Paul Graham and the Changing Economics of Building a Startup

In a controlled experiment conducted by researchers at Microsoft, GitHub, and the Massachusetts Institute of Technology, software developers assigned to build a standalone HTTP server in JavaScript finished the task 55.8 percent faster when working with an AI coding assistant than peers working without one. Subsequent randomized field experiments across 4,867 engineers at Microsoft, Accenture, and a Fortune 100 electronics manufacturer recorded a 26.08 percent increase in weekly completed pull requests, with the largest relative gains concentrated among junior and shorter-tenure programmers.

Yet when researchers at the evaluation institute METR ran a randomized controlled trial in 2025 with 16 experienced open-source maintainers completing 246 real issues inside their own mature repositories—codebases averaging more than one million lines of code where the participants averaged five years of prior experience—the result inverted. Developers permitted to use frontier AI coding environments took 19 percent longer to complete assigned tasks than when working unaided, even though those same developers had forecast a 24 percent speedup beforehand and estimated afterward that the tools had sped them up by 20 percent.

Those divergent findings capture the central economic paradox confronting technology founders and investors. A two-person team can now draft user interfaces, wire up application logic, and ship a working prototype in days rather than quarters. Building an enduring software business, however, has not become proportionally easier. When the initial artifact is cheap to produce, the bottleneck in company building migrates.

Nowhere is that shift more consequential than in the framework guiding early-stage venture investing. Through his essays and his role in co-founding Y Combinator in 2005, Paul Graham helped codify the modern seed-stage playbook: back small teams of technical founders, spend as little money as possible, launch quickly, iterate directly with users, and treat rapid growth as the defining test of a startup.

As Lazy Tensor launches Capital and its inaugural series, The Investors, Graham’s published body of work—from his foundational 2005 essays to his September 2026 reflection, "Making Startups Powerful"—offers a revealing lens for examining the changing economics of startup formation. If artificial intelligence compresses the marginal cost of writing code, which parts of the classic Silicon Valley doctrine become more potent, and which assumptions face a structural test?

The investment philosophy that helped shape Silicon Valley

To understand why Graham’s arguments reshaped early-stage investing, consider the capital environment he was reacting against. When Graham and Robert Morris founded Viaweb in 1995 to let merchants build online stores through a web browser (before Yahoo acquired the company in 1998), software startups were expected to raise multi-million-dollar Series A rounds, lease commercial hardware, and hire professional executives before proving customer demand.

By March 2005, when Graham co-founded Y Combinator in Cambridge, Massachusetts, with Jessica Livingston, Robert Morris, and Trevor Blackwell using $200,000 in initial capital, the cost of computing had already begun an earlier structural decline. Open-source software, commodity Linux servers, and browser delivery were replacing proprietary hardware and enterprise licensing. In his March 2005 essay "How to Start a Startup", adapted from a talk at the Harvard Computer Society, Graham distilled the new arithmetic into three requirements: "to start with good people, to make something customers actually want, and to spend as little money as possible."

That third clause was an economic argument about control and survival. In his July 2009 essay "Ramen Profitable", Graham observed that a startup earning just enough to cover its founders’ modest living expenses no longer depended on investors’ permission to stay alive while searching for a scalable business model. Capital efficiency bought time, allowing small technical teams to outlast heavier competitors.

Alongside frugality sat two other pillars of Graham’s philosophy. First was a strict analytical distinction between a venture-scale company and a conventional small business. In "Startup = Growth", published in September 2012, he argued that a startup is defined neither by being newly founded nor by working on technology, but by a single property: "A startup is a company designed to grow fast."

Second was an insistence that early traction rarely happens automatically. In his July 2013 essay "Do Things that Don't Scale", Graham challenged the engineering belief that building a superior mousetrap is sufficient to attract users. Startups take off, he wrote, because founders manually crank the engine: recruiting early customers one by one, configuring software on their behalf (a practice he illustrated with Stripe’s "Collison installation"), and delivering an attentively managed experience before automated distribution takes over.

Graham’s individual essays are distinct from the broader institutional evolution of Y Combinator, which expanded after he stepped down as president in 2014 into a global organization funding thousands of companies across software, hardware, and biotechnology. Yet his recent writing has returned to the structural mechanics of founder control and market leverage. In his September 2024 essay "Founder Mode", he argued that scaling companies often deteriorate when founders hand off operational command to conventional management layers. And in his September 2026 essay "Making Startups Powerful", Graham described a change in his own office-hours advice: rather than asking how a company could make more money—which tends to elicit incremental tactics—he began asking what would make the company more powerful.

What happens when the cost of building falls?

The first wave of startup cost reduction in the 2000s lowered the price of infrastructure as cloud hosting replaced server rooms and open-source frameworks replaced commercial middleware. Writing application software, however, remained a labor-intensive craft requiring scarce engineering hours.

