Over a trillion dollars has been wiped from software stocks since September 2025. The IGV software sector ETF is down roughly 30%. Forward earnings multiples for software companies have collapsed from 39x to 21x — a 46% compression in months, not years. Morgan Stanley published what they diplomatically called a “gut check” on the sector, identifying a “trinity of software fears” that amounts to: AI isn’t just disrupting software companies, it’s eating the concept of software as a business.
There’s no venture-backed SaaS IPO filing on the horizon. Klarna ripped out Salesforce and built its own CRM with AI — though even Klarna’s CEO later said he doubts most companies can replicate the move. Every time Anthropic launches a new product — cybersecurity tools, legal copilots — a cluster of stocks in that vertical sells off like someone yelled fire in a theater.
The honest version of every LP meeting happening right now is four words: what the fuck now?
I’ve raised $108.5 million over ten years. Seed to Series B, angels to Autodesk to Lockheed Martin to the CIA’s venture arm to the British government. I have over a hundred investors on my cap table spanning every category of capital except public markets and private equity. Between 2019 and 2022, I took roughly a thousand VC meetings to keep my company alive through Covid. I’ve been skiing with these people. I’ve been on holiday with them. I’ve read their books, made thousands of versions of my investor deck, and sat through more pitch meetings than I can count. I’ve also sat on the other side of the table — nearly sixty angel investments and counting. I’ve been the founder begging for checks and the check-writer deciding who gets one. The view is different from each side. The herd looks the same from both. I’m not writing this as an analyst. I’m writing it as someone who’s spent a decade watching this from the inside.
The Herd
Most VCs lose money.
The Kauffman Foundation — one of America’s largest institutional investors and a major LP — published a landmark study in 2012, “We Have Met the Enemy... And He Is Us.” They analyzed 99 VC funds from 1989 to 2011 and found that 62 of 99 failed to exceed returns available from small-cap public equities after fees. Not “slightly trailed the S&P.” Beaten by a boring index fund with zero management fees.
The power law in venture capital is real, but it’s even more extreme than most people realize. A tiny fraction of funds capture nearly all of the industry’s returns. AngelList’s 2024 Access Fund analysis — the work of Abraham Othman and team on access versus selection — showed that broad market-rate exposure to the startup ecosystem is virtually assured of capturing return-driving companies. The individual stock-picker, statistically, is not.
This matters because of what it reveals about how VCs actually make decisions. If you’re a fund manager and you know — consciously or not — that your odds of picking winners are low, what’s your rational strategy? It’s not to develop genuinely contrarian theses. It’s to stay close to consensus. If you invest in what everyone else is investing in and it fails, you were unlucky along with the market. If you go contrarian and fail, you’re an idiot.
Career risk dominates investment risk. (Marc Andreessen has made this observation, and it’s one of the more underappreciated truths about institutional capital allocation.)
The incentive structure of venture capital — 2% management fees on committed capital regardless of returns — means that staying in the herd is individually rational even when it’s collectively stupid. Steven Sinofsky, writing on his Substack, put it plainly: “Wall Street is filled with investors of all types... they tend to run in herds. The past couple of weeks have definitely seen the herd collectively conclude that somehow software is dead.”
Sequoia — arguably the most successful venture firm in history — has published its “wake up” memo three times. “R.I.P. Good Times” in October 2008. “Adapting to Endure” in May 2022. Their generative AI memo in 2024. The pattern itself is the data point. Every cycle, the herd stampedes in one direction, the smartest firm in the room publishes a correction, the landscape reshuffles, and the herd finds a new direction to stampede.
So when every VC in San Francisco is saying “software is dead, AI is everything, atoms are the new bits” — and when every VC in London is saying “see, we told you hardware matters” — the first question to ask isn’t whether they’re right.
It’s whether this is analysis or another stampede.
The Track Record
The PC was supposed to kill the mainframe. E-commerce was supposed to kill retail. Streaming was supposed to kill Hollywood. In each case, the threatened industry shed its weakest players while the overall activity expanded. U.S. e-commerce is still under 20% of total retail sales a quarter century after Amazon went public. Sinofsky’s core argument is that every disruption plays out WAY slower than predicted, on 25+ year career timescales rather than the 2-3 year windows panicking investors imagine. He’s broadly right. The historical track record of “X is dead” predictions is abysmal.
But sometimes the herd is directionally right even when it’s wrong on timing.
