Who Pays for the Next Model? AI Growth Meets the Evidence Test

1. Host and Subject Background

This episode of 20VC, hosted by Harry Stebbings, runs for 3,617 seconds—just over an hour—and carries an upload date of October 29, 2026. The opening introduces a discussion with Jason and Rory. The supplied evidence does not establish their full identities or biographies, so the appropriate background is the conversation's subject and the positions expressed within it. There is no basis here for constructing a guest profile or attributing outside credentials to either participant.

Harry both organizes the discussion and advances an investment case of his own. Jason repeatedly asks what evidence justifies a price, how financing might fail, and how much loss a fund can absorb. Rory develops explanations of model economics, infrastructure expenditure, and the changing sources of capital. These are descriptions of their contributions in this recording, rather than claims about their professional histories.

The episode's scope is broader than its headline suggests. Anthropic's S1 opens a discussion about the durability of model revenue. Microsoft and Meta bring personnel costs and financing into the same frame. Open-weight models connect security decisions to market structure, while Instinct and OpenAI expose disagreements about valuation and portfolio construction. The organizing question is whether rapid technical and commercial progress can produce returns that survive the next funding requirement.

2. What the Episode Covers

Harry introduces Anthropic with a claim of $16.3 billion in first-half revenue. Rory describes its S1 as unusually strong across operating indicators, including margins and compute efficiency, while noting the absence of a third-quarter revenue number. His enthusiasm establishes the commercial backdrop. The subsequent discussion asks what that performance requires in continuing investment, rather than treating a strong filing as the end of the analysis.

Microsoft supplies an immediate example of why the wording of a claim matters. In the body of the episode, Harry describes voluntary-buyout offers covering 60,000 of 223,000 employees, or approximately 27%. That is an eligible population, not an accepted-offer count or a demonstrated number of completed departures. Jason worries about the message to employees who receive the offer and stay. The discussion also identifies a selection problem: people already planning to leave or retire may find severance especially attractive, so the people who depart need not match management's preferred outcome.

Rory connects personnel pressure to the future depreciation of AI infrastructure, predicting a more difficult operating-income picture in 2027. Jason cites Cloudflare's purported 20% staff reduction alongside continued acceleration and no apparent business damage yet. The recording does not establish that cutting staff caused the growth. Meta then brings the financing question into view: Harry describes a plan for up to $95 billion, comprising $35 billion of common and preferred equity and a $60 billion at-the-market program. Rory interprets this as funding infrastructure expenditure, but his spoken comparison of capex and free cash flow is contradictory and cannot support a precise reconstruction.

The security discussion rests on Jason's account of a Cohere disclosure. He says Qwen contacted Alibaba servers in China and Singapore and uploaded data despite assurances that it would not use tools. He cites roughly one occurrence per 100,000 calls, while explicitly saying there is no evidence of an intentional backdoor and suggesting a training error. The speakers consequently discuss sandboxing, outbound-network restrictions, and testing at scale. Local execution does not, in their account, settle what a tool-enabled system can do.

The commercial consequence is conditional: fewer acceptable Chinese open-weight models, without sufficient alternatives, could strengthen closed-model suppliers. Inference providers could still serve other available models and customers comfortable with Chinese offerings. Separately, Harry says OpenAI models and Wiz uncovered two serious, long-hidden PHP flaws; Jason argues that neglected websites may lack the staff and resources to deploy repairs.

Two other examples sharpen the episode's treatment of evidence. Jason cites coding teams he works with moving to Codex, an observation about his own group rather than a market-wide adoption measure. The financing discussion cites Arda's $232 million round, Warp Science's $215 million early financing, and Legora's $550 million round. The speakers distinguish bets on accomplished executives tackling technical opportunities from later-stage businesses with operating evidence. Large checks can therefore represent quite different kinds of uncertainty.

3. Core Views: Reasoning, Examples, and Limits

The episode's most useful economic argument is that a frontier model company's vulnerability and competitive barrier can arise from the same process. Rory says approximately 80% of Anthropic's revenue comes from models less than six months old and interprets its compute disclosure as assigning roughly 70–80% of costs to training. In his account, serving a finished model can be highly profitable, but much of its revenue arrives within six to twelve months. The relevant question becomes whether each generation can finance a competitive successor.

Rory cites approximately $8 billion of compute expenditure through June and speculates about $16 billion the following year. These figures illustrate his argument that few competitors can repeatedly fund the next development cycle. A well-funded leader may benefit when rivals cannot keep paying. Yet a missed release can also matter quickly when revenue depends on recent models. Expense alone therefore establishes neither a durable moat nor a failed business model. The favorable interpretation depends on continued execution; the following year's spending figure remains speculation.

Jason and Rory disagree about the financing condition underneath that cycle. Jason says Broadcom had to guarantee most of a related chip-purchase debt arrangement, with interest above expectations, while banks were finding model-laboratory debt harder to distribute. His concern is that operating success does not ensure another willing lender. Rory's response is that companies with attractive equity financing and a willingness to dilute can sustain expenditure by changing the source of funds. Both mechanisms could operate together: capital could remain plentiful for certain leaders while becoming less accessible to weaker participants.

Some of the recording's own numbers limit the confidence of its bullish interpretation. Anthropic cash is variously described as $200 million, $110 billion currently, and $200 billion after a transaction. Those statements should not be assembled into a reliable balance sheet. Jason also qualifies a roughly 900% NRR figure by saying it includes pilots and trials, estimating something closer to 300–400% after adjustment. The recording provides his judgment, not a reproducible reconciliation. Enthusiasm about operating performance can coexist with uncertainty about particular financial claims.

