"if you spent the last 30 years making money in software is software dead feels kind of like a personal attack." - Rory O'Driscoll [01:42]
"we are spending in AI capex just among the hyperscalers $688 billion this year... what's coming out the other side is about 110 billion in actual revenues." - Rory O'Driscoll [05:35]
Disclaimer: Orignal content owned by or sourced from third parties. It does not represent the views of 'Nuggets' platform or it's team. AI is used extensively across this platform including for summaries. Accuracy is not guaranteed, there can be mistakes. Any info or content on this platform is not a financial, legal, or investment advice. Do your own research. Refer for complete disclosures:- Terms of Use · Full Disclaimer
"When things are in invest mode when you're spending more money than you're taking in at any point you could have a hiccup." - Rory O'Driscoll [07:34]
"the modern version of is software dead... is how does economic value get allocated between the foundation models and modern software companies." - Rory O'Driscoll [14:08]
"about 10% of our portfolio was literally DOA when chat GPT comes out because they'd solved an AI problem that was just so much easier to solve with ChatGPT that just wasn't fixable." - Rory O'Driscoll [23:15]
"I think that AI is the new sales and marketing but that has to be true. If it's not true and you have 700 people in sales and marketing and 50% [compute COGS], then you're right, you either have compute that sells itself or people who sells it but you can't do both." - Rory O'Driscoll [36:43]
Speakers & Credentials
Rory O'Driscoll: Partner at Scale Venture Partners. A veteran software and venture capital investor with 30+ years of experience navigating the transitions from client-server, to cloud computing, to modern AI foundation models.
Jason Lemkin: Host, Founder of SaaStr. An influential operator and investor in the B2B SaaS ecosystem.
1. Executive Summary
The foundational economics of SaaS have structurally broken, with average industry growth rates plummeting from historical highs of 30% to a stagnant sub-10% environment.
The entire technology sector is operating in a dangerously imbalanced "Invest Mode," deploying an estimated $688 billion in AI infrastructure capital expenditures to chase merely $110 billion in current ecosystem revenues.
Venture capital returns are now predicated on an existential battle for value capture: determining whether foundation models (the "apex predators") will capture all enterprise value, or if an application layer of defensible, moat-driven software can survive.
Legacy SaaS companies face immediate extinction threats; ~10% of portfolios were wiped out instantly by LLMs, forcing surviving entities to adopt "agent-first" architectures or risk becoming obsolete "plain vanilla" workflow wrappers.
Moving forward, software defensibility requires avoiding compute-heavy commodity battles; successful AI apps will either leverage physical-world sensor integration, deeply entrenched network-effect marketplaces, or pivot entirely into full-stack service businesses to insulate themselves from OpenAI and Anthropic.
2. Chronological Table of Contents
[00:00] Introduction: The Collapse of SaaS Multiples & Growth
[02:29] Expansive vs. Extinctive AI & Technological Predictability
[05:11] The Great Imbalance: AI "Invest Mode" Capex vs. Revenue
[09:08] Mapping the AI Stack: Making AI vs. Using AI
[13:22] The Trillion Dollar Question: Economic Value Allocation
[15:59] Technical & Business Moats in the Generative AI Era
[22:49] The Pre-2022 SaaS Extinction Event & Mortality Analysis
[28:36] Q&A: Pitching Agentic SaaS, T2D3 Growth, and Compute COGS
3. Detailed Thematic Summary
The Macro Collapse of SaaS and the "Invest Mode" Reality
SaaS multiples did not collapse arbitrarily; they collapsed fundamentally because average industry growth rates plummeted from robust 30% benchmarks down to sub-10% [00:16]. Public markets are heavily punishing SaaS companies lacking category conviction and high-velocity growth.
The broader tech ecosystem is heavily functioning in an AI "invest mode," a highly volatile state where capital outlay massively outstrips generated revenue [05:11].
Hyperscalers are deploying roughly $688 billion in total AI capital expenditure this year, dwarfing the roughly $110 billion in actual market revenues returning from those investments [05:35].