Generative AI and agentic coding tools alter that labor equation, though the empirical evidence is far more nuanced than vendor marketing suggests. Closer inspection of the three major productivity studies shows that the returns to AI assistance depend heavily on task complexity and codebase maturity. In the 2023 Peng et al. experiment, 95 freelance programmers were asked to implement a greenfield HTTP server in Node.js against a pre-written test suite—an isolated, well-specified problem with no legacy architecture or long-term maintenance horizon (and while completion time fell 55.8 percent, overall task success rates did not differ significantly). In the 2024 Cui et al. enterprise field experiments, the 26.08 percent rise in pull requests measured output volume rather than system complexity, with gains concentrated among less experienced staff. By contrast, METR’s 2025 trial tested veteran maintainers resolving live issues in million-line repositories, where prompting an AI, reviewing subtle errors, and enforcing implicit architectural invariants outweighed raw typing speed.

Longitudinal repository data points to the same tension between rapid creation and long-term maintenance. An analysis by GitClear examining more than 211 million changed lines of code between 2020 and 2024 found that as AI coding tools spread, the share of refactored code fell sharply while duplicated code blocks and short-term churn—lines modified or reverted within two weeks—rose significantly.

At the macro level, lower friction at the starting line is already visible in company-formation and payment data. In its 2024 Annual Letter and 2025 Annual Letter, payment processor Stripe reported that code pushes to GitHub surged 41 percent between the third quarters of 2024 and 2025, startup formations on Stripe Atlas rose 41 percent in 2025, and the share of Atlas startups billing their first customer within 30 days climbed to 20 percent, up from 8 percent in 2020. Stripe also noted that the number of startups on its platform reaching $10 million in annualized recurring revenue (ARR) within three months of launch doubled between 2024 and 2025, citing company-reported and platform-tracked examples such as AI application builders Lovable ($17 million ARR in three months) and Bolt ($20 million ARR in two months).

Individual breakout companies illustrate both the early leverage of small technical teams and the subsequent capital and headcount required to defend a market. Anysphere, the maker of the AI code editor Cursor, reported in January 2025 that it had surpassed $100 million in recurring revenue while describing itself as "a small, talent-dense group," before announcing in June 2025 that it had exceeded $500 million in ARR and in November 2025 that it had crossed $1 billion in annualized revenue while expanding its team past 300 engineers, researchers, and operators alongside $2.3 billion in Series D financing. Similarly, AI audio developer ElevenLabs reported in January 2025 that employees at more than 60 percent of Fortune 500 companies had adopted its tools within two years of launch as it raised a $180 million Series C round at a $3.3 billion valuation. Because revenue and headcount figures for private startups are almost always company-reported or drawn from payment-processor cohorts rather than audited financial statements—and because early ARR spikes in AI tools can reflect experimental monthly subscriptions with high churn—investors caution against extrapolating industry-wide durability from outlier trajectories.

The new scarcity

When anyone can generate a passable workflow application over a weekend, a working prototype ceases to be a reliable signal of technical defensibility. Across software markets, Lazy Tensor’s analysis suggests that lasting economic value is migrating toward four assets that cannot be synthesized in a prompt window.

The first is distribution and ownership of the customer relationship. In "Making Startups Powerful", Graham argues that a company supplying a technical component to another business occupies a structurally weaker position than the company that owns the end-customer relationship, makes money flow through its platform, or goes "full stack" by delivering the complete service itself. In a notable footnote to that essay, Graham applies the same logic to artificial intelligence: startups that route tokens and user interactions own the customer relationship in theory, turning foundation-model labs into component suppliers—yet face the constant strategic question of how easily those model providers might engulf them or their customers. As Lazy Tensor examined in From Dark Factories to Zero-Page Websites, as digital interfaces reorganize around automated agents, owning the trusted point of customer intent and transaction execution matters far more than owning a thin layer of presentation code.

The second scarcity is proprietary context and compounding network effects. Software that merely wraps a general-purpose language model around public data invites immediate price competition. Graham notes in "Making Startups Powerful" that startups gain structural power when they create network effects—for instance, when users opt in to training specialized systems on domain interactions, or when an agent utility evolves into a multi-sided marketplace. Systems anchored in verified institutional records, regulated workflows, or closed-loop operational data accumulate context that fast followers cannot trivially clone.

The third scarcity is deterministic operational infrastructure. An AI-assisted prototype may demonstrate a feature in minutes, but commercial software still requires rock-solid foundations: authentication, role-based authorization, durable storage, usage-based billing, payments, security auditing, regulatory compliance, and real-time monitoring. As AI multiplies the number of newly launched software services, demand rises for dependable underlying infrastructure; enterprise buyers will not tolerate leaked credentials, billing errors, or unmonitored agent actions.

