The Twenty Minute VCDavid Cahn: Why Servers, Steel and Power Are the Pillars Powering the Future of AI | E1186
CHAPTERS
- 0:00 – 0:56
Cold open: Why data centers matter most — “servers, steel, and power”
David Cahn opens with a provocative claim: frontier models won’t be trained twice on the same data center because hardware and facility constraints move too fast. He frames AI as an industrial revolution where the most important assets are physical: servers, steel, and power.
- •Frontier training quickly outdates GPUs and even entire data centers
- •Scaling laws push ever-larger clusters and facilities
- •Data centers become a core strategic asset, not just “compute”
- •AI’s next phase looks like an industrial buildout
- •Thesis: servers, steel, and power are the pillars
- 0:56 – 4:48
Personal foundations: being a twin, competitiveness, and family heritage
Harry and David discuss formative experiences—twin dynamics, non-conformism, competitiveness, and the immigrant/escape-from-persecution history of David’s family. These themes set up his drive, intensity, and long-term orientation.
- •Twinhood as a “license” to be non-conformist
- •Hypercompetitiveness from always being close to an equal
- •Family history (Nazi Germany escape, Syrian immigration) as motivation
- •A culture of ‘what’s next?’ over celebrating wins
- •How early pressure and opportunity shape ambition
- 4:48 – 8:09
Two separate questions: Will AI change the world vs. is near-term AI CapEx rational?
David separates belief in AI’s transformative impact from whether today’s massive capital spending will pay back soon. He argues Silicon Valley often conflates the two and avoids the hard math behind the investment wave.
- •AI can be world-changing while near-term CapEx still overreaches
- •The ecosystem is deploying ‘hundreds of billions’ in a short window
- •His context: years investing across W&B, Runway, Hugging Face
- •“$600B question” is about economic payback, not AI faith
- •Clarifying hype vs. financial reality helps everyone plan better
- 8:09 – 12:02
Conscious overspending and oligopoly game theory: why the spending likely continues
They explore how Big Tech views AI buildout as a strategic arms race rather than a clean ROI project. David argues the cloud oligopoly is historically powerful and will spend aggressively to defend its position, even acknowledging risk.
- •Zuck/Sundar acknowledge risk, but still must invest
- •Spending is strategic defense of the cloud’s “golden goose”
- •Big Tech is not guaranteeing rapid AI revenue payback
- •Oligopoly power (Azure/AWS/Google) enables aggressive spend
- •Barrier-to-entry is intentionally raised via capital intensity
- 12:02 – 17:35
Reconciling ‘compute gets cheaper’ with ‘compute is the future’: the physical reality of data centers
David explains the tension between falling compute prices and compute’s long-term importance by grounding “compute” in physical infrastructure. Data centers are hard, slow, and prone to being built “wrong” relative to fast-moving chip and cooling needs.
- •“Compute” is a euphemism for physical data centers + GPUs + cooling
- •Data centers take ~2 years and cost ~$2B each (order-of-magnitude)
- •Chips and cooling evolve, making facilities obsolete or mismatched
- •Upgrades require real-world retrofits, not just software changes
- •The buildout isn’t ‘15 years of compute in 2 years’—it’s messy
- 17:35 – 21:38
Will model breakthroughs outpace data center construction? Scaling laws vs. efficiency leaps
They debate whether research improvements (reasoning, efficiency) can outrun the pace of building larger clusters. David sees current evidence favoring scaling laws but hopes both scaling and breakthroughs accelerate progress.
- •Core question: research breakthroughs vs. infrastructure build speed
- •Two schools: ‘scale is all that matters’ vs. ‘efficiency will win’
- •Scaling laws currently appear more supported by evidence
- •Bigger clusters (100k → 300k GPUs) reshape facility architecture
- •Model progress can itself force data center redesigns
- 21:38 – 28:51
Vertical integration in AI: why models and data centers must be tightly coupled
David argues the key integration is not “everyone builds chips,” but that model development and data center operations need to be unified. He highlights Meta and xAI-style approaches and flags structural tension for partnerships like Microsoft/OpenAI and Amazon/Anthropic.
- •Hard to ‘separate’ data center ops from frontier model teams
- •Meta and xAI benefit from tight coupling of infra and model work
- •Don’t bet against Jensen/NVIDIA, but infra integration still matters
- •Startups need a “cash machine” to compete in frontier models
- •Big partnerships face coordination challenges across org boundaries
- 28:51 – 31:56
What’s the real bottleneck? From compute/data/algorithms to ‘servers, steel, and power’
David reframes the bottleneck debate away from abstract inputs and toward industrial constraints. He argues leading labs are converging on similar approaches, making physical supply chains and energy the differentiators.
- •Data advantage is hard to claim among top model labs
- •Compute is increasingly a purchasable commodity
- •“Secret sauce” in models shrinks if scaling laws dominate
- •Reframe: servers (chips), steel (construction/supply chain), power (energy)
- •AI as an industrial revolution more than a software moment
- 31:56 – 32:26
Chip pools and the coming chip wars: NVIDIA’s roadmap, challengers, and geopolitics
They discuss NVIDIA’s continued performance gains and the market dynamics that attract competition. David also touches on geopolitical concentration risk around Taiwan and the likelihood of greater US supply-chain investment.
