Will foundation models subsume application-layer AI startups built on workflow automation?

Des Traynor · Nov 15, 2023

The best LLM for customer support would require hundreds of millions of customer support conversations to train, which would advantage a company like Intercom that holds that data.

Domain-specific conversational volume is the input that would make a uniquely better support model.

Scope: Framed as the alternative scenario to commoditization

14:07 20VC: How to Survive and Thrive in a World of OpenAI, Are LLMs Being Commoditised, Where Does the Value Lie; Infrastructure or Application Layer, How Apple Could Win in a World of AI, How Amazon Could Threaten OpenAI and Why Google Struggle with Des Trayn

David Frankel · Oct 14, 2024

Incumbent vertical and horizontal SaaS companies that own large proprietary data sets are the natural winners of the AI wave and it is theirs to lose

They hold truckloads of customer data; layering AI tools onto that data creates an enormous opportunity, whereas failing to enrich the data is what puts them out of business

Scope: applies to data owners specifically

63:52 20VC: Investing Lessons from FC Seeding Uber, Airtable and Coupang | Why Pro Rata is the Original Sin in VC | Why Liquidity Has Died in 2024 | Why LPs are Pissed with VCs | The Hard Truth About Seed Fund Economics with David Frankel @ Founder Collective

Jake Saper · Mar 10, 2025

B2B software vendors retain an important role even when tools like Bolt and Cursor make coding cheap or free, because customers are buying an opinionated perspective and proprietary data, not just code.

A vendor who has dedicated their life to a problem across many use cases has insight and proprietary data that no off-the-shelf closed-source model provides.

Scope: acknowledges the take is self-serving as an investor in B2B software vendors

47:35 20VC: Lessons from Investing $2BN and Returning $8BN in Cash | Why Most Venture Partnerships are Broken | We Sold Salesforce Early and Lost Out on Billions | Are The Best Deals Always Expensive and Competitive with Jake Saper @ Emergence Capital

Nabeel Hyatt · Apr 4, 2025

An AI-native product with the best users in its field accumulates expert data exhaust that compounds into better products and better models

Descript doesn't just own a reinvented interface, it sits on every edit every user made turning rough cuts into polished production — internalizing the wisdom of experts

Scope: requires being a top-of-market product with the best users

51:54 20VC: Why To Win in AI, Investors Need to Change Their Approach | Why VC is Run by Principals and Associates and is a Broken System | The Bull Case for Anthropic & Whether Deepseek Changes Their Strategy with Nabeel Hyatt @ Spark Capital

Harry Stebbings · Apr 14, 2026

Vertical model companies built on non-proprietary data get commoditized by the frontier labs, as Cursor has been, while companies generating proprietary physical data like Periodic Labs will not be

Periodic's data is produced physically in its own lab and therefore can't be replicated by a general model

Scope: stated as a premise while asking how to tell the two apart in advance

10:55 20VC: Anj Midha on Investing $300M into Anthropic | The Early Days of Anthropic & How 21 of 22 VCs Turned it Down | The Four Bottlenecks to Compute | What the China Has Smashed and Why We Should Be Worried

Becca Lindquist · May 2, 2026 · hedged

When choosing which AI company to join, look for defensibility beyond 'we automate this workflow' — such as Clay's data marketplace — because pure automation businesses are vulnerable to being subsumed by foundation models.

Many AI companies trigger the 'Claude spookies' — the question of whether Claude will just do this and put you out of business; a data marketplace built from ~180 providers is genuinely hard to replicate.

Scope: admits she can't name another company clearly doing this; 'the real answer is I don't really know'

22:35 20VC: Inside Clay's Sales Playbook Scaling to $100M ARR | How to Set Sales Comp Plans | How to Read Sales Talent Linkedin Profiles | What Profiles to Hire & Fire | How to Increase Performance and Speed in Sales Teams with Becca Lindquist

Shiv Rao · May 16, 2026

Defensibility against foundation models comes from going miles deep in a regulated industry with proprietary data sets built into very specific workflows, plus reaching scale fast enough to fuel mid-training and post-training on user edits

Scale gives you daily learning from user edits across every type of doctor, care setting and spoken language, which is hard to replicate

Scope: especially on the enterprise side of healthcare

16:39 20VC: Lessons from Jensen Huang on "Founder Mode" | How to Know if OpenAI or Anthropic Will Kill your Company | How USV Liking Music Made Them $1BN on an Investment | The Five Year Desert to Product Market Fit & a $5.3BN Valuation with Shiv Rao @ Abridge

Mike Mignano · Jul 6, 2026

Accumulated context inside an organization is the key defensibility for AI products, because an enterprise won't give up a rich history of information that makes it work better

Once everyone in an organization uses the product and accumulates valuable context, that store of information becomes something the enterprise doesn't want to abandon

Scope: offered as a counterweight to a bundling threat he admits is real

40:02 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Mati Staniszewski · Sep 8, 2025

Even if OpenAI matches the research, the durable advantage sits in the product layer — creative workflow depth for narration/voiceover, and knowledge bases, integrations, functions, deployment, testing, evaluation and monitoring for voice agents — and OpenAI is not investing there

making voice output perfect or making an agent production-ready requires many additional steps that only a dedicated platform builds

Scope: OpenAI 'could' invest there but currently isn't

14:10 20VC: ElevenLabs Hits $200M ARR: The Untold Story of Europe's Fastest Growing AI Startup | The Real Cost of AI from Talent to Data Centres | How US VCs are in a Different League to Europeans | The Future of Foundation Models with Mati Staniszewski

Eran Zinman · Mar 2, 2026

The AWS precedent shows infrastructure providers don't capture application value: predictions that Amazon would take all enterprise software value were wrong, and instead a boom of companies built on top of AWS

