DSTL - Audio in. Alpha out.

dstl

TERMINAL

LIBRARY

Block's 40% RIF: Restructuring a Public Company Around AI-Native Development

Join

Block's 40% RIF: Restructuring a Public Company Around AI-Native Development

Apr 2, 2026

a16z

27:03

16K Views

THESIS

Block's Owen Jennings argues the correlation between headcount and output permanently broke in December 2024, forcing a 40% workforce reduction as AI agents replace linear engineering teams.

ASSET CLASS

SECULAR

CONVICTION

HIGH

TIME HORIZON

6 to 18 months

PREMISE

The decades-long correlation between employee headcount and company output has structurally broken

For decades, companies scaled output by scaling headcount — more engineers, more designers, more PMs meant more product. Block observed that beginning in late November to early December 2024, a binary shift occurred in foundational model capability. Specifically, models like Opus 4.6 and Codex 5.3 crossed a threshold where they moved from being effective at writing code for greenfield projects to being extraordinarily capable of working with existing, complex codebases. This meant that one or two engineers using AI tools could achieve 10x, 20x, or even 100x the productivity of prior configurations. The traditional feature team of 14 people could be replaced by a squad of four people plus unlimited token access. Block spent Q1 2025 pressure-testing this observation across its executive team and concluded the shift was permanent and accelerating.

MECHANISM

Forced restructuring around AI-native squads creates a compounding productivity loop that competitors cannot replicate incrementally

The mechanism operates on multiple reinforcing layers. First, Block executed a greater-than-40% reduction in force concentrated heavily on the development side — not operations, compliance, or sales — signaling this was a technology-driven restructuring, not a cost-cutting exercise. Second, the company collapsed organizational layers by 50-60%, enabling faster information flow and eliminating coordination overhead. Third, Block built proprietary agentic infrastructure — Goose (an open-source, model-agnostic agent harness), Builderbot (an autonomous PR-merging and feature-building tool), and G2 (an internal agentic operating system for automating deterministic workflows). This infrastructure means remaining employees operate fundamentally differently: instead of linear workflows where one engineer writes one PR, a single person now manages 8-14 concurrent AI agents building PRs simultaneously, context-switching between them as an editor rather than an author. Designers and PMs are shipping PRs directly. The cycle time from idea to product in the hands of a million customers has compressed from months to weeks. Critically, Jennings argues that companies attempting incremental 15% reductions will suffer culturally from perpetual riff anxiety without achieving the forcing function that compels full workflow transformation. Block's decisive cut forced immediate adoption at scale.

OUTCOME

Block repositions as an intelligent system where defensibility lies in proprietary signal understanding and speed of agentic iteration, not headcount

The expected outcome is that Block transforms from a traditional fintech company into what Jennings describes as an 'intelligent system' — a company defined not by its workforce size but by its deep, proprietary understanding of how sellers and buyers participate in the economy, coupled with the ability to iterate on that understanding hundreds or thousands of times per day through agentic loops. On the product side, this manifests as generative UI (Cash App and Square interfaces that are dynamically created per user by AI rather than statically coded), proactive intelligence features like Moneybot and ManagerBot that prompt customers with personalized financial actions, and the ability to build entirely custom applications on the fly within the Square ecosystem. On the financial side, Block was already performing at second-quintile gross profit per employee before the restructuring. The implication is a dramatic step-function improvement in operating leverage. Jennings explicitly frames long-term defensibility not around traditional moats like distribution, network effects, regulatory licenses, or hardware — though those exist in the near and medium term — but around which companies understand something deeply that is hard for others to understand and can run the build-measure-learn loop fastest through agentic systems.

NECESSARY CONDITION

Regulatory frameworks must remain permissive to innovation (avoiding the 'European' model) and open source development must remain unencumbered by downstream liability.

I think that basically broke the first week of December and what we were seeing is that one or two engineers or a designer and an engineer who was on the tools quote unquote as we say is able to be 10 20 100x more productive.

