videos in real-time, and its Oasis physical AI model is designed to simulate environments to train robots and self-driving vehicles.

👋 That’s it for today! Thank you for reading, and have a relaxing Sunday! And if you enjoyed this newsletter, invite your friends and colleagues to sign up:

Subscribe now

Good morning & happy Friday! Today, we’re diving into Anthropic's historic $2 trillion IPO, which would be the largest listing ever (we can finally price it thanks to a leaked Coatue deck, why the multiple isn't the scary part here & one number that decides whether $2T truly holds + bonus deep dives into The New Enterprise AI Go-To-Market Playbook & The Age of the One-Person Company inside), and Tempo Earn, Stripe's new stablecoin yield play that pays up to 4% a year after Congress banned exactly that (how the GENIUS Act workaround actually works, why Deel's 1.5M contractors are the perfect first customers, & what’s next + bonus dives into Stripe & the ultimate list of stables resources inside). So let’s just jump straight into the market-changing stuff 🌶️

Anthropic’s historic $2 trillion IPO hands Coatue its 2030 exit four years early 🤯💸

Following the money 💸 Six Anthropic backers told the Financial Times this week that they expect the Claude maker to go public in October 2026 at a valuation of $2 trillion or more, which would make it the largest IPO in history, ahead of SpaceX’s $1.77 trillion June debut.

A leaked Coatue deck, built on December 2025 company data, shows exactly what one of Anthropic’s most aggressive investors privately underwrote to get in.

The fund’s entry price is in there. So is its modeled 2030 exit, and so is the 35% IRR that October’s rumored offer would hand back four years ahead of schedule. Not bad at all… 👀

Let’s take a closer look at this and unpack Coatue’s math in detail, the one number in the deck that I think decides whether $2 trillion truly holds, and the price where I’d buy. Hint: that price is a lot lower.

Read more

Subscribe now

For over a decade, “solo founder” was startup shorthand for “couldn’t find a co-founder.” Investors read it as a risk flag. Accelerators nudged applicants to pair up. The entire venture playbook assumed that building something valuable took a team, because a team was the only way to cover everything a company needs done: code, sales, design, support, books.

That assumption is now measurably wrong. Solo-founded startups grew from 23.7% of new US ventures in 2019 to 36.3% by mid-2025 (Forbes). On Stripe Atlas, solo founders accounted for 63% of new C corporations formed in the second quarter of 2026, an all-time high.

Carta puts solo founders at 36% of all startups incorporated on its platform in 2025, roughly double the share of a decade earlier.

Different platforms, different populations, same trajectory.

The solopreneur economy is therefore a clear structural change in how companies get built, and it changes what founders should be optimizing for over the next few years.

Here’s what we’re getting into today: the $80 million solo exit and the venture funds now backing one-person companies; the formation data across the US and Europe showing this is not another pandemic-style blip; the 61x revenue gap between top-decile solo founders and the median, and the specific habits behind it; why AI-native solo startups generate nearly twice the revenue of their peers by year 2; where the one-person model breaks down; and the 3 places the next opportunity is actually forming.

To make this even more valuable and actionable, and so that you can go straight from reading to building, we’re also sharing the playbooks that pair with it:

• The full AI OS to run your startup with Claude

• Anthropic, which just told AI founders exactly what to build in 2026 (full founder playbook + where Anthropic’s own products will & won’t compete)

• The new rules of AI startup strategy: how to pick a wedge you can dominate and build moats that strengthen as models improve

• The distribution playbook for a market where building is cheap and being found is the scarce resource

• The complete guide to QM, Y Combinator’s open-source AI agent harness

• The top 100 AI and tech investors of 2026, and how to reach them

One person with the right stack can now outcompete a team of 20. Everything below, the full analysis and every playbook above, exists to make sure that person is you.