Before AWS the hard part was servers, storage and uptime, so people assumed Amazon would own the value; in fact software grew exponentially on top of it

12:22 20VC: Monday.com CEO on Is SaaS Dead: Will Everything Be Vibe Coded | Will Systems of Record Become Valueless Databases in an Agentic World | Will LLMs Own the Value in the Application Layer with Eran Zinman

Aaron Levie · Apr 20, 2026

There will still be a lot of money to be made by companies that apply frontier lab innovation to real-world enterprise work, whether as vertical AI or new enterprise tooling

The innovation happening in the labs still needs to be translated into the actual work that happens inside enterprises

Scope: framed as a generic statement rather than a specific pick

48:13 20VC: Everyone is Wrong; We Will Have More Developers in Five Years | Why Frontier Labs Will Be Way More Valuable Than They Are Today | Are SaaS Companies Cooked: Which Thrive & Which Die with Aaron Levie, Founder at Box

Aravind Srinivas · Jun 15, 2026

A pure reseller of model tokens has no business, because models will be commoditized — even model builders have no business at that layer

Models get commoditized; only the infra layer captures value on serving raw output tokens, while application layers need grounding, orchestration and connectors to create value

Scope: infra layer retains some business serving output tokens

13:02 20VC: Micron Will Be More Valuable Than Meta | How Export Controls Helped Not Hurt China | Power is the Bottleneck to AI | Why Dario Has Done a Disservice to AI with his Labour Replacement Messaging with Aravind Srinivas, Founder @ Perplexity

Arvind Jain · Jul 11, 2026

Founders building outside frontier model training should treat the model labs as an asset rather than a competitor and not worry about being encroached on

The labs' and open source's progress enables products these companies could never have delivered on their own; founders should focus on solving problems while tracking lab capabilities

Scope: applies to AI companies not doing frontier model training; still need to anticipate what labs will do

13:14 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Lin Qiao · Jul 20, 2026

Frontier labs are building valuable foundational infrastructure — a power line distributing intelligence others build on — but that infrastructure will not replace everything companies do

Like electricity, it enables appliances and branded, taste-encoding products rather than substituting for them

11:59 20VC: Are OpenAI and Anthropic Overvalued? The Open-Source AI Reality | How Token Costs Will Fall 10x And Usage Will Explode 100x | The Future Is Not One AGI; It's Millions of Specialised Models with Lin Qiao, Founder and CEO @ Fireworks

Des Traynor · Nov 15, 2023

Thin AI integrations layered on top of a third-party model do not create disruption, because the same capability is equally available to incumbents and startups

If you're shelling out for your differentiation to a third party, that differentiation is available 'on top' for Mailchimp and for the new startup alike; real disruption comes only when you'd build the entire thing differently post-AI

46:40 20VC: How to Survive and Thrive in a World of OpenAI, Are LLMs Being Commoditised, Where Does the Value Lie; Infrastructure or Application Layer, How Apple Could Win in a World of AI, How Amazon Could Threaten OpenAI and Why Google Struggle with Des Trayn

Delian Asparouhov · Jul 29, 2024

Software's zero marginal distribution cost is matched by zero marginal returns, because there is no moat — a couple of engineers with today's AI can replicate a vertical SaaS product in a weekend

With modern AI, replicating a known software product is trivially cheap, so no defensibility accrues

53:49 20VC: Twitter's Most Controversial VC Delian Asparouhov on Inside the Walls of Founders Fund: What the World Does Not See | Why Western Europe Will Be Like the Third World | Why SaaS as an Industry Might Be Dead

Delian Asparouhov · Jul 29, 2024

Deep software workflows can be replicated very quickly today by a team of ten strong engineers using AI, once the workflows are a known outcome rather than something you must invent from customer discovery

Inventing workflows for the first time is the hard part; copying an already-proven workflow is fast with AI tooling

Scope: applies to replication of known workflows, not first-time invention; distribution and branding advantages are real and not replicable this way

55:13 20VC: Twitter's Most Controversial VC Delian Asparouhov on Inside the Walls of Founders Fund: What the World Does Not See | Why Western Europe Will Be Like the Third World | Why SaaS as an Industry Might Be Dead

Nabeel Hyatt · Apr 4, 2025

Most vertical agent startups are near-term arbitrage that will take their own market to zero, because dozens of competitors will build the same call agent into that vertical tomorrow

If you meet the market where it is today with the models of today and your innovation ends there, you'll be awash with 50 other people out of other incubators doing the same thing and your marginal benefit to the world is zero

Scope: applies to companies built for today's models with no second, third or fourth act

54:00 20VC: Why To Win in AI, Investors Need to Change Their Approach | Why VC is Run by Principals and Associates and is a Broken System | The Bull Case for Anthropic & Whether Deepseek Changes Their Strategy with Nabeel Hyatt @ Spark Capital

Brendan Foody · Jun 1, 2026

Within the next twelve months models will be able to clone products like Slack end to end, which is very bad for companies betting on software moats

2025 was the year of getting a model to open a PR in a codebase; 2026 is the year of cloning a product end to end, and Mercor's eval sets measuring agents building full SaaS applications show those capabilities arriving

Scope: next twelve months

36:30 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility, The Model is the Product | Token Spend Will Exceed Headcount Spend in 5 Years | The True Cost of Hiring AI Researchers in the Valley Today with Brendan Foody

Lin Qiao · Jul 20, 2026

Implementation is no longer a moat in software — the application development lifecycle has collapsed from tens of engineers over multiple quarters to one person in a few weeks, so competition must be redefined

General coding intelligence has collapsed the timeline and resources needed to go from idea to production scale, and ideas alone don't differentiate since many people have similar ones

22:34 20VC: Are OpenAI and Anthropic Overvalued? The Open-Source AI Reality | How Token Costs Will Fall 10x And Usage Will Explode 100x | The Future Is Not One AGI; It's Millions of Specialised Models with Lin Qiao, Founder and CEO @ Fireworks

