CHAPTERS
- 0:00 – 1:43
Goldman Sachs “Calling the Top on AI” report: what it gets wrong
Sarah opens by unpacking a widely circulated Goldman Sachs report that argues AI’s economic impact will be limited and that current AI capex is largely wasted. She frames why it’s worth engaging with the critique while strongly disagreeing with its premises.
- •Report’s central claim: AI impacts <5% of tasks; model-training capex is wasteful
- •Contrasts AI vs. internet adoption economics and value capture
- •Motivation to decompose the arguments rather than dismiss them outright
- 1:43 – 3:33
Old-school ML mental models vs. modern foundation models
Sarah and Elad argue that many critics are applying outdated “traditional ML” assumptions to transformer-era systems. They emphasize that scale and generalization properties have already demonstrated meaningful capability gains.
- •Misconception: more data/scale won’t improve real-world tasks (e.g., customer support)
- •Transformers/foundation models enable qualitatively different functionality vs. classic ML
- •Data matters, but model architecture + application design are equally critical
- 3:33 – 4:07
Why “AI won’t improve much over 10 years” is a bad bet
They challenge the idea that AI progress will be slow or marginal, noting how quickly costs and capabilities have been shifting. The discussion highlights how even builders of the tech are hesitant to make long-range predictions—making confident pessimism suspect.
- •Second flawed assumption: AI won’t get better fast / cheaper fast
- •Difficulty of forecasting even 2 years ahead in this environment
- •Confidence in near-term: broader applications and rapid exploitation ahead
- 4:07 – 4:52
Enterprise adoption is still early—and the biggest wave hasn’t hit
Elad lays out how large enterprises are approaching AI in stages: buying vendor tools, retrofitting internal systems, and eventually shipping AI-native customer products. They argue most enterprise impact is still ahead because planning cycles and organizational frictions slow deployment.
- •Three enterprise tracks: vendor procurement, internal tooling, external AI products
- •So far: mostly vendor buys; product-side adoption is still early
- •Prosumer AI companies lead current scale; enterprise wave is forthcoming
- 4:52 – 6:09
Tech history repeating: value capture won’t be evenly “competed away”
Sarah counters the claim that AI efficiency gains will be fully competed away, pointing to internet-era winners and losers. Elad reinforces that skepticism about transformative tech is recurring—and often wrong in hindsight.
- •Analogy: “everyone uses it so nobody wins” ignores execution differences
- •Internet example: Amazon vs. Borders as divergent outcomes
- •Historical skepticism: incumbents dismissed internet; winners became megacaps
- 6:09 – 8:47
New AI markets: capabilities + APIs + mandates reopen buyer behavior
Elad describes why AI is expanding addressable markets in unexpected ways. He attributes it to new cross-domain capabilities, easy global access via APIs, and organizational mandates that force adoption—similar to early internet corporate behavior.
- •Drivers: new general capabilities, API accessibility, enterprise AI mandates
- •General models replace many bespoke mini-models
- •Mandates shift budgets and openness to new products/approaches
- 8:47 – 10:56
AI-backed buyouts: shortcutting distribution, adoption, and change management
They explore AI-driven acquisitions as a way to transform services-heavy businesses by directly controlling operations and customer relationships. Klarna’s customer support automation serves as a concrete example of cost structure and quality improvements.
- •Buyouts shortcut slow tech adoption and internal change management
- •Owning the asset enables full workflow redesign around AI
- •Klarna example: headcount reduction + 24/7 coverage + multilingual + improved metrics
- •Best fit: “email jobs” and repeatable service workflows that can be automated
- 10:56 – 13:04
When AI buyouts work (and when it’s just PE arbitrage)
Elad contrasts AI buyouts with earlier “tech-enabled rollups” that often relied on multiple arbitrage rather than real operational transformation. They discuss the need for hybrid teams combining software/AI depth with private equity and operational expertise.
- •Past rollups: thin software veneer + multiple arbitrage often collapsed later
- •AI leverage can be real, but only with genuine operational redesign
- •Team composition matters: software founder + strong PE/ops talent (or vice versa)
- •Rationale: sometimes you’re buying distribution/customer relationships directly
- 13:04 – 16:11
AI incubation: why it usually fails—and why now may be different
Sarah and Elad discuss incubation as historically low-odds, working best when there’s deep vertical expertise, proprietary access, or a captive customer base. Elad argues AI has created unusual whitespace and customer pull, making selective incubation more attractive.
- •Incubation tends to work with deep vertical specialization + relationships
- •Startup ecosystem normally relies on massive parallel exploration by founders
- •AI era: some categories are overcrowded while others are wide open
- •Customer/market pull for AI makes targeted incubation more viable
- 16:11 – 18:16
What to incubate: bridging the technologist–domain expert gap
Sarah notes a common mismatch: people who understand AI often don’t understand the domain, and vice versa—creating opportunity for curated founding teams. Elad shares example efforts (e.g., enterprise AI adoption tooling) and areas he’s actively exploring, including healthcare and enterprise assets with built-in customers.
- •Opportunity comes from pairing AI capability insight with domain problem knowledge
- •Risk: technologists searching for use cases can misfire; right match is powerful
- •Enterprise “asset-backed” incubation can provide instant customers
- •Focus areas mentioned: healthcare, services, and specialized models for key verticals
- 18:16 – 22:33
AI and public markets: identifying winners, durability, and AI indexing
They shift to how AI changes investing and competitive moats in later-stage companies. Sarah highlights risk for companies too slow to reinvent; Elad outlines a framework spanning next ‘Magnificent Seven,’ incumbents benefiting from upgrade cycles, AI-resistant durable businesses, and an AI infrastructure index beyond Nvidia.
- •Risk: mid/late-stage companies without speed/will to adapt may suffer
- •Framework: next mega-compounders (new ‘Magnificent Seven’)
- •Incumbent upside: device/compute upgrade cycles (e.g., AI-driven phone refresh)
- •Durability bucket: businesses where AI isn’t a meaningful competitive factor
- •AI index: exposure beyond Nvidia to the broader AI buildout
- 22:33 – 29:12
Staffing AI companies: small teams, velocity, and founder-led product taste
The conversation closes on how AI reshapes team structure and execution cadence. They discuss the push for high revenue per employee, which functions may shrink first (support/SDRs), and the need to preserve product velocity—plus how model “aesthetics” and founder taste increasingly shape AI-native products.
- •Trend: founders aim for smaller teams and higher revenue per employee
- •Near-term limits: tools boost efficiency but not yet full “5x engineer” impact
- •Likely early headcount impact: support and SDR/sales development at scale
- •Operating principle: prioritize velocity amid rapid capability shifts
- •AI products reflect founder taste via data/model-output “design” (e.g., creative tools)