The music industry really did collapse. U.S. recorded music revenue (per the RIAA) fell from its peak of roughly $14.6 billion in 1999 to $6.7 billion in 2014. It recovered through streaming — but the recovery took 15 years and the value accrued to entirely different companies. If you were a VC invested in a traditional label in 2001, “the music industry will eventually be fine” was cold comfort. It was fine. You were still broke.
Print media really did collapse. Not “evolved” — collapsed. U.S. newspaper advertising revenue fell from roughly $47 billion in 2006 to $9 billion by 2020, per Pew Research Center. The total amount of written content produced exploded, exactly as optimists predicted. But the economic value migrated to platforms the newspaper industry never owned and couldn’t build.
The pattern isn’t “nothing ever dies.” The pattern is: the activity expands, but the economic value migrates to new structures, and the migration takes longer than pessimists fear but is more total than optimists expect.
Is software in the “retail” bucket — disrupted but fundamentally sound? Or the “newspapers” bucket — activity explodes, value migrates, incumbents die?
What’s Actually Different This Time
The cost of building software is approaching zero. Not declining. Not getting cheaper. Approaching zero. Lex Zhao at One Way Ventures says “the barriers to entry for creating software are so low now thanks to coding agents that the build versus buy decision is shifting toward build.” This isn’t aspirational. It’s descriptive. Klarna built its own CRM. AI-native startups — Cursor, Lovable, Bolt among them — have reached staggering revenue milestones at speeds that would’ve been physically impossible three years ago.
This is genuinely new. In every previous technology disruption, the cost of the new thing dropped while the cost of the existing thing held. PCs got cheap; mainframes stayed expensive. Streaming got cheap; producing prestige TV stayed expensive. This time, the cost of building the very thing venture capital has been funding for 30 years — software — is the thing that’s cratering.
The per-seat model is breaking. SaaS economics were built on a beautiful flywheel: more employees at the customer means more seats, more seats means more revenue, revenue grows with the customer’s headcount. But if AI agents do the work humans used to do, the customer’s headcount shrinks — and their software subscriptions shrink proportionally. Morgan Stanley’s “trinity of fears” identified this: if software lets a company cut staff by half, it cuts software subscriptions by half too.
The revenue model and the value proposition are in direct conflict.
Some companies are adapting. Sierra, Bret Taylor’s AI customer service company, reportedly hit $100 million ARR in under two years using outcome-based pricing — you pay for results, not seats. But this isn’t a minor pricing adjustment. It’s a fundamental restructuring of how software value is captured, and most existing SaaS companies have their entire go-to-market, sales compensation, and financial projections built on the old model.
The foundation model layer is commoditizing faster than anyone expected. Sequoia flagged this in their late-2024 memo: the “one model to rule them all” hypothesis was wrong. GPT-4-level token pricing has dropped over 90% in under two years (compare OpenAI’s pricing page from March 2023 to today’s frontier model rates). The model layer is a knife fight between Microsoft/OpenAI, AWS/Anthropic, Meta’s open-source play, and Google DeepMind. Foundation models largely failed to become application-layer products. The value isn’t in the model. It’s in what you build on top of it.
Here’s where it gets interesting. Sinofsky argues that “AI changes what we build and who builds it, but not how much needs to be built.” Demis Hassabis has suggested the shakeout will take a decade or more before anything like “radical abundance” materializes. An F-Prime investor told TechCrunch: “This may be the first time in history that the terminal value of software is being fundamentally questioned.”
Sinofsky is probably right about volume. Hassabis is probably right about timescale. And F-Prime has identified the crux: it’s not that software stops existing. It’s that software stops being a business in the way venture capital has understood that word for three decades.
If more software than ever gets built (Sinofsky), but building it costs nothing (observable), and per-seat monetization is broken (observable), and the shakeout takes a decade-plus (Hassabis) — then: the volume of software will explode while the venture-backable economics of software implode.
More software. Fewer software businesses worth funding.
That’s not “software is dead.” It’s more precise and, for VCs, more unsettling: software is becoming a commodity input rather than a venture-scale opportunity. Like bandwidth. Like cloud compute. Essential, ubiquitous, and not where the margin lives.
Where the Logic Leads
If the venture-scale margin isn’t in software, where is it?
