Software Stocks Implode: From 'When' to 'If' on SaaS Survival

Anthropic is three for three in tanking market sectors in a single month. A legal plugin, a security tool, and a COBOL modernizer, each triggering double-digit stock drops across their respective industries. Then a fictional Substack post about a 2028 economic collapse went mega-viral at 28 million views and moved actual markets on Monday. The All-In crew uses these events as a lens to examine something deeper: whether the market’s relationship with SaaS has fundamentally changed, and what the AI productivity explosion means for jobs, data center policy, and American politics.

The “When” to “If” Shift

Chamath offers the most structurally interesting framing of the SaaS selloff. The market has historically debated when a company’s cash flows would be disrupted. Now it’s debating if those cash flows will exist at all.

The mechanics: when markets shift into “if” mode, holders demand a massive margin of safety. P/E multiples that were 40x should trade at 20x. Revenue multiples drop from 10x to 3x. The weighted average cost of capital (WACC) gets jacked up from 6% to 12-13%, reflecting the market’s way of saying “I’m now debating if these things will even exist, so I need a huge buffer to own this stuff.”

This isn’t just about pricing; it’s about the entire category losing its defining trait: predictability. Sacks explains that SaaS used to be modeled as a “growth annuity.” You had annually recurring revenue, net dollar retention of 120%, clear metrics, clear outcomes. Companies were bought at 13x ARR with confidence in the underlying durability. Now you have to factor in the possibility that AI disrupts the entire market, changes pricing models, or eats into growth opportunity. Even if it doesn’t eliminate Salesforce, the unknowns shatter the annuity thesis.

“We have moved away from a when to now an if.”

Chamath identifies two forces at work: a tactical one (smart money hedge funds degrossing, trimming positions, reducing risk across the board) and a structural one (the “when to if” shift). He can’t assign percentages between the two, but believes the structural change is the more important one.

The ripple effects hit recruiting and retention hardest. Stock-based compensation, the primary tool tech companies use to attract talent, depends on stock price confidence. When that confidence evaporates, the entire talent flywheel starts grinding.

The Citrini Report: Science Fiction That Moved Markets

A Substack post from Citrini Research, set in a fictional 2028 “Global Intelligence Crisis,” went viral at 28 million views on X. Its thesis: companies embrace AI, cut staff, margins go up, then they lose their customer base because consumers have no discretionary spending, creating a death spiral. The piece also speculated that AI agents would eliminate credit card interchange fees by moving transactions to stablecoins. Financial stocks got hammered: Amex down 8%, Capital One 8%, Mastercard 6%, Visa 4%.

Sacks raises a credibility issue. The article’s co-authorship was amended after publication to include a managing partner of a $262 million SEC-registered hedge fund who confirmed short positions in the companies the report named. The question of whether it went genuinely viral or was amplified by interested parties remains open.

The best response, according to Sacks, came from Derek Thompson’s piece “Nobody Knows Anything” (a reference to William Goldman’s famous line about Hollywood). Thompson argues that the conversation about AI is “a marketplace of competing science fiction narratives” because “the level of uncertainty is so high and the quality and supply of real-world real-time information about AI’s macroeconomic effects so paltry that very serious conversations about AI are often more literary than genuinely analytical.”

A Kalshi prediction market shows only about 12% believe the Citrini scenario will actually play out. But the damage to stock prices was real.

Jevons’ Paradox in Action

Sacks presents the counterargument with data. Anthropic is hiring software engineers at $570,000. Citadel Securities’ report shows software engineer job postings rising roughly 10% year-over-year. Company formation is also rapidly expanding, possibly because AI is lowering the barrier to starting a business.

The framework comes from Aaron Levie: when you lower the cost of something that was previously supply-constrained, demand for that thing goes up. Software engineering has always faced chronic shortages, even among Silicon Valley startups. Fortune 500 companies have had it worse. If you 10x or 100x productivity per engineer, the unfilled demand could absorb all that new capacity without mass layoffs.

Sacks puts a number on it: the average Fortune 500 business spends about 5% of its cost structure on all IT. But as Elon frames it, companies are “cybernetic organisms that are part software, part human.” If a company is currently 1-2% software, maybe it should be 50%. The market for software was so constrained by supply that the demand headroom is enormous.

Chamath agrees on the structure: he expects OPEX as a percentage of revenue to fall off a cliff, but within that OPEX, the percentage allocated to technology goes way up from where it is today.