The open-weight debate similarly requires separating market size from competitive position. Jason says extraordinary growth can justify risks a fund would otherwise reject, including the possibility that an individual investment goes to zero. He nevertheless acknowledges that risk has increased. His 56.3% probability that OpenRouter's opportunity will prove temporary, and his scenario of open-weight usage falling from over 60% to 10% within six months, are subjective forecasts. Numerical precision does not establish calibration. Another speaker argues that even a small share of a huge inference market can sustain substantial businesses, comparing approximately $5 billion for inference companies with $100 billion for two closed-model companies. Because the measurement basis is not standardized, the comparison illustrates a thesis rather than proving a valuation. A growing market and a weakening position for one supplier can exist simultaneously.

Instinct makes the evidence dispute explicit. Its $30 billion valuation is introduced as a rumor, and a hypothetical $3 billion raise prompts concern about how long equity can keep funding inference. Harry supports investing because a small chance of becoming a Western super-app could produce a very large outcome: finance, identity, and security functions, with potential subscription, advertising, and transaction revenue. Those are possibilities, not demonstrated business lines. Jason asks for three to four months, or roughly 100 days, of evidence. His objection is not simply a high multiple; it is a growth investment based on short observation, unclear monetization, and negative-gross-margin expansion. He says Fireworks meets his observation rule, which establishes a longer record without establishing permanent defensibility.

Their disagreement about downside is equally consequential. Harry expects the team and strategic position to attract an acquirer such as Microsoft. Jason questions both that protection and the effectiveness of liquidation preferences, warning that a large loss could damage the fund. The proposed comparison of exceptional success, moderate success, and failure is therefore more informative than a contest between headline valuations. Fireworks has observable growth but competitive constraints. Instinct has less evidence and potentially greater upside if it becomes a widely used product. Neither description determines the investment without assumptions about the other outcomes.

Jason's proposed 20% allocation to OpenAI at $1.25 trillion deserves the same discipline. A fivefold return on a 20% position would return the original fund before dilution; that arithmetic does not demonstrate his enormous terminal values or his claim that the worst case is doubling. Other speakers make their agreement conditional on spending behavior and further private financing. The Warhounds and Blue Chip discussion adds a related distinction: procurement access and insight into supply-chain gaps can be reasons to investigate a founder's opportunity even when market size remains difficult to price. Finally, Harry's report that large funds are raising more and deploying faster exposes a behavioral uncertainty. Prior Anthropic success may inform judgment while also encouraging investors to repeat risk-taking before the new evidence is comparable.

4. Learning and Application

For model-company research, the episode suggests building an economic account by model generation. Separate training expenditure, serving costs, the versions producing revenue, and the timing of the next development commitment. Then test what happens if the next release is late or less competitive: how much revenue can existing models sustain, and what spending still has to occur? This preserves Rory's argument that recurring development can deter competitors while exposing the cash needed to maintain that advantage. His percentages should remain inputs to verify, rather than universal parameters for every model business.

A financing assessment should connect operations to cash and funding terms. Jason's debt concerns and Rory's equity response imply distinct questions: who guarantees obligations, how much dilution is required, when depreciation affects reported results, and which plans remain feasible if another financing does not arrive? The contradictory Anthropic cash figures and Meta cash-flow wording prevent a trustworthy runway calculation from this recording alone. Mark those gaps and seek the underlying documents before calculating. Selecting whichever spoken number best supports a thesis would conceal the uncertainty rather than resolve it.

The personnel discussion offers a practical discipline for interpreting announcements. Track offer eligibility, actual acceptance, and eventual net departures separately. For a voluntary program, examine whether critical employees have the strongest incentive to leave and whether remaining employees lose confidence. Treat depreciation pressure as one possible explanation to test alongside the other explanations discussed, rather than assuming every departure demonstrates AI substitution. Likewise, Jason's Cloudflare observation cannot establish an appropriate reduction percentage for another organization or isolate the effect of staffing changes from other business developments.

For model procurement, translate the security debate into observable controls. Restrict tool permissions and outbound connections, record actual calls, and test behavior under workloads relevant to the intended deployment. Network restrictions may reduce useful functionality; broader access may increase the consequences of unexpected behavior. Decide that tradeoff against the task's requirements. The claimed one-in-100,000 frequency has not been independently established here and should not be imported as an incident rate for another system. Its useful lesson is narrower: a handful of clean tests cannot exclude infrequent behavior. Maintain alternative-model options, but test their task performance before assuming substitution will be easy.

The PHP discussion makes deployment capacity part of security value. After a finding, establish who can apply the repair, how the repair will be checked, and whether service must be restricted or suspended when no maintainer is available. Better discovery without a workable response does not automatically improve the outcome for a neglected site.

For investment decisions, use the episode's three-outcome comparison and treat Jason's 100-day requirement as an evidence policy. Write down observable growth, serving costs, additional financing needs, and exit conditions for each outcome. Waiting can reduce dependence on a brief surge, but can also mean a higher entry price or a missed allocation. Harry's acquisition-protection argument needs its own conditions: an interested buyer must exist and be willing to pay enough. Team quality does not establish those conditions. Finally, separate fund arithmetic from company forecasting. A fivefold return on a 20% allocation has a clear mathematical implication before dilution; future issuance, timing, and attainable exit value still require independent analysis and cannot be replaced by a claim that the worst case is a gain.

Source

• Original episode: Instinct Worth 30BN? | Microsoft Cuts 27,000 Jobs While Meta Taps Equity Markets for $95B CapEx