The Twenty Minute VCGroq Founder, Jonathan Ross: OpenAI & Anthropic Will Build Their Own Chips & Will NVIDIA Hit $10TRN
CHAPTERS
- 0:00 – 1:09
Compute + energy as the real AI power: setting the frame
Ross opens with the core thesis that whoever controls compute will control AI—and compute is inseparable from energy. The conversation tees up why hardware capacity (and the ability to power it) matters as much as models.
- •Compute is the scarce strategic resource for AI leadership
- •Energy availability is a prerequisite for compute at scale
- •AI progress is constrained less by ideas than by deployable capacity
- 1:09 – 3:54
Is AI a bubble? Follow what the “smart money” is doing
Instead of debating whether valuations are inflated, Ross argues you should watch hyperscalers and governments: they’re increasing spend aggressively. He likens today’s market to early oil drilling—many dry holes, a few gushers, and highly concentrated revenue.
- •Reframe: don’t ask 'bubble?'—ask what major buyers are doing
- •Hyperscalers repeatedly raise capex guidance for AI
- •Revenue/token spend is extremely concentrated and 'lumpy'
- •Early-market analogy: oil drilling before it became a predictable science
- 3:54 – 8:08
Why hyperscalers keep spending: fear of being locked out + real ROI examples
Ross explains hyperscalers aren’t spending purely for near-term ROI; they’re protecting leadership and avoiding existential disruption. He illustrates tangible AI returns with an internal 'vibe coding' example where a feature shipped to production in hours via prompting.
- •Spending is driven by strategic survival, not just finance math
- •Staying in the top tier (Mag 7/10) incentivizes outsized investment
- •AI already produces meaningful business value, but unevenly
- •Example: prompt-driven development compressed a feature cycle to hours
- 8:08 – 10:26
The 'infinite money loop' and why compute scarcity makes revenue scale fast
Harry challenges the idea of money cycling from NVIDIA to AI labs and back; Ross argues much spend flows into real infrastructure and suppliers. He claims major labs are compute-limited, so additional inference capacity would translate quickly into higher usage and revenue.
- •AI capex isn’t pure round-tripping; a large share funds infrastructure buildout
- •Compute scarcity is the binding constraint for leading labs
- •More inference capacity can lift revenue via fewer rate limits and higher engagement
- •Product quality and monetization are directly shaped by available compute
- 10:26 – 12:35
Latency is a feature: why speed drives engagement and conversion
Ross pushes back on the idea that users will tolerate slow, background AI. Drawing from consumer psychology and web history, he argues faster responses increase brand affinity and conversion—making inference speed strategically decisive.
- •Speed correlates with engagement and brand affinity (CPG analogy)
- •Historical lesson: Google/Facebook won by obsessing over latency
- •Even 'faster than you can read' can matter for user behavior
- •Latency becomes a growth lever, not just an engineering metric
- 12:35 – 16:46
Why OpenAI/Anthropic (and hyperscalers) will build chips—and why it’s hard
Ross predicts OpenAI, Anthropic, and hyperscalers will build chips, but cautions chip success is rare and software is the real difficulty. He explains the deeper motive: controlling allocation and reducing dependency when GPU supply is constrained.
- •Chip building is difficult; many efforts fail or get canceled
- •The hard parts evolve: hardware → software → staying current with the frontier
- •Key motivation is 'control over destiny' (allocation leverage), not perfection
- •Small performance edges can dominate because system-level value multiplies
- 16:46 – 22:51
HBM bottlenecks, long lead times, and the capex vs opex reality of chips
The discussion goes deep on why GPUs are supply constrained: HBM and packaging capacity, not just compute die production. Ross also reframes amortization: after deployment, the key question becomes whether a chip remains profitable versus operating costs as new generations arrive.
- •NVIDIA’s effective leverage comes from HBM scarcity and pre-committed supply
- •New entrants face lead-time realities: checks written years in advance
- •Amortization differs for buying vs keeping hardware running (capex vs opex)
- •Risk: newer chips can push older chips below profitable operating thresholds
- 22:51 – 29:16
Groq’s pitch: faster supply chain loops + world-scale system optimization
Ross argues customers start by asking for speed but quickly fixate on guaranteed capacity. He positions Groq’s advantage as a shorter supply chain and the ability to optimize not just per-chip, but across data centers globally based on geographic demand patterns.
