Steve Blank on the Startup Whose Market Moved While It Built Tech

While the tech was being built, the customers and competitors changed

Steve Blank wrote on his blog on March 17 about meeting “Chris,” a founder he had invested in six years earlier. Chris had been developing autonomous operation technology and integrating it into an existing aviation platform. But over the five years he spent building the tech, the number of competitors solving similar problems grew, and new demand emerged in an adjacent defense market. Blank’s view: the technical competitiveness¹ Chris had built up still mattered, but the business model he’d originally drawn up needed a full re-examination.

I recently met with a few GPs running funds and some angel investors. I heard a lot of the same story — reviewing investments from scratch, or noting that the flow of startup investment itself has changed. One fund manager told me they’d gotten limited partner (LP) approval to sell off, at a loss, stakes in companies they judged unlikely to raise follow-on rounds, and reinvest the proceeds in AI companies. To be clear, these are cases I heard directly from people I met, not findings from a survey of the broader fund landscape.

Blank is a figure who, through his customer development methodology, heavily influenced the lean startup² approach. He warns that if more than two years have passed since a company’s founding, there’s a good chance many of its underlying assumptions have shifted. I read this not as a call to abandon the existing business entirely, but as a prompt to check what’s changed — in investment, in development, and in the customer.

Investment Money Is Concentrating in AI Companies

In an analysis released in February, the OECD found that AI companies accounted for 61% of global venture investment by value in 2025—$258.7 billion. That’s a sharp jump from 30% in 2022. About 73% of AI investment came from large deals exceeding $100 million, and deals over $1 billion alone made up roughly half of total AI investment. This means a handful of massive deals carry substantial weight in the overall figures.

The same concentration shows up domestically. According to TheVC’s tally, AI’s share by value rose from 9.4% in 2022 to 23.6% in 2025, then to over 45% in the first quarter of 2026. This category includes AI semiconductors as well as applied services that use AI as their core technology. Rather than mixing classifications or reporting periods from different sources, it’s better to track change within a single consistent dataset.

Companies raising capital outside AI also need to examine how investors’ benchmarks have shifted. That said, a rising AI share of investment doesn’t mean funding fell in every non-AI sector, nor that every AI company finds it easy to raise money.

In the case of Chris, whom Blank profiled, the problem was missing shifts in an adjacent market. As the war in Ukraine reshaped demand and competition for autonomous drones and unmanned vehicles, Chris stayed focused on developing technology for the existing market and failed to track these changes closely enough.

Blank saw potential for Chris’s technology in resupply logistics and medical evacuation in conflict zones. That’s an investor’s judgment that new customers might be found there—it isn’t verification that the actual performance requirements, operating conditions, and procurement processes on the ground would match. It can be understood as a suggestion to keep the integration technology with existing aviation platforms while revisiting which markets to apply it to.

Time to focus on technology development is necessary. At the same time, founders need to periodically check whether customer requirements, competing products, and purchasing budgets have shifted. The longer development takes, the more important it is to keep checking whether the original market assumptions still hold.

We Need to Recalculate Development Costs and Validation Plans

The second point is how AI coding tools affect the software development process. Blank argues that as the cost and time of drafting code fall, teams need to reconsider both how they’re structured and how they validate products.

Vibe coding³, introduced by Karpathy in early 2025, drew attention as a method of describing what you want in natural language and letting AI generate the code. It’s expanded the range of options for writing code or building prototypes with AI. But a large volume of generated code doesn’t necessarily mean development costs have dropped proportionally, or that product quality has been validated.

Blank claims that some MVPs⁴ can now be built in days or hours instead of months. For simple CRUD apps⁵ focused on registering and retrieving data, this can indeed help produce prototypes quickly. But shipping to actual customers still requires security, data management, integration with existing systems, and bug fixes. Not every product’s development time shrinks at the same rate.

If multiple prototypes can be built easily, you might also prepare pricing options or screens for comparison faster. Blank cites this possibility as a reason parallel experiments can increase in Agile⁶ development. But Agile was never a methodology that banned parallel work in the first place—testing multiple hypotheses at once was possible even before AI.

The speed of building experiment variants and the speed of gathering sufficient evidence from customers are also different things. Making ten pricing options doesn’t mean customer reactions get validated instantly. You still need to decide which hypothesis to test and what outcome will count as proof.

Team size can also be recalculated based on these conditions. That means checking which tasks have actually gotten faster, how much review and maintenance cost, and then adjusting headcount and schedules accordingly. The mere existence of AI tools isn’t enough to conclude that an existing development team or its technology investments are all excessive.

You Need to Know Exactly What Outcome Customers Are Paying For

The third factor is how much of the work software does on the customer’s behalf. Blank believes that beyond simply displaying information, the ability to actually finish the task will become the differentiator. Existing software has automated plenty already, but as AI agents take on more of the work, customer expectations could shift too.

AI agents⁷ can call tools to handle inquiries, schedule appointments, triage sales leads, or prepare inventory orders. How much gets executed automatically depends on the system’s accuracy, its permissions, and the human approval steps built around it.

Blank expects pricing to shift accordingly — away from per-seat billing and toward pricing based on outcomes delivered. You’d charge per inquiry resolved, or per meeting booked, rather than per user. But not every service will convert to this model. You still need to agree with the customer on what counts as “resolved,” and how to settle up when something goes wrong.

He frames this as AI Agent/Customer Outcome Fit — checking whether AI is actually producing the outcome the customer wants. He’s also proposed the term MPO⁸ in place of MVP. I read this less as a declaration that product-market fit is dead, and more as a call to go beyond simply demonstrating a product — to actually measure how much of the customer’s job got done.

