BusinessIssue #89

Is Your 2-Year-Old Startup Plan Still Valid?

AI hasn't just changed the technology — it's rewritten the entire startup formula

Oswarld (Kwangseob Ahn)· Writer & editor of INLEVEL9

Is Your 2-Year-Old Startup Plan Still Valid?

Opening

Dear subscriber, imagine meeting a startup founder you invested in 6 years ago. This founder spent 5 years writing code to solve the hard technical problem of autonomous driving. The business model was unique, and the technical moat1 was solid. But then, preparing for a funding round, they opened up the investor deck — and found the world had completely changed.

I’ve recently been talking with several people who run GP funds and do angel investing, and I keep hearing that they’re completely redoing their investment reviews, or that startup investment trends have fundamentally shifted. I’ve even heard that some funds have already gotten LP approval to pull out of their weakest portfolio companies at the next round — taking the loss — and redeploy that capital into AI companies. This isn’t just something I’ve heard secondhand — it’s also a real story that Lean Startup2’s founder Steve Blank recently shared on his blog. His conclusion is blunt: “Most startups over 2 years old already have an invalidated business plan.” Today, let’s talk about why this is happening — and what we should take away from it.

🌊 VC Money Is Flooding Into AI

Let’s start with the numbers. According to a report the OECD published this past February, 61% of global venture capital investment in 2025 — $258.7 billion (about ₩370 trillion) — went to AI companies. That’s more than double the 30% share in 2022, in just 3 years. What’s even more striking is the concentration within that figure. Mega-deals (investments of $100 million or more) accounted for 73% of total AI investment, and deals of $1 billion or more alone made up nearly half. Massive rounds like OpenAI’s $40 billion round and Anthropic’s $13 billion round are what’s pulling up the overall numbers.

Korea is no different. According to The VC’s data, the share of domestic AI investment expanded from 9.4% in 2022 to 23.6% in 2025 — and surpassed 45% in Q1 2026. Data from Hyeoksin-ui-sup (“Innovation Forest”), a Korean startup-data platform, points the same direction. According to Innovation Forest, AI’s share of domestic startup investment rose from 27.9% in 2024 to 31.7% as of August 2025, and by deal count, AI/deep-tech/blockchain has taken the No. 1 or No. 2 spot almost every month. Total investment in 2024 was 1,416 deals worth about ₩6.7564 trillion, and even as the total number of deals falls, the tilt toward AI keeps intensifying. Fewer deals but more money means capital is concentrating on a handful of AI companies while everyone else struggles more.

These numbers tell a clear story. Startups with nothing to do with AI now have to compete over an ever-shrinking slice of the pie. This is exactly the situation facing “Chris,” the autonomous-drone startup founder Blank writes about. While Chris kept his head down building technology for 5 years, the war in Ukraine caused the autonomous drone market to explode. According to PitchBook data, VC investment in defense-tech startups reached $49.1 billion in 2025, nearly double the $27.2 billion of the year before. 10 new unicorns emerged in this space in 2025 alone, and Anduril raised an additional $2.5 billion at a $30.5 billion valuation.

Chris’s product was a perfect fit for medical evacuation or supply transport in conflict zones — but he didn’t even know this opportunity existed. As Blank puts it, the aerial-platform integration technology Chris built is still competitive. But the business model surrounding that technology needs to be rewritten from scratch.

The era when “keeping your head down and staying focused” was a startup virtue is ending. Now, keeping your head up and scanning your surroundings is a condition for survival.

⚡ The Entire Cost Equation of Software Has Changed

The second shift Blank points to is that the underlying economics of software development have collapsed.

In early 2025, Andrej Karpathy coined the term “vibe coding”3 — describing a workflow where you explain your intent in natural language and AI generates the code. It had enough cultural impact that Collins Dictionary named it Word of the Year for 2025. And now it’s not an experiment — it’s standard practice. 25% of Y Combinator’s Winter 2025 batch had AI generate more than 95% of their codebase, and Cursor, an AI code editor, surpassed $2 billion in annual revenue as of early 2026. One survey found that 92% of developers worldwide use AI coding tools at least once a month.

