Eric Vishria - A Decade of Lessons Investing in Software & Hardware - [Invest Like the Best, EP.486]

Episode Summary

Executive Summary: Patrick O’Shaughnessy and his guest argue that AI is not a winner-take-all market: cloud, infrastructure, apps, chips, and even new CPUs can all produce major winners. The key shift is that AI changes the competitive frontier, rewards technical depth and adaptability, and forces companies to rethink sales, product, and capital allocation in real time.

Main Topics: AI as an oligopoly, not a monopoly (Priority: 5/5): The guest compares AI to cloud: early predictions that one player would dominate were wrong because the market was too large. He expects multiple winners across labs, infrastructure, neoclouds, chips, edge inference, and applications. Cloud history as a template for AI (Priority: 5/5): AWS did not consume the entire market; instead, Azure, GCP, Cloudflare, and several software companies emerged alongside it. The guest uses that history to argue AI will also create durable, multi-layered winners rather than a single dominant company. The new competitive frontier in software (Priority: 5/5): AI changes what matters in product development and company strategy. Winners now need taste, customer understanding, and deep awareness of the jagged edge of model capability rather than traditional SaaS operating playbooks. Model progress, adoption, and enterprise change (Priority: 4/5): Enterprises are adopting AI with more urgency than cloud, but adoption barriers remain. The guest emphasizes the role of companies acting as 'AI Sherpas' to bridge enterprise reality and model capability. Hardware, compute, and energy bottlenecks (Priority: 4/5): Cerebras is used to illustrate how hard AI hardware is to build and run, and why energy may become the main bottleneck for AI scale. The guest sees compute converting into intelligence, constrained by power generation. Robotics and autonomous systems (Priority: 4/5): Robotics will require high-quality data, vertical integration, and AI pretraining/post-training loops similar to LLMs. Household robotics is viewed as emerging through data collection and evaluation rigor, not just task selection. Venture, board work, and higher-stakes capital (Priority: 4/5): The guest explains why the firm raised a growth fund: the range of high-IRR opportunities now extends beyond seed-stage investing. He also stresses chemistry, conviction, and willingness to be a real partner on the board.

Key Arguments: AI markets are too large and too distributed for one company to 'eat everything'; the likely outcome is a set of oligopolistic winners across layers. Cloud history proves zero-sum thinking is often wrong: early skepticism about AWS ignored the emergence of Azure, GCP, Cloudflare, and new application winners. Traditional SaaS playbooks break in AI because the frontier shifts faster; companies must iterate around model capability, not around fixed product requirements. Winners increasingly need three traits: understanding customer problems, taste, and knowledge of the jagged edge of AI capabilities. Enterprise adoption of AI is real but uneven; the best companies will help customers bridge the gap rather than sell abstract technology. Sales models built on quota capacity can fail when products feel like 'magic' and demand is pulled by capability rather than pushed by reps. AI hardware is extraordinarily difficult; software investors often underestimate the physics, supply chain, and bring-up complexity. Energy, not chips alone, may become the primary binding constraint on AI growth because compute ultimately needs power. Robotics needs an internet-like data flywheel; vertical integration and high-quality data collection are prerequisites for meaningful autonomy. The firm’s investment criteria now include whether the opportunity could plausibly be someone’s life’s work and whether the investor-founder relationship has true chemistry.

Data Points: AWS launch year: 2006 - Used as the starting point for the cloud analogy and the long arc of market skepticism to eventual dominance. AWS investor sentiment in 2007: 0 for 30 - Guest estimates that even very smart investors would have assigned near-zero odds of AWS being a durable high-margin business. Cloud market share split: 40/30/20 - Guest describes today’s cloud oligopoly as roughly AWS/Azure/GCP by scale (approximate framing). Fireworks performance advantage: 5x - Guest says Fireworks can run open source models about 5x faster than cloud providers, with additional throughput advantages. AI company growth rates: 10x, 20x, 30x million ARR - Guest says some AI sales reps are seeing annual recurring revenue per rep far above classic SaaS quota models. Traditional SaaS quota model: $1.2M-$2.5M per rep - Referenced as the old sales planning framework that breaks in AI-native selling. SaaS public-company valuation: 30x revenue (2021) vs ~6x revenue later - Used to show how multiple compression changed the meaning of growth and investor expectations. Cerebras funding: $500M - Guest recalls the scale of capital required and the fear that the project was 'melty' during the buildout. Cerebras core count: 450,000 cores - Described as the wafer-scale design goal to maximize on-chip parallelism. Cerebras on-chip memory: 20 GB SRAM - Cited as a way to avoid going off-chip for memory access. Cerebras process node: 7 nm - Guest mentions the first chip was around 7 nm in the early implementation. Robert benchmark on venture investing: 1-2 companies per year - Guest says he and his partner each invest in roughly one to two companies annually. Total personal investments: 18 companies in 12 years - Guest gives his own cumulative investing volume as evidence of high-conviction selection. Ramp savings claim: 5% annually on average - Sponsor ad claim about finance automation savings. Ramp growth claim: 3.2x faster - Sponsor ad claim about customers growing revenue faster than the average American business. WorkOS enterprise features: SSO, SCIM, RBAC, audit logs - Sponsor ad describes the core capabilities needed for enterprise adoption. Vanta vendor assessment reduction: up to 50% - Sponsor ad claim about compliance automation. Sierra company evolution: customer service to broader Horizon agents - Describes the move from automating support toward longer-running agent workflows.

Pivotal Quotes: "What if it all works? It all works." — Guest: Argument against zero-sum thinking in AI infrastructure and applications. "You need to check everything at the door. Check it all." — Guest: Advice on evaluating founders and opportunities in a fast-shifting AI environment. "We used to be building castles. Now we're building sand castles that are walking away." — Guest: Brett Taylor quote used to describe how quickly AI product assumptions become obsolete.

Implications: AI will likely produce many durable winners, but only for teams that adapt quickly, own their data and distribution, and treat energy, hardware, and model capability as strategic constraints. Old SaaS playbooks are not enough.