The headcount-to-revenue ratio just broke

Read more

Good morning & happy Wednesday! Today, we’re diving into NVIDIA, which just convinced six of Wall Street's biggest names to turn compute into an asset class and itself into the bank of AI (what the $500B financing deal with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs & KKR actually is, why fell on the news, & one thing nobody added up yet + a bonus deep dives into The AI Monopoly Playbook Nvidia is now running at a scale, and Who Actually Makes Money When AI Eats the World inside), and Grok Bot, SpaceXAI's new AI agent OS that drives the UI like a human employee would (what Grok Bot actually is, how it stacks up against OpenAI's ChatGPT Work, Anthropic's Claude Cowork & Google's Gemini Spark, and what this means for the future of AI + bonus deep dives into Y Combinators Open-Source AI Agent Harness, and How to Build an Agentic OS with Claude Fable 5 inside). So let’s just jump straight into the fascinating stuff 🌶️

Compute is an Asset Class now, and NVIDIA is the Bank of AI 🤖🏦

The BIG News 🗞️ $5.26 trillion tech giant Nvidia’s $500 billion new financing agreement with Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs and KKR turns GPUs into loan collateral for the AI build-out, and the price of Wall Street’s participation is that Nvidia must guarantee what its own chips will be worth secondhand.

Jensen Huang just agreed to those terms, which makes the world’s most valuable chipmaker something closer to a bank than a semiconductor company 👀

Let’s take a closer look at this to see why the stock fell on the news, who actually gets the cheap money, the number to hunt for in the first prospectus, the contingent liability nobody has added up yet, and why the whole deal is a page from our AI Monopoly Playbook.

Read more

Subscribe now

Y Combinator, arguably the world’s top startup accelerator, just open-sourced QM (short for Quartermaster). The repo describes it as “a multiplayer agent harness for work. In Slack and on the web”.

YC has been running it internally for months across accounting, legal, events, and engineering, including building QM with QM, and recently released the whole thing under an MIT license. In other words, any team can now deploy it for the cost of cloud hosting and model tokens.

Interestingly, it arrived just 10 days after Block open-sourced Buzz, the Nostr-based agent workspace we recently covered in our newsletter, and the two now make opposite bets:Buzz wants to replace Slack and GitHub outright, while QM plugs into the Slack you already run.

The launch did big numbers: 2 million+ views on the announcement thread, 7,000 GitHub stars inside three days (above 13,000 now), a Hacker News thread that finished over 600 points. But the more interesting claim came from inside YC. Eve Bouffard, the accelerator’s head of design and QM’s self-described number-one power user, says she watched partners and staff “use it to 1000x themselves,” and credits the tool with part of how a small team gets so much out the door.

Of course, you should take the multiplier as primarily marketing here, yet the question underneath it is real, and every startup is asking it now: how do you get AI agents out of one engineer’s terminal and into the whole company without losing control of them? QM is YC’s answer.

Unsurprisingly, almost nobody is deploying it well yet.

QM still sits at version 0.1.x, YC’s announcement warns it is “an experiment. It’s early and has bugs,” and the docs assume you already know your way around cloud infrastructure.

That said, whether a team ends up with compounding value or an expensive transcript generator mostly comes down to setup decisions, prompts, and workflow habits the official docs never spell out. This guide was written exactly to fill that gap.

We read the QM source and everything substantive written about it since launch, and kept what holds up:

• A deployment path checked against the published code

• 15+ copy-paste prompts for the 9 highest-impact use cases

• A first-week rollout plan

• The security tradeoffs QM’s own SECURITY.md admits, & most teams skip

• A 30-day pilot framework that ends in an actual adoption decision.

Let’s dive in.