Building an MVP4​ used to take months. Now it takes days, sometimes hours. Reports of prototyping speeds increasing 3-5x are common, and for simple CRUD apps5​, there are cases of speedups exceeding 10x. Blank makes a sharp point here: “The MVP is no longer proof of a team’s capability.” Because anyone can build one now.

This raises a fundamental question for agile6 development methodology too. Blank points out that agile was, at its core, a serial process — test one hypothesis at a time, look at the results, move to the next stage. But with AI agents, you can test 5 pricing models, 10 messages, and 20 UX flows simultaneously. Sequential development is turning into parallel development. In Blank’s words, the bottleneck isn’t “can you build and ship it” but “do you know what to test”. That means judgment, customer insight, and distribution — not engineering — are becoming the core competencies.

Work that used to require a 10-person dev team can now be done by 2-3 people, sometimes even one person alone. When development costs fall, barriers to entry fall too. And that means the tech stack, team size, and roadmap you built 2 years ago are now all over-investment.

🤖 AI Agents Are Redefining What Software Even Is

The third shift is even more fundamental. Blank frames it this way: until now, software has been a tool for showing users information. Dashboards, notifications, workflows — all of it exists to tell you “what to do next.” But customers don’t buy software to look at a dashboard. They buy it to get the job done.

AI agents7 execute that “next step” directly. They resolve customer inquiries, schedule meetings, qualify leads, and order inventory. Blank’s phrase is striking: the next generation of apps won’t display information on a screen — they’ll act like employees. They resolve support tickets, book meetings, screen leads, and reorder stock.

Once that happens, the pricing structure changes too — from “charge per seat” to “charge per outcome.” A fixed price per resolved ticket, a fixed price per booked meeting. When software shifts from being an “interface” to being a “result,” what customers pay for shifts from “access rights” to “achieved outcomes.”

Blank sums this up in one line: the era of Product/Market Fit is ending, and the era of AI Agent/Customer Outcome Fit is arriving. The MVP (Minimum Viable Product) is becoming the MPO (Minimum Productive Outcome)8​. If your product is still popping up “do this next” on a screen while a competitor’s product just goes ahead and does it for the customer, your product isn’t really a competing product anymore.

Hardware startups aren’t exempt either. Blank notes that before building a physical prototype, you can now use AI to simulate far more design variants, build digital twins9, and stress-test your assumptions earlier and more cheaply. And once AI gets embedded in the backend of hardware, a camera becomes a surveillance system, a vibration sensor becomes a machine-failure prediction system, and a robot becomes a factory worker. The moat is no longer the hardware itself — it’s the combination of what the hardware senses and what AI decides and does with that data.

🧱 The Sunk-Cost Trap

This is where it gets hardest. Like Blank’s “Chris,” you have years of accumulated code, a team, a roadmap. There are plenty of reasons this is hard to abandon: “the VC invested in this idea,” “the customer still wants this UI,” “the team believes in this roadmap.” And here’s the key point Blank makes — all of these become reasons not to pivot.

Blank splits sunk costs into two categories.

What’s still an asset: deep domain knowledge, customer relationships, proprietary data, hard-won regulatory approvals, physical integration. For Chris, this includes the system integration with the aerial platform. These are things AI can’t replace overnight.

What’s already become a liability: a large engineering team built around slow development cycles, a seat-based pricing model, a feature-driven product roadmap. Blank calls this the “dead moose on the table” — something so obviously wrong that nobody wants to point it out.

The core question Blank poses is this: “If you were starting this company from scratch today, with today’s tools, in today’s market, what would you build?” This question is uncomfortable when you already have a funded thesis. But Blank warns it’s less uncomfortable than hearing an investor say, at the next round, “We’re not funding this anymore.”

Oz’s Lens

Honestly, reading Blank’s piece, I kept thinking about Korea’s startup ecosystem.

I’ve worked on GTM strategy, and there’s a pattern I see over and over in the Korean market: building the technology first, then looking for a market. The question shouldn’t be “our technology is this good, where can we sell it” — it should start from “how do we produce this customer’s outcome.” When Blank talks about the shift from Product/Market Fit to Agent/Customer Outcome Fit, he’s saying this order needs to be fundamentally reversed.

