The Twenty Minute VCSpaceX's Financials Leaked: Is it Worth $2TN | Meta Debuts Muse Spark: Are They Back in the AI Race?
CHAPTERS
- 0:00 – 1:14
Anthropic’s Mythos withheld: what “too good at hacking” really means
The hosts react to Anthropic unveiling Mythos and choosing not to release it publicly due to its hacking capability. They debate whether the fear was justified, and what Mythos implies for the speed and autonomy of security exploitation.
- •Mythos reportedly finds large numbers of vulnerabilities and can operate agentically across big codebases
- •Debate: real safety concern vs. marketing/compute-constraint theater
- •Why speed/scale (agentic scanning) changes the threat profile vs. older models
- •Initial market reaction included broad fear and knock-on stock moves
- 1:14 – 4:31
Why Mythos is a step-change in cyber risk: “rifle vs machine gun”
Rory argues the key difference isn’t whether older models can find bugs, but how quickly and autonomously Mythos can do it. The chapter frames agentic capability as a quantitative leap that turns vulnerability discovery into mass exploitation.
- •Older models can be guided to similar findings, but require more human steering
- •Agentic automation turns vulnerability hunting into continuous, scalable discovery
- •Metaphor: machine-gun volume changes outcomes even if individual “bullets” are similar
- •Implication: defenders must assume missed bugs will be found quickly
- 4:31 – 7:35
The coming breach wave: AI-built apps, thin auth, and a worsening transition period
Jason predicts a near-term deterioration in security as more software is built quickly (often by AI) and shipped with basic mistakes. He shares an example breach and argues the ecosystem will eventually adapt, but not before a painful ramp in attacks.
- •AI accelerates both app creation and attacker capability, increasing exploit surface area
- •Example: newly acquired consumer app breached quickly due to missing/weak authentication setup
- •Expect a transition phase where security gets worse before it gets better
- •Bad actors may operationalize these tools broadly as costs drop
- 7:35 – 13:21
Cyber stocks selling off makes little sense: defense spend should rise
Rory challenges the idea that better offensive AI should hurt cyber vendors’ valuations. If attackers get “machine guns,” enterprises should buy more defenses, and new workflows/tools will be needed to operationalize AI-based code scanning and hardening.
- •AI-based pre-deploy scanning becomes a standard part of software release processes
- •Security vendors can build frameworks to harness models for defense
- •Arms race logic suggests increased demand for security, not less
- •Key shift: assume exploit discovery is near-certain, not probabilistic
- 13:21 – 17:39
Jason vs Dario: “boy who cried wolf,” doom messaging, and what’s inspiring
Jason says he’s fatigued by repeated AI-doom rhetoric and distrusts the framing around withholding Mythos. Rory pushes back that grand narratives can be sincerely held and culturally powerful even when the literal predictions are wrong.
- •Jason: constant job-loss/doomer messaging is demotivating and loses credibility over time
- •Rory: grandiosity can be a unifying rallying cry (Mars/Oppenheimer analogies)
- •Marketing value can coexist with sincere belief in risk
- •Tension: safety messaging vs enterprise/customer appetite for optimism and utility
- 17:39 – 22:02
Amazon’s $20B Trainium story: denting NVIDIA without “merchant silicon”
They unpack claims that Mythos was trained “on Trainium” and what Amazon’s chip business actually is. The conclusion: AWS is substituting some NVIDIA demand internally via bundled cloud services, which matters at the margin but isn’t a direct chip-market challenger.
- •AWS isn’t a large standalone merchant chip seller; it’s using Trainium for its own cloud offerings
- •Some Anthropic workloads may run on Trainium because AWS provides the compute
- •$20B annualized is meaningful displacement vs NVIDIA, though not an existential threat
- •Broader bear case: more custom silicon and inference optimization could compress NVIDIA multiples
- 22:02 – 24:58
Anthropic moves toward app-building: threat to Lovable/Replit/Cursor and the “maiming” effect
Discussion turns to Anthropic competing more directly with vibe-coding and developer tooling startups. Even partial product moves by a frontier model provider could pressure smaller players, without needing to fully replicate their stacks.
- •Debate over announcement vs shipped product; still, competitors assume it’s coming
- •Anthropic may not need full hosting/DB/OAuth/support to impact the market
- •Partial solutions could be enough to “maim” prosumer/dev-tool companies
- •Competitive pressure expands from models into application and workflow layers
- 24:58 – 41:50
Public SaaS “60% death spiral”: why agents must be monetizable on their own
Jason argues most incumbents are building underpowered, cost-constrained AI features that customers won’t pay extra for. Rory adopts this as a core valuation test: if you can’t charge independently for AI agents, you won’t re-accelerate growth and will be valued as a mature cash-flow name.