Vince Hankes · May 3, 2023

The wave of AI copywriting and content startups were thin user experiences layered on top of the genuinely valuable thing, the underlying model

Spending time with many of these companies, the team kept concluding the differentiation sat in the model, which almost no one was discussing in the mainstream at the time

14:32 20VC: The OpenAI Memo: Why Invest? Is it too Late to Catch OpenAI? Are OpenAI's Models Truly Defensible? Does the Value in AI Accrue to Incumbemts or Startups - Application Layer/Infrastructure? What Happens with Regulation? with Vince Hankes @ Thrive

Emad Mostaque · May 17, 2023

Most AI startups today are good ideas rather than businesses — surface-level wrapper layers with no thinking about distribution or data, which are the only real moats

Innovation at the core is not a business; it becomes one when innovation becomes product and distribution and has a data advantage — as with Stability going to Amazon Bedrock for 100,000 SageMaker SMEs, or OpenAI using Microsoft for distribution

20:26 20VC: Why the AI Bubble Will Be Bigger Than The Dot Com Bubble, Why AI Will Have a Bigger Impact Than COVID, Why No Models Used Today Will Be Used in a Year, Why All Models are Biased and How AI Kills Traditional Media with Emad Mostaque, Founder & CEO @

Harry Stebbings · May 17, 2023

Almost all of the new AI companies being pitched today lack a defensible moat — they are thin application layers on top of someone else's model with no proprietary data

He has met 50+ AI companies in the last month and the feedback on them is consistently the same

Scope: based on the last month of meetings; '99% of the feedback'

32:53 20VC: Why the AI Bubble Will Be Bigger Than The Dot Com Bubble, Why AI Will Have a Bigger Impact Than COVID, Why No Models Used Today Will Be Used in a Year, Why All Models are Biased and How AI Kills Traditional Media with Emad Mostaque, Founder & CEO @

Harry Stebbings · May 22, 2023

Investors generally dismiss consumer subscription companies built on OpenAI models as uninteresting thin layers on someone else's model with no proprietary data.

40:54 20VC: Why Your Fund Model Should Not Rely on $10BN+ Outcomes, Why the Large Funds Got Too Large, The Rise of Solo GP's; The Pros and Cons & Is Consumer Subscription Even a Good Sector to Invest in with Nico Wittenborn @ Adjacent

Harry Stebbings · Nov 22, 2023

Around 90% of the AI companies he sees are shallow 'AI for X' products that were clearly not hard to build

They are only three weeks old when pitched

Scope: based on his own deal flow

25:43 20VC: Why OpenAI Will Become an Infrastructure Play, Why Apple Will Win in an AI World, Why Google is the Most Vulnerable Incumbent, Will LLMs Be Commoditised, Which Startups Are Thin vs Thick Wrappers on Top of LLMs with Jeff Seibert, Founder @ Digits

Anish Acharya · Feb 9, 2026

Models are specializing (Gemini for front end, Codex for back end; Midjourney/Krea aesthetically opinionated, Ideogram deliberately not), so users need an app layer to orchestrate multiple models rather than switching between CLIs

Vibe coders want both front-end and back-end strengths without switching tools, and creatives at big companies need both opinionated and unopinionated image models depending on the job

12:48 20VC: Is SaaS Dead in a World of AI | Do Margins Matter Anymore | Is Triple, Triple, Double, Double Dead Today? | Who Wins the Dev Market: Cursor or Claude Code | Why We Are Not in an AI Bubble with Anish Acharya @ a16z

Eran Zinman · Mar 2, 2026

There is a large horizontal opportunity to be the default workspace that orchestrates collaboration between humans and agents, alongside the vertical AI tools companies will adopt

Agents will output tables, docs and files that humans review and build on, and there will be a long period where humans and agents must work together

Scope: agents doing 100% of the work is possible in the far future but he doubts it

29:52 20VC: Monday.com CEO on Is SaaS Dead: Will Everything Be Vibe Coded | Will Systems of Record Become Valueless Databases in an Agentic World | Will LLMs Own the Value in the Application Layer with Eran Zinman

Aravind Srinivas · Jun 15, 2026

Whoever solves orchestration and maximizes token value per user will capture the most economic value in AI long term, even if short term the exponential revenue growth appears to belong to the labs

Maximizing token value for the user is the one objective that truly matters over the long run

Scope: short term other labs' revenue growth may look more impressive

27:42 20VC: Micron Will Be More Valuable Than Meta | How Export Controls Helped Not Hurt China | Power is the Bottleneck to AI | Why Dario Has Done a Disservice to AI with his Labour Replacement Messaging with Aravind Srinivas, Founder @ Perplexity

Aravind Srinivas · Jun 15, 2026

Perplexity is best positioned to win the orchestrator role because its product improves whenever any layer of the AI stack improves, so it doesn't depend on any single player winning

It has no incentive to token-max, only to deliver user value; Anthropic's model progress and OpenAI's price competition both improved Perplexity's product and burn, and revenue more than tripled since the start of the year

29:02 20VC: Micron Will Be More Valuable Than Meta | How Export Controls Helped Not Hurt China | Power is the Bottleneck to AI | Why Dario Has Done a Disservice to AI with his Labour Replacement Messaging with Aravind Srinivas, Founder @ Perplexity

Des Traynor · Nov 15, 2023

OpenAI will only ever build to a 'good enough for everyone' minimum viable level in any given vertical, so deep vertical products remain viable while shallow ones die.

Analogous to iOS shipping a basic notes app and basic camera while people still buy Bear or Photoroom; OpenAI will never put five engineers deep on wealth management and banking integrations.

