Don't Mistake a Month of Hype for the End of Software

Gil sees real meta-trends behind the market panic but many wrong conclusions about specific companies. The premise that per-seat software will vanish because anyone can vibe-code a Salesforce replacement strikes him as short-sighted in the near term, even if the long run is anyone's guess. His example is Samsara, a fleet-management company in which he still holds shares: nobody is going to vibe-code the app, the in-cab camera sensors, the enterprise sales and the support agents to match.

There is a genuine shift underneath, he adds. Companies such as Decagon and Sierra are moving customer support from per-seat pricing to utilization-based agents, which may hurt an earlier wave of per-seat vendors. But that is not every SaaS company, and Gil calls the sell-off a massive correction around something that is only partly true and only for some firms.

Guo finds the idea of "vibe enterprise sales" hilarious. She points to AI-native portfolio companies with hundreds of millions in revenue that run on as much token usage as possible and very few engineers, yet still went from zero to nearly 100 salespeople very quickly. Humans, in other words, are still doing the selling.

Both hosts argue that the extrapolation comes from engineers' own picture of the world. A small technical team that builds a custom CRM instead of using a spreadsheet is fine, and impressive, but a Fortune 100 company will not replace its CRM with something built over a weekend. Guo adds that the engineer annoyed about paying $10 a seat for Jira would not want to run change management at Bank of America, handle security reviews and maintain the system afterwards.

Gil says people underestimate how much demand there is for software of every kind, including AI. With so little engineering supply relative to that demand, a large productivity boost simply gets absorbed because there is so much more to build. He notes that startup teams are not giving up on hiring engineers; the nature of the work is shifting instead.

That shift will be uncomfortable for some. Gil distinguishes engineers who value bespoke craftsmanship, aesthetics and code quality from those who write code as a utility to build product. He compares the first group to indie game developers versus people who work at a large studio like EA. As coding tools accelerate, he expects many craft-focused engineers to be unhappy at larger companies, while utility-minded builders will feel freed.

Guo adds that an engineering identity built on a status ranking of difficult work is fragile, because the kinds of engineering seen as impressive can be less hard for agents than expected. She recommends the advice from a founder at Applied Intuition, who wrote an essay arguing to "keep your identity small" and so stay adaptable during this period.

Guo says agent-first engineering management and code quality are unsolved problems in all of her teams. If an enormous amount of code is generated and no one reads it, nobody deeply understands the code base and fragility grows. It is the slop problem, she says, but inside your production code base rather than on throwaway websites for non-technical users.

She thinks ticketing systems like Jira could be at risk, but the bigger opening is for a company that works out how to manage human attention in engineering. Candidate approaches include testing, smarter review, letting agents do the review, and formal verification. Her verdict is that it is "open season" around this very large problem.

The hosts also push back on the claim that agents are already choosing software vendors. Gil argues that what is really happening is that a partnership, with a coding-agent company for example, leads an agent to spin up a basic instance of specific tools, just as Airtable users unknowingly spin up AWS instances. Real agent commerce, where agents understand your preferences, is a long-run story.

Gil calls the past month a month of hype in which mainstream media believed overstated stories, including agents supposedly pushing humans out of their own forum, when some of the posts were likely human-generated or planted for marketing. Guo adds that demos differ from the full software people need, and that equity researchers seem to have forgotten that competitive advantage does not disappear overnight. Both stay excited about the long-run effect of cheaper software, which lets many more points of view get expressed in code.

Gil shares a chart that a colleague, Jared, built from Capital IQ data showing how many years companies took to go from $1 billion to $10 billion in revenue. ADP and Adobe needed 20-plus years, Salesforce and SAP eight or nine, Microsoft about seven or eight, and Google, Meta and AWS three to five. The AI labs did it in roughly one year.

Public projections for the jump from $10 billion to $100 billion look similar, he says: about 27 years for Microsoft, over a decade for Google, AWS and Meta, but only three to five years for the labs. He attributes this to a global online pool of customers, easy distribution, higher GDP and capabilities that create value at huge scale. In his view people are ignoring the revenue and usage side of the equation.

The second chart covers token pricing. For GPT-4-level capability, a million tokens went from roughly $37 to 25 cents in 21 months, about a 150x drop. For o1-equivalent models, the price fell from about $26 in December 2024 to about 30 cents in November 2025, an 88x drop in 11 months. Cost is collapsing while usage and revenue ramp, and Guo adds that inference-cloud consumption is growing around a thousand-fold.

The two then riff on the human brain, which runs on roughly 12 to 20 watts, as a reminder of how much efficiency is still to be won in models. Gil stresses he is bullish on all of it, and his point is only that the numbers should be weighed against over-extrapolation of what the systems can actually do today.

