AI Has Redrawn the Startup Map: What the Numbers Mean for Builders

A Forbes analysis by product leader and venture investor Michael Wee argues that two forces have fundamentally reshaped what it takes to start a company: the cost of querying an AI model has fallen 280 times in two years, and software can now see, listen, reason, and write. AI-native teams are running at 20 times the median SaaS revenue per employee, seed rounds are getting fewer but larger, and the length of tasks an AI agent can handle unsupervised is doubling every four months. The logistics of starting a company have changed, even if the foundational questions have not.

What happened

Data point	Detail
Inference cost drop	$20 to $0.07 per million tokens (GPT-3.5 level), Nov 2022 to Oct 2024
Cost decline rate	280x in roughly two years; hardware costs falling ~30% annually
YC Winter 2025 AI codebases	~95% AI-generated for one quarter of the batch
YC cohort growth	~10% week-over-week, highest in YC history
Cursor ARR	$2B annualised revenue, February 2026
AI-native revenue per employee	20x the median SaaS benchmark
Q1 2026 venture concentration	Four AI rounds absorbed two-thirds of global venture dollars
Agent task-length doubling time	Every 7 months since 2019; recently accelerating to every 4 months

The Stanford AI Index, cited in the piece, tracked the collapse in inference pricing. That kind of deflation is faster than almost any commodity in the history of technology, according to the analysis. Hardware efficiency is improving around 40% per year on top of the price declines, meaning the trend has multiple compounding drivers, not just competition between model providers.

On the capability side, the metric Wee focuses on is “task length”: how many consecutive steps an AI agent can complete before it needs a human to intervene. At a 50% success rate, that frontier has been doubling every seven months since 2019. The more recent measurements put the doubling rate closer to four months. If that holds, agents capable of running unsupervised for an entire month on complex tasks could arrive within a few years.

Why it matters

The company architecture is changing

When software gets cheaper to produce and agents handle more of the execution, teams get smaller and revenue per head gets much larger. Cursor, Lovable, and Midjourney are cited as examples of companies scaling to significant revenue on tiny teams. Cursor’s $2 billion annualised revenue figure as of February 2026 is the clearest data point. The analysis puts AI-native companies at 20 times the median SaaS revenue per employee, which is a structural shift, not an efficiency tweak.

Venture capital is concentrating fast

The funding picture described is K-shaped. A small number of very large AI rounds are absorbing most of the capital. Four deals took two-thirds of all global venture dollars in Q1 2026. At the same time, seed deals are getting fewer and larger. Early-stage founders are competing for less capital against more sophisticated teams that are already shipping AI products. That raises the bar for what a seed-stage company needs to demonstrate.

Where value is actually accruing

The analysis identifies three layers where durable value is building up: foundation models themselves, applications that own specific workflows and accumulated user taste, and AI-enabled services businesses that roll up previously human-labor-intensive work. For most founders, layer two and three are the realistic playing field. Owning a workflow means the product gets stickier as models improve rather than becoming redundant.

This connects to a broader pattern we’re tracking in the agent breakouts of mid-2026: the businesses with lasting positions are those that built distribution and workflow lock-in, not just model wrappers.

Our take

The framing here is honest and useful. The 280x inference cost drop is not a projection. It is a measured, documented change that has already happened. The YC cohort data is similarly concrete. What the analysis gets right is the distinction between features that compensate for today’s model weaknesses (liabilities with a short half-life) and workflows, datasets, and distribution advantages that compound as models improve (assets).

The harder practical question is how a small business or agency applies this. Our read: the window to build a workflow-owning product is narrowing, not widening. Every month that passes, better-funded teams ship more of those workflows. The advantage for smaller operators right now is domain knowledge in niches that large teams have not prioritized. A paralegal tool for one specific jurisdiction. An estimating tool for one trade. A customer-intake automation for one service vertical. Narrow beats broad right now.

If you are trying to figure out where AI integration fits into your own business before competitors move, the cost floor has dropped far enough that experimentation is genuinely cheap. The bottleneck is judgment about which workflows are worth automating, not the cost of the models themselves.

What to do about it

1. Audit your current workflows and identify which ones are high-repetition and low-exception. These are the first candidates for AI automation at current model capability.
2. Separate features you are building to work around today’s model limits from the core workflow you want to own long-term. Deprioritize the former when possible.
3. Track task-length progress in the models you use. When an agent can handle a task end-to-end that previously needed human checkpoints, that is a signal to revisit your product architecture.
4. If you are an early-stage founder raising, prepare for a harder seed environment. Two-thirds of Q1 2026 venture capital went to four deals. The bar for demonstrating traction before raising has risen.
5. Focus on distribution and accumulated data over raw AI capability. According to this analysis, those are the assets that compound as models improve, rather than being erased by the next model release.

The cost of intelligence is still falling. Build the part that does not deflate with it.

Source: Bing News · Midjourney