The AI Code Slop: Risk or Opportunity?

2026-02-19 - 41 min - source - Read full transcript

Sarah Guo (host)Elad Gil (host)

Key insights

The 'SaaS is dying, everything gets vibe coded' narrative is overstated in the near term.

Enterprise software isn't just code - it requires distribution, enterprise sales, security review, and change management across large organizations (e.g. Bank of America) that a five-person startup's internally vibe-coded CRM doesn't have to solve. Extrapolating small-team behavior onto the Fortune 100 misreads the moment; in the long run, some categories (per-seat to usage-based support, e.g. Decagon/Sierra) will shift, but not every SaaS company.

saas-disruption-narrative

Claims that AI agents are already autonomously choosing which vendors/software to buy are overstated.

What looks like an agent 'deciding' to use a specific tool (e.g. spinning up an AWS instance via Airtable) is usually just a pre-negotiated partnership executing in the background, not genuine autonomous purchasing behavior. Real agent-driven purchasing decisions - understanding a company's persona and needs - are a longer-run development, not a present-day reality.

saas-disruption-narrative

Abundant AI-generated code creates a new 'attention' bottleneck, not just a production one.

Once code generation is cheap, the risk shifts to nobody deeply reading, reviewing, or understanding the resulting codebase - more fragility from 'vibe coding slop' in real production systems, for every engineer, not just non-technical hobbyists. The open problem is managing human attention to engineering quality, via testing, smart review, or formal verification - and no one has solved it yet.

code-slop-and-engineering-culture

Engineers with a craftsmanship-based identity will react to AI coding tools very differently than utility-focused engineers.

Some engineers value the bespoke, artisanal act of writing code itself (like indie game developers), while others just want code as a means to build product (like an EA studio). As coding tools accelerate, the craftsmanship-identity group is likely to become unhappy at larger companies because agentic coding erodes the parts of the work they valued; utility-focused engineers experience the same shift as freeing.

code-slop-and-engineering-culture

AI labs are hitting revenue milestones at a pace with no historical precedent.

Internal analysis (via Capital IQ data) shows companies took 20+ years (ADP, Adobe) down to 8-9 years (Salesforce, SAP) to 3-5 years (Google, Meta, AWS) to grow from $1B to $10B in revenue - while the current wave of AI labs did it in roughly a year, and public projections put the $10B-to-$100B jump at 3-5 years versus ~27 years for Microsoft and over a decade for Google/AWS/Meta.

ai-revenue-and-pricing-curves

Token pricing for equivalent-capability AI models has collapsed by two orders of magnitude in under two years, even as usage and revenue explode simultaneously.

The cost of a GPT-4-equivalent model dropped roughly 150x in 21 months (from ~$37 to ~$0.25 per million tokens); an o1-equivalent model dropped roughly 88x in just 11 months (from ~$26 to ~$0.30 per million tokens). This combination of collapsing unit cost and explosive usage growth is framed as the strongest under-discussed signal in the AI market, versus focusing only on hype-driven anecdotes.

ai-revenue-and-pricing-curves

Tech's share of the US economy and stock market has roughly tripled in two decades, and AI is expected to accelerate that further.

Tech went from about 4% of US GDP and roughly 10% of S&P 500 value in 2005 to about 12% of GDP and over 50% of S&P value today (the top 8 tech companies alone represent about $23 trillion in market cap). Different assumptions about how much of GDP AI converts into 'tech spend' produce projections ranging from 15% to 30% of GDP by 2035.

ai-revenue-and-pricing-curves

Market value in tech remains sharply power-law distributed, and AI is unlikely to change that concentration.

Despite theories like 'the long tail' predicting value would spread out, most value (e.g. Google's ad revenue, YC's overall returns) concentrates in a small head-and-torso group of companies - a pattern that has held across cycles. What changes with AI is likely the absolute count of very large ($100B+) companies as the addressable surface area of software grows, not the underlying concentration ratio.

power-law-market-concentration

Historically, platform owners forward-integrate into the most valuable applications built on top of them, and AI labs are likely to repeat this pattern.

Microsoft's OS forward-integrated into Office, Excel, and Word by killing or acquiring competitors; Google's search platform forward-integrated into travel, local, and other verticals. The same dynamic is already visible with AI labs and coding, raising the open question of which application categories are durable and defensible versus likely to be absorbed by the labs themselves.

power-law-market-concentration

The current AI moment more closely resembles the volatile, fast-turnover dot-com internet wave than the slower, more 'protected' SaaS/cloud era.