Jason’s OpenClaw Experiment: 30 Days > 10 Years of SaaS

Calacanis provides the most vivid ground-level evidence. His 20-person firm brought 15 people in for a weekend training session on OpenClaw (the AI agent platform). In 30 days, his team built every piece of software they had wanted to buy or build over the last 10 years.

Concrete examples:

• Podcast ad sales agent: Scans top 100 podcasts, identifies advertisers from transcripts, cross-references with their CRM (Pipedrive), checks last contact date, and feeds it into the sales pipeline. This was an SDR job they wanted to fill.
• Productivity tracking: An “Ultron” agent with root access to Gmail, Calendar, Zoom, Notion, and Slack that generates weekly summaries for each employee: emails sent, meetings taken, contacts made, threads involved.
• Self-improving thumbnail agent: Every week, it autonomously searches for the latest YouTube thumbnail optimization techniques. Found an article about Mr. Beast’s team using heat maps, added it to its skill file, and now applies those learnings whenever they create thumbnails. The system is recursive: it gets better every week without human intervention.
• Podcast clipping: Takes an episode, identifies the three best moments, makes clips, adds subtitles, and posts them to Slack. Was going to be a full-time job.

The result: same headcount, 10-20% more efficient every week, compounding. The team isn’t growing; they’re not adding people.

“Every piece of software that we wanted to buy or build over the last 10 years that we never got to, my people are building in the last 30 days.”

The most telling moment: their agent suggested replacing Slack with an open-source alternative called Mattermost, offering to spin it up over a weekend and export their data. Calacanis declined (only spending $6-10K/year on Slack), but the implication is clear: when you renegotiate with any SaaS vendor, you can credibly say “we could roll our own.”

Ryan Peterson posted that Claude for legal “seems to work just as well as Harvey,” extending the SaaS apocalypse to private companies too.

Why Doomer Narratives Win

Sacks identifies three heuristic biases that make doomer AI narratives inherently more appealing:

1. The seen vs. the unseen: It’s far easier to see existing jobs that could be obsoleted than to imagine new jobs and business models that haven’t been created yet. The creation takes a genius innovator; seeing potential destruction takes no creativity.

2. The fixed pie fallacy: Most people think of the economy as zero-sum. If someone’s getting rich, it must be at someone else’s expense. But as one article put it: “The economy is not a pie, it’s a garden. And technology is rain.”

3. Dystopia sells: Most sci-fi movies are dystopian, not utopian. Doom narratives are inherently more compelling and shareable.

Sacks also flags real-world constraints that make both hyper-utopian and hyper-dystopian narratives unlikely in the near term: token supply limitations, energy production, land, power, chip production. There isn’t enough time in the next few years for the whole economy to change as dramatically as either extreme predicts.

Chamath expects a 10x increase in token demand but also a 90% price reduction in output tokens by year-end, creating an enormous upswell of demand. His firm (8090) has already added AI token costs as a line item in fully burdened employee cost models because some engineers are “racking up ginormous bills.”

The Consumptive Capacity Question

Friedberg raises a deeper structural question that the others don’t fully engage with. His framework: humans need roughly 10% income improvement per year to feel happy. That’s the floor of consumptive capacity. But is there an upper limit?

If AI creates such profound productivity leverage that the ability to make stuff exceeds the capacity to consume stuff, we’re in uncharted territory. General economic models, productivity models, and social models all start to break.

“Knowledge work in general may also be a transitory phenomenon that only existed between the foundation of computing tools and the existence of AI.”

If 100x’d productivity has no consumer on the other end, the models collapse. This is the question nobody has a good answer to yet.

Data Centers: Build Here or Lose to the Middle East

Chamath presents data showing 100 data center projects currently face local opposition. In 2025, about 5 gigawatts of projects were cancelled. Using OpenAI CFO Sarah Frier’s estimate that every gigawatt equals roughly $10 billion in revenue, that’s $50 billion in lost revenue for 2025. If 7 gigawatts get cancelled in 2026, that’s another $70 billion, roughly $130 billion in lost revenue over two years.

Sacks explains the Ratepayer Protection Pledge announced in the State of the Union: big tech companies provide their own power for AI data centers so residential consumers don’t see rate increases. The approach includes behind-the-meter power generation where data centers don’t even connect to the grid. When they do connect and have excess capacity, they can actually bring down consumer electricity prices by increasing grid scale and reducing per-unit fixed costs.