Within that $110B revenue pool, a staggering $89 billion (on a GAAP basis) is estimated to be captured exclusively by the two major foundation models, leaving a disproportionately tiny slice for the rest of the application ecosystem [05:54].
Market forecasts project it will take until at least 2031 or 2032 for the cumulative revenue from the foundation models to actually surpass the cumulative total capex spending [06:53].
For hyperscale investments to ultimately pay off, AI technology must directly capture 15-17% of all knowledge worker dollars in the US economy. In specialized roles like software engineering, this requires capturing north of 25% of human capital spend, equating to roughly $50,000 in token spend for every $200,000 human developer [08:30].
Predicting the Unpredictable & The Value Stack
Despite alarmist charts forecasting everything from absolute post-scarcity abundance to literal human extinction, the actual trajectory of AI technology was highly knowable through publicly documented milestones: ImageNet in 2012, Transformers in 2017, OpenAI Scaling Laws in 2020, ChatGPT in 2022, and Leopold Aschenbrenner's Situational Awareness paper in 2024 [03:54].
Using Jensen Huang's architectural model of the tech stack, the ecosystem fundamentally splits into two operational zones: "Making AI" (energy, chips, infrastructure) and "Using AI" (models and end-user applications) [09:28].
The "Making AI" layer is currently consuming over 80% of total ecosystem capital, mostly funneled into non-venture-backed infrastructure, with Nvidia serving as the primary beneficiary as global capex surged from $200B to over $600B in four years [09:53].
Foundation models operate as the "apex predators" of value distribution, creating an existential dilemma: how does the remaining economic value get allocated between monolithic AI providers and specialized venture-backed software companies like Sierra, which the transcript notes recently raised at a $15B valuation? [11:31].
Defensibility, Moats, and Surviving the LLM "Rollover"
Application architecture is rapidly converging into an AI equivalent of the old LAMP stack. While LLM-based apps will structurally look identical on the backend ("The Harness"), their true defensibility will stem from how they instantiate in verticalized physical or digital environments [17:17].
Absolute defensibility is found in physical-world integrations, specifically by pairing generative reasoning models with hardware sensors. Foundation models like OpenAI and Anthropic have no intention of shipping hardware, protecting these vertical integrations from direct competition [19:34].
Marketplaces and multi-sided network effects (such as Paraform in recruiting) remain highly defensible because their value is derived from user incentive alignment and network gravity, not just pure intelligence generation which models could mimic [19:57].
The "Full-Stack" business model offers a massive moat. Instead of selling SaaS to legacy wealth managers, AI-native companies like Range are using LLMs to become the wealth manager themselves, effectively bypassing the software layer to compete directly in the services economy [21:00].
The SaaS Extinction Event & Business Operating Metrics
The release of ChatGPT triggered an immediate extinction event in venture portfolios. Approximately 10% of Scale Venture Partners' existing portfolio was rendered literally dead on arrival (DOA), as highly complex, $30M R&D machine learning solutions were instantly commoditized by a $5-per-million-token generalized prompt [23:15].
The surviving 90% of the pre-2022 SaaS ecosystem fractured into distinct thirds: 1/3 insulated (orthogonal data networks unaffected by text generation), 1/3 additive (where generative AI perfectly enhances the core workflow, like DroneDeploy's mapping tools), and 1/3 threatened (plain vanilla SaaS wrappers facing imminent irrelevance) [23:34].
Compute costs (COGS) vary wildly based on defensibility. Base-layer foundation models allocate 70-80% of revenue to compute; coding copilot companies run at 50-60%; but highly defensible, moat-driven application layer companies manage to spend only ~10% of revenue on LLMs, capturing 90% as value-add margin [35:05].
Traditional B2B SaaS companies adding AI features must target 8-20% for compute COGS. If an app runs highly compute-intensive features, the AI product must essentially act as the new sales and marketing engine—"selling itself"—because unit economics structurally cannot support both 50% compute costs and a legacy 700-person enterprise sales team [36:22].