In the SaaS era, growth was incremental and market leadership felt durable once achieved. In the dot-com wave, roughly 900 companies went public in 1999-2000 and only a dozen or two are still relevant today, because leaders like Friendster, Netscape, and AOL got displaced fast by new distribution, performance, and business models. The speakers argue AI is triggering the same kind of rapid leaderboard resets (e.g. 'is the next model jump from the labs going to reset my position?').

startup-exit-and-defense-strategy

Founders should treat exit timing as a routine, pre-scheduled board discussion rather than an emotional, reactive decision.

Because a company's peak value window can be as short as about 12 months, and past 'unassailable' leaders (Lotus 1-2-3 before Excel, dot-com winners before the crash) have collapsed even after explosive early growth, the recommended practice (credited to Ben Horowitz's approach at Opsware) is to put exit discussion on the board calendar once or twice a year so it becomes a normal check-in ('still not the time') rather than a crisis conversation.

startup-exit-and-defense-strategy

Multi-product bundling is framed primarily as a defensive strategy against AI-driven commoditization, not an offensive growth tactic.

A company that cross-sells five or ten integrated products into the same customer becomes a default part of the workflow and is much harder to clone or displace than a single-feature product. This is presented as a reversal of the old SaaS-era advice to 'do one thing well,' which the speakers argue was already flawed advice specific to a slower-moving era.

startup-exit-and-defense-strategy

Books referenced

The Long Tail - Chris Anderson - Referenced when arguing that, contrary to the book's thesis, value (e.g. Google's ad revenue) actually concentrates in the head and torso rather than the long tail - and that this pattern won't change in the AI era.

Media referenced

Mark Zuckerberg's post on AI eating the world - article - Cited approvingly as thoughtful, forward-looking commentary on how much unmet demand for software exists relative to engineering supply.
Keep Your Identity Small (essay, attributed to an Applied Intuition co-founder's blog post) - article - Invoked as advice for staying adaptable as AI reshapes what kinds of engineering work are considered high-status or difficult.

Companies

Samsara - Used as an example of a durable, hardware-plus-software business that vibe coding cannot easily replicate or displace.
Decagon - Cited as a real example of the shift from per-seat software to utilization-based customer support agents.
Sierra - Cited alongside Decagon as evidence of a genuine per-seat-to-usage-based pricing shift in customer support.
Jira (Atlassian) - Used repeatedly as the example of a ticketing tool people complain about but wouldn't actually want to rebuild or replace internally.
Airtable - Used as an example of how 'agents choosing vendors' really means a pre-negotiated partnership spinning up infrastructure like AWS under the hood, not genuine autonomous purchasing.
OpenAI / Anthropic - Used in the Capital IQ revenue-growth comparison as the AI labs that went from $1B to $10B in revenue in roughly a year, versus decades for prior generations of companies.
ADP, Adobe, Salesforce, SAP, Microsoft, Google, Meta, AWS - Comparison set for how many years each took to grow revenue from $1B to $10B (20+ years down to 3-5 years), illustrating the AI labs' unprecedented speed.
Baseten, Fireworks - Named as inference clouds where a large share of AI token inference is happening, alongside the large model providers directly.
Nvidia - Cited as an example of market cap exploding from tens of billions to trillions of dollars in a few years.
Cognition, Claude/Anthropic - Used as examples of coding-agent partnerships where a company spins up specific tools/infrastructure as part of a vendor relationship, not autonomous agent purchasing.
Stripe, PayPal, Braintree - Payments history example: everyone said payments was already won by PayPal/Braintree, then Stripe still became a winner in a fragmented market.
Friendster, MySpace, Facebook - Social networking history example of 'unassailable' leaders being displaced by a later entrant.
AOL, Yahoo, Netscape, Internet Explorer, eBay, Time Warner - 1990s dot-com era case studies used to argue today's AI era looks more like the fast, high-turnover internet wave than the slower SaaS/cloud era; AOL's sale to Time Warner at its peak valuation is cited as good exit timing.
Lotus / Lotus 1-2-3, IBM, Microsoft Excel - 1980s example: Lotus 1-2-3 grew explosively as the killer spreadsheet app, then collapsed and was absorbed by IBM once Microsoft launched Excel and took the market.
Mark Cuban's company (sold to Yahoo) - Cited as a rare example of a founder exiting and hedging (collaring Yahoo stock) at exactly the market's high-water mark, preserving the value as the stock later collapsed.
Notion - Used as an example of a software product that embeds a specific point of view on how to work, illustrating how cheaper software production lets more distinct points of view get expressed as products.

