ExpertsFM20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founde
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch · June 13, 2026
5 Selected Lessons:
reasoning_pattern · 3 heuristic · 2
Value accrual as a time-dependent, contested phenomenon
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founde (June 13, 2026) · at 12:51 ▸
When this applies
when multiple layers of a stack (suppliers, platforms, integrators) are all racing to commoditize each other
The takeaway
In a multi-layer value chain, assume value capture rotates between layers over time rather than settling permanently with one; position your company to benefit from commoditizing adjacent layers.
Why it works
Each layer in a stack benefits when the other layers become interchangeable, so all players rationally try to commoditize their neighbors, and whoever currently holds pricing power shifts over time.
One level up
Durable value only accrues at the layer that remains scarce; everything else trends toward commoditization under competitive pressure.
Where it stops holding
This stops working when one layer achieves a durable technical or regulatory monopoly that prevents the usual commoditization dynamic (e.g., if one model provider becomes drastically superior to all others).
Detail
approach: Map which layer currently has pricing power, recognize that every layer is trying to commoditize the others, and identify the bear case (one player becoming so dominant it breaks the equilibrium).
signals:
- Everyone publicly claims the other layers are the ones becoming commoditized
- A single model/provider separating itself clearly from the pack in capability
- Customers struggling to track which model is 'best' week to week, favoring aggregators
rationale: Factory's own survival thesis depends on no single model provider running away with dominance; if one did, it would create pressure to standardize entirely with them, which is bad for the broader ecosystem too.
In their words
The reality is, value accrual is a time-dependent phenomenon. It's not like there is one person whose steady state gets all of the value. That's not how it works.
Token spend should track individual leverage, not a uniform org-wide ratio
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founde (June 13, 2026) · at 21:29 ▸
When this applies
when allocating a new variable cost (compute, tooling spend, expense budget) across a heterogeneous workforce
The takeaway
Set resource budgets per the actual value-creation mechanism of the role/person, not as a flat percentage applied uniformly across a function.
Why it works
Different individuals create value through fundamentally different activities (token-heavy parallel delegation vs. face-to-face relationship work), so a single blanket ratio either starves high-leverage users or wastes spend on low-leverage ones.
One level up
Uniform policies optimize for administrative simplicity at the cost of matching resources to where they actually create value.
Where it stops holding
This stops working when the organization lacks the visibility or trust to differentiate fairly between individuals, making uniform rules the only politically feasible option.
Detail
approach: Ask what the individual's actual value-creation mechanism is before assigning a token/resource budget.
signals:
- An org mandates one standard token-spend percentage for every engineer regardless of role
- Some individuals' highest-value work has nothing to do with the metered resource (e.g., face-to-face sales)
- Others delegate to many parallel agents and need outsized budgets
rationale: If your org has a standard number applied to everyone, you're probably painting with way too wide a brush; the median might land near salary-scale over three years but the variance across individuals will be enormous (0% to thousands of percent).
In their words
In fact, I would argue that if your org has a standard number where it's like we want every engineer to be at this percent of their salary and token use, you're probably painting with way too wide a brush.
Credentialism as a crutch for uncertainty in hiring
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founde (June 13, 2026) · at 28:23 ▸
When this applies
when decision-makers face irreducible uncertainty about future outcomes and reach for legible credentials as a substitute for judgment
The takeaway
Recognize credential-based filters as a response to your own uncertainty, not as a reliable predictor; replace them, where possible, with evidence of ownership and end-to-end agency.
Why it works
In the absence of certainty about what will actually matter (e.g., which AI labs will dominate), legible, validated signals like competition wins function as a psychological crutch even when they're weakly correlated with the trait actually needed.
One level up
The more uncertain an environment, the more organizations reach for false certainty via proxies — spot this pattern and discount it accordingly.
Where it stops holding
This stops working when the domain truly does require the deep, narrow technical skill that the credential was built to measure (e.g., some competition-specific skill is the actual job).
Detail
approach: Ask whether a credential measures ownership/agency over outcomes, or just narrow technical competition performance under a funnel someone else built for you.
signals:
- Reliance on a signal increases specifically when future uncertainty is high (e.g., 'what will OpenAI/Anthropic do to the application layer')
- Candidates who pursued competitions because a prescribed 'path to success' funnel told them to, versus those who sought it out independently from an unlikely background
rationale: The actual trait Factory looks for — owning outcomes end-to-end, demonstrating agency — doesn't reliably correlate with formal competition pedigree, and can even anti-correlate when the credential reflects following a funnel rather than self-directed choice.
In their words
In a world where we desperately seek certainty, we look for and the validators of someone being a math Olympiad, or you name it, whatever that is, that validation, in the wake of not having other certainty, that serves as a good crutch
Cost/quality/speed routing beats frontier-model default
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founde (June 13, 2026) · at 16:26 ▸ · at 23:49 ▸
When this applies
when allocating a scarce or expensive high-end resource across tasks of widely varying importance
The takeaway
Segment tasks by how much they actually need the premium resource, and spend disproportionately only on the small fraction of 'decision-making' tasks that matter most.
Why it works
Most tasks don't require peak intelligence/capability; routing tasks to the cheapest adequate resource and reserving the premium resource for the few high-stakes decision points captures most of the value at a fraction of the cost.
One level up
Resource allocation should mirror the actual distribution of stakes, not the ego or convenience of using the best tool everywhere.
Where it stops holding
This stops working when the cost of mis-routing a task to a weaker resource is catastrophic and hard to detect in advance, removing the safety margin for experimentation.
Detail
rule: Reserve frontier/expensive models for the small set of high-stakes planning/decision tokens; route the bulk of execution tasks to cheaper open models.
threshold: Roughly 80-90% of tasks today using frontier models could be done with open-source models, per Grinberg's estimate.
rationale: Mirrors how human orgs work: leadership spends few hours on irreversible strategic decisions while most hours go to lower-stakes implementation.
In their words
What percent of tasks today using Frontier models could be done with open source models? Probably 80 to 90%. It's typically the planning that really needs the frontier models.
Avoid customer concentration risk despite short-term convenience
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Model War: OpenAI, Anthropic or Open-Source | Token Maxing, AI Hangovers & The Coming ROI Reckoning | Labour Displacement Fears are BS & Overblown | From Physicist to Sequoia Founde (June 13, 2026) · at 1:07:59 ▸
When this applies
when any company evaluates the tradeoff between revenue growth from one dominant customer versus revenue diversification
The takeaway
Treat heavy customer concentration as an elevated risk factor for investors/stakeholders, independent of whether current unit economics look favorable.
Why it works
Concentrated revenue creates a single point of failure; losing or renegotiating with that customer can be existential, raising risk for investors even if the current steady-state economics look fine.
One level up
Risk assessment must weigh tail/failure scenarios, not just steady-state profitability.
Where it stops holding
This stops working when the concentrated customer relationship is contractually locked in for very long terms with strong switching costs in the company's favor, reducing the tail risk.
Detail
rule: View ~90% single-customer revenue concentration as bad primarily from an investor risk perspective, even if a steady state can exist.
threshold: ~90% of revenue from one customer, as cited in the conversation
rationale: Grinberg describes it as a 'sword of Damocles' — survivable but structurally risky.
In their words
It's bad if you're an investor in one of those companies because it's a little riskier. But I think you can find a steady state. It's just scary. You just know there's kind of a sword of Damocles above your head.