The Reference Vault
4. Data & Figures
Data Point
Value
Context
Timestamp
Historical Avg SaaS Growth
30%
The historical baseline growth rate that justified high SaaS venture multiples over the past decade.
The "Invest Mode" Risk Curve [07:34]
Operating in "invest mode" dictates a market reality where capital deployment intentionally and massively eclipses near-term revenue generation. The strategic irony of this framework is that while long-term technological dominance demands staggering infrastructure spend ($688B), it creates acute short-term fragility. A sudden macroeconomic shock or investor hesitation can trigger an immediate ecosystem collapse if capital markets decide to "slow down" before the technology bridges the $1 Trillion revenue gap required by 2032.
Jensen's AI Value Stack (Making vs. Using AI) [09:28]
Sourced from Jensen Huang's mental model, this framework cleanly bifurcates the generative AI ecosystem into two distinct operational layers. The bottom tier ("Making AI") consists of energy grids, hardware chips, and foundational infrastructure—absorbing over 80% of all capital but offering little room for traditional software venture capital. The top tier ("Using AI") contains the foundation models and end-user applications. Understanding this stack forces investors to recognize that most of the capital is flowing into industrial hardware, not software, entirely shifting how B2B moats must be underwritten.
The LLM LAMP Stack ("The Harness") [17:17]
Historically, the LAMP stack (Linux, Apache, MySQL, PHP) standardized the backend of Web 2.0 SaaS applications; similarly, "The Harness" will standardize the integration of LLMs. Applications will soon utilize identical architectural frameworks for context injection, action execution, and logging. The strategic implication is that software wrappers are inherently commoditized. Therefore, differentiation will not come from unique engineering architecture, but entirely from the specific, verticalized business context in which the "Harness" is deployed.
The Full-Stack Services Wedge [21:00]
Traditional SaaS operates by selling horizontal workflow tools to existing service providers (e.g., selling accounting software to an accountant). The generative AI era flips this paradigm via the Full-Stack Services Wedge. Instead of empowering legacy professionals, modern AI startups use models to entirely bypass the middleman and become the service provider (e.g., operating directly as an AI-driven wealth management firm). This acts as an ultimate defensive moat because generalist foundation models like OpenAI are fundamentally unwilling and legally unequipped to vertically integrate into specialized professional services.
The "AI as Sales & Marketing" Tradeoff [36:43]
This financial modeling framework exposes a brutal reality regarding SaaS unit economics. If an AI product demands high compute intensity (costing upwards of 50-60% of revenue in LLM tokens), traditional SaaS scaling math breaks. A company cannot simultaneously float massive operational compute costs and maintain a bloated, 700-person enterprise sales team. Consequently, heavy AI compute is only justifiable if the product experience is so exceptionally frictionless that it acts as its own GTM engine, effectively replacing human sales and marketing overhead with sheer product velocity.
6. Anecdotes
The Predictability of the Extinction Event [03:54]
O'Driscoll highlights a timeline of public research—from ImageNet in 2012, Transformers in 2017, and Scaling Laws in 2020—to brutally undercut the narrative that the AI explosion was an "unpredictable black swan." He tells this story to chide investors and operators who feign shock at the current market turbulence; the data demonstrating that infinite compute yields infinite capability was openly published for a decade, proving that those facing extinction simply weren't paying attention.
The $30M Feature vs. the $5 Token Model [23:15]
To explicitly define the carnage within venture portfolios, O'Driscoll recounts how ~10% of his firm's investments were immediately labeled literally dead on arrival post-ChatGPT. He tells the story of companies that had painstakingly raised and spent $30 million developing highly complex, proprietary natural language models, only to watch a user achieve the exact same functionality via a $5-per-million-token API call to OpenAI. The anecdote serves as a terrifying warning against building "shallow" technology moats that a foundation model can absorb in a single update.