Techniques and frameworks

Pre-scheduled, non-emotional exit-timing board meetings - Recommended founder practice (attributed to Ben Horowitz's approach at Opsware): put exit discussion on the board calendar once or twice a year in advance so it's a routine, unemotional check-in rather than a reactive, anxiety-laden decision.
Multi-product bundling as a defensive strategy - Framed as the best defense against AI-driven commoditization: becoming embedded across many parts of a customer's workflow makes a company harder to displace than a single, easily-cloned feature.
Keep your identity small - Applied to engineers: tying self-worth to a specific type of technical craftsmanship makes the AI-driven shift in what's 'hard' or high-status more threatening; staying identity-flexible makes people more adaptable.

Summary

Sarah Guo and Elad Gil use this episode to push back on what they call a month of overstated "SaaS apocalypse" narratives - the idea that per-seat enterprise software is about to be wholesale replaced by vibe-coded internal tools. They argue this claim conflates what a five-person AI-native startup can do (skip Salesforce, build a CRM over a weekend) with what a Fortune 100 company actually needs: distribution, enterprise sales, security review, and change management at scale. Real shifts are happening in specific categories - Decagon and Sierra moving customer support from per-seat to usage-based pricing - but this isn't a universal replacement of software vendors. They also debunk the claim that AI agents are already making autonomous vendor-purchase decisions, arguing that what looks like agentic purchasing is usually a pre-negotiated partnership (a coding agent spinning up AWS infrastructure) rather than genuine independent choice.

The conversation turns to the flip side of abundant AI-generated code: the "slop" problem. If code production is no longer the bottleneck, the new bottleneck is human attention - nobody reviews or deeply understands the resulting codebase, and fragility rises. They frame this as an open, unsolved space (testing, smart review, formal verification are partial ideas) and note that engineers will react very differently depending on whether their identity is rooted in bespoke craftsmanship (who may become unhappy as AI erodes the parts of coding they valued) or in using code as a utility to ship product (who will find the shift freeing).

They then pivot to data meant to counter the hype-driven narrative with harder signals: AI labs reached $1B to $10B in revenue in roughly a year, versus 20+ years for ADP and Adobe, 8-9 years for Salesforce and SAP, and 3-5 years for Google, Meta, and AWS. Token pricing for equivalent-capability models has simultaneously collapsed - about 150x in 21 months for GPT-4-class models and 88x in 11 months for o1-class models - even as usage and revenue explode. They connect this to a broader macro trend: tech's share of US GDP has roughly tripled since 2005 (4% to 12%), and the top eight tech companies now represent over half of S&P 500 value, prompting a discussion of how much further tech's share of GDP could grow by 2035.

A significant thread addresses market structure: despite theories like "the long tail," they argue value in tech remains sharply power-law distributed (a handful of companies capture most value, as with Google's ad revenue or YC's overall returns), and AI is unlikely to change that underlying concentration - though the total number of very large companies may increase as the addressable market grows. They also revisit the historical pattern of platforms forward-integrating into their most valuable vertical applications (Microsoft into Office, Google into travel/local search), suggesting AI labs will do the same with coding and other categories, which raises the open question of which startup categories are actually durable.

The episode closes on founder-facing advice drawn from historical parallels - the dot-com wave (roughly 900 IPOs in 1999-2000, only a dozen or two still relevant today), Lotus 1-2-3's collapse once Microsoft launched Excel, and Mark Cuban's well-timed sale to Yahoo. Gil and Guo argue the current AI moment resembles the volatile internet era more than the protected SaaS era, and recommend founders pre-schedule non-emotional board discussions about exit timing (crediting this practice to Ben Horowitz at Opsware) since a company's peak-value window can be as short as roughly 12 months. They also frame multi-product bundling as fundamentally a defensive move - becoming embedded across many parts of a customer's workflow - rather than the old SaaS-era advice to "do one thing well."

Notable Quotes

Note: the source transcript (YouTube auto-captions) carries no reliable speaker diarization - turn markers are sparse and several speaker changes happen with no marker at all (confirmed by tracing the raw caption stream). Attributions below are best-effort, based on self-identifying context (e.g. references to "my team," a specific portfolio holding, or one speaker addressing the other by name) rather than verified audio, and should be treated accordingly.

"It's like the slop problem, but instead of it being like vibe coding slop for random websites for non-technical people, it's vibe coding slop in my actual production code base for every lazy engineer, which is every engineer." - Elad Gil

"I mean, I think the idea of vibe enterprise sales is hilarious, because we have portfolio companies with hundreds of millions of dollars of revenue who are very committed to as much token usage as we can, as few great people as we can have." - Sarah Guo

"Does that mean that a Fortune 100 company is going to displace their CRM with some internal thing they got vibe coded over the weekend? Probably not." - Elad Gil

"We're seeing the fastest time to real massive revenue that we've ever seen in the history of software." - Elad Gil

"There's about a 12-month window where your company is the most valuable it'll ever be, and then it crashes out." - Elad Gil