Guo relays a conversation with a purchasing lead at a large traditional enterprise who argued that incumbents have the money to fight back. She counters with reflexivity in markets: if the best AI labs and application companies grow to a billion in run rate and their valuations rise with that, incumbents may lack the market cap to compete as acquirers. Gil sees signs of the cash influx in San Francisco housing, where lab tender offers are creating new wealth.

Gil's team asked what proportion of GDP is tech. In 2005 Google was worth about $100 billion and Exxon was the most valuable company at $400 billion; Apple became the first trillion-dollar company only in 2018, when tech was about 30% of the S&P. Today the top eight tech companies are worth about $23 trillion and make up well over half of the S&P, while tech went from about 4% of GDP in 2005 to about 12%.

AI accelerates that by augmenting services and jobs and converting them into software or tech spend. Depending on assumptions, Gil says tech could reach anywhere from 15-20% to 30% of GDP by 2035, which would make today's market caps look small. He sees this as a useful lens on the terminal values of these companies a decade out.

Guo asks what all this means for investing. If everything gets much bigger, a billion dollars is no longer late stage, and Gil notes some companies already raise at a billion-dollar valuation on tiny revenue. He recalls a blog post from about a decade ago on how hard it was to build a sustainable $5 billion company, when the largest market caps were in the low hundreds of billions; since then values have grown tenfold, and the open question is how many trillion-dollar companies there can be.

He argues that every platform shift has seen the platform forward-integrate into the most important vertical applications: Microsoft folded Office into its OS franchise, and Google built travel and local on top of search. The labs will likely do the same, as they already have with code, so founders must ask which verticals are defensible. Gil leans toward the view that the tail grows, because the addressable surface area is expanding so quickly.

Guo cautions that value stays power-law concentrated in the head and torso, as with Google's ad revenue or Paul Graham's observation that most of YC's value comes from a handful of companies. The two broadly agree there are more hundred-billion-dollar businesses now even if the distribution is just as skewed. Guo adds that category leadership once looked unassailable and no longer does, since a company can reach a billion in revenue and still face the question of who the winner will be.

Gil points to precedents: roughly a thousand to two thousand companies went public in the internet wave and only a couple of dozen stayed relevant. Friendster and MySpace lost to Facebook, and when he invested in Stripe people asked why anyone would back another payments company with Braintree and PayPal around. Lotus 1-2-3 grew explosively in the 1980s, then fell into IBM's arms while Microsoft's Excel took the market.

Gil says founders should consider two questions: how durable is the business, and when should they exit if they are going to. Often there is a window of about 12 months when a company is the most valuable it will ever be before it crashes. For a very small handful of companies the answer is never sell, but for most the point is to sell when the timing is right.

His practical tip is to pre-schedule a board meeting once or twice a year about exits, which turns it into a non-emotional discussion. He attributes a version of the advice to Ben Horowitz from his Opsware days. The meeting can conclude that it is not yet time, or that a competitor shift or an unusually high offer means now is the moment, so that years of hard work are not lost.

Guo thinks AI looks more like the internet era, with growth at wild speed as the default, than like the incremental SaaS era. She cites AOL, Yahoo, Netscape and eBay as the internet's early control points, and Gil notes that AOL sold to Time Warner at its peak. Gil adds Mark Cuban's sale to Yahoo, which he hedged by collaring the stock so he kept the value as Yahoo's price collapsed.

Guo says the useful question is whether the next major capability jump from the labs will reset the leaderboard, or whether shifts such as agents versus IDEs or voice as a default will reallocate power. Gil adds that with changes this fast, the exit question is worth revisiting every six months.

Gil's main defensive advice is to build a bundle: a multi-product surface area that cross-sells into the same organization so you become a default part of the workflow. A company used for five or ten parts of a vertical is much harder to displace than a single easy-to-clone product. He notes that bundles are usually seen as offensive but are underrated for defense, and expects them to be the way to win long term in many vertical applications.

Guo notes that the SaaS-era mantra was to do one thing well, which Gil calls bad advice that only became popular because SaaS technology moved slowly. Earlier waves of software companies were acquisitive and multi-product. With AI, the velocity of change is so high that a displacement cycle that took a decade now happens in a year or two, driven by scaling laws, reasoning and post-training advances.

Founders therefore need a new mindset, Gil says: if every two years is like ten, you must react to change quickly. Guo closes by listing defenses that do not rest on incremental features, namely platforms, ecosystems, networks, bundles and even hardware. Her advice is not to over-rotate on the last month, to be honest about your market position, and to identify your real control points.