Topics Pioneers Insight Method Author
Back

What Happens When the AI Boom Slows Down

·Invest Like the Best

Rendering the long image…

Topics Battle for Frontier AI Moats Trillion-Dollar AI Capex and Bubble Risks SaaS Valuation Reset Under AI Disruption AI Venture Funding and Bubble Risks Coding Agents and Future Software Engineering Autonomous Agents Reshaping Enterprise SaaS
Pioneers Martin Casado

Summary

  • Frontier AI has pricing power, but model leadership may be unusually perishable. Casado says the best models have captured 80%+ of dollar-weighted demand over three years, and estimates an individual model’s commercial lifespan at three to six months. Continued market growth favors the frontier; saturation could send value toward open source, and “nobody knows” how that transition would work.
  • Saturated coding workflows may slow token growth without making cheaper models competitive. Automating 95% of tasks does not establish that a business can accept less intelligence: competing startups “would always pay 10 times more for the smartest model,” Casado argues. Growth would then require higher prices or broader adoption beyond coding.
  • AI makes capital productive at a new scale, while financing and physical constraints favor large operators. Twenty people can now use $1 billion productively, although Casado does not know whether that produces $500 million or $2 billion. He sees debt becoming harder to obtain for smaller providers, while chips, power, and materials remain the binding constraints on physical expansion.
  • Established infrastructure businesses may offer a different risk profile from frontier AI. Casado describes profitable companies growing 40–60% at $300–500 million ARR that struggle to access public markets. His argument is operational: “more software means you need more infra,” and reproducing a database’s features, performance, and reliability remains difficult.
  • Casado accepts competitive market prices and concentrates diligence on technology and risk. He believes private markets are efficient, with his work identifying a less risky asset rather than a bargain. Applying mature revenue metrics too early can exclude disruptive technology because “business models always lag disruptive technologies.”
  • The Cursor deal connected compute supply with users, data, and model-building capability. Casado says that, if Cursor built models, it would need more funding and compute; he guesses Mythos may have cost roughly $3 billion to make. “SpaceX had the compute,” while Cursor supplied complementary assets. His broader M&A lesson is that trust and strategic alignment create deals; leaks can destabilize them.
  • Agents reopen infrastructure design, but Casado’s confidence stops at the digital world. Removing human-use constraints could enable different databases, development tools, and execution environments. He believes sufficiently funded reinforcement learning can address anything done on a computer; whether robotics, science, and medicine follow the same scaling laws remains unresolved.

Deep dive

1. Disruptive technology requires a coherent investing framework

  • After a decade and hundreds of deals, Casado distinguishes four mindsets. Financial investors use markets and financial metrics as proxies for business health. Productive investors underwrite a value-creating asset, asking how its technical approach will evolve, create a market, and eventually capture the value it produces.

  • Frontier investors trust exceptional founders pursuing something outsiders cannot yet evaluate. Casado describes this as the strong form of productive investing: “YOLO” trust that founders may know better than VCs. He recalls Chris Dixon investing in nootropics and names Soylent, Oculus, Coinbase, and crypto as science-fiction-like bets. Speculative investors instead time prices, sometimes knowing little about the underlying asset. Mixing these approaches lets investors “justify whatever personal bias you have.”

  • His infrastructure diligence examines leading companies’ internal systems, adjacent budgets, customer interest, and willingness to pay before a complete business model exists: “partial indicators, not full indicators.” Technical expertise helps assess how a product fits platforms, spending, and defensibility; it does not mean knowing the technology better than its founder.

  • “Business models always lag disruptive technologies,” Casado insists. The internet’s monetization was initially unclear, and shifting from perpetual licenses to recurring revenue took years. Each investing approach can be legitimate, but claiming to underwrite disruption while demanding established TAM and strong NRR creates an incoherent selection process.

2. A moat’s shape matters more than its mechanism

  • O’Shaughnessy offers the familiar definition of network effects: a product improves with each marginal user. Casado finds that almost tautological—users also report bugs and provide feedback. It says little about defensibility. The useful distinction separates the mechanism creating an advantage from how that advantage grows with scale.

  • His test is competitive: give a leader and challenger $5 million each. A superlinear advantage lets the larger company grow faster from the same investment; a sublinear advantage lets the challenger catch up because the leader must spend progressively more. Competitors’ capital and growth also affect the outcome. Casado says classic telephone, communications, and transit networks can gain connections as users or destinations join, so interactions—and value—grow faster than linearly.

  • Genuine interaction-driven network effects seem rare to him; he names telcos, LinkedIn, social networking, and GitHub as examples. These durable, high-value companies are built around interconnections. A moat can also come from brand, scale, or a two-sided marketplace. A scale effect can improve unit economics through pricing power or lower CAC; a two-sided marketplace is self-reinforcing supply and demand.

  • Uber looks more like a marketplace and funding-scale advantage: drivers and passengers can use competing services. A self-reinforcing flywheel alone does not establish that its benefits grow superlinearly.