What QM actually is

Read more

Good morning & happy Monday! Today, we’re diving into Cloudflare, which just shipped the full stack for AI agents that spend money (what Kitesurf & Cloudflare OS actually are, why Cloudflare just made agent browsing 3-4x cheaper than Chromium, how the full stack locks in when you add Wallets and the Monetization Gateway & what this means for agentic commerce and fintech + bonus deep dives into Claude Managed Agents & The AI Monopoly Playbook inside), and Shopify's latest financials to see why a monster quarter sent the stock soaring 23%, & why we're still not buying (breaking down the most important facts & figures from Shopify's Q2 2026, including the 1.48 cents of gross profit per GMV dollar that stayed frozen through a year of AI launches, payments wins and enterprise logos, to see whether is still worth your time & money + bonus deep dives into the latest financials of Coinbase, Robinhood, and How to Build and Agentic OS with Claude Fable 5 inside). So let’s just jump straight into the finnovative stuff 🌶️

Cloudflare just shipped the full stack for AI Agents that spend money 🤖💸

The news 🗞️ Cloudflare skipped the AI model wars last week and shipped something harder to replicate. Its Agents Week 2026 produced a vertically integrated infrastructure for agents that browse and transact under enterprise governance.

Two flagship launches, the Kitesurf headless browser and the open-source Cloudflare OS workspace, join the Wallets and Monetization Gateway announced days earlier to close the loop from agent compute to agent spend.

The lock-in economics underneath are exactly what matter for agentic commerce and fintech. Let’s take a closer look at this.

Read more

If you’re not a subscriber, here’s what you missed this week:

1. Are We Bullish Enough on Revolut? 🦄💸 [deep dive into Bullhound Capital's 56-page report on Revolut, breaking down whether the report is actually bullish about the right things, the 3 options it leaves completely unpriced, including PRAGMA, the in-house AI foundation model quietly becoming the bank's operating system, plus the 2 numbers most likely to break the $400B path, & actionable takeaways for founders, operators & investors + bonus deep dives into the AI Monopoly Playbook, the latest financials of Monzo & Starling Bank, and The New Enterprise AI Go-To-Market Playbook inside]

2. The New Enterprise AI Go-To-Market Playbook 📚 [The a16z framework that decides which AI startups win the enterprise, and how founders can use it in 2026]

3. How to Build an AI Monopoly: The New Rules of Startup Strategy 🦄 [Anyone can build an AI product now. Almost no one can build an AI company. Here is how to escape competition, build moats that compound, and capture the value you create]

4. The Ultimate Guide to Kimi K3 🤖 [Moonshot’s 2.8T open model is now #1 in the world for frontend coding and AI agents at a third of Claude’s price. What it changes for founders, builders, and investors, and exactly how to operate it]

5. Turn Claude Opus 5 Into a Financial Analyst That Never Sleeps 📊 [Claude Opus 5 just took #1 on the leading independent finance benchmark. Here’s the framework that turns it into an AI analyst that works for you 24/7]

6. How to Build an Agentic OS with Claude Fable 5 🤖 [The complete guide to building an autonomous AI agent operating system with Claude Fable 5 — architecture, working code, cost control, security, and the economics of agentic labor]

7. Coinbase’s Q2 2026: a record 10.3% market share, a $359M loss & why COIN stock fell 13% anyway 🥇📉 [breaking down the most important facts & figures from Coinbase's Q2 2026 to see whether is still worth your time & money + bonus deep dive into Robinhood’s latest financials, & How to Turn Claude Opus 5 Into a Financial Analyst That Never Sleeps inside]

8. Cloudflare gave AI Agents a stablecoin wallet. Now it needs them to spend 🪙👛 [what Cloudflare Wallets actually do & why the 14-month two-sided infrastructure stack they complete matters, how Google already won the consumer protocol race but Cloudflare is playing a completely different game with x402 + bonus deep dives into Building an Agentic OS with Claude Fable 5 & the ultimate list of stables resources inside]

9. Anthropic Just Told AI Founders Exactly What to Build in 2026 🦄 [1 million conversations. 9 consumer AI domains. A full founder playbook - plus where Anthropic’s own products will and won’t compete]

10. Loops at Scale: The Governance Layer Nobody Built Yet 🔁 [Why founders running AI agent loops in 2026 need harnesses, outer-loop ownership, and swarm governance, not just better prompts]

As for today, here are the 3 fascinating FinTech stories that are changing the world of financial technology as we know it. This was yet another wild week in the financial technology space, so make sure to check all the above stories.