Revisiting the Silicon Valley Hardware Eras [26:55]
O'Driscoll nostalgically compares today's capital-intensive AI ecosystem to his early days investing 30 years ago when deep tech products were heavily hardware-focused, requiring massive manufacturing bets in places like Milpitas and Santa Clara. He uses this historical parallel to remind modern SaaS founders—who are used to asset-light, purely digital margins—that technology has cyclically returned to a phase of massive, heavy-industrial capital intensity, fundamentally changing the DNA of early-stage venture capital.
Intercom's Open-Heart Surgery [29:54]
In response to an audience member asking if legacy SaaS can successfully pivot, O'Driscoll references his investment in Intercom and their aggressive development of an AI customer support agent called "Finn." He tells this story to prove that while pivot survival is theoretically possible, it requires what he terms "open heart surgery." It demands abandoning safe legacy subscription models, radically restructuring architecture toward agentic outcomes, and adopting usage-based pricing—a transition so severe that founders must be willing to risk destroying their current business to save it.
7. References & Recommendations
People
Jensen Huang: CEO of Nvidia. Referenced for his architectural mental model of the AI value stack (Making AI vs Using AI), which dictates how trillions in capital are currently flowing. [09:28]
Leopold Aschenbrenner: Author of the "Situational Awareness" paper. Cited as a definitive blueprint predicting the exact trajectory of AGI and the necessity of massive infrastructure investments. [04:45]
Maynard Keynes: Economist. Paraphrased ("I'd rather be roughly right than exactly wrong") to justify using directional, macro estimates for the massive $688B capex figures without getting bogged down in granular accounting semantics. [05:28]
Harry Stebbings: Investor and podcast host. Mentioned by Jason Lemkin as having scoffed at the traditional T2D3 venture model in a prior episode. [32:47]
Companies & Products
OpenAI & Anthropic: The dominant, venture-backed "foundation models" referred to repeatedly as the "apex predators" destined to capture the vast majority of the $110B industry revenue. [11:31]
Sierra: A modern application-layer AI company mentioned as raising at a $15 billion valuation, proving that massive enterprise value can still be created outside of foundation models. [14:34]
Paraform: A recruiting marketplace company in Scale's portfolio used as an example of a defensible "network effect" business that AI cannot easily commoditize. [19:57]
Intercom (Finn): Legacy SaaS company cited for executing a successful but brutal "open heart surgery" turnaround to build an agent-first customer support paradigm. [29:54]
DroneDeploy: A portfolio company utilizing AI in an "additive" manner, proving that AI can safely enhance core horizontal workflow software rather than destroy it. [24:03]
Range: An AI-native wealth management company referenced to show the power of full-stack delivery models. [21:00]
Cerebras: Hardware acceleration venture architecture cited as a major upcoming exit victory in the VC space. [09:58]
Safe Superintelligence (SSI) / Thinking Investments: Next-generation foundational architectures built to circumvent core LLM model ceilings. [25:40]
Eleven Labs: Cited as a highly specific, capital-efficient speech foundational layer modeling company that successfully built a massive enterprise software workflow business. [26:02]
Conceptual & Historical Frameworks
LAMP Stack: The historical Linux, Apache, MySQL, PHP framework used to illustrate how AI application infrastructure will standardize, pushing defensibility to the application edge. [17:17]
T2D3 Model: The legacy SaaS growth benchmark (Triple, Triple, Double, Double, Double) questioned by the audience regarding whether it is still valid in an era of compressed multiples. [30:27]
ImageNet (2012) & Transformers (2017): Foundational deep learning milestones cited as proof that the AI explosion was mathematically predictable years before ChatGPT's consumer launch. [03:54]
Jul 18, 2026
China’s EV makers are already reshaping global auto markets | 17 Jul 2026 | Strategic Alternatives Podcast
1. Executive Briefing TL;DR The Core Thesis: China’s automotive original equipment manufacturers OEMs are fundamentally shifting from domestic dominance to rapid global localization. Backed by entrenched control of the battery supply chain…
Foundation Model Revenue
$89 Billion
The portion of total industry revenue captured specifically by the top two foundation models (OpenAI and Anthropic).