  • More provocatively, Casado thinks moat debates usually receive too much attention. Growing markets create white space that companies often prefer to pursue rather than compete over; shrinking markets often consolidate anyway. O’Shaughnessy asks whether an Anthropic S-1 buyer needs stronger defensibility analysis. Casado’s response remains conditional: if penetration is only 0.1%, substantial value creation could still lie ahead, and liquid investors can move in and out.

3. Agents reopen the infrastructure stack

  • Casado defines infrastructure by technical differentiation, organized around compute, networking, storage, and databases. Models also count as infrastructure, resembling ASICs more than conventional software in capital requirements and competitive pressure. He says they are disrupting the application stack above them, though it is not yet clear whether they will layer. He calls this “the biggest refresh of computer science infrastructure we’ve seen maybe since the internet, maybe ever.”

  • He separates software produced by agents from systems with AI at their core. The former still needs ordinary databases and compute. The latter changes design requirements: infrastructure previously simplified for human ergonomics can become more complicated if agents handle it, potentially changing trade-offs around safety and reliability.

  • OLTP databases illustrate the opportunity: new designs need not overcome the same programmer-retraining barrier if agents already know how to use them. Networking’s large clusters and rigid performance requirements are reviving entrants, optical technologies, and materials changes after years of stagnation. Compute is reconsidering whether agents belong in containers, VMs, or durable workflows such as Temporal.

4. Frontier leadership pays, but individual models decay quickly

  • Across three years of rapid growth, Casado says the best frontier models have captured 80%+ of dollar-weighted demand. Being even “an epsilon better” provides tremendous pricing power. He cites estimates of 3% enterprise penetration but roughly 1% or less of budgets, leaving considerable expansion possible.

  • The overlooked risk is continuity. Casado estimates a model’s useful commercial lifespan at three to six months. If Anthropic failed to release another model for six months—or a technical, regulatory, or security problem caused a three-month slip—it could lose the entire market. Conventional software’s retained-revenue tail is a poor template.

  • Continued growth gives him a familiar answer: “the frontier wins.” Saturation produces a question he cannot resolve. If use cases stop needing additional intelligence or demand slows for another reason, there’s “a very real world where the value comes crashing back down into open source.” He does not claim to know the outcome.

5. Coding saturation does not settle demand for intelligence

  • O’Shaughnessy’s pushback is concrete: people describe a high share of coding tasks as saturated, yet still spend aggressively on frontier tokens. Casado separates task coverage from model substitution. Automating 95% may limit additional token consumption, but competing startups still want better software faster: “You would always pay 10 times more for the smartest model.”

  • That leaves providers needing higher prices or broader use beyond coding. Casado expects a broadening into general knowledge work. His own Grok bot handles personal finances and modular synthesizer questions; his newest project asks it to become the “living AI in my house.” It bought a few hundred dollars of obscure home-automation equipment, some unfamiliar even to him, toward a vintage Star Trek vision.

  • World Labs’ Atlas, which he describes as generating dynamic 3D scenes, prompts a different possibility: entering Spirited Away or experiencing an immersive production like Sleep No More in VR. These are visions of what could be built. His framing is that “the biggest constraint is no longer the tech”—it is creativity constrained by expectations formed around frustrating software.

6. Digital confidence does not extend automatically to the physical world

  • Casado’s strongest technical claim is categorical: “Anything you do with a computer, AI will be able to do.” Digital activities can supply reinforcement-learning environments, whether code, Blender, or general computer use. Some capabilities may not work today, but he sees a training path that sufficient capital can fund.

  • He explicitly withholds the same confidence from robotics, science, and medicine. Physical experiments can be constructed, but he does not know whether their scaling curves or distributions resemble digital environments. “I have no idea” whether the same laws apply.

  • On security, he recalls Windows 95 machines becoming infected before installation finished. New computing paradigms bring years of mayhem before defenses mature. If defenders can outspend attackers, as he says is typically true, he expects technical solutions. Exposed vulnerabilities can represent progress, and he thinks AI’s defensive role ultimately makes software “much more secure.”

7. Durable infrastructure businesses are stranded between market categories

  • O’Shaughnessy observes that AI’s apparent returns make previously attractive software investments harder to fund. Casado sits on boards of profitable companies growing 40–60% at $300–500 million ARR that remain too small to IPO. He would welcome personal exposure to a diversified basket: nobody knows what AI businesses look like after growth slows.

  • The host challenges their disruption exposure. Casado answers from infrastructure: “more software means you need more infra.” Using Snowflake as an example, he says that if growth is 30–40% with good economics, AI does not look like its principal disruption threat; Databricks might. Replicating long-tail features and performance, and operating these systems, is difficult. SPAC approaches suggest financial markets are seeking ways to package these assets.

8. AI changes the relationship between money and engineering

  • The Mythical Man-Month captures the old constraint: adding engineers could slow delivery through coordination overhead. Giving five founders $1 billion might once have left money idle or encouraged destructive hiring. Now 20 people can use that sum productively. Casado does not know whether they generate $500 million or $2 billion; the ability to deploy it is the change.