Scope: Concedes shallow single-purpose bots like a 'wealth management advisor bot' are indeed dead

12:27 20VC: How to Survive and Thrive in a World of OpenAI, Are LLMs Being Commoditised, Where Does the Value Lie; Infrastructure or Application Layer, How Apple Could Win in a World of AI, How Amazon Could Threaten OpenAI and Why Google Struggle with Des Trayn

Max Junestrand · Aug 15, 2025

For application-layer builders, the floor is always ChatGPT, and because ChatGPT keeps improving quickly application companies must keep getting better than it

The general platform product sets the baseline that any application must beat

Scope: from the perspective of an application builder

76:16 20VC: 15 Term Sheets in 7 Days and Choosing Benchmark | Harvey vs Legora: Who Wins Legal and How to Play When You Have $600M Less Funding | Are AI Models Plateauing Today | Building a 9-9-6 Culture From Stockholm with Max Junestrand

Everett Randle · Nov 10, 2025

The labs set the baseline customer experience for AI app companies, so an app must be sufficiently differentiated from what a $20-200/month ChatGPT subscription delivers

the labs build apps on top of their own models and ship them directly to users, and app companies need to charge far more than $20/month to have a sustainable B2B business

Scope: applies to a lot of categories, not all

29:36 20VC: Benchmark's Newest General Partner Ev Randle on Why Margins Matter Less in AI | Why Mega Funds Will Not Produce Good Returns | OpenAI vs Anthropic: What Happens and Who Wins Coding | Investing Lessons from Peter Thiel and Mamoon Hamid

Winston Weinberg · Jan 19, 2026

The biggest existential threat for application-layer AI companies is not moving fast enough on product, because model improvement erodes product value unless there is a massive delta versus an Enterprise GPT license

The labs don't need to target legal or tax specifically — just improving their products and models lowers the value of yours, so you must reach escape velocity on product moat

Scope: applies to application-layer companies generally

21:39 20VC: How Model Performance is Plateauing | Two Key Rules for Effective Deal-Making | Company Building Lessons from Keith Rabois, Brian Halligan and Pat Grady | Why Enterprise AI Adoption is Years Off with Harvey CEO Winston Weinberg

Harry Stebbings · Mar 10, 2025

Being deeply embedded in a specific profession's core cross-team workflows is what makes an AI company's retention durable

Cites Solve as an example — deeply entrenched with patent lawyers and core to workflows across teams

44:30 20VC: Lessons from Investing $2BN and Returning $8BN in Cash | Why Most Venture Partnerships are Broken | We Sold Salesforce Early and Lost Out on Billions | Are The Best Deals Always Expensive and Competitive with Jake Saper @ Emergence Capital

Eran Zinman · Mar 2, 2026

OpenAI, Anthropic and Google will not win the human-agent collaboration workspace because their products are personal tools, not multiplayer systems where thousands of people share information

Selling a tool for individual use is a fundamentally different product and different size of product than building collaboration between agents and teams

31:08 20VC: Monday.com CEO on Is SaaS Dead: Will Everything Be Vibe Coded | Will Systems of Record Become Valueless Databases in an Agentic World | Will LLMs Own the Value in the Application Layer with Eran Zinman

Eran Zinman · Mar 2, 2026

OpenAI, Anthropic and Google will not pursue the human-agent collaboration opportunity because it requires a completely different top-down enterprise sales process and tool usage

The sales motion and product usage are too different from what the labs do, while Monday's advantage is already being the best cross-organizational work platform

36:45 20VC: Monday.com CEO on Is SaaS Dead: Will Everything Be Vibe Coded | Will Systems of Record Become Valueless Databases in an Agentic World | Will LLMs Own the Value in the Application Layer with Eran Zinman

Matt Murphy · Jul 27, 2026

Legora is defensible against foundation models because legal work crosses organizational boundaries with multiple constituents, requiring context and workflows a model can't just walk in and provide

Corporate lawyers, law firms, clients and multiple firms on a case create a near n-squared coordination problem, and Legora has lawyers and FTEs embedded understanding those workflows

30:01 20VC: Leading Anthropic's First Ever Round | Will Open Source Threaten Anthropic's Business | Do Margins Matter in a World of AI | Why Triple, Triple, Double, Double is Not Good Enough Today | Why Series A is Hard Today with Matt Murphy @ Menlo

Nico Wittenborn · May 22, 2023 · hedged

Most software value over the last twenty years has been captured in the application layer that verticalizes breakthroughs for a specific use case or customer demographic and monetizes them, so lacking a proprietary model is not disqualifying.

The value historically accrued to whoever encoded and monetized the breakthrough for a specific workflow, and AI is now taking that automation to the next level in both enterprise and consumer applications.

41:09 20VC: Why Your Fund Model Should Not Rely on $10BN+ Outcomes, Why the Large Funds Got Too Large, The Rise of Solo GP's; The Pros and Cons & Is Consumer Subscription Even a Good Sector to Invest in with Nico Wittenborn @ Adjacent

Sarah Tavel · May 6, 2024

Much of the defensible value in AI companies sits beyond the model itself — in workflow, integration and customer-specific customization that model labs won't build

DeepL started as a copy-paste translation site built on its own model but gradually layered workflow, local vocabulary lists and seamless integration on top, and an OpenAI is not going to build that same depth of workflow

31:25 20VC: Benchmark's Sarah Tavel on Are Foundation Models Commoditising | Why Frontier Models Will Be Closed Source | Why the Value is in the Application Layer | The Future of AI is "Selling the Work" Not the Tools

Tom Hulme · May 8, 2024

At the application layer a company must have something proprietary in data or distribution to generate sustainable value

Without proprietary data or distribution the business is as ephemeral as the foundation models; Synthesia works because it built an end-to-end enterprise-ready solution and go-to-market in a space with no real incumbent