Friedberg makes the geopolitical argument: data moves at the speed of light, so data centers can go anywhere. Saudi Arabia, UAE, and Qatar are 10x-ing their data center builds. If America doesn’t build them, the economic value (construction jobs, energy installation, second and third-order industries) accrues elsewhere. “All the data centers in the world fit under the tip of a pin” in terms of physical footprint relative to economic value.

Chamath adds a nuance: even if data centers fund their own power, utility prices could still rise because of how the utility business model works. Utilities present capex plans to public utility commissions that allow them to invest for a return. Residential electricity consumption is also going up from more devices, EVs, and other draws independent of data centers.

The most egregious example: Micron’s $100 billion mega-fab in New York took 1,200 days from announcement to groundbreaking, with 612 days on the environmental impact study, and was challenged by just six citizens. Calacanis contrasts this with Texas, where Elon built the Gigafactory in under 18 months.

Science Corner: Yamanaka Factors Enter Human Trials

Life Biosciences (co-founded by Harvard’s David Sinclair, the controversial resveratrol/aging researcher who sold a company to GSK for $720 million that didn’t pan out) has reached an agreement with the FDA for the first human trial of Yamanaka factors. These four proteins, discovered in 2006, can reverse the age of mammalian cells by resetting the epigenome: the markers on top of DNA that turn genes on and off.

The first application: injecting Yamanaka factors into the vitreous fluid of the eye to rejuvenate the retina and restore vision in people blinded by glaucoma or stroke-like eye diseases. The delivery mechanism uses AAV viruses carrying DNA that produces the proteins, with a switch mechanism controlled by the antibiotic doxycycline. Take the antibiotic and protein production turns on; stop and it turns off.

If it works (expected based on animal models), it’s not just a blindness breakthrough. It’s the first human application of cellular age reversal. Over a dozen startups are now pursuing similar approaches. Next targets could include joints, arthritis, and skin (monkey studies show wrinkles disappearing). Friedberg calls it “the beginning of a wave of what I think will be the most extraordinary revolution in human therapeutics.”

SCOTUS, Tariffs, and the 150-Day Clock

The Supreme Court voted 6-3 against Trump’s emergency powers tariffs (three conservatives joining three liberals), the biggest rebuke of executive policy in 91 years. About $175 billion in tariffs collected, with potentially 50% subject to refund. 2,000 importers have already filed.

But Sacks argues the tariffs aren’t going away. Kavanaugh’s 70-page dissent provided a roadmap with multiple alternative legal bases: Section 122 of the Trade Act of 1974 enables temporary 150-day tariffs up to 15% for balance-of-payments issues (already invoked by Trump), while Section 301 and Section 338 provide longer-term authority once the administration substantiates its case through studies and agency reviews.

Chamath considers the tariff experiment successful: it smoked out structural trade imbalances that existed not because they made economic sense for America but because of what he calls “a hodgepodge of globalist drivel.”

Friedberg focuses on the institutional significance: a court with a majority of politically aligned appointees ruling against the president demonstrates the system works. Calacanis makes a plea for bipartisan collaboration, noting Congress could simply ratify the existing tariffs while establishing a framework for future ones.

Afterthoughts

This episode captures a market in genuine identity crisis. The SaaS category didn’t just have a bad week; it lost the narrative that made it investable for two decades: predictability.

• Chamath’s “when to if” framework is the sharpest articulation of what’s happening. When investors can’t model the next 5 years of a business with confidence, the entire valuation methodology breaks down. It doesn’t matter if AI ultimately doesn’t kill Salesforce; the uncertainty alone demands a massive discount.
• Calacanis’s OpenClaw experiment is more revealing than any market analysis. When a 20-person media company can replicate a decade of desired SaaS purchases in 30 days with AI agents, and those agents recursively self-improve (finding new techniques, updating their own skill files, applying them automatically), the category’s value proposition is genuinely under threat.
• The Citrini report saga (viral science fiction by a hedge fund shorting the names it bearishly analyzed) is the kind of thing markets will need to develop antibodies against as AI narratives become the primary mechanism for moving stocks.
• The buried lede may be Friedberg’s consumptive capacity question: if production can outstrip consumption, we’re not just in a SaaS repricing. We’re in a philosophical crisis about what economic growth means.