Finally, execution speed changes character. When code generation is cheap, raw feature velocity matters less than diagnostic judgment: knowing which customer problem is worth solving, watching for what Graham calls "tails that wag the dog" when users adapt a peripheral feature in unexpected ways, and designing a system whose unit economics remain positive when inference bills arrive.

Does venture capital become less necessary?

This shifting cost structure has prompted a recurring debate in Silicon Valley: if two founders can build and launch a software product on modest personal savings or a small angel check, does traditional venture capital lose its leverage?

At the formation stage, founders gain genuine optionality. Lower prototyping costs make Graham’s concept of ramen profitability achievable faster and with fewer people, letting a focused vertical software business reach cash-flow sustainability without surrendering 20 percent of its equity in an institutional seed round.

Yet venture capital data tells a very different story for companies pursuing category leadership. According to the PitchBook-NVCA Venture Monitor, artificial intelligence and machine learning companies captured 46.4 percent of all US venture capital deal value in 2024 and exceeded half of total US deal value in 2025. Rather than shrinking, venture investment has concentrated aggressively into fewer, larger financings.

Three economic forces explain why lower coding costs do not automatically reduce capital requirements across a company’s lifecycle:

1. Compressed gross margins and variable compute costs. Traditional software-as-a-service businesses historically enjoyed gross margins of 75 to 85 percent because serving an additional software seat carried near-zero marginal cost. By contrast, investment analyses by Andreessen Horowitz document that generative AI application startups frequently operate with gross margins between 50 and 60 percent—and significantly lower in compute-heavy workflows—because every customer query triggers variable model-inference, token, or GPU hosting expenses.
2. Go-to-market and compliance expenses dominate the income statement. Initial software development is only one line item. As companies move upmarket into enterprise accounts, sales, marketing, security certifications (such as SOC 2 and ISO standards), legal indemnification, and customer support routinely consume a far larger share of revenue than research and development.
3. Crowded categories inflate customer acquisition costs. When lower entry barriers allow dozens of startups to ship competing tools simultaneously, buyer attention becomes the bottleneck. Rivals must spend heavily on distribution partnerships, enterprise sales teams, and brand credibility to win and retain accounts.

Starting a software company requires less capital than ever. Winning a contested market still requires substantial capital—deployed toward compute capacity, data rights, compliance, and enterprise distribution rather than large rosters of junior application programmers.

What Graham’s philosophy gets right—and what AI may challenge

Measured against these economic realities, much of Paul Graham’s published philosophy looks remarkably durable, often for reasons that become clearer once the mystique of code generation is stripped away.

His insistence in "Do Things that Don't Scale" that founders manually recruit and serve early users becomes even more decisive in an AI-saturated market. When anyone can automate a product launch or generate synthetic outbound emails, personal onboarding, deep workflow customization, and direct founder accountability become primary signals of trustworthiness. The laborious crank required to start the engine cannot be outsourced to an autocomplete model.

Similarly, his core test from "How to Start a Startup"—making something customers actually want while spending as little money as possible—aligns with his September 2026 conclusion in "Making Startups Powerful": "Newly founded startups are too weak to force anything on anyone. The only way they can become powerful is to make customers' lives better." And his argument in "Founder Mode" fits the organizational geometry of AI-leveraged companies: when small teams can manage wider operational surface areas, founders can stay directly engaged with architecture and customer detail far longer before adding administrative layers.

Where Lazy Tensor’s analysis departs from traditional 2000s seed orthodoxy is on two assumptions that artificial intelligence complicates. First, early shipping velocity has become a noisier proxy for technical capability. For two decades, accelerator partners could treat a functional prototype built in a few weeks as strong evidence of engineering talent; today, when AI coding assistants let average builders assemble convincing demos in days, investors must evaluate deeper indicators: architectural coherence, security posture, gross-margin structure, and domain authority. Second, iterating in public on thin software carries higher competitive risk. When fast followers and incumbent platforms can replicate an interface and prompt pipeline almost immediately, launching an undifferentiated minimum viable product without a distribution, workflow, or data wedge can simply educate the market for better-capitalized rivals.

Building is not the same as winning

Every major drop in the cost of technology creation triggers predictions that traditional company-building disciplines are obsolete. Open-source software, cloud computing, and mobile app stores each democratized production while raising the bar for commercial defensibility.

Artificial intelligence extends that trajectory to programming itself. By compressing the time and capital required to turn an idea into executable code, it removes a historical tax on entrepreneurship. More ideas will be tested, more niche workflows will be automated, and more small teams will reach initial revenue without institutional permission.

Yet the central lesson of Paul Graham’s essays, read as economic analysis rather than startup folklore, is that software was never valuable merely because it was difficult to type. Companies endure because their founders identify a consequential customer need, earn the trust required to serve it reliably, and construct a position of structural power that competitors cannot trivially copy. As the cost of building falls, the value of judgment, distribution, operational rigor, and customer insight does not shrink. It becomes the whole game.