Scope: application layer specifically

50:14 20VC: GV's Tom Hulme on Why Investing in Foundation Models is like Investing in "Power Stations", The Conventional Wisdom in VC that is BS & Lessons from a 24x Angel Track Record, 255x on Robinhood and Making Billions on Uber

Mike Krieger · Mar 3, 2025 · hedged

Durable value in AI will accrue to companies that own at least one, ideally two or three, of differentiated go-to-market, deep industry-specific knowledge, or uniquely accessible data

Those assets are hard to replicate, let you pull on what's great from foundation models while getting better the longer you're deployed in the domain, and they can't be built in an accelerator or a short sprint

Scope: he says he has no perfect answer because it's hard to have a crystal ball; cites financial, legal and healthcare sectors as examples

4:28 20VC: Anthropic CPO Mike Krieger: Where Will Value Be Created in a World of AI | Have Foundation Models Commoditized | When Do Model Providers Become Application Providers | What Anthropic Learned from Deepseek

Maor Shlomo · Nov 24, 2025

Lovable, Replit and Bolt are not the competitors worth worrying about; the real risk is market dynamics among the model providers

There is a durable advantage to players who work across multiple model providers, but if one provider decisively wins the model race, its next logical move is to conquer vibe coding as the largest software category

Scope: 'in my nonobjective eyes' on vibe coding being the largest software category

22:52 20VC: Base44's Maor Shlomo on How Vibe Coding Will Kill SaaS and Salesforce | Why it is BS that Vibe Coding Platforms Do Not Have Defensibility and Bad Margins | Why He Worries About Google, Not Replit and Lovable | Why Long Anthropic, Not OpenAI?

Anish Acharya · Feb 9, 2026

For categories that need multi-model access and rich feature surface, being an apps company is the better position

Model companies have ambitions in too many directions to prioritize opinionated UIs for niches like the legal community, and OpenAI, Anthropic and Google will each only ever serve their own models

Scope: Concedes bundling an 80%-feature version into an existing solution may capture the majority of users

18:53 20VC: Is SaaS Dead in a World of AI | Do Margins Matter Anymore | Is Triple, Triple, Double, Double Dead Today? | Who Wins the Dev Market: Cursor or Claude Code | Why We Are Not in an AI Bubble with Anish Acharya @ a16z

Miles Clements · Mar 9, 2026

Being multi-model makes Cursor effectively an index on AI innovation, giving it a compounding product flywheel.

The product improves both from Cursor's own feature work and from every improvement in the underlying models

9:08 20VC: Inside Accel's $4BN Growth Investing Machine | Cursor is Dead is Total BS: Here is Why | What Missing Rippling and ElevenLabs Taught Us | Are $2BN-$10BN IPOs Dead | Why Now is a Great Time to be Thoma Bravo with Miles Clements

Peter Singlehurst · Mar 19, 2025

Companies from the pre-AI era can still be exceptional where their product involves foundational difficulties AI cannot easily replicate, such as managing regulation in financial technology

AI will have an impact, but some foundational problems in building those products won't be blown apart by AI tools

Scope: concedes Harry's concern is a possibility; fintech/regulation given as the illustrative area

19:11 20VC: The 10 Question Framework a $217BN Manager Uses to Make Investment Decisions | Lessons from Turning Down Stripe, Coinbase and Losing Money on Northvault | The Bull Case for Bytedance | How Anduril Could Be a $200BN Company with Peter Singlehurst

Mike Mignano · Jul 6, 2026

In highly regulated industries like healthcare, years of relationships, partnerships and regulatory clearance constitute a real moat against labs and hyperscalers walking in

Abridge spent close to ten years clearing those hurdles before hitting its inflection point, and you can't just show up and say you're doing healthcare now

Scope: specific to highly regulated industries

34:17 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Fred Turner · Jul 18, 2026

Complex, regulated industries will capture the benefits of frontier models without facing competition from Anthropic or OpenAI

The model providers aren't going to start an insurance company or enter regulated complex industries

Scope: applies to regulated, complex industries

67:58 20VC: $5BN in Revenue, 7 to 7,000 Employees in 9 Months, 206,000 Tests in a Single Day: The Craziest Story in Startups: Curative with Fred Turner

Bucky Moore · May 5, 2025

The current model-provider expansion into applications will play out like the rise of AWS, Azure and Google Cloud: most application categories will remain defensible for focused startups

Many software categories require unique understanding of customer pain, dedicated resourcing and insights that only come from focus, which platform providers cannot replicate across every category

10:03 20VC Exclusive: Why Mega Platforms Will Win in VC | Why You Cannot Do VC If You Do Not Do Pre-Seed | Why Market Sizing is BS | Where Will Foundation Models Build/Buy Apps vs Where Will They Not with Bucky Moore

Harry Stebbings · Jul 21, 2025

The application layer will not be absorbed by the model layer

There is infinite product breadth for labs to go after and they cannot cover everything

57:26 20VC: Scaling to $1BN+ in Revenue with No Funding: Surge AI | The Most Insane Scaling Story in Tech |

David George · Dec 15, 2025

Model labs shipping application products doesn't threaten independent app companies, because each is one of thirty things in their AI division

Analogous to AWS offering a service for basically everything in cloud while many independent infrastructure companies still thrive

56:06 20VC: a16z's David George on How $BN Funds Can 5×, Do Margins & Revenue Matter in AI & the Most Controversial Bet at a16z

Alex Lebrun · Jun 19, 2023

It is not fair to call generative AI applications a thin, valueless layer on top of foundational models

The same was said of C in 1972 and of databases in the eighties, and it was untrue; LLMs are a new kind of infrastructure that is genuinely hard to build on — non-deterministic, hallucinating, hard to configure and control