- •A 60% solution can’t be monetized; it becomes an included/free checkbox feature
- •Token costs and feature limitations push incumbents into lagging, locked-down offerings
- •Without re-acceleration, public SaaS moves into a different valuation bucket (value vs growth)
- •Moats trap existing customers but don’t create excitement or net-new growth in an agentic era
- 41:50 – 47:06
Meta’s Muse Spark and Alex Wang’s Super Intelligence Labs: back in the model race?
They assess Meta’s Muse Spark as “good enough” to signal a return to competitiveness after earlier disappointments. The strategic case is existential: Meta doesn’t want to be dependent on external model vendors, especially while its ad business is thriving.
- •Muse Spark seen as competent but not leading-edge on newest capabilities
- •Meta’s shift toward more closed approaches has ecosystem implications
- •Owning models may be strategic insurance for Meta vs buying tokens from rivals
- •AI didn’t kill Meta/Google—both ad businesses remain strong, supporting aggressive investment
- 47:06 – 57:11
OpenAI’s ad monetization plan: $100B ads still may not be enough
The hosts discuss leaked/projected OpenAI ad numbers and why ads feel inevitable for a consumer product at ChatGPT’s scale. Rory argues even a massive ad business may be insufficient relative to OpenAI’s implied scale, making enterprise revenue critical.
- •Ads are framed as obvious/inevitable given precedent from Google/Meta/Amazon
- •Projected ramp: small pilot traction scaling to very large multi-year targets
- •Even ~$100B ads could be only ~10% of global ad spend and may not justify valuation alone
- •Implication: OpenAI likely needs a second major pillar—enterprise—to match ambition/burn
- 57:11 – 1:02:45
Enterprise power shift: compute scarcity, CIO “token maxing,” and vendor standardization
They explore how enterprise AI buying may move from developer-led experimentation to CIO-controlled budgets and token allocations. This could change which vendors win (packaging, procurement, sales motion), and raises the stakes of OpenAI’s relationship with Microsoft.
- •Compute scarcity will drive throttling and price-based allocation of tokens
- •CIOs are setting fixed token/dollar budgets and forcing internal prioritization (“token maxing”)
- •Top-down standardization may favor vendors with enterprise-grade sales/procurement motions
- •If enterprise is the majority of AI spend, OpenAI–Microsoft alignment becomes strategically crucial
- 1:02:45 – 1:13:59
SpaceX leaked financials and the $2T IPO math: the “Elon discount rate”
They analyze SpaceX’s reported revenue and losses and what it implies for a potential $2T valuation. The discussion centers on how much of the valuation depends on future optionality (direct-to-cell, space data centers, etc.) and the assumptions investors make about timing and probability of success.
- •Headline multiple: ~$2T on ~$18.5B revenue implies ~108x revenue
- •Accounting nuance: losses may reflect acquired businesses and partial periods; true run-rate unclear
- •Bull case relies on multiple future initiatives and massive TAM expansion
- •“Elon discount rate”: markets may price future wins as certain and immediate (0% failure probability, 0% discounting)
- 1:13:59 – 1:23:43
Private markets stress: Thoma Bravo exits growth equity and PE’s AI transformation challenge
They interpret Thoma Bravo shutting its growth equity effort as a retreat to core control-buyout strategy amid stress in mature software. The larger question: can PE owners transform slow-growth SaaS into AI-upside assets, or are many headed toward value traps and leverage pain?
- •Growth equity is a distinct business from control LBOs; in stress, firms refocus on core
- •Many mature SaaS assets bought at high multiples now face lower comps and debt pressure
- •PE “transformation” must move beyond cost-cutting to building monetizable (not 60%) agents
- •Upsell to installed base could save outcomes, but cultural/product execution risk is high
- 1:23:43 – 1:30:30
Who IPOs first and leadership alignment: Anthropic vs OpenAI, CFO/CEO sync
They close with predictions on IPO ordering and discuss the operational importance of tight executive alignment for public-market readiness. The conversation highlights how leaks, reporting structure, and internal discord can undermine roadshow credibility.
- •Prediction: SpaceX first, then Anthropic, then OpenAI
- •Anthropic board additions/readiness moves are interpreted as IPO preparation signals
- •For OpenAI, CEO–CFO alignment and clean reporting lines matter for market confidence
- •Media/political-style drama increases scrutiny; internal discipline becomes a competitive edge