Visa just paid $2.4 billion for BioCatch for the right to authenticate Your AI Agent 🤖🔐

Following the money 💸 AI agents are about to start spending money autonomously, and nobody in payments has a credible answer for how to verify they’re authorized to do it. Global losses from account takeovers and AI-driven scams already top $1 trillion a year, and agentic commerce hasn’t even scaled yet.

Visa’s $2.4 billion all-cash bid for BioCatch is its answer.

Most will call this a fraud acquisition. What Visa actually bought is ownership of the trust layer for machine-initiated commerce, with consequences for every network, issuer, and fintech building toward an agentic future.

Let’s unpack this.

More on this 👉 BioCatch doesn’t catch fraudsters after they act. It continuously reads thousands of behavioral signals (keystroke cadence, touch pressure, device handling, session navigation) to determine whether the entity interacting with a bank is who it claims to be, across roughly 19 billion sessions a month.

The company already protects 760 million users at more than 350 banks. Its models already score for AI-agent usage and jailbroken devices.

Zoom out 🔎 That scoring capability is what makes this more than a fraud deal. Visa has spent the past year building “Visa Intelligent Commerce”: APIs, a Model Context Protocol server, agent toolkits, token provisioning for AI agents, and live pilots with banks across Europe and Asia-Pacific.

The infrastructure for agents to browse, select, and buy on your behalf is nearly ready. What’s been missing is a continuous way to verify that an agent acting on your behalf is actually yours, operating within your constraints, and not compromised.

BioCatch fills that gap. Integrated into Visa’s risk engine and agent-token lifecycle, behavioral scoring becomes persistent authentication for agent sessions, one that checks the session’s behavioral signature against the authorized pattern continuously, before money moves, rather than relying on a one-time password at login or a device fingerprint a sophisticated attacker can spoof.

THE TAKEAWAY ✈️

What’s next? 🤔 First and foremost, we must note that the acquisition fits two other recent deals: Featurespace (late 2024) for transaction-level AI risk scoring, and Prisma/Newpay (February 2026) for Argentine payment rails with built-in tokenization. Visa is clearly assembling a full-stack trust platform. Fraud and risk tools are already among its fastest-growing revenue lines, and pushing BioCatch-grade intelligence across 14,500-plus institutions turns every transaction into a higher-margin one. Readers of our AI Monopoly Playbook will recognize the pattern: build the moat from the data flywheel, and make sure it improves when AI models improve. Every new agent session feeding BioCatch’s network thus makes the behavioral models harder to replicate and the trust signal more valuable, which is exactly the kind of compounding defensibility that separates an acquisition from an asset. Zooming out, the second-order effect is competitive. Visa is raising the cost of entry for agentic commerce. Any network routing agent-initiated transactions will need equivalent behavioral depth, and the pool of acquirable behavioral biometrics firms just shrank. Mastercard bought Recorded Future in 2024, but threat intelligence and continuous behavioral authentication solve different problems at different layers. Looking ahead, watch for Visa to bundle BioCatch scoring into its value-added services for issuers, thus making behavioral authentication a default network feature. When that happens, the moat stops being about fraud rates and becomes about which rails AI agents are built to trust first.