12:04 20VC: Why No Models Today Will Be Used in a Year, Why Open Will Always Beat Closed in AI, Why Proprietary Data is Less Important Than Ever And Why EU AI Regulation is a Disaster with Alex Lebrun, Founder & CEO @ Nabla

Richard Socher · Aug 18, 2023

While some AI companies really are thin wrappers with little moat, people underestimate the complexity involved in other cases, such as the search backend you.com had to build that knows when to retrieve which facts from the internet

To be factual, up to date, and citable, you must build a complex search backend and feed the right facts into the model rather than relying on its memory

14:16 20VC: Does Value Accrue to Incumbents or Startups in the AI Race, Why Model Size Matters More Than Data Size, Why Artificial General Intelligence is Far Away, Why Carpenters Will Be Paid More Than Software Engineers & Future of Jobs with Richard Socher

Des Traynor · Nov 15, 2023

The VC 'thin wrapper' critique is only valid against actual thin wrappers; there is a massive gap between a half-day chatbot demo and a product a large company will deploy.

A deployable product requires ingesting knowledge bases, refreshes, self-reporting, CSAT reporting, parsing multiple conversations, customer context, premium-vs-free handling, permissioning and targeting — layers of work beyond calling an LLM.

Scope: Concedes the argument holds if the product genuinely is a thin wrapper

9:28 20VC: How to Survive and Thrive in a World of OpenAI, Are LLMs Being Commoditised, Where Does the Value Lie; Infrastructure or Application Layer, How Apple Could Win in a World of AI, How Amazon Could Threaten OpenAI and Why Google Struggle with Des Trayn

Harry Stebbings · Jul 6, 2026

Even when labs enter a category the outcome isn't binary — multiple players can win, and his own expectation of a runaway winner in dev tools was a mistake

Anthropic dedicated a team to design against Figma yet Figma still does multiple billions in revenue, and Lovable, Cognition, Replit, Claude Code and Codex have all scaled simultaneously

Scope: walks back his earlier expectation of a single runaway winner

35:50 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Arvind Jain · Jul 11, 2026

Anthropic's vertical products for design, legal and finance are shallow and are expanding the market to non-experts rather than cannibalizing incumbents like Figma

He doesn't know of anyone moving workloads entirely off Figma to Anthropic; designers still use Figma while non-designers use Claude's design tooling

Scope: based on what he observes with customers

11:03 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Alex Atallah · Aug 10, 2026 · hedged

Claude Design has not yet shown meaningful repeat usage or impact on Figma

Designers including his own tried it but he hasn't heard repeat-use stories or much chatter, and Figma's earnings were very strong

Scope: admits he has not talked to many designers about it

24:46 20VC: Will OpenRouter Sell for $10BN to Stripe? | Why Chinese Open Models Are Beating America—and What Happens Next | Why Enterprises Are More Fearful of Anthropic and OpenAI Than China | Is the Routing Layer Becoming a Commodity with Alex Atallah

Mike Mignano · Jul 6, 2026

Competitive intrusion by model providers is a genuine worry even though startups can still win, but it's the same worry that has always existed in startups

OpenAI launched a product directly competitive with Granola, and Suno faces massive incumbents

Scope: holds both sides: still believes one company can't do everything

37:53 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Mike Mignano · Jul 6, 2026

Foundation model providers will not be able to do everything; no single company or handful of companies will absorb the entire application layer.

He previously believed one to five model companies would do everything, but was reminded that even the biggest companies can't do everything — the same fear existed about Google and Apple ten to fifteen years ago and it never materialized.

Scope: explicitly a reversal of the view he held until recently

64:13 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Harry Stebbings · Jul 6, 2026

Incumbents sometimes succeed in expanding far outside their core (AWS being the exemplar), but they cannot do everything.

Scope: concedes some adjacency bets do work

65:16 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Des Traynor · Nov 15, 2023

Companies that deeply understand and build for the full user workflow will win AI applications, not those who point out that a capability has been academically possible for months.

Once someone builds the full workflow — feedback, collaboration, approvals, asset management, hosting, embedding, usage reporting — the thin-wrapper competitor has an enormous amount left to build; it becomes a classic B2B arms race and people buy real software for real work.

11:01 20VC: How to Survive and Thrive in a World of OpenAI, Are LLMs Being Commoditised, Where Does the Value Lie; Infrastructure or Application Layer, How Apple Could Win in a World of AI, How Amazon Could Threaten OpenAI and Why Google Struggle with Des Trayn

Max Junestrand · Jan 26, 2026

Roughly 80% of the value in an applied AI product comes from building normal enterprise-grade software around the models — scaffolding, interaction surfaces — not from the models themselves

The value pyramid has base models at the bottom, the vertical's legal interpretation of those models in the middle, and enabling clients to do differentiated things at the top; most of that middle layer is ordinary software engineering

Scope: framed around Legora's own product responsibility

13:56 20VC: From Only OpenAI to Die-Hard Anthropic: The Downfall of OpenAI in Enterprise | Harvey vs Legora: Legal AI is a Winner Take All | $7M ARR in a Single Day and Raising $200M Across 3 Rounds with No Deck with Max Junestrand, CEO @ Legora

Harry Stebbings · Jun 27, 2026

Paul Graham now asks every YC batch member how to 'AI-proof' their product by adding non-AI features that build defensibility

26:40 20VC: How We Got Fred Wilson, Benchmark and Index to Invest $94M | Why Robinhood's Strategy is Wrong | Why 1-1s are BS and What Every Founder Gets Wrong About Equity | Why Taste Beats AI But How AI Kills Org Charts with Paul Erlanger, CEO @ fomo

Harry Stebbings · Jan 15, 2025

Investors he speaks to say Synthesia's space is the most obvious area for OpenAI to move into, alongside customer service.

