ExpertsFM20VC: Who Wins the Inference Market | Has Open-Source LLM Inference Peaked? | What Happens to the Cost of Compute | The Real Limits of Post-Training and RL with Erik Bernhardsson, Modal CEO
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch · November 2, 2026
5 Selected Lessons:
heuristic · 2 priority · 1 reasoning_pattern · 1 framework · 1
Three determinants of startup defensibility
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Inference Market | Has Open-Source LLM Inference Peaked? | What Happens to the Cost of Compute | The Real Limits of Post-Training and RL with Erik Bernhardsson, Modal CEO (November 2, 2026) · at 8:51 ▸ · at 11:04 ▸
When this applies
when assessing whether any young company has real staying power against competitors
The takeaway
Rank team quality, product velocity, and brand above classical moat theory when judging early-stage defensibility.
Why it works
Traditional moats (cornered resources, market power) take a decade to build; early-stage defensibility instead comes from execution speed and perception.
One level up
Static frameworks (like '7 Powers') describe end-state advantages, not the dynamic capabilities that create them.
Where it stops holding
This stops working once a company matures past ~10 years and actual structural moats (scale, switching costs, network effects) become the dominant source of defensibility.
Detail
hierarchy:
- Top 1% team
- Speed of product/feature shipping
- Brand
rationale: Classic moats (per Seven Powers) don't exist in early-stage companies; speed and perception substitute until true structural advantages can form.
In their words
I actually think like the three things that matter today when I invest is like, one, team. Are they top 1% team? Yeah. Two, speed of product decision making
Lead-time-based exponential demand forecasting for capacity purchasing
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Inference Market | Has Open-Source LLM Inference Peaked? | What Happens to the Cost of Compute | The Real Limits of Post-Training and RL with Erik Bernhardsson, Modal CEO (November 2, 2026) · at 20:19 ▸
When this applies
when purchasing any long-lead-time scarce input against fast-growing, compounding demand
The takeaway
When input lead time is fixed and growth is consistent, project current month-over-month growth forward by the lead-time window to size purchases.
Why it works
With a known growth rate and known lead time, extrapolating current growth forward covers the gap between order and delivery, converting a forecasting problem into simple exponential projection.
One level up
Supply decisions under lead-time lag require planning against projected future demand, not current demand.
Where it stops holding
This stops working when growth rates are volatile or non-exponential, making straight-line/exponential extrapolation unreliable.
Detail
procedure:
- Measure aggregate demand growth rate across all customers (30-40%/month observed)
- Apply straight-line or exponential extrapolation forward by the lead time (3-4 months)
- Resulting multiplier (~2-3x) determines how much capacity to buy now
- Be willing to pay above 'market' price to secure capacity since customers will also pay more
benchmarks: {"lead_time":"3-4 months","growth_rate":"30-40% month-on-month for prior 12 months","buy_multiple":"2-3x current capacity"}
In their words
The GPU lead times right now is about three to four months. So we basically do a straight line or exponential line extrapolation
Gingerbread man strategy (outrun, don't out-moat)
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Inference Market | Has Open-Source LLM Inference Peaked? | What Happens to the Cost of Compute | The Real Limits of Post-Training and RL with Erik Bernhardsson, Modal CEO (November 2, 2026) · at 8:51 ▸
When this applies
when a smaller player competes against a dominant incumbent that could theoretically clone any feature
The takeaway
When you can't build a moat, outrun your competitors instead of trying to defend ground.
Why it works
If an incumbent has no structural barrier to copying you, your only defense is being unreachable in speed, not protected by moat.
One level up
Speed can substitute for structural advantage when none yet exists.
Where it stops holding
This stops working once the company matures and speed advantages erode, requiring a transition to genuine structural moats.
Detail
approach: Accept you have no moat; focus entirely on outpacing competitors in execution speed.
signals:
- Competitor could technically replicate your product
- No regulatory, capital, or resource lock you have
rationale: Snap's only strategy against Facebook was moving faster, not building defenses Facebook couldn't copy.
In their words
Snap doesn't have a mode, they just run faster. Like, they're the gingerbread man. And I actually always come back to that. I think that's so true.
Breadth of product avoids commoditization in infrastructure markets
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Inference Market | Has Open-Source LLM Inference Peaked? | What Happens to the Cost of Compute | The Real Limits of Post-Training and RL with Erik Bernhardsson, Modal CEO (November 2, 2026) · at 13:04 ▸ · at 15:09 ▸
When this applies
when selling a unit-priced resource that competitors can match on price alone
The takeaway
Bundle a commodity core with adjacent, hard-to-replicate products to create switching costs and escape pure price competition.
Why it works
Single-product, unit-priced offerings have zero switching costs — buyers can swap providers instantly, driving a race to the bottom. Bundling adjacent capabilities creates integration costs that make switching painful.
One level up
Differentiation in commodity markets comes from ecosystem breadth, not from the commodity itself.
Where it stops holding
This stops working when the adjacent products are themselves easy to replicate or customers only want the core commodity unbundled.
Detail
rule: Don't sell a single-trick commodity; bundle it with a portfolio of adjacent tools customers need.
rationale: If you're just selling tokens for dollars, there's no switching cost and it's a race to the bottom; a portfolio creates staying power because the infrastructure behind it is hard to copy.
In their words
Yeah, because otherwise you're just selling commodity product. And if you're just selling tokens for dollars, there's no switching cost. People can just switch to different provider. It's just a race to the bottom.
Post-training/RL works best where feedback loops are short
Source
The Twenty Minute VC (20VC): Venture Capital | Startup Funding | The Pitch — 20VC: Who Wins the Inference Market | Has Open-Source LLM Inference Peaked? | What Happens to the Cost of Compute | The Real Limits of Post-Training and RL with Erik Bernhardsson, Modal CEO (November 2, 2026) · at 41:33 ▸
When this applies
when deciding whether a task/process is a good candidate for a tight feedback-loop optimization method (training, agentic automation, iterative experimentation)
The takeaway
A process is a good fit for rapid iterative optimization only if it has a fast, well-defined feedback signal; longer feedback loops resist this approach until new mechanisms are found.
Why it works
RL requires a measurable, fast feedback signal to iterate against; tasks like coding (testable in sandboxes) or math (verifiable proofs) have this, while tasks with multi-month human feedback loops don't yet have an analogous mechanism.
One level up
The feedback loop length of a task is often a better predictor of automatability than its apparent complexity.
Where it stops holding
This stops working when new techniques emerge that can simulate or compress long feedback loops (e.g., proxy metrics, simulation environments) that didn't exist before — Bernhardsson himself flags he's 'been wrong for so long' on these predictions.
Detail
rule: Post-training/RL works well for tasks with fast, finite feedback loops (coding, cybersecurity, math); longer feedback-loop tasks remain out of reach for now.
rationale: You can run code in sandboxes and verify success quickly; tasks with multi-week/month human feedback loops lack an equivalent verification mechanism yet.
In their words
I don't know. There's a lot of tasks out there that have way longer feedback loops that I don't know if post-training quite yet is within reach.