  • He calls Anthropic and OpenAI mechanisms for “turning money into growth,” observing that their aggregate fundraising exceeds the downstream ecosystem’s by a large amount. This makes capital allocation more directly connected to technical progress than in his own decade operating a startup, when engineering capacity resisted scaling.

  • Debt is a separate concern, and he stresses that he is not a debt specialist. On the ground, financing is harder to obtain and likely to favor investment-grade companies. Smaller providers may need prepayment, requiring their customers to prepay in turn. Available cash therefore increasingly influences who can secure infrastructure.

9. Neoclouds face physical bottlenecks and a fragmented remainder

  • Casado defines a neocloud as physical infrastructure ending in tokens: GPUs, fine-tuning, or inference services. All want to become token factories and verticalize. A neolab instead seeks new model architectures and training breakthroughs; owning GPUs does not make infrastructure its product. Some effectively propose that solving ASI will make the business model straightforward.

  • At physical buildout, he identifies hardware supply, power, and materials as the actual limiting factors. Under those constraints, he expects 80% to go to scaled providers and 20% to niches or hard-to-reach opportunities. His discussion of financing does not eliminate these supply constraints.

  • The fragmented remainder can still support many companies. His example is a GPU marketplace sourcing capacity from converted chicken sheds on farms. A provider such as SpaceX is unlikely to pursue every such pocket; the overall market is large enough for those awkward opportunities to matter.

  • O’Shaughnessy recounts Larry Page predicting “five nines” of compute could eventually be orbital. Casado says space infrastructure pencils out, but emphasizes regulation: he cites 500 proposed US measures restricting data-center construction, including in Texas. If projects cannot break ground, builders must look internationally or into space.

10. Efficient prices and an expanding private market can coexist

  • O’Shaughnessy presses the valuation objection: even a valuable disruptive asset can ruin returns at the wrong price. Casado’s answer is unusually direct—he believes private markets are efficient and accepts competitive prices. His diligence seeks lower risk for himself: “I’m not trying to find a better price, I’m just trying to find a less risky asset.”

  • The host asks why that does not imply indexing. Casado would happily own 5% of every competitive deal backed by legitimate investors, but competition, capacity, and board commitments prevent it. Instead, his team selects the strongest company in each credible sector. O’Shaughnessy recognizes the approach as “stratified sampling.”

  • Casado rejects the identity of the lone investor uncovering bargains everyone else misses. His successful portfolio companies were generally highly competitive investments. Going against consensus can also jeopardize follow-on capital. He sees investor ego in the belief that superior “truffle hunt” abilities are the essential source of returns.

  • He also rejects the claim that only 20 startups matter and additional venture funding merely overfunds a fixed opportunity set. Capital can expand investable TAM, support larger private companies, and let businesses remain private longer. Especially with AI, his premise is that capital enables technology to create additional value.

11. Strategic acquisitions buy market options through trust

  • Casado frames strategic M&A as buying an option on a large market. His $1.26 billion Nicira sale to VMware illustrates roughly a $1 billion option on a $40 billion market. The proof required both a base case—selling into networking budgets—and an inductive step: attaching the product to VMware’s virtual machines could expand distribution.

  • These discussions resemble high-stakes enterprise sales more than spreadsheet-driven asset auctions. Founders and strategically knowledgeable investors must explain the combined opportunity. Deep personal alignment frequently matters; shorter technology timelines could produce more strategic exits. A banker can help, but Casado dismisses the idea of outplaying sophisticated acquirers through poker-like tactics.

  • Cursor was preparing to raise a few billion dollars when Mythos came out. Casado guesses it may have cost roughly $3 billion to make; if Cursor built models, he says, it would need more funding and compute. “SpaceX had the compute,” while Cursor had users, data, and a team that had built models successfully. The available financing also helped establish valuation.

  • His lesson from Cursor, World Labs, and OpenRouter is good-faith engagement around an understood value. World Labs did not leak because only a few people knew; OpenRouter leaked repeatedly. Efforts to manufacture competing bids through press leaks can destroy trust: “That destabilizes deals.” He has seen large transactions collapse after investors encouraged such tactics.

12. Investors need stage-specific skills and better listening

  • Private companies now rival public companies of earlier decades, making traditional finance expertise increasingly relevant. Casado objects when those intuitions reach too far downstage. He recalls PE investors on Series B boards reasoning primarily through burn, comparing the mismatch to putting him on a mining-company board. Productive early-stage VC requires a different toolkit.

  • His own improvement is learning that “the primary job is just to listen.” Founder meetings often become performative questioning or attempts by former operators to fix a business. The task is to understand what founders believe and ask questions that change the investment decision.

  • Casado likens the shift to an Industrial Revolution-level change from a zero-sum to a non-zero-sum world of newly created value. Digital problem-solving is becoming a capital-resourcing problem, putting allocators in a more direct position to drive innovation.