ICYMI: Visa and Mastercard’s Agentic AI payment platforms have a volume problem 🤖📊 [what it’s all about, why it matters & why payment titans have a volume problem here + bonus reads into other Visa & Mastercard’s agentic AI moves, & how Google wants to be the OS for all commerce inside]

Visa’s Stablecoin Platform launches with Open USD, sending Circle down 6% 😳🪙 [why Visa followed the same pattern Anthropic set with Claude Managed Agents, & what’s the bigger play here + bonus deep dives into Coinbase, Circle, and the Ultimate List of Stables Resources inside]

OpenAI fires back at Apple’s trade secret lawsuit, but the evidence cuts both ways ⚔️⚖️

The news 🗞️ OpenAI broke litigation protocol on August 3 by publishing a blog post, internal iMessages, and email correspondence to publicly rebut Apple’s trade secret claims. The company’s central argument: Apple’s lawsuit rests on sloppy lawyering, mischaracterized facts, and Apple’s own failure to revoke system access when employees leave.

The move is unusual for a defendant in active federal litigation and reads less like a legal strategy than a pre-IPO reputation play. The published emails reveal an embarrassing blunder by Apple’s outside counsel that weakens the original complaint.

But the iMessages OpenAI released to exonerate Chang Liu tell a more complicated story.

More on this 👉 At the core, the email chain hurts Apple more than it helps OpenAI. OpenAI published it showing that Apple’s outside counsel at Weil Gotshal sent the February warning letter to the wrong person after confusing two people with Asian last names, then falsely claimed in a follow-up email to have spoken with OpenAI’s General Counsel by phone. OpenAI’s GC, Che Chang, emailed Apple’s in-house team directly calling out the fabrication.

Weil’s lawyer later admitted the mix-up. For a company that built part of its case on the claim that OpenAI ignored a good-faith warning, that’s a real problem.

The iMessages are more complicated. OpenAI published them to show that Apple employees repeatedly reached out to Chang Liu after he left, asking for help locating files and technical details. On their face, the texts do show Apple employees initiating: requesting airdrop help on Liu’s last day, asking technical questions about redacted product decisions in February and March 2026. But read the same thread from Apple’s perspective and a different picture forms. A former employee, now at a direct hardware competitor, was freely answering detailed technical questions about unreleased products and directing colleagues to specific internal file repositories and named contacts. On March 5, an Apple manager flagged one such exchange as “highly irregular” and shut it down. In trade secret law, who started the conversation matters less than what information moved and where it ended up.

Zoom out 🔎 On Tang Tan, OpenAI’s defense is a single sentence: he “has always been clear with the team that we do not want, and must not use, any confidential information.” No documentation, no process specifics. Against Apple’s claim that Tan ran structured interviews designed to extract trade secrets from candidates, that’s thin 🤷‍♂️

Then there’s what OpenAI didn’t address. The “LOL, I found out I can access the [server]” text that anchored Apple’s original complaint is absent from the response entirely. So is any rebuttal of the allegation that OpenAI misled a manufacturing partner into demonstrating Apple’s proprietary metal-finishing process.

THE TAKEAWAY ✈️

What’s next? 🤔 All in all, the strategic logic of going public is clearer here than the legal logic. OpenAI has an S-1 on file targeting a Q4 2026 listing at roughly $850 billion. Every week the dominant narrative stays at “OpenAI stole Apple’s hardware secrets” makes that valuation harder to defend. A public counter-narrative, even an imperfect one, gives bankers and prospective investors something to cite. Other than that, it’s clear that the February email chain genuinely wounds Apple’s pre-suit conduct story. But the iMessages, published to exonerate Liu, may end up in Apple’s next filing as evidence of exactly what they alleged: that confidential product information continued flowing to an OpenAI employee long after he should have been cut off. Thus, going public shaped the OpenAI IPO narrative. It may also have given Apple’s lawyers more to work with. Wild times 🍿

ICYMI:

Kimi turned a credit card into an AI distribution channel 🤖💳

The news 🗞️ AI giant Moonshot AI found a way to make a bank pay for its user acquisition.

Let’s take a quick look at this.

