The Twenty Minute VCNVIDIA Crushes Quarter | OpenAI Cuts Off Cursor | Instinct Hits $2.5B Valuation
CHAPTERS
- 0:00 – 2:38
NVIDIA’s record quarter: why a miss looks unlikely (for now)
The group unpacks NVIDIA’s blowout quarter and why the company appears structurally insulated from near-term disappointment while it remains supply constrained. They outline the handful of things that could still go wrong—chiefly demand softening downstream or competitive share loss—and what investors should watch.
- •NVIDIA demand is described as fully supply constrained, making near-term revenue misses improbable
- •2027/2028 guidance implies demand for compute stays intense at least another 12+ months
- •Core risk isn’t hyperscaler spend; it’s eventual end-user demand not scaling as forecasts assume
- •Competitive threats are acknowledged as a secondary risk: someone taking meaningful share
- •Macro implication: hyperscaler CapEx keeps extending the ‘normalization’ timeline for cash flows
- 2:38 – 7:36
CapEx flywheel and ‘round-tripping’ concerns: why markets are still green
They discuss how cash, financing structures, and massive forward bookings keep the AI infrastructure flywheel spinning. The panel argues that as long as end-user demand keeps expanding, the circularity of deals and financing is less concerning—until growth slows and the math stops working.
- •CFO pushback: stop over-fixating on round-trip/financing structures if demand holds
- •Only a significant end-user demand slowdown would propagate trouble through the stack
- •Investor posture: if NVIDIA is strong, broader AI ecosystem strength follows
- •CapEx projections keep getting revised up, pushing out the ‘return to normal’ narrative
- •A future yellow light would be any NVIDIA hiccup, even if explainable
- 7:36 – 10:44
Why NVIDIA would buy Hugging Face: compute economics, open source, and control
The conversation shifts to NVIDIA’s rumored/advancing acquisition of Hugging Face and the strategic logic behind it. They frame open-source and open weights as margin-compressing forces that are great for GPU sellers—and as part of NVIDIA’s push to “win every segment.”
- •Simple logic: make compute cheaper/easier to use so more compute gets sold
- •Open-source model ecosystems can redirect value from high-margin model providers to GPUs
- •NVIDIA’s incentives favor lower margins for everyone else, higher GPU capture
- •Hugging Face price vs ARR is debated; strategic value outweighs standalone fundamentals
- •Jensen’s visible commitment to open weights signals an ‘own the whole stack’ endgame
- 10:44 – 15:29
OpenAI cuts off Cursor: competition, lawsuits, and distillation fears
They analyze OpenAI restricting Cursor access and why it’s both personal-drama adjacent and strategically rational. The key issues are competitive collision, IP/distillation risk, and the reality that you typically don’t keep deep partnerships with entities tied to active litigation.
- •Cursor/OpenAI were on a competitive collision course even without personal animus
- •Distillation/IP leakage risk: providing models may accelerate a competitor’s model building
- •Elon/OpenAI legal conflict makes continued business cooperation fragile
- •‘Only 5% traffic’ claim is questioned; practical impact mitigated by bring-your-own-key options
- •Pattern match: similar cutoffs have happened before (e.g., Anthropic/Windsurf)
- 15:29 – 21:54
Agent ‘swarms’ hacking: don’t anthropomorphize, but panic about cyber
A deep dive into reports of many agents cooperating in sophisticated attacks, with a strong warning against anthropomorphizing agent behavior. Despite the narrative framing, they argue the real takeaway is that persistent, goal-seeking agents drastically reduce the cost of attacks and amplify cyber risk.
- •Anthropomorphizing misleads: ‘civilizations’ and ‘sacrifice’ are narrative, not mechanism
- •LLMs are goal-seeking systems; loosen guardrails + long runtimes = exploit discovery
- •Persistent agents can hammer complexity 24/7, coordinating at scale with cheap compute
- •Cyber is a near-term, bounded but severe risk area where AI meaningfully escalates threats
- •Rogue actors (states/organized crime) will copy techniques quickly; defenders have months, not years
- 21:54 – 31:55
Instinct and personal agents: credit cards, email access, and reward hacking
The discussion moves from enterprise hacking to consumer risk: should anyone trust personal agents with payments and inboxes? They debate whether reward hacking is solvable today, what guardrails fail to do, and why hard locks/limits may be the only workable short-term control.
