The New Logistics Of Startup Ideas In The AI Age

The Two Drivers

Two forces explain nearly everything happening to startups: the cost of developing software has collapsed, and software has acquired completely new capabilities.

Let’s start with cost. The Stanford AI Index found that the price of querying a model at GPT-3.5-level performance fell from $20 per million tokens in November 2022 to $0.07 by October 2024, a 280x decline in roughly two years, with hardware costs falling about 30% annually and energy efficiency improving about 40% per year. Intelligence as an input cost is deflating faster than almost any commodity in the history of technology.

The labor side of the equation has moved just as fast. In early 2025, Y Combinator reported that for a quarter of its Winter 2025 batch, roughly 95% of the codebase was AI-generated, notable for highly technical founders who would have written the software themselves a year earlier. That batch also had the highest growth rate in YC’s history, with the entire cohort compounding around 10% week-over-week.

The second driver is new software capabilities. Software can now see, listen, speak, reason, and write. Entire product categories that were barely imaginable in 2020 are now weekend projects such as a tutor that converses in Mandarin, a paralegal that drafts and redlines, and a designer that produces forty variations before lunch. The capability curve that matters most is task length, how long an agent can work before it needs human review, and it is the field’s closest thing to a Moore’s Law: the length of software tasks frontier models can complete (at a 50% success rate) doubled roughly every seven months since 2019. This pace is accelerating, with more recent measurements putting the doubling closer to every four months. To put that into context, in a few years we may see AI agents that can perform intensive tasks without oversight for an entire month uninterrupted.

Cheaper inputs accelerate speed and margins while new capabilities expand markets. When both happen simultaneously, the frontier of viable startups expands explosively.

Build for the Future, Optimize the Present

The state of the art, as of mid-2026, showcases agentic AI that works but needs manual guidance. Complex domains like distributed systems, state machines, and features requiring exacting precision still cause models to struggle without a human providing product direction, engineering architecture, and real thought. Over time, the models will gracefully absorb more complex requirements. The frontier of what needs a human recedes every quarter, and the companies that know exactly where the line sits have a real edge.

This creates a strategic discipline I’d summarize as: build for the future while optimizing the present.

Optimizing the present means shipping on today’s models: designing around their failure modes and adding guardrails and human checkpoints. Building for the future means architecting your product so that it inherits the next model generation rather than being rendered obsolete by it. Every feature you build to compensate for a model’s current weakness is a liability with a half-life. Every workflow, dataset, and distribution advantage you build is an asset that compounds as models improve. The best AI-era founders hold both timeframes in their head at once: they ask not only “does this work today?” but also “does this get better or worse when the models get smarter?”

The New Architecture of Companies

If software is cheaper to produce and agents carry more of the load, the architecture of companies dramatically shifts. The evidence is already in the revenue per employee data, which is eclipsing every historical benchmark.

Anysphere’s Cursor crossed $2 billion in annualized revenue in February 2026, roughly three years from founding, and by June, reporting put the figure near $4 billion. Lovable reached $100 million in annualized revenue within eight months of launch and passed $500 million by June 2026; at the $400 million mark, the company had 146 employees, roughly $2.7 million of revenue per head. Midjourney, self-funded from day one, is estimated to generate around half a billion dollars annually with a team in the low hundreds. For context, the median private SaaS company generates about $141,000 of revenue per employee, while the leading AI-native companies are running at twenty times that or more. Gartner now predicts that by 2030, AI-native development platforms will lead 80% of organizations to evolve large software engineering teams into smaller, AI-augmented ones, with the share of organizations running smaller engineering teams at scale reaching 60% as soon as 2029, up from 15% in 2026.

Smaller teams of high-agency engineers and builders are managing fleets of agents rather than executing every task by hand. The org chart flattens because the coordination overhead that justified middle layers, translating between product, builders, and reviewers, is increasingly absorbed by AI tools. Founders must consider how they want to architect their org chart and absorb the downstream effects on company culture, including a renewed emphasis on cohesion, speed, coordination, meticulousness, autonomy, and clear communication to execute at full capacity.

The Old Maze, and the New One

Today, prototyping and de-risking an idea are cheaper and faster than ever, which inverts the logistics maze for testing startup ideas. Consider how a software startup used to get built. A founding team would form and set out to validate the idea. Fundraising followed to raise just enough fuel to reach the next milestone. Seed money bought you a prototype and early customers; those proof points unlocked the Series A, which bought product-market fit, which unlocked the Series B, which bought scale. Each milestone demanded evidence you’d made it from the last one, and running out of fuel between milestones was the most common cause of startup death.

The old model assumed what is no longer true: building the initial product was the expensive part. In the age of AI, validation now largely precedes capital rather than depending on it. A motivated founder can stand up a working product, put it in front of paying users, and arrive at the first milestone with revenue instead of a pitch deck.