03:15

RISK

Steel Man Counter-Thesis

Block has executed a massive, irreversible organizational restructuring based on approximately 4-5 months of observed AI productivity gains in a period of unusually rapid model improvement. The counter-thesis is that Block has confused a step-function improvement in AI coding capability with a permanent, compounding productivity revolution, and has locked in a structurally fragile organization that is optimized for the current AI capability frontier but dangerously exposed to plateau, regression, or unexpected failure modes. Specifically: (1) Historical precedent from every prior wave of developer productivity tools — from IDEs to cloud infrastructure to DevOps to low-code — shows initial 5-10x claims that normalize to 1.5-3x sustained gains as complexity compounds, edge cases accumulate, and maintenance burdens grow. Block has extrapolated from the steepest part of the adoption S-curve. (2) The company is simultaneously reducing headcount AND increasing product surface area (generative UI, Moneybot, ManagerBot, cross-ecosystem features), creating a situation where fewer humans are responsible for more complexity — a fragility pattern that historically produces cascading failures in financial infrastructure. (3) The regulatory environment for AI in financial services is tightening, not loosening. Block's trajectory toward removing human-in-the-loop oversight in compliance, risk decisioning, and customer-facing financial tools runs directly into an incoming wave of AI-specific financial regulation (EU AI Act, CFPB AI guidance, OCC model risk management expectations). A single enforcement action could force costly re-hiring and operational restructuring. (4) The competitive moat thesis — proprietary data understanding — is undermined by the very AI democratization the speaker celebrates. If AI makes building and iterating trivial, it does so for Block's competitors too, and Block's transaction data is not uniquely irreplicable in a market with Stripe, PayPal, Adyen, and banking-as-a-service platforms sitting on comparable datasets. The net result: Block may have achieved a short-term efficiency gain at the cost of long-term organizational resilience, regulatory risk, and a moat that is narrower than the narrative suggests. The stock's seven-year flatness may reflect not market irrationality but a legitimate market assessment that Block's competitive position does not justify a premium multiple — and a 40% RIF based on unproven AI assumptions does not change that fundamental calculus.

RISK 01

AI Productivity Gains Are Assumed But Unproven at Durable Scale

THESIS

The entire thesis rests on the claim that 1-2 engineers with AI tools can replace teams of 14, delivering 10-100x productivity gains. However, the evidence cited is extremely recent (post-December 2024) and limited to a few months of operation. Software development productivity is notoriously difficult to measure, and initial gains from AI coding tools often accrue to greenfield or well-scoped tasks. The harder, compounding challenges — debugging complex system interactions, maintaining large codebases over years, handling edge cases in regulated financial infrastructure, and managing technical debt — may not scale with the same efficiency. Block is essentially running a live experiment on its production systems with 40% fewer people. If the productivity multiplier proves to be 3-5x rather than 10-100x, the company is critically understaffed for its roadmap, and rehiring in a competitive talent market will be costly and slow. The speaker's own admission that Builderbot achieves 85-90% completion requiring human finishing work suggests the last-mile problem remains real and labor-intensive.

DEFENSE

The speaker asserts the productivity gains as self-evident ('pretty clear,' 'very very clear') but provides no quantitative metrics, controlled comparisons, or time-series data to validate the 10-100x claim. The entire argument rests on anecdotal observation over roughly 4-5 months. There is no discussion of what happens if the productivity thesis disappoints — no contingency hiring plan, no scenario analysis for a partial rather than complete paradigm shift. The 85-90% Builderbot completion rate is mentioned casually without acknowledging that for regulated financial products, the final 10-15% often represents the most complex and liability-laden work.

RISK 02

Non-Deterministic AI Outputs in Regulated Financial Products Create Systemic Compliance and Liability Risk

THESIS

The speaker explicitly describes generative UI where AI dynamically creates visualizations, charts, and even entire applications (ManagerBot building scheduling apps) that are not in the source code pushed to app stores. For a regulated financial services company handling payments, banking, lending (Afterpay), and money transmission, this introduces a category of risk that is fundamentally different from traditional software QA. Non-deterministic outputs interacting with customer financial data could produce misleading financial information, violate disclosure requirements, or create fair lending/fair treatment issues at scale. The speaker acknowledges this is 'potentially a nightmare from a QA perspective' but frames it as a solvable engineering problem rather than a deep regulatory and legal exposure. Regulators (CFPB, OCC, state money transmitter regulators) have been explicitly flagging AI-generated financial advice and non-transparent algorithmic decisioning. A single high-profile incident — incorrect financial advice from Moneybot, discriminatory outcomes in AI-driven risk operations, or a compliance failure in automated decisioning — could trigger enforcement actions, consent orders, or loss of critical licenses.

DEFENSE

The speaker partially addresses this by noting that the compliance team and compliance technology team were essentially untouched in the RIF, and that 'human in the loop' is currently maintained as a principle. However, the defense is weak because: (1) the speaker frames human-in-the-loop as a transitional state that will be phased out ('over time it's pretty obvious these systems are going to be so much better'), (2) the QA challenge for non-deterministic outputs serving tens of millions of financial customers is acknowledged but not resolved, and (3) the speaker's own trajectory points toward reducing human oversight while simultaneously increasing AI autonomy in regulated domains — a direction that runs counter to current regulatory posture.