AI makes software nearly free to produce. The bottleneck — the scarce thing, and therefore the valuable thing — shifts to whatever AI and software can’t easily replicate. Physical systems. Regulatory moats. Domain expertise forged through years of real-world operations. Supply chains. Manufacturing capability. Anything where the constraint isn’t “can we write the code?” but “can we move atoms, navigate regulations, or accumulate irreplaceable operating knowledge?”
Not a novel observation. But the reasoning that gets you here matters, because it reveals the scale.
Every previous wave of information technology ultimately drove demand for physical things. The PC revolution created demand for monitors, printers, cables, data centers. The internet created demand for fiber optics, cell towers, server farms, delivery logistics. Mobile created demand for rare earth mining, precision glass manufacturing, semiconductor fabs. Each wave of bits eventually crashed into the shore of atoms, and when it did, the companies at that interface captured enormous value.
The AI wave will be no different, except the interface is broader. When AI can write software but can’t machine a turbine blade or build a semiconductor fab or navigate FDA approval, capital should flow toward companies that combine AI with physical capability. Not AI or atoms — AI applied to atoms.
But here’s what this thesis gets wrong when stated too simply: not all “atoms” businesses are venture-backable, and pretending otherwise is how you lose a billion dollars.
The graveyard is real. Rethink Robotics — Rod Brooks’ company, with a pedigree that should’ve been unkillable — shut down in 2018. Veo Robotics: fire-sold for pennies on the dollar. Sight Machine raised north of $100 million and flatlined. The “AI + physical world” thesis has been tried repeatedly, particularly in manufacturing, and the default outcome has been failure.
Why? Three reasons that the current hype wave conveniently ignores. First, selling to manufacturers is brutal — long sales cycles, conservative buyers, fragmented markets, no virality. Second, manufacturing data is heterogeneous garbage: every shop floor is different, the data is owned by the customer, and building training datasets requires site-by-site integration that doesn’t scale like scraping the internet does. Third, the economics were always fighting gravity — manufacturing technology companies historically trade at 2-4x revenue, not the 10-15x SaaS commands, because the margins are lower and the capital requirements are higher.
What’s structurally different now? The foundation model layer. Previous “AI + atoms” companies had to build their own ML from scratch, which meant each vertical application required years of R&D before producing a useful product. Today’s foundation models provide a massive head start — the reasoning, vision, and language capabilities come pre-built, and the application layer can focus on domain integration rather than reinventing perception. That doesn’t eliminate the sales cycle problem or the data problem. But it compresses the R&D timeline from “five years to first value” to “six months to first value,” and that changes the venture math.
The second shift: orchestration infrastructure. Multi-agent systems can now ingest unstructured data and navigate undefined workflows — previous enterprise AI required clean data and months of integration. That changes the addressable market from “companies with clean data” to “everyone.” Sequoia calls these “domain-specific cognitive architectures.” The hardest, most defensible ones touch physical reality, because they require operational knowledge, regulatory compliance, and domain-specific data that can’t be scraped from the internet.
But you have to disaggregate. Lumping manufacturing, biotech, defense, and energy into one “atoms” bucket is the kind of lazy pattern-matching the essay’s first section warns against. These sectors have radically different market structures:
Defense and aerospace have genuine winner-take-all dynamics — classified clearances, program lock-in, multi-year contracts, and switching costs measured in decades. Once you’re on a platform, you’re on it. Anduril is the obvious example of venture economics working here, because the market structure rewards scale and lock-in.
Manufacturing technology mostly lacks network effects. Each customer deployment is bespoke. The winners will be companies that crack horizontal applicability across shop types — but that’s been the grail for two decades and no one’s found it yet. The venture-favorable play is software-layer manufacturing (AI that makes existing machines smarter) rather than vertically-integrated factories, because the former has software-like margins applied to an atoms problem.
Energy and climate have regulatory moats and massive capital requirements. These are project-finance businesses dressed up as startups. Some will produce huge outcomes; most require patient capital structures that don’t fit a 10-year fund.
Biotech has always been its own animal — binary outcomes, regulatory gates, and a well-established (if brutal) venture model. AI is genuinely accelerating drug discovery timelines, but the FDA approval bottleneck hasn’t changed.
The point: “go invest in atoms” isn’t an investment thesis. It’s a sector rotation. An investment thesis would say which specific market structures within the atoms world produce venture-favorable dynamics, and why AI changes the historical math.
The Complement
There’s a pattern in technology commoditization that’s been true for thirty years and that most VCs are about to re-learn the hard way.