Scope: reporting investor views from the last eighteen months

31:07 20VC: Why Scaling Laws Will Not Continue | OpenAI vs Anthropic vs X.ai: Who Wins and Why | How Far Will Model Providers Go Into the Application Layer | The End State for Models: Many Specialised or Few Generalised with Victor Riparbelli @ Synthesia

Arvind Jain · Jul 11, 2026

Anthropic already competes with Glean in enterprise deals, and started competing with them earlier than with others, because question answering and information seeking is the largest AI use case in the world today

Claude Desktop's primary use case has always been question answering, and MCP lets it connect to enterprise systems, so Glean must explain why building real context is hard

11:50 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Alex Atallah · Aug 10, 2026

Model labs will eventually go after application companies built on them, and startups serving teams that are strategic to the labs face the most near-term threat

Labs have an incentive to get multiple teams inside customer companies dependent on them; Claude Design was strategic not for revenue but to make design teams care about Anthropic models

Scope: thin go-to-market wrappers are safe if labs don't care about that market, and there will be many such markets

22:57 20VC: Will OpenRouter Sell for $10BN to Stripe? | Why Chinese Open Models Are Beating America—and What Happens Next | Why Enterprises Are More Fearful of Anthropic and OpenAI Than China | Is the Routing Layer Becoming a Commodity with Alex Atallah

Aaron Levie · May 22, 2024

Startups should avoid building anything a horizontal chat interface could instantly subsume, and instead build the deep workflows a human would need to run a full business process

Functionality that a universal chat assistant can absorb offers no defensible position

0:00 20VC: Box's Aaron Levie on Predictions for the Next Wave of AI: Will Foundation Models Be Commoditised | How the Business Model of SaaS Changes Forever | Startups vs Incumbents: Who Wins | App vs Infrastructure Layer: Where is the Value?

Aaron Levie · May 22, 2024

The right heuristic for AI startups is to avoid anything that a horizontal chat interface could instantly subsume or that is one training run away from being absorbed by a better model, and instead build the unglamorous end-to-end workflows a human needs to actually run a business process

OpenAI has told us it will become a universal assistant interface plus an API business handling audio, video and text with complete intelligence, so those areas are pre-claimed; the remaining defensible work is business process workflow, which most people find boring

Scope: "to 80% of people it sounds like the most boring part of software"

26:42 20VC: Box's Aaron Levie on Predictions for the Next Wave of AI: Will Foundation Models Be Commoditised | How the Business Model of SaaS Changes Forever | Startups vs Incumbents: Who Wins | App vs Infrastructure Layer: Where is the Value?

Mike Krieger · Mar 3, 2025

Application companies in seemingly horizontal categories can stay differentiated because deep workflow knowledge is valuable — the defensible value sits in professional use cases, while consumer and prosumer needs get served well enough by base models

Eleven Labs' console and Descript's design are clearly built for people voicing or editing hours of content day in and day out; whereas charging a consumer $10/month just to translate feels iffy because the models are already good at that

Scope: consumer/prosumer side is 'good enough' from a basic AI product perspective; conceded partly in response to Harry's power-user vs casual-user distinction

40:35 20VC: Anthropic CPO Mike Krieger: Where Will Value Be Created in a World of AI | Have Foundation Models Commoditized | When Do Model Providers Become Application Providers | What Anthropic Learned from Deepseek

Mike Krieger · Mar 3, 2025

A model provider should only move into applications where the product is generalizable; Anthropic will not build many bespoke verticalized experiences tied to a specific workflow

Its product team is small relative to the surface area it already supports (API, Claude AI, Claude for Work, Claude Code), so bets must be general purpose with specialization only at the user level

Scope: may still pick a persona or vertical to target; specialization at the user level is acceptable

39:18 20VC: Anthropic CPO Mike Krieger: Where Will Value Be Created in a World of AI | Have Foundation Models Commoditized | When Do Model Providers Become Application Providers | What Anthropic Learned from Deepseek

Harry Stebbings · Jun 1, 2026

Vertical application-layer companies like Legora have real defensibility against foundation labs, because their products are deeply tailored to specific professional workflows and a lab like Anthropic would need to build entirely separate product, GTM, customer success and adoption organizations to compete.

The product is deep and specifically fitted to lawyer workflows, and Anthropic would have to stand up entirely separate product divisions, go-to-market, customer success and adoption teams — effectively a different company

Scope: stated as an investor in such companies

35:52 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility, The Model is the Product | Token Spend Will Exceed Headcount Spend in 5 Years | The True Cost of Hiring AI Researchers in the Valley Today with Brendan Foody

Harry Stebbings · Jul 27, 2026

Anthropic is not a real threat to Legora because Legora is a heavy go-to-market business built on lawyer relationships and legal deployments, which is a completely different business from a model company

The work is relationship- and deployment-driven, not model-driven

Scope: stated in opposition to what 'everyone tells me'

29:25 20VC: Leading Anthropic's First Ever Round | Will Open Source Threaten Anthropic's Business | Do Margins Matter in a World of AI | Why Triple, Triple, Double, Double is Not Good Enough Today | Why Series A is Hard Today with Matt Murphy @ Menlo

Eric Vishria · Sep 25, 2024

AI infrastructure companies are attractive investments but carry the persistent risk that foundation model providers move up the stack into their territory

Infrastructure companies are growing astoundingly quickly, but entrepreneurs must constantly update their mental models as the foundation model layer expands

30:12 20VC: Benchmark's Eric Vishria on Where is the Value in AI: Chips, Models or Apps | Why Nvidia Will Not Be The Only Game in Town | The Commoditisation of Foundation Models | Which AI Apps Have Sustaining Value vs Hype and Short Term Revenue

Sridhar Ramaswamy · Feb 10, 2025

It is terrifying to build a startup on top of OpenAI, because the model providers can move into the application layer at any time