Hardware is following a similar path: AI can compare design alternatives or run tests using digital twins⁹. The examples Blank cites include analyzing camera footage and predicting equipment failure from sensor data. But this doesn’t eliminate the time required for actual manufacturing, physical testing, safety validation, and supply chains. You still have to verify what performance you actually get when hardware and AI are combined.

Sort what you’ve already invested in: what to keep, what to rethink

Once you have working code, a team, and a product roadmap in place, it’s hard to change direction. There are also promises made to investors and customers. Blank points out that spending a lot of money and time in the past doesn’t automatically justify the choices you make now.

He suggests separating what you’ve already built into two buckets: what keeps helping you, and what needs a fresh look.

Worth keeping: knowledge of your customers and industry, customer relationships, proprietary data, regulatory approvals, and experience integrating with physical systems. Chris’s integration with airline platforms is a good example. The question is whether capabilities built up over a long time can also serve a new line of business.

Worth reconsidering: team size relative to current workload, pricing relative to the value customers actually get, and priorities in product development. If your existing approach is genuinely delivering results, there’s reason to keep it. But if you’re still spending money just because of an old plan, it needs adjusting. Here, “asset” and “liability” aren’t accounting categories — they’re shorthand for whether something is still useful to the business.

Blank’s question is this: if you started the company today, with today’s market and today’s tools, what would you build? Revisiting hypotheses that already have investment behind them can feel uncomfortable. Still, it’s better to check before the next round of funding or the next big development spend.

Oswarld’s Lens

Reading Blank’s piece, I found myself thinking back to conversations I’ve had with Korean startups.

Working on go-to-market strategy, I’ve often seen founders build the technology first and only afterward figure out who to sell it to. In those cases, it’s worth going back to first principles: what job is the customer actually trying to get done? Whether you’re using AI or conventional software, the customer needs a real reason to spend the time and money to adopt it.

TheVC’s tally for Q1 2026 shows deal count down 17.4% year-over-year, even as total funding rose 55.4%. But the report notes that the upward trend in funding holds even if you strip out the ₩640 billion (~$463 million) mega-deal. Reading this as “nobody gets funded except a handful of AI companies” risks missing real differences across the market. What matters is checking which companies in your own sector actually raised capital, and what their investors saw in them.

What struck me most was framing accumulated industry knowledge and customer relationships as assets worth preserving. In fields like semiconductors, shipbuilding, or automotive—where understanding actual processes and operations matters—knowing the ground truth of the work may count for as much as bolting on AI. But that experience only becomes a real business strength if you can articulate exactly which customer problem it solves.

Closing

I think there are three things worth checking after reading this piece: whether the criteria investors use have changed, how development and operating costs shift once you adopt new tools, and what outcome customers are actually paying for.

I’d suggest revisiting these questions every quarter, comparing your original plan against what’s actually happened. You don’t need to throw everything out just because the plan is two years old. What’s needed is keeping the assumptions that still hold, and fixing the spending and timelines that were built on assumptions that no longer do.

What struck me was that Blank—who has spent years insisting founders get out of the building to test hypotheses with customers—is now saying the method of testing itself needs to be reconsidered in light of new tools. I’ve come to read this piece not as a call to discard startup methodology, but as an invitation to pause what you’re doing and take stock.

References & Further Reading

Primary sources

• Steve Blank, “Your Startup Is Probably Dead On Arrival,” 2026.03.17. Blank’s take on Chris’s case and revisiting business assumptions.
• OECD, “AI firms capture 61% of global venture capital in 2025,” 2026.02.17. An analysis explaining AI’s share of investment and the concentration of mega-deals.

Background

• TheVC, “2025 Korea Startup Investment Statistics,” 2026.01.01. Breaks down AI investment categories along with changes in deal value and count.
• TheVC, “2026 Q1 Korea Startup Investment Statistics,” 2026.04.01. Also covers the impact of mega-deals and the trend when they’re excluded.

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

Footnotes

1. Technical competitiveness: a business advantage that arises because competitors can’t easily replicate the same level of performance or operational experience. In investing circles this is sometimes called a “moat.” ↩

2. Lean Startup: a startup methodology systematized by Eric Ries, influenced by Steve Blank’s customer development methodology, among others. It’s an approach of testing hypotheses with customers and adjusting the product and business based on the results. ↩

3. Vibe Coding: a term introduced by Andrej Karpathy in 2025. It refers to making requests in natural language and using the code AI produces, including proceeding without closely understanding the generated code in detail. It isn’t synonymous with AI-assisted development as a whole. ↩

4. MVP (Minimum Viable Product): the minimum product or experiment needed to test and learn from assumptions about customers. It doesn’t simply mean a finished product with fewer features. ↩

5. CRUD: refers to Create, Read, Update, and Delete operations on data. ↩

6. Agile: an approach that develops in short cycles and adjusts based on feedback and change. ↩

7. AI Agent: an AI system that calls tools and performs tasks according to a given goal. The scope of what it can do depends on the permissions granted and the tools connected to it. ↩

8. MPO (Minimum Productive Outcome): a term Blank proposes in this piece, meaning to create and validate an outcome that’s useful to the customer. It isn’t a widely agreed-upon standard term. ↩

9. Digital Twin: a digital model that reflects the characteristics and state of a physical object or system as data. It’s used for analysis and simulation, among other things, but isn’t a perfect replica of the real thing. ↩