When bandwidth became free in the late ’90s, telecom investors lost over $2 trillion. WorldCom went from $150 billion to $150 million. The fiber they overpaid to lay became Google’s free ride. Sequoia turned $12.5 million into $30 billion on the complement — applications that consumed the bandwidth someone else had bled money to build. When compute became free through cloud, Sun Microsystems got absorbed into Oracle for a fraction of its dot-com peak while Sutter Hill turned $200 million in Snowflake into $12.6 billion at IPO. Same pattern, every time.
The commodity layer gets destroyed. Not “returns are modest.” Destroyed. And the complement layer — the thing that becomes more valuable because the commodity is cheap — captures value at 10 to 100 times the scale.
Genome sequencing is the live analog. Illumina won the commodity race, drove costs from $100 million per genome to under $100, and has lost 75% of its market cap since peak. 23andMe — consumer genomics, the obvious complement bet — went bankrupt. The real complement, therapeutics and precision medicine, hasn’t produced its Google yet. The obvious bet is almost never the right one.
AI inference is next. The commodity — cost per token — has dropped over 90% in under two years. The firms stacking chips in the commodity layer (foundation model API wrappers, undifferentiated AI tooling) are the 2025 equivalent of the companies laying redundant fiber in 1999. The platform players — OpenAI, Anthropic, the hyperscalers — are making the AWS play: capturing value between the commodity and the application. Some of them will succeed. Most of the value will still accrue above them.
The question is what the complement of near-free intelligence looks like. And the data, once you pull it apart sector by sector, points somewhere specific.
AI collapses the knowledge bottleneck. Every time. In every sector. The constraint then migrates to physical execution.
In manufacturing, AI is making CNC programming nearly free — Siemens NX, Autodesk Fusion, and a generation of AI-CAM tools are compressing what took a senior machinist hours into minutes. The new binding constraint: there aren’t enough machines, materials, or bodies to run them. The U.S. has 2.1 million manufacturing jobs it can’t fill, projected through 2030, and the number grows as AI removes the programming bottleneck because you can now program more work than you can physically produce.
In defense, Anduril’s $1 billion Ohio factory is the thesis made concrete. AI handles sensor fusion, autonomous targeting, battlefield awareness. The constraint is manufacturing autonomous systems at scale — drones, munitions, naval vessels. Ukraine demonstrated that the West’s munitions production capacity is structurally inadequate for a peer conflict. Software-defined hardware, produced at volume, is the binding constraint.
The same pattern repeats in biotech (AI accelerates drug discovery; FDA trials and manufacturing capacity don’t move), energy (hundreds of gigawatts waiting years for grid connection that AI can’t build), and construction (500,000 workers short for data centers alone).
The pattern is the same in every case: AI makes the thinking free and the doing becomes the bottleneck. The venture opportunity isn’t in the thinking layer — that’s commoditizing. And it isn’t in the doing layer alone — pure physical capacity is capital-intensive with low margins. The opportunity is at the interface: companies that translate AI’s infinite digital output into physical-world execution, with network effects at the boundary.
But the VC portfolio data reveals how few firms have actually repositioned.
Sequoia — the firm that literally published the memo on AI as the defining platform shift — is backing OpenAI, Anthropic, and xAI simultaneously. Based on their publicly disclosed portfolio, there’s virtually no atoms exposure — no defense, no manufacturing, no energy. The visible portfolio is overwhelmingly software and AI infrastructure. They appear to have made an explicit bet that AI software captures more value than AI applied to physical industries.
Then look at Founders Fund. Thiel has been saying “we wanted flying cars” for a decade. And the portfolio matches: led Anduril’s $2.5 billion round, co-led Hadrian’s $260 million for defense manufacturing, backs Varda for space manufacturing, SpaceX, Radiant for nuclear microreactors. Now raising a $6 billion growth fund — their largest ever. Lux Capital ($7 billion AUM, seed investor in Anduril when nobody would touch defense) and Eclipse (ninety percent physical-industry focus) are similar: stated thesis and revealed portfolio actually align. They’re the exceptions.
The category that’s actually growing fastest isn’t “atoms” — it’s AI+atoms. Anduril at $84 billion secondary valuation. Shield AI at $5 billion-plus. Applied Intuition at $15 billion for autonomous vehicle simulation. Figure AI for humanoid robots. Helsing, the European defense AI company, raised €600 million at a €12 billion valuation — from Spotify founder Daniel Ek, no less. Defense tech alone is now a $100 billion-plus private market. Anduril’s secondary valuation exceeds L3Harris’s market cap.