More on this 👉 Kimi, the Chinese AI lab whose consumer app has been losing ground to Doubao, Qwen, and DeepSeek, launched a co-branded credit card with Agricultural Bank of China and American Express. The reward currency is inference compute, not cashback or miles. Every swipe earns tokens redeemable for Agent runs, Kimi Code capacity, multi-agent clusters, and priority model access. The official pitch frames AI as infrastructure “like water, electricity, and the network,” and the card wraps a familiar consumer meter around that infrastructure.

The product looks like fintech, but the logic is pure distribution. Kimi’s app rankings have slipped in one of the most brutal AI consumer markets on earth, and paid subscriptions alone can’t cover the GPU bill, especially after repeated capacity crunches following model launches like K3. A physical card reaches high-frequency spenders who would never download an AI app, locks earned credits inside Kimi’s ecosystem so every cardholder becomes a retained user by default, and lets a state bank’s interchange economics subsidize what would otherwise be free-trial compute costs.

Zoom out 🔎 The timing matters too. Moonshot was in the middle of a valuation sprint from roughly $20 billion to $35 billion, with IPO preparation underway. “AI plus consumer finance” tells a fuller story to capital markets than “another chatbot,” and the company has already telegraphed what comes next, including agent-initiated payments, bidirectional token-to-points conversion, and credit lines tied to token balances.

THE TAKEAWAY ✈️

What’s next? 🤔 First and foremost, whether the program works long-term depends on where reward redemption concentrates. If earned compute pulls users into Kimi’s highest-margin features (agents, coding, long-running workflows), the card becomes a real retention engine. But if credits get diluted into generic chat, it’s a marketing stunt with a bank’s logo on it 🤷‍♂️ On the supply side, China Merchants Bank is already experimenting with MiniMax tokens, so the pattern is replicating domestically. Looking ahead, OpenAI or Anthropic partnering with a Visa or Mastercard issuer isn’t hard to imagine, but privacy constraints around combining spending data with AI prompts will slow any version outside China considerably.

ICYMI:

🧠 What else I’m watching

• BBVA Deploys AI Agents 🤖 BBVA’s call center teams in Italy and Germany have developed AI assistants that cut response times for common customer inquiries by over 15%, sifting through internal data to deliver tailored answers in natural language. Built on BBVA’s corporate AI platform, the tools handled 40,000 messages in June alone, showcasing how frontline teams can drive impactful innovation. ICYMI:

• Robinhood UK Crypto Approved ✅ Robinhood’s UK arm is now registered with the FCA to offer cryptocurrency services, meeting AML requirements and joining over 50 approved firms, including Ripple and BlackRock. The approval arrives just ahead of the UK’s new crypto regulatory framework, set to launch in October, with a brief registration window for firms. ICYMI: Robinhood’s Q2 2026: a 44% EPS beat, a $106M gain that’s already gone & why HOOD stock fell anyway 🏹📉 [breaking down the most important facts & figures from Robinhood’s Q2 2026, including the $106M gain that vanished before the earnings call even started, to see whether is still worth your time & money + bonus guide on how to Turn Claude Opus 5 Into a Financial Analyst That Never Sleeps inside]

• Plaid and Sierra Partner Up 🤝 Plaid and Sierra have partnered to enable AI agents to access financial data securely via Plaid within Sierra’s Horizon platform, transforming one-off conversations into multi-step workflows like loan refinancing or insurance claims. The integration prioritizes security, consumer control, and compliance, allowing agents to orchestrate complex tasks without breaking the conversation. It will be interesting to see whether this redefines how we interact with financial services. ICYMI:

💸 Following the Money

• Capital markets tokenization platform Licuido has secured a strategic investment from Ripple to scale its operations and support the development of digital capital markets infrastructure on the XRP Ledger (XRPL).

• Amex Ventures invests in SMB growth platform Pie.

• Pan-African payments fintech Moment raises $22M.

👋 That’s it for today! Thank you for reading, and have a relaxing Sunday! And if you enjoyed this newsletter, invite your friends and colleagues to sign up:

Subscribe now