The Twenty Minute VCAnthropic Raises $45B but Falls Short on Compute & Thoma Bravo Hand Back Medallia Keys to Creditors
CHAPTERS
- 0:00 – 3:41
OpenAI misses growth targets: late reaction vs real competitive signal
The group debates whether the market is overreacting to reports that OpenAI missed user and revenue growth targets. They argue much of this reflects last year’s model-quality gap versus Anthropic, and that sentiment often lags what power users already see in real-time.
- •OpenAI’s perceived miss is framed as a “late-dropping shoe” from last year’s model underperformance
- •Anthropic gained share largely through coding strength; OpenAI responds by improving Codex
- •AI Twitter narrative shifts quickly—Anthropic now criticized for reliability/throughput
- •Public-market reactions (CoreWeave/Oracle down) may reflect second-order expectations about demand
- 3:41 – 6:35
The rise of AI agents: why humans stop choosing models and tools
Jason argues the next competitive battleground is not which model humans prefer, but what autonomous agents choose on behalf of users and companies. That shift changes how moats form and could reset perceived advantages among model providers.
- •Agents will select vendors/tools/LLMs based on performance and workflow fit, not human preference
- •Human-facing advantages (e.g., Claude’s usability) may not translate to agent preferences
- •A transition is coming from human-led AI workflows to agent-led workflows (2026–2027 timeframe)
- •Implication: today’s model “winner” in human UX may not be tomorrow’s winner in agent ecosystems
- 6:35 – 10:05
Agent-driven procurement: oligopoly dynamics and the fight for the agent layer
Rory presses on whether this becomes a stable three-player oligopoly (OpenAI/Anthropic/Gemini) and what it takes to be ‘chosen’ by agents. The conversation converges on the strategic value of owning the agentic layer—because the agent layer can create distribution and lock-in.
- •Agents tend to recommend market leaders with momentum; legacy tools may be bypassed
- •Jason’s “agentic API grader” example: leaders favored; Stripe graded exceptionally well
- •Owning the agent layer matters as much as (or more than) owning the base model
- •If one company wins agents, it can influence downstream model choice and create lock-in
- 10:05 – 18:48
Enterprise software in an agent world: multi-year deals, churn masking, and ‘systems of record’
The hosts examine whether long enterprise contracts still matter if agents reduce the need for traditional SaaS tooling. They argue multi-year deals can delay—but not eliminate—churn, and markets will increasingly value evidence of agent usage and agent-driven revenue.
- •Multi-year contracts can ‘mask decay’ by deferring churn rather than preventing it
- •Systems of record may persist but face growth ceilings if agents don’t build on them
- •ServiceNow earnings questions signal investor focus: is ‘agent revenue’ real or bundled?
- •Framework emerges: melting iceberg (terminal value erodes), system of record (stable/no growth), agent platform (growth re-accelerates)
- 18:48 – 24:01
Anthropic’s $45B hyperscaler-backed financing: the compute constraint story
Discussion turns to Anthropic’s massive round and why hyperscalers (Google/Amazon) are effectively underwriting its scaling needs. Rory frames the core operating challenge for foundation model companies as synchronizing model quality with compute availability—Anthropic ‘won’ on models but ran short on compute.
- •Two hard jobs: build great models and secure enough compute to meet demand
- •OpenAI had compute but lagged on models; Anthropic had models but lacked compute capacity
- •Anthropic’s success created demand shock; forecasting compute needs is extremely difficult
- •Hyperscaler financing aligns with bundling infrastructure, capacity, and strategic control
- 24:01 – 27:10
Compute ≠ revenue: capital intensity, forecasting risk, and the first crack in the AI model
Rory and Jason unpack the danger in the simplistic thesis that ‘compute equals revenue.’ They argue compute is necessary but not sufficient; model quality creates demand, and the required CapEx is massive, long-lead, and error-prone—creating stranded-capacity risk or lost-revenue risk.
- •Reframing: no compute → no revenue, but compute + weak model → still no revenue
- •Scaling scenario shows huge CapEx needs: multiple dollars of infrastructure per $1 of run-rate revenue
- •Two-year lead times make demand forecasting extremely hard; mistakes swing from ‘too much compute’ to ‘too little’
- •Agents may drive 50–100x token usage vs human workflows, increasing medium-term demand but not removing volatility
- 27:10 – 28:59
Secondary markets for compute: subletting capacity and how hyperscalers arbitrate supply
Harry asks which is easier: selling excess compute or acquiring emergency compute. The group suggests a ‘compute sublet’ market can emerge, mediated by hyperscalers who hold the capacity and can re-route it between customers as utilization shifts.
- •Foundation model firms often contract for compute while hyperscalers ‘own’ the supply
- •Excess capacity could be reallocated to competitors via hyperscalers (like a sublet)
- •Historical analogy: Samsung selling components to competitors
- •Macro reality: compute allocation is already shifting in real time (data centers reassigned between buyers)
- 28:59 – 34:58
Why Google may be the biggest winner: capacity, cash flow, and chips as leverage (plus Nvidia implications)
Jason argues Google wins regardless of whether customers use Gemini or Anthropic if it supplies infrastructure and capital. Rory adds that Google and Amazon also push their own chips (TPUs/Trainium) via bundling—potentially pressuring Nvidia’s share and margins in certain workloads.