- •In practice, capacity availability becomes the decisive buying criterion
- •Groq claims a significantly shorter procurement-to-deployment window
- •The 'Hardware Lottery': models get designed around incumbent hardware
- •Operational edge: global load balancing and geography-specific optimizations
- 29:16 – 34:37
China, DeepSeek, and the 'home vs away game' for AI geopolitics
Ross challenges the belief that leading Chinese models are cheaper to run, arguing price got confused with cost and that some are more expensive at inference. He frames competition as China’s ability to win domestically via subsidies/energy buildout versus the US advantage exporting efficient compute to allies.
- •Cost vs price confusion in perceptions of Chinese model economics
- •Training efficiency vs inference efficiency trade-offs
- •China can subsidize at home; allies face hard energy constraints
- •US advantage: better chips matter more in the 'away game' across partners
- 34:37 – 48:17
Europe’s path to compete: energy first (renewables + nuclear) and permitting reform
Ross contends Europe can still compete if it acts quickly, primarily by unlocking abundant energy and colocating compute where power is cheap. He argues bureaucracy and permitting—especially around nuclear—are the real blockers, and warns model sovereignty without compute won’t be enough.
- •Europe can compete by mobilizing energy assets and colocating compute
- •Renewables (e.g., Norway wind + hydro) are presented as underused leverage
- •Nuclear is safe/viable, but fear and permitting dominate costs and timelines
- •Model sovereignty is insufficient without large-scale compute capacity
- 48:17 – 51:54
AI’s macro impact: deflation, opting out of work, and new labor shortages
Ross predicts AI will be massively deflationary, lowering costs across supply chains via automation and optimization. Paradoxically, he expects labor shortages as people work less and new industries emerge that are hard to imagine today.
- •Deflationary pressure reduces cost of goods and services broadly
- •People may work fewer hours/years as living costs fall
- •New industries and jobs will emerge (historical analogy: agriculture shift)
- •Net effect could be labor shortages, not mass unemployment
- 51:54 – 53:40
Vibe coding becomes literacy: coding spreads to every function
Ross argues 'vibe coding' will mirror the shift from literacy as a profession (scribes) to literacy as a baseline skill. He expects non-engineers across roles to build tools via prompting, with software creation becoming ubiquitous and iterative through real-world feedback.
- •Coding shifts from specialized career skill to universal workplace literacy
- •Non-technical operators can build and iterate tools via prompting
- •Edge cases and user feedback still drive iteration—just faster
- •Tool choice will keep changing, but overall capability diffusion accelerates
- 53:40 – 58:47
Why AI companies should aim for low margins—and what that means for competition
Ross says margin is mainly insurance against volatility, but high margins invite competition and misalign with customers. His preference is low margins plus volume growth (Jevons paradox), building customer trust and expanding usage as costs fall.
- •Margins provide stability, but create entry incentives for rivals
- •Lower margins can build customer 'equity' and trust over time
- •Compute demand expands as cost drops (Jevons paradox)
- •Focus on solving customer problems over optimizing the bottom line early
- 58:47 – 1:21:49
Where markets go next: Mag 7 concentration, undervalued AI labs, and the chip landscape in 5 years
Ross separates 'weighing machine' value from 'popularity contest' pricing, arguing AI is delivering real measurable value despite market choppiness risks. He calls OpenAI and Anthropic undervalued in a market that expands with R&D, then forecasts NVIDIA retaining most revenue share while shipping a minority of chips as alternatives proliferate.
- •AI value is tangible (e.g., PE demand) even if multiples fluctuate
- •Concentration risk can amplify downturns, but also births great companies
- •OpenAI/Anthropic can grow alongside incumbents as the market expands
- •5-year chip view: NVIDIA keeps >50% revenue via brand/margins, but fewer units sold
- 1:21:49 – 1:31:19
Quick-fire: CUDA moat debate, why it’s 'too late' to start a chip company, and what’s next
In rapid Q&A, Ross argues CUDA is less of a moat for inference than people think and highlights Groq’s developer traction. He explains why he wouldn’t start a new chip company today due to long cycles, reflects on focus vs optionality, and closes with an optimistic 'telescope of the mind' view of LLMs.
- •Misconception: NVIDIA software moat is weaker for inference than training
- •Starting a new chip company now is disadvantaged by multi-year cycles
- •Business lesson: shifting from preserving optionality to relentless focus
- •Long-term optimism: LLMs expand human understanding like telescopes did