The capital markets have reorganized around this reality, and the result is a K-shaped market. At the top, record sums concentrate into foundation models and infrastructure: in the first quarter of 2026, the four largest venture rounds ever recorded (investments in OpenAI, Anthropic, xAI, and Waymo) closed within a single quarter and absorbed roughly two-thirds of global venture dollars, with AI overall taking about 80%. At the bottom, seed has become richer per deal and rarer: deal counts are down roughly a quarter from a year earlier, and more than half of seed dollars now flow into rounds of $10 million or more. The checks have become fewer but larger for founders who show up with traction. The “raise on a napkin” era is over, and, for capable builders, largely unnecessary. Motivated, talented individuals (which is to say, founders) can go much further with far fewer resources than at any point in the history of software.

The new maze also poses deeper questions beyond how to fund the journey: what to build, and where value will accrue in the AI age. Three theses dominate the debate. First is the model layer, where the two leading labs achieved a combined annualized revenue run rate north of $70 billion by mid-2026 and each release absorbs the thin applications around it. Next is the application layer, where Cursor and Lovable became some of the fastest-growing software companies ever by owning what the models don’t: workflows, distribution, proprietary data, and taste. Finally, AI-enabled services rollups are emerging that, rather than selling software to slow-moving industries, simply acquire, consolidate, and add AI, building in the $16 trillion services economy. If software ate the world by selling to it, AI may finish the job by buying it.

The Buildout Continues

The pace is not slowing down anytime soon. The four largest US hyperscalers alone issued guidance of roughly $725 billion of combined capital expenditure in 2026, up about 77% from 2025. Goldman Sachs projects on the order of $5 trillion of capital expenditures from these companies through 2030 across compute, data centers, and power. Add in Oracle and sovereign-scale projects like Stargate, and the buildout starts to resemble a national infrastructure program more than a corporate capex cycle.

For founders, the implication is simple: the platform beneath you compounds on someone else’s balance sheet. Researchers will have access to ever more supercharged models, and consumers of models will keep receiving more capable, cheaper intelligence. Two curves define the era: the cost of intelligence falling, and the capability of intelligence rising. Every startup in the world is long both.

Zoom all the way out and this moment snaps into a familiar frame. The economist Carlota Perez, in Technological Revolutions and Financial Capital, showed that every major technological revolution over 250 years from canals and railways to steel and information technology follows the same arc: an installation phase, in which speculative financial capital floods in and builds the infrastructure amid a frenzy; a turning point, usually a financial correction; and then a deployment phase in which the technology diffuses through the whole economy and the enduring institutions get built.

Following Perez’s map, this AI wave is now deep in the installation phase. The capex figures, concentration of capital, and exuberant valuations are the signatures of every frenzy phase in history, and by analogy history tells us that a correction of some kind is more likely than not. But Perez’s deeper lesson is one founders should internalize: that the frenzy is not wasted. The railway manias left behind functioning rail networks, and the dot-com bubble left behind server and network infrastructure that made the following two decades of internet companies cheap to build. Today’s buildout is leaving behind compute, power, and models. The companies that came to define the internet era did their scaling in the deployment phase, on top of infrastructure the installation phase paid for. For founders, the tactical translation is to keep your capitalization honest, build products that benefit from compute getting cheaper, and remember that the turning point, if it comes, is often when the greatest companies of the era get started, not when they end.

Navigating at Speed

If the pace is not slowing down, the question becomes how emerging founders can keep up. The answer begins with a paradox: the faster the terrain shifts, the more the foundations matter. When tools, costs, and capabilities are repriced every quarter, the durable factors such as why you are building, your understanding of the problem domain, and your ability to drive execution and product become the instruments you navigate by. You can follow the shoreline only while landmarks hold their positions; when the map is being redrawn, you navigate by the stars. The foundations aren’t the stepping stones to race day, rather they are what make the speed controllable.

Founders must simultaneously stay close to the customer, their purpose, and the frontier. They should imagine the confluence of the capability curves and a problem domain they want to reimagine, holding both in mind and asking what can be built when they meet. Founders must keep exploring the frontier, treating comfort with rapid change as a compounding skill: the ones with the intuition for model strengths and limitations use them daily, building, breaking, and playing. Finally, founders may dare to stop speculating about the future and start inventing it. The best founders don’t keep up with change; they are the reason the rest of us have to.

The Logistics Maze, Revisited

The foundational questions from 2020 still come first. Why do you want to build? What are your strengths? How do you define success? Only the founder can answer these questions.

But the logistics around them have inverted. In 2020, the maze was constrained by capital, headcount, and the cost of discovering whether an idea was any good. In 2026, ideas are easy to validate, software is cheap to build, and a single high-agency founder with taste can move rapidly with capabilities that once required a funded team. The scarce resources now are judgment, distribution, and the discipline to build for where the models will be, not just where they are.

In my first essay, I wrote that if you paddle a boat for ten years, a one-degree difference in launch trajectory means thousands of miles of separation in the fullness of time. The analogy holds, except that your boat has grown an engine, and the engine upgrades compound every few months. Aim carefully. Then go much faster than you think you can.