How AI is Changing the Way We Build Companies (ft. Guy Wuollet and Noah Citron)

Guy Wuollet argues that token spend can now be thought of like headcount because the ability to scale output has become more linear with money spent. Noah Citron adds that their own token spend has grown from a few thousand dollars a month to a few hundred thousand, with potential to reach a few million, and that this shift toward linear scalability has only become apparent in the past three months. Six months to a year ago, usage stayed within subscription limits and required no serious API budget management.

Engineering teams now face a P&L-like dynamic where token spend represents the cost side and attributed work output represents the return. An A/B testing pipeline run by an autonomous agent system could allow businesses to measure token spend against concrete metrics like retention rate and churn, and the ability for a product to improve its LTV ratio in proportion to token spend makes software engineering legible to capital in a new and more quantitative way. Traditional tech companies struggle to A/B test individual contributor ROI because many are network-effect businesses where isolating one person's impact is difficult.

The pod shop analogy has emerged as one framework for where this leads. A pod shop is a hedge fund structure where an overarching firm allocates capital to individual portfolio managers who each run their own teams and generate alpha, receiving roughly half the returns they produce, while the firm provides data infrastructure, talent pipeline, and a balance sheet. Small engineering teams may increasingly own a P&L for a specific product and receive a token budget to improve CAC to LTV ratios in a similar structure. Unlike headcount budgets, token budgets can be drastically increased or reduced overnight, making the model more Darwinian. The key limitation is that engineering organizations working on a single product with many components make it very hard to measure each pod's contribution to revenue, meaning most businesses will remain reliant on qualitative performance assessment.

The consulting analogy may be equally or more apt. Citron argues that tokens function as variable labor in the same way consultants do, and that AI models represent a more general intelligence than human consultants, expanding the addressable market for consulting-style work. One speaker noted that a core reason management consultants exist is to provide third-party justification for decisions the internal team already wants to make, functioning as a highly paid consensus mechanism, while a second reason is talent laundering, where high-status firms attract credentialed people who would not join lower-status businesses directly. In a world where models can be used directly, that recruiting arbitrage becomes less necessary, though one speaker pushed back, arguing that achieving organizational consensus is genuinely and incredibly valuable regardless of the underlying motive.

Wuollet draws a historical parallel to the origins of private equity, which he says emerged in the late 1970s and early 1980s partly because spreadsheets allowed one person with a computer to analyze a business that previously required a floor of people and two weeks. He argues LLMs will make a new set of businesses legible to acquisition or formation in an analogous way, and that the modern equivalent of an MBA hire for business transformation is someone who can articulate return on tokens and deploy them effectively. He also identifies an emerging space for smaller companies that need variable headcount and AI-driven token budgets but do not fit the venture capital model because the success case is not large enough for typical VC outcomes, with convertible debt rather than venture capital as one financing model being discussed.

Wuollet predicts that high-velocity AI agent economic activity will by default migrate to stablecoins and blockchain infrastructure, with users thinking of it as AI transformation rather than crypto. Stablecoins are increasingly being chosen as financial infrastructure not because they uniquely enable a business but because integrating a blockchain and stablecoins is now simply easier than establishing banking partnerships and navigating ACH rails. New entrants are expected to skip legacy systems entirely, mirroring how developing economies leapfrogged older infrastructure.

There is described as relative consensus that a form of AGI was reached at the end of 2025, yet macroeconomic data has not yet reflected meaningful AI-driven productivity gains. Wuollet predicts those gains will appear in GDP data around Q4 2027 or Q1 2028, though he acknowledges that prediction is vibes-based. One explanation attributed to Andrej Karpathy is that the roughly 2 percent year-over-year growth rate visible in US data going back to 1850 reflects the pace at which society can integrate new technology rather than the pace of invention, making the constraint human and organizational rather than technological. Large corporations are also seen as unlikely to make significant headcount reductions because professional managerial class incentives punish CEOs who cut staff and then have bad quarters.

Wuollet notes that AI models score above the 125 to 130 IQ threshold at which the correlation between IQ and income disappears, with one model reportedly measured at 136. He argues that raw intelligence above that level is not the key economic differentiator and that grit, tenacity, and agency are more strongly correlated with high earning potential. AI models are described as lacking agency because they lack desires and are fundamentally just math.