- •Instinct-style agents require high-trust permissions (email, calendar, credit cards) to be useful
- •Reward hacking/goal-seeking can cause harmful actions even without malicious intent
- •Guardrails can conflict; agents may override constraints to satisfy higher-priority goals
- •Hard caps/lock-and-key mechanisms (strict spend limits, permissions) are proposed as the only near-term fix
- •Category question: will these become enduring workflows or a short-lived fad?
- 31:55 – 37:09
Are agent assistants investable or quickly commoditized?
They address whether Instinct is defensible given many clones and whether incumbents will embed similar capabilities. The panel argues that early versions are easy to copy, but moats form as products accrete complexity—security, reliability, integrations, and real-world edge cases.
- •Every app will add autonomous features; overlap pressures standalone agent assistants
- •Early clones are easy, but sustained leadership depends on compounding capabilities over time
- •Moat drivers: security hardening, integrations, long-tail edge-case handling, operational reliability
- •Venture behavior: ‘get exposure’ to hot categories even before winners are clear
- •Agents framed as the defining venture theme of 2025–2026, driving premium valuations
- 37:09 – 41:33
Coding agents and TAM shock: Cognition at $46B and software output exploding
They react to Cognition’s reported valuation and revenue pace, arguing the coding market is bigger than expected. The key twist isn’t just replacing developer labor; it’s that teams are building vastly more software than before, expanding the effective market for tools and automation.
- •Cognition’s scale despite not being ‘top 2’ suggests the market supports multiple massive players
- •Top-down TAM anchored in global software labor; debate over % converting to AI spend
- •New insight: code production volume is exploding (they cite “100x more software”)
- •Long-running autonomous coding agents create distinct subsegments vs IDE copilots or no-code builders
- •‘Postmates effect’: even #3 in a huge category can become a major outcome
- 41:33 – 47:47
Why every startup becomes a compound company: shipping speed as survival
The panel argues AI-driven development collapses traditional category boundaries, forcing companies to expand product surface area fast. Winners will be those that ship relentlessly and integrate broader suites, while point solutions risk irrelevance as competitors compound into adjacent modules.
- •AI accelerates feature delivery, driving competitive convergence across B2B apps
- •Founders must out-ship rivals; ‘heat maps’ of safe niches disappear (‘all lava’)
- •Economic tension: software spend won’t 10x, so competition shifts to suite value and consolidation
- •Compound strategy: integrated 10-module suite can win even at 4x price, killing point products
- •Execution risks: product bloat and ‘product slop’ demand strong product leadership
- 47:47 – 52:45
Europe vs US: capital intensity, headcount paradox, and the kingmaking debate
They discuss whether European startups are disadvantaged due to smaller rounds in a world where speed and compute spend matter. This transitions into a broader debate on VC “kingmaking”—how capital plus brand attracts talent and customers, accelerating compounding for already-fast growers.
- •European companies may struggle to match US pace due to funding and token costs
- •Iconiq data cited: hypergrowth companies are increasing headcount aggressively (doing more with more)
- •Paradox: AI should reduce engineering needs, but competitive dynamics encourage hiring + expansion
- •VC brand effects: top firms can amplify talent attraction, customer validation, and fundraising ease
- •Practical investor model: identify winners early, ‘stuff capital’ in, then stay out of the way
- 52:45 – 54:27
Andreessen expands growth fund: ‘everything is green’ capital allocation logic
A short segment on why a16z increased growth capital: strong deployment opportunities and intense competition for allocation in breakout companies. The panel frames it as a signal of perceived momentum and urgency in the current growth environment.