RISK 03

Proprietary Data Moat Thesis Is Weaker Than Claimed — Commoditization Risk From Foundation Model Convergence

THESIS

The speaker's long-term defensibility thesis reduces to: companies that 'understand something super hard for other people to understand' will survive, anchored in Block's proprietary signal about 'how sellers and buyers participate in the economy.' This thesis has a critical vulnerability: as foundation models become more capable and commoditized, the ability to extract insight from transactional data becomes less differentiated. Stripe, Adyen, Toast, PayPal, and dozens of fintech companies sit on comparable or overlapping transaction data. Moreover, the speaker describes Block's AI infrastructure (Goose) as model-agnostic, running across 120+ models — which means the orchestration layer itself is not deeply moated since any competitor can build similar agent harnesses. If the key differentiator is data understanding plus rapid iteration via AI tools, and both the AI tools and the data patterns are increasingly accessible, then Block's moat may be narrower than presented. The Jevons paradox the speaker references could work against Block specifically: if AI makes it trivial to build fintech products, the proliferation of competitors erodes Block's pricing power and distribution advantage simultaneously.

DEFENSE

The speaker mentions distribution, network effects (50-60M MAUs), regulatory licenses, and hardware as near-to-medium term moats but explicitly concedes these may not hold long-term given the rate of change. The deeper 'proprietary understanding' moat is asserted philosophically but never substantiated with specifics about what Block uniquely understands that competitors cannot replicate. The speaker even acknowledges 'anyone can create a peer-to-peer app in a week' — the defense is that no one can vibe-code 50M MAUs, but this is a distribution argument, not an intelligence argument. The 'world model' concept described is compelling but early-stage and unvalidated. No evidence is provided that Block's data yields insights materially different from what competitors with similar transaction data could derive.

ASYMMETRIC SKEW

The downside is structurally underappreciated relative to the upside. Upside scenario: AI productivity gains are real and durable, Block operates at dramatically higher margins with fewer employees, and the market re-rates the stock as an AI-native financial platform — potentially 50-100% upside over 2-3 years. Downside scenario: productivity gains plateau at 2-3x rather than 10-100x, the reduced workforce cannot maintain system reliability and product velocity across regulated financial products, a compliance or outage incident triggers regulatory scrutiny, and Block must re-hire in a tight market at premium costs while competitors who retained talent continue executing — potentially 30-50% downside plus reputational damage. The skew is unfavorable because the downside is a compounding operational crisis (fewer people + more complexity + regulatory exposure = fragility), while the upside requires sustained, unprecedented AI capability improvement and flawless execution with a skeleton crew. The bet is asymmetrically risky: Block has already absorbed the full cost and disruption of the RIF, but the productivity benefits remain unproven beyond initial months.

ALPHA

NOISE

The Consensus

The market consensus on Block (SQ) is that the company is a mature, mid-tier fintech that has underperformed its stock price potential for 6-7 years despite growing the underlying business. The market views AI-driven workforce reductions in tech as largely a continuation of post-2021 overhiring corrections dressed up in AI narrative, and remains skeptical that AI productivity gains will translate into durable margin expansion or reacceleration of growth. The broader market consensus on AI in enterprise is that it improves productivity at the margin but that transformative, headcount-eliminating impact is still 2-3 years away for most companies, and that large-scale RIFs are risky operationally and culturally.

The market's causal logic is: (1) Block overhired during 2020-2021, (2) AI narrative provides cover for necessary cost cuts, (3) AI productivity gains are real but incremental and will take years to compound into meaningful financial impact, (4) workforce reductions of this magnitude carry execution risk (outages, talent loss, cultural damage), and (5) the competitive landscape in payments/fintech is intensifying with or without AI. The stock remains flat because revenue growth has decelerated and margins haven't yet proven the RIF's value.

SIGNAL

The Variant

Jennings believes a genuine binary discontinuity occurred in the first week of December 2024 — not a gradual improvement but a phase transition — when frontier models (Claude Opus 4.6, Codex 5.3) crossed a capability threshold enabling them to work with existing complex codebases, not just greenfield projects. This means the historical correlation between headcount and output is permanently broken, not cyclically correctable. Block's 40%+ RIF is therefore not a cost optimization or overhiring correction but a structural reorganization around a fundamentally different production function. He believes the market is mispricing Block because it is treating this as a typical tech layoff cycle rather than recognizing the company is 12-18 months ahead of peers in building the agentic infrastructure (Goose, G2, Builderbot) required to operate in this new paradigm. The implication is that Block's gross profit per employee will diverge dramatically from peers, and that the flat stock price represents a weighing-machine lag, not a signal problem.