- •Moore’s Law-like improvements continue in price/performance
- •NVIDIA’s roadmap (e.g., next-gen chips) sustains momentum
- •High gross margins invite AMD, Broadcom, and startup challengers
- •Taiwan/TSMC concentration adds political and supply risk
- •On-shoring and policy become increasingly relevant to AI hardware
- 32:26 – 35:38
‘Steel’ as a catchall: supply chains, generators, labor, and convincing factories to scale
David uses “steel” to represent the full industrial stack behind data centers—construction, generators, batteries, and labor. A key challenge is persuading manufacturers to expand capacity when they fear demand may prove temporary.
- •Data center buildouts cascade through developers, contractors, and trades
- •Labor (e.g., electricians) is a major cost and a real bottleneck
- •Cloud giants push suppliers with multi-year capacity commitments
- •Factory owners hesitate to build idle plants on speculative demand
- •Scale and credibility let Big Tech effectively pre-buy capacity
- 35:38 – 37:48
Power constraints and the energy buildout: AI as the catalyst for an energy revolution
They dive into whether the grid can support AI growth and why demand may force faster energy innovation than policy alone. David highlights large incumbents like NextEra and argues the transition may be steady rather than a sudden crisis.
- •Power demand is rising faster than current supply capacity
- •Capitalism and AI demand can accelerate solar, storage, and generation
- •IRA helped, but AI may be the bigger forcing function
- •Utilities and major builders (e.g., NextEra) are underappreciated in SV
- •Energy transition may be gradual ‘plotting our way there,’ not panic
- 37:48 – 47:04
Open vs. closed models, AGI temperature checks, and China’s progress
David downshifts the fear-based AGI framing, preferring a pragmatic view: it’s good to have both open and closed options, and near-term catastrophe is unlikely. They also discuss China’s likely catch-up and the importance of not underestimating competitors.
- •Open vs. closed debate often correlates with near-term AGI beliefs
- •David is less worried about imminent AGI; prefers balanced ecosystem
- •Critique: Silicon Valley ‘replaces religion with AGI anxiety’
- •China will progress; the US has strong structural advantages
- •Best posture: assume competitors execute well and keep improving
- 47:04 – 51:12
From AI macro to venture craft: lessons from Databricks/Snowflake/UiPath and real customer behavior
The conversation pivots to David’s investing apprenticeship and what he learned from iconic software winners. A key lesson: customers often complain loudly but keep paying—watch actions over words to understand true value delivery.
- •He supported (not led) major deals; learned through deep diligence work
- •Customers say ‘we’ll rip it out’ far more often than they actually do
- •Expensive tools that retain usage usually deliver real value
- •Competitive alternatives (Redshift/BigQuery, etc.) didn’t prevent winners
- •Focus on value delivered, not stated intent or theoretical churn
- 51:12 – 57:39
Sequoia’s ‘slugger’ standard: conviction, constraints, and putting your neck on the line
David explains Sequoia’s culture: success is defined by producing billion-dollar-plus gains, and the partnership structure pressures real conviction. Constraints—few bets, very high bar—force sharper selection and long-term company-building focus.
- •Single success metric: generating outsized, billion-dollar gains
- •Hardest part isn’t identifying good companies; it’s committing publicly
- •Partnership meetings test conviction under scrutiny
- •Constraints (few deals/year) increase decisiveness and selectivity
- •Sequoia “breaks down and rebuilds” investors around company-building
- 57:39 – 1:00:57
Founder assessment: a 2x2 framework (science vs. intuition) × (technology vs. humans)
David shares a structured approach to evaluating founders, inspired by studying biographies of historic innovators. He argues generational, world-class founders combine technical rigor, personal “hardcore” optimization, product intuition, and human leadership.
- •Axis 1: science vs. intuition; Axis 2: technology vs. humans
- •Science+tech: engineering excellence; Science+human: hardcore self-maximization
- •Intuition+tech: product vision (e.g., Notion-style taste); Intuition+human: leadership
- •Great founders often need multiple quadrants; all four for the biggest outcomes
- •Skills can be developed over time (growth mindset)
- 1:00:57 – 1:13:01
Operating as a young investor: routines, deal-winning tactics, first deal, and quick-fire
David discusses how youth changes his value proposition to founders, his intense morning routine, and creative outreach methods that helped him win meetings. He closes with the story of his first memorable deal (Starburst Data) and a rapid Q&A on beliefs and investing philosophy.
- •Youth advantage: grow alongside founders; different value prop than veterans
- •Routine: early gym, biking to Sand Hill, morning founder meetings
- •Creative tactics: Cameo gifts, daily Looms (“Emo Loom”) to earn attention
- •First ‘I found and won it’ deal: Starburst Data; $30M check
- •Quick-fire: belief in God, hiring advice, OpenAI durability, learning to drive