- •Fund expansion interpreted as confidence that growth opportunities are unusually abundant
- •Contrast with firms that shrink funds when they can’t deploy at target quality
- •‘Need more money now’ mindset: hot deals (e.g., model infrastructure and AI apps) soak capital quickly
- •LP behavior mirrors VC behavior: allocate more to what’s working
- •Momentum and allocation scarcity drive fund-size changes as much as fundamentals
- 54:27 – 1:04:25
Salesforce + Anthropic (‘ClaudeForce’): headless surfaces and outcome pricing
They dissect Salesforce’s partnership messaging with Anthropic, separating marketing from deeper strategic shifts. The meaningful changes are Salesforce embracing multi-surface access (including headless/LLM-driven interfaces) and moving toward outcome-based pricing—both existential adaptations for systems of record.
- •Surface-level ‘skills’ are seen as commoditized; partnership value is more strategic than technical
- •Salesforce embracing multi-surface usage (Slack, Claude, headless) despite UI lock-in threats
- •Outcome-based pricing framed as a major shift in enterprise buying behavior
- •Salesforce token spend vs engineering spend (e.g., $300M vs $6B) suggests room for growth in AI budgets
- •Valuation discussion: system-of-record stickiness must translate into measurable outcomes to endure
- 1:04:25 – 1:06:22
M&A and valuation watch: Stripe–PayPal dance, plus Flock backlash and privacy tradeoffs
They briefly cover the Stripe–PayPal deal falling apart (likely price), noting such negotiations often cycle before resolution. They close with Flock camera controversies: product misuse, errors, and surveillance concerns driving political backlash even in law-and-order regions.
- •Stripe–PayPal seen as a pricing mismatch; deal dynamics may revive after ‘falling apart’
- •Boardroom ‘dance’ and market price movements complicate M&A outcomes
- •Flock issues: misidentification leading to overreliance on AI outputs in policing
- •Misuse risks: officers using surveillance tools for personal reasons (exes, favors)
- •Core tension: public safety benefits vs surveillance costs; need for stronger controls and messaging
- 1:06:22 – 1:11:20
Why Clay might be ‘cheap’ at $7B: agentic GTM and tools agents prefer
Clay is revisited as a surprising potential winner as GTM becomes agent-driven. Jason argues agents increasingly select the tools they want to use, and Clay’s agent-friendliness and output quality can translate into outsized usage growth—even if revenue lags initially.
- •Shift from ‘checkbox AI tool’ skepticism to practical dependency in agent workflows
- •Agentic GTM could massively expand usage (24/7 campaigns, analysis, orchestration)
- •Key wedge: agents choosing tools based on API utility and results, not human persuasion
- •Bull case framed as category exposure plus breakout leadership; path to very large outcomes posited
- •Theme: AEO/GEO can be gamed; agent tool-choice is harder to game and more merit-based
- 1:11:20 – 1:20:06
Linear at $2.5B: the case for agent-friendly systems to manage 100x software output
They evaluate Linear’s growth and valuation through the lens of teams building far more software with agents. The bull case: as feature/issue volume explodes, humans need an agent-compatible system of record to coordinate work—making Linear a potential beneficiary despite the ‘old’ project management category.
- •Traditional project management seen as stagnating, but Linear positioned differently for agent-era workflows
- •Value proposition: manage overwhelming task/issue volume generated by agent-accelerated development
- •Agent-friendliness matters: external agents can reliably use the product to coordinate work
- •Valuation framed as reasonable if agent-driven usage eventually becomes monetized
- •Broader implication: infrastructure (e.g., databases like ClickHouse) may see query/usage explosions from agents
- 1:20:06 – 1:20:28
Closing IC recap: investing discipline and ‘sometimes no is productive’
Harry wraps the episode as an “investment committee,” jokingly tallying calls made during the discussion. The closing reinforces the broader thesis: in fast-moving AI markets, selective conviction matters, and restraint can be a decision too.
- •Episode framing as an IC highlights how narratives translate into investing posture
- •Humor aside, they emphasize capital allocation toward compounding winners
- •Acknowledgment that not every week produces a ‘new buy’—saying no is a choice
- •Reinforces themes: agents, cyber risk, and compounding products reshape software markets
- •Ends with a light recap of who ‘made investments’ in the conversation