- •Google benefits from owning both a model option (Gemini) and a key partner (Anthropic)
- •Google’s surplus capacity and cash flow let it flex compute allocation strategically
- •Bundling in-house chips can capture gross margin otherwise flowing to Nvidia
- •Debate: Nvidia vs Google as investments—Nvidia as pure AI beta, Google as ‘multiple ways to win’
- 34:58 – 42:27
China blocks Meta’s $2B Manus deal: investors get paid, humans face risk, and a broader geopolitical message
The conversation shifts to China’s effort to unwind Meta’s acquisition of Manus. They emphasize that investors who already received distributions are unlikely to return funds, while the real leverage may be technology access and the safety/mobility of team members still in China.
- •Investors are unlikely to repay distributed proceeds; enforcement is limited outside China
- •The pressure point is Meta and any ongoing China exposure, not venture investors directly
- •Human stakes: employees/founders in China may face exit-visa or legal constraints
- •Viewed as part of a broader US–China AI competition shaped by chip sanctions and tech controls
- 42:27 – 48:00
Thoma Bravo hands Medallia to creditors: overpaying, debt service limits, and vendor consolidation
They dissect Medallia’s wipeout as a rare large-scale PE equity loss and clarify it wasn’t necessarily ‘too much leverage’ so much as ‘too high a price’ plus deteriorating fundamentals. AI-driven vendor consolidation makes Medallia an easy cut for CIOs, worsening retention and making the debt untenable.
- •Key insight: not only leverage kills deals—overpaying can be fatal even with moderate leverage
- •Debt service becomes impossible on low-growth software lacking a credible AI rewrite path
- •Medallia is not a hard system-of-record; switching is feasible and AI-first products are emerging
- •Vendor consolidation redirects budget to AI; ‘nice-to-have’ tools get cut before Workday/Salesforce
- 48:00 – 52:38
Private equity exit routes in trouble: refinancing cliffs, LP shock, and PE as ‘buyer of last resort’ disappears
From Medallia, the discussion expands to a broader PE stress narrative: debt trading below par, PIK toggles, and eventual refinancing cliffs. Rory argues losses in the ‘safe’ PE bucket can reduce risk appetite, and the disappearance of PE as a dependable exit route reshapes venture outcomes.
- •Distress indicators: loans trading below par → PIK features → refinancing cliff → recap/restructuring
- •LPs expected venture pain in 2021 vintages; PE losses feel more destabilizing because PE was the ‘stable’ allocator sleeve
- •PE as a routine exit/buyer-of-last-resort fades except at much lower prices
- •Older SaaS companies (2015–2022) without AI stories face fewer paths to liquidity
- 52:38 – 1:06:53
The venture knock-on effect: fewer but bigger winners, higher IPO bar, and founder-to-founder ‘key handoffs’
They argue the IPO window isn’t closed, but the bar is higher—requiring larger scale and faster growth, while strategics remain selective. Jason predicts more “give the keys to my friend” outcomes—mergers or leadership handoffs—when founders can’t see a viable exit path in the new funnel.
- •Exit funnel narrows: IPO must be big; strategics are targeted; PE less available
- •Portfolio construction implication: many companies wither, a few massive winners return the fund
- •IPO expectations rise (debate: 400M revenue vs 1B+ with strong growth)
- •Emerging pattern: founders merge/hand control to stronger operators to pursue consolidation and cash flow compounding
- 1:06:53 – 1:10:50
Seed-stage integrity: Gary Tan calls out ‘bullshit ARR’ and YC as market maker
The group discusses Gary Tan’s guidance on revenue definitions and why ambiguous metrics are becoming problematic. Rory frames YC’s move as both principled and strategic: as a dominant seed market maker, YC must enforce trust and standardization to keep the ecosystem liquid and credible.
- •Multiple conflicting ARR definitions can mislead investors and border on fraud
- •Transparency matters even at seed stage; unclear revenue harms long-term outcomes
- •YC’s guidelines may become a shorthand ‘seal of approval’ for revenue reporting
- •Market-maker logic: like public exchanges, a platform must police integrity to maintain trust
- 1:10:50 – 1:28:18
Retail access to private tech: Thrive Eternal, AngelList USVC, and the true cost of ‘private-market exposure’
They compare new vehicles that offer broader access to late-stage private companies and alternative ‘eternal’ assets. The discussion highlights fee drag (e.g., 3.61% annually) versus traditional VC fee+carry economics and questions whether returns at today’s valuations justify the costs.
- •Thrive Eternal is positioned more as ‘enduring non-AI-disrupted assets’ than Sequoia-style evergreen venture
- •AngelList USVC offers symbolic access to marquee names but raises valuation and fee questions
- •Fee debate: 3.61% feels huge in public markets but can resemble VC net fee drag when carry is included
- •Core question: can late-stage private mega-companies still compound enough to overcome the fee load?
