I believe the corporate world is witnessing a structural break that mirrors the internet revolution in its magnitude. We seem to have entered the period of the “Second Great Compression,” a term that describes the radical collapse in the time, capital, and headcount required to launch and scale a new venture. At the heart of this change is a new organisational archetype: the “20x Company.” Coined by the founders of GigaML, the term describes tiny, expert-dense teams that use autonomous AI agents to outperform incumbents twenty times their size. These companies are rebuilding the nature of the firm itself, instead of simply using AI to do the same things faster.
Defining the Second Great Compression
Traditionally, a SaaS startup would have a timeline of 12 to 18 months and a capital investment of over $1 million just to reach product-market fit. But with agentic AI, these baselines have evaporated. I’m seeing companies moving beyond “AI-interested” (those bolting chatbots onto legacy structures) and into AI-native organisations where intelligence is the foundational architectural principle.
The 20x company represents the pinnacle of this change. These firms, often comprising fewer than twenty people, are closing Fortune 500 accounts and posting nine-figure revenues in record time. I believe the window for competitive advantage is closing for those taking the safer path of incremental adoption. Organisations that merely renovate legacy structures will be outmanoeuvred by those who rebuild with AI as the central nervous system.
The Economics of Compression: Time, Capital, and Labour
The engine of this compression is the transition through three programming paradigms, as described by Andrej Karpathy: Software 1.0 (human code), Software 2.0 (neural network weights), and the current phase of Software 3.0, where large language models are programmable via English prompts. This birthed ‘vibe coding,’ a term viralised by Karpathy to describe using natural language to prototype functional software in hours or days rather than weeks.
But Karpathy has since moved the goalposts. He now argues that vibe coding, while excellent for rapid prototyping, breaks down once real software development is on the table. He has introduced a successor term for what comes next: agentic engineering.
Vibe coding means accepting whatever the model produces, and agentic engineering brings human-in-the-loop oversight, verifiable evaluation loops, and a strict architectural quality bar back into the process. It is less a new tool and more a change in posture, one that forces people to stop acting like casual hobbyists chasing a working demo, and start thinking like professional software architects and systems designers responsible for what they ship.
When done right, this change can bring massive capital savings. Research shows that AI-native startups reach Series A milestones with 80% less capital and in timelines that are 20% to 40% shorter than their non-AI peers. This is possible only when a company practises agentic engineering and not just vibe coding.
You need structured oversight so your speed survives the real world and doesn’t collapse under technical debt once paying customers arrive. This efficiency also requires a top-heavy talent strategy. AI-native firms employ 15% fewer entry-level workers and 15% fewer managers. Instead, they hire senior, technical talent who can act as “intelligence managers,” orchestrating systems of agents rather than executing manual tasks.
The Blueprint of a 20x Organisation
Flatter and expert-dense
The internal architecture of a 20x company is different from the 20th-century hierarchy. Because an intelligence layer can route information more efficiently than human middleware, these firms are half a seniority level flatter than traditional organisations. According to researchers Hyunjin Kim (INSEAD) and Rembrand Koning (Harvard Business School), this organisational pattern is driven by two channels:
- The Process Channel: Using AI internally (e.g., agentic coding, automated sales) to augment or automate work.
- The Product Channel: Embedding AI directly into the offering so the software itself performs the knowledge work that previously required large human teams.
Startups like Gamma and Legion Health exemplify the product channel. Gamma allows a team of 30 to serve millions of users by turning presentation deck requests into product interactions rather than organising internal human workflows for each deck.
To manage this kind of work, leaders like Jack Dorsey at Block suggest three new employee archetypes: the Individual Contributor (builder-operator), the Directly Responsible Individual (outcome-focused), and the AI Founder (leading by example through tinkering). I think this framing is useful because it names something most leaders might be already doing without a title for it.
The 20x engine: relentless internal automation
A 20x company operates as an ‘intelligent closed loop,’ a self-regulating system that continuously monitors its output and adjusts its processes to improve toward a goal. For this to work, the entire organisation must be “queryable,” meaning every meeting, decision, and process produces a digital artefact that the central intelligence layer can learn from. Legion Health, which is building an AI native psychiatry network, built an internal interface that became an AI-integrated source of truth that gave employees instant context across the entire system.
We are also seeing the emergence of software factories, where humans provide the specifications and tests, and AI agents generate the implementation and iterate until the tests pass. Some AI-native teams have reached a point where their repositories contain no handwritten code, only specs and test harnesses. This has enabled the age of the 1000x engineer, where one individual, surrounded by agentic systems, can build what previously required entire departments.
‘Token maxing’ over headcount
The most critical strategic change in a 20x company is the move from headcount-centric growth to token maxing. This refers to the radical decision to run an uncomfortably high API bill, because it is much cheaper and more scalable than an inflated human headcount. These firms prioritise API and token usage as their primary driver of growth, allowing them to keep their human teams small and culture-dense.
Hiring for this environment is less about checking for skills and more about checking for willingness. AI-savvy companies build one-pizza teams – small, cross-functional groups where humans and AI agents collaborate in real time. This shifts hiring priorities towards candidates who can communicate clearly, delegate work effectively, and treat AI agents as genuine teammates with defined responsibilities. As Andrej Karpathy has observed, English is becoming the ultimate programming language, making communication just as important as coding.
The Competitive Threat to Incumbents
The rise of the 20x company is a direct threat to the existence of incumbents, not just a trend confined to startups. The Intelligence Moat is real: as more small firms use agentic systems to solve ‘unsolvable’ manual tasks, they build proprietary workflow knowledge that becomes a compounding advantage.
I believe the future requires an always-beta mindset, as championed by Deloitte, where organisations embrace perpetual evolution, and software behaves like a living system. In the next few years, I’m confident that market leadership will be defined by who has the most scalable, queryable intelligence.
References:
- Kim, Hyunjin and Koning, Rembrand, AI-Native Firms * (June 09, 2026). Available at SSRN: https://ssrn.com/abstract=
- HBR – How Agentic AI Supercharges Startups and Threatens Incumbents, July/Aug 2026
- A New Way to Build a Startup – Y Combinator
- Andrej Karpathy: Software Is Changing (Again) – Y Combinator
- McKinsey – The seven operating truths of AI-native companies, June 2026
- Forbes – Exactly What Is An AI-First Company? June 2025
- How To Build A Company With AI From the Ground Up – Y Combinator