These aren’t atoms companies. They aren’t software companies. They’re companies that use AI to solve physical-world problems at software-like margins, with lock-in dynamics that pure SaaS can only dream of. Once Anduril’s Lattice is the operating system for a theater of operations, the switching costs aren’t measured in contract terms. They’re measured in lives.
That’s the complement. Not chatbots. Not coding assistants. Not another SaaS vertical with an AI copilot bolted on. AI applied to the physical world — where the bottleneck just moved, where the scarcity lives, and where the market structures reward the kind of lock-in that produces venture-scale returns.
The non-obvious complement wins every time. That’s what the thirty years of data show. And the non-obvious complement of near-free intelligence isn’t more intelligence. It’s the physical world that intelligence alone can’t touch.
The Honest Answer
So: WTF should VCs do now?
Software isn’t dying. It’s becoming air. Everywhere, essential, and not something you can build a toll booth on. The volume of software will explode; the number of venture-backable software businesses will contract. That’s not a cycle. It’s a phase transition. And the VCs treating it as a cycle — “fears are only temporary,” as most VCs told TechCrunch — are the ones the Kauffman study will cite in its next edition.
Capital should flow to AI applied to physical systems — but not indiscriminately. The graveyard of Rethink Robotics and Veo Robotics is a warning. The companies that will capture this transition bring software-like margins to atoms-world problems, using AI to compress the timeline from “decade of R&D” to “year of integration.” Defense companies with platform lock-in. Manufacturing software that makes existing infrastructure intelligent. Biotech where AI shortens the discovery cycle inside regulatory moats.
Follow the atoms all the way down and you find specific interfaces where AI has already collapsed the knowledge bottleneck and the constraint has migrated to physical execution. Manufacturing data platforms that accumulate machining knowledge across thousands of shops — network effects that compound with every part programmed. Grid interconnection software that navigates the 2,300-gigawatt queue bottleneck as data center demand outpaces grid expansion. Clinical trial acceleration tools — adaptive design, AI patient matching, synthetic control arms — that capture value from every AI-accelerated drug discovery pipeline upstream. These are venture-native: software margins, regulatory lock-in, compounding data moats. The adjacent opportunities in modular construction and autonomous systems manufacturing are real but capital-intensive — closer to growth equity than seed-stage venture, and the firms funding them need fund structures to match.
The time horizon matters enormously. If Hassabis is right about a decade-plus, and Sinofsky’s historical parallels suggest even longer, then this transition is an investing opportunity, not a reason to freeze. The best returns in any regime change go to the people who reallocate early — before the consensus reprices the opportunity away. But “early” in a 15-year transition means an LP committing to a 10-year fund focused on atoms may not see returns in-life. The fund structure itself may need to change.
And the herd? The herd is doing what it always does. FOBO — fear of becoming obsolete — is driving a huge amount of current VC behavior. Firms are piling into AI deals not because the analysis supports it but because not being in AI feels existentially dangerous. That’s not investing. That’s panic buying. And panic buying — whether it’s toilet paper in 2020 or AI startups in 2026 — never generates good returns.
Here’s what would prove this thesis wrong. The strongest counterargument is the Jevons Paradox: when the cost of a resource collapses, total consumption can explode so much that total spending actually increases. It happened in music — recorded music revenue recovered to over $17 billion by 2023, higher than the 1999 peak, despite per-unit prices collapsing to near zero. More music, more revenue, different companies. If total software spend increases even as per-unit costs crater — because near-free software means vastly more software gets built — then the thesis breaks. SaaS doesn’t die; it multiplies. The other falsifiers: if per-seat pricing stabilizes because AI augments workers rather than replacing them — headcount doesn’t shrink, output per head just grows. If foundation model commoditization reverses and one player achieves durable monopoly pricing. If AI makes designing and manufacturing physical systems just as cheap as it makes software, eliminating the scarcity advantage of atoms. Any of those would break the logic chain. They’re not impossible. They’re just not what the data currently shows.
The firms that will be on the right side of the power law are the ones reallocating now — not to “atoms” as a vague category, but to specific sectors with venture-favorable dynamics where AI changes the historical math. The ones still chasing the last wave will be in the next Kauffman dataset too.
Just on the other side.