And one more thing. Looking at The VC and Innovation Forest data together, the tilt toward AI investment is unmistakable domestically too. By The VC’s numbers, the number of deals in Q1 2026 fell 17% while the total amount invested rose 55%, and in Innovation Forest’s monthly rankings, AI/deep-tech/blockchain never loses its No. 1 spot by deal count. That means capital is concentrating in a handful of mega-deals. This is exactly the same global pattern Blank describes. If you’re running a non-AI startup, you need to be able to answer, right now: “why can’t an AI-native competitor replace us?”

What I find especially notable is what Blank puts on the list of “sunk costs that are still assets”: domain knowledge, customer relationships, regulatory approvals — these are areas where Korean companies have traditionally been strong. Combining the domain expertise accumulated in semiconductors, shipbuilding, and automotive with AI — that’s the “defensible moat” Blank is talking about. It’s not about how much code you’ve stacked up. It’s about whether you have something code can’t replace.

Closing

Blank’s message boils down to three points.

First, the playbook from before 2024 doesn’t work in 2026. Fundraising, tech stack, business model — all of it has changed. Second, defensible moats still exist, but their nature has changed. Things like proprietary data, deep understanding of customer outcomes, regulatory barriers. A few tens of thousands of lines of code is no longer a moat. Third, you need to ask yourself, every single quarter: “if I were starting over today, what would I build?”

What makes Blank’s piece land so heavily is that he’s the person who created the concept of Lean Startup in the first place. He’s the one telling us that even that methodology needs an update now.

References & Further Reading

Primary sources

Background

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis — business data management and business analytics — while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master’s from Korea University’s Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

  1. Moat: a business defense wall that competitors can’t easily replicate. Originally referring to the water-filled ditch surrounding a castle, it’s a term Warren Buffett is fond of using in investing. Patents, proprietary data, and network effects can all serve as moats. ↩

  2. Lean Startup: a startup methodology systematized by Steve Blank and Eric Ries. Instead of building a perfect product before launching, you quickly build a Minimum Viable Product (MVP), test customer reactions, validate hypotheses, and iterate. ↩

  3. Vibe Coding: a concept OpenAI co-founder Andrej Karpathy proposed in early 2025. Instead of writing code directly, you describe your intent in natural language and AI generates the code. Collins Dictionary even named it Word of the Year for 2025. ↩

  4. MVP (Minimum Viable Product): a minimal version of a product with only its core features. Built to test market reaction before creating a polished, complete product. ↩

  5. A CRUD app is a basic application that performs the core database operations: Create, Read, Update, and Delete. ↩

  6. Agile: a development approach where software isn’t completed all at once but built incrementally in short cycles (usually two weeks), incorporating feedback along the way. It spread widely from the early 2000s as an alternative to the Waterfall model. ↩

  7. AI Agent: an AI system that takes instructions from a person and autonomously judges and executes tasks. While earlier AI would just say “here’s the result,” an agent goes further — it says “I looked at the result and carried out the next step too.” ↩

  8. MPO (Minimum Productive Outcome): a concept Blank newly proposes in this piece. Where MVP was about building a “minimal product,” MPO is about producing a “minimal customer outcome” — a validation standard suited to the age of AI agents. ↩

  9. Digital Twin: an exact virtual replica of a physical product or system. Used to test virtually before building the real thing, or to monitor the real-time state of a system already in operation. ↩

Your take shapes the next issue

What resonated most in this issue, or where has your experience been different?

Any registered reader can comment for free.

Related issues

Issue #87Business

Middle Managers Aren't Vanishing. Something Else Is

Titles are changing, but the essence of management isn't going anywhere

Issue #86Business

Bloomberg Terminal Is Ugly and Clunky—Everyone Still Uses It

Iran's parliament speaker tweeted a Bloomberg command—by accident?

Issue #84Business

Apple's Caviar Store Forgot How to Sell Cheese

Vision Pro's failure wasn't about the product — it was the cost of gutting retail's soul in the name of 'efficiency.'