Jennings's causal chain is materially different: (1) Block invested in agentic infrastructure starting in early 2024 (Goose launched as open-source agent harness), building institutional capability before the December 2024 model capability jump, (2) the December discontinuity made the old organizational model — hierarchical, linear, team-of-14 feature development — structurally obsolete overnight, (3) small squads of 1-6 people plus unlimited token access and multiple parallel AI agents now produce output equivalent to or exceeding prior teams of 14+, (4) the cuts were disproportionately on the development side precisely because that is where AI capability leapt, not on operational/sales teams where it hasn't yet, proving this is technology-driven not cost-driven, (5) the rebuilt organization has 50-60% fewer layers, 70-80% fewer meetings, and fluid squad allocation across products, and (6) products like Moneybot and ManagerBot built on the Goose platform represent a shift from static UI to generative UI that will fundamentally change engagement and monetization within 6 months. The compounding effect is that Block's proprietary signal — deep understanding of how sellers and buyers participate in the economy — combined with agentic tooling creates an iterative loop that accelerates product development from months to days, and eventually to hundreds of iterations per day.

SOURCE OF THE EDGE

Jennings's claimed edge rests on two pillars: (1) operational experience — he has been at Block for 12 years, ran Cash App during its scaling period, and directly oversaw the RIF and reorganization, giving him ground-truth visibility into before/after productivity metrics, and (2) proprietary tooling — Block built Goose, Builderbot, and G2 internally, meaning he has direct knowledge of what these tools can actually do versus what external observers speculate. This is a genuine structural informational advantage. He is not theorizing about AI productivity; he is reporting observed results — features built to 85-100% completion by autonomous agents, designers and PMs shipping PRs, 10-20x productivity multipliers for individual engineers. The specific details he provides (squad sizes, token budgets of $2,000, 14 parallel agent instances, Builderbot autonomously merging PRs) are operationally granular in a way that is very difficult to fabricate or narrativize. However, there is a meaningful credibility discount to apply: he is a sitting executive at a public company that just executed a controversial 40%+ RIF and needs to justify this decision to investors, employees, and the market. His incentive to frame this as visionary rather than risky is enormous. The productivity claims (10-20-100x) are extraordinary and unverified by third-party data. The 'binary change in December' narrative is convenient for justifying the timing. The edge is likely real but probably overstated in magnitude — the direction of travel is credible, the implied speed and completeness of transformation should be discounted by perhaps 30-40%.

CONVICTION DETECTED

• That basically broke • We're not writing code by hand anymore. That's over. That's done. • It was just there was a binary change • Pretty clear that in the future you'll be able to run that loop hundreds, thousands of times a day • I think that's like very very clear • It's like pretty obvious that these systems are just going to be so much better than like having a thousand humans • That's going to fundamentally change in the next like six months • I do believe that fundamentally for a given product or for a given road map, you're going to need fewer engineers, fewer designers, fewer PMs • I find Jack to be generally right and generally early • This was obviously a decision to go in a different direction • Generative UI is is is here

HEDGE DETECTED

• I'm not here to predict the future. I'm focused on block. • Maybe the humans are more like editors. • Maybe not. • I don't know what to expect. • That doesn't necessarily mean that there's going to be fewer engineers, designers, and PMs in the world. • It's also potentially a nightmare from like a QA perspective. • Right now I think it's critical that we have a human in the loop. • Markets are markets are cyclical and there's all sorts of things that are happening. The ratio of conviction to hedging is heavily skewed toward conviction. Jennings hedges on macro predictions about the broader industry and labor market — areas where hedging is rational and expected from a public company executive — but he is almost entirely unhedged on Block-specific operational claims about AI productivity, the permanence of the organizational shift, and the obsolescence of traditional software development. This pattern suggests genuine internal confidence about what he has observed operationally, combined with disciplined caution about extrapolating beyond his direct control. The hedges feel authentic rather than performative, which increases the credibility of the conviction statements. However, the absolute language — 'that's over, that's done,' 'binary change,' '100x' — is notably strong even for a true believer, and the absence of any hedging about whether the productivity gains will sustain or whether the RIF might have cut too deep suggests either very high internal confidence or an executive who cannot afford to express doubt publicly given the magnitude of the decision he just made. Weight the thesis seriously but recognize that the most extreme claims (100x productivity, features built to 100% autonomously) likely represent peak-case outcomes being presented as typical.