The line between infrastructure provider and application provider is now blurry, and there is no guarantee OpenAI, Anthropic, Microsoft or Google won't build what you're building

0:00 20VC: Why Model Providers Will Kill Many Startups Moving into the Application Layer | Why Deepseek is not a Threat to OpenAI & Why OpenAI Beats Anthropic | Apps vs Models vs Infrastructure: Where is Value in AI with Sridhar Ramaswamy, Snowflake CEO

Harry Stebbings · Apr 4, 2025 · hedged

The sustainability of value in startups has almost completely eroded, because a single new foundation model release can wipe out a company overnight

OpenAI could release a model that kills a product like Granola, or a foundation model provider could roll a data-provider's core market into its own product

30:38 20VC: Why To Win in AI, Investors Need to Change Their Approach | Why VC is Run by Principals and Associates and is a Broken System | The Bull Case for Anthropic & Whether Deepseek Changes Their Strategy with Nabeel Hyatt @ Spark Capital

Maor Shlomo · Nov 24, 2025

Model providers are moving up the software stack, building tools at every level of the value chain rather than only selling APIs and end-user chat

All of them now ship CLI coding tools and consumer chat products, and Gemini is being embedded across Gmail and workflow automation, driving consumption across the stack

26:28 20VC: Base44's Maor Shlomo on How Vibe Coding Will Kill SaaS and Salesforce | Why it is BS that Vibe Coding Platforms Do Not Have Defensibility and Bad Margins | Why He Worries About Google, Not Replit and Lovable | Why Long Anthropic, Not OpenAI?

Arvind Jain · Jul 11, 2026

Anthropic should be considered an application-layer company, not just a model company

An ecosystem is forming on top of its platform — developers building automations, skills and MCP servers connecting internal systems to Claude

Scope: OpenAI and Anthropic are fundamentally different businesses; OpenAI's strength is consumer product

17:18 20VC: Why OpenAI and Anthropic Won't Win the App Layer | Why Teams Will Get Bigger Not Smaller in a World of AI | Why AI Removes Incumbents Advantage of Bundling | China vs America: Who Wins the AI War with Arvind Jain, Co-Founder @ Glean

Richard Socher · Apr 18, 2025

OpenAI shipping features will not kill the generation of specialized consumer AI app companies overnight; specialized players that go deeper and integrate into workflows will still build real revenue

Speech recognition is a commodity with abundant open source, yet several companies still make tens or hundreds of millions doing the last little bit perfectly; companies want custom solutions and pixel-perfect, workflow-integrated output that a clever early adopter couldn't produce alone

47:39 20VC: Foundation Models: Who Wins & Who Loses | How Economies and Labour Markets Need to Change in a World of AI | China vs the US in an AI Race: What You Need to Know | Rich Socher, Founder @ You.com

Anish Acharya · Feb 9, 2026

Model companies will often recreate an app's core primitive and do product marketing around it, but in markets that demand a lot of feature surface the labs are less set up to prioritize building it

Granola's meeting transcription primitive was copied everywhere including by OpenAI, but Granola's vision is a full productivity suite; the labs lack the prioritization, resources and ambition to build all the feature surface around a primitive

Scope: Applies to markets demanding lots of feature surface; He is assuming, not stating, Granola's roadmap

17:52 20VC: Is SaaS Dead in a World of AI | Do Margins Matter Anymore | Is Triple, Triple, Double, Double Dead Today? | Who Wins the Dev Market: Cursor or Claude Code | Why We Are Not in an AI Bubble with Anish Acharya @ a16z

Shiv Rao · May 16, 2026

OpenAI's and Anthropic's moves into forward-deployed engineers and partnerships with large private equity groups are a strong signal that vertical AI has an enormous opportunity for the foreseeable future.

If the foundation model labs are deploying engineers into portfolio companies, it shows the application layer in specific verticals is where value is being unlocked.

Scope: for the foreseeable future

0:00 20VC: Lessons from Jensen Huang on "Founder Mode" | How to Know if OpenAI or Anthropic Will Kill your Company | How USV Liking Music Made Them $1BN on an Investment | The Five Year Desert to Product Market Fit & a $5.3BN Valuation with Shiv Rao @ Abridge

Shiv Rao · May 16, 2026

OpenAI and Anthropic announcing forward deployed engineers and partnerships with private equity groups is proof of an incredible ongoing opportunity for vertical AI companies

Getting into an enterprise, accessing and cleaning and organizing the data, integrating into specific workflows, and doing it compliantly — HITRUST, cyber, treating behavioral health data differently from primary care data — takes enormous effort and a scalable machine to build

Scope: for the foreseeable future

23:58 20VC: Lessons from Jensen Huang on "Founder Mode" | How to Know if OpenAI or Anthropic Will Kill your Company | How USV Liking Music Made Them $1BN on an Investment | The Five Year Desert to Product Market Fit & a $5.3BN Valuation with Shiv Rao @ Abridge

Harry Stebbings · Aug 3, 2026

Vulnerability to the labs depends on go-to-market: self-serve tools like design software are easy for labs to take, while relationship-heavy enterprise sales and deployment (legal) are defensible

Designers can pick up a tool themselves, but selling to law firm partners and forcing deployment on resistant junior lawyers is heavy real-world lifting the labs won't do

52:20 20VC: 70% of Neolabs Will Die | There Will be a $100BN US Open-Source Model | Data is a Trillion $ Market | Governments Cannot Regulate Models: It is Too Late | The Cyber Attacks to Come Will be Insane with Anastasios Angelopoulos @ Arena

Anastasios Angelopoulos · Aug 3, 2026

Network-effect and operations-heavy businesses like system integrators are safer from lab competition and more likely to adopt the labs, while scalable software products like legal chatbots or insurance claims automation are vulnerable

Labs are less likely to compete with an Infosys-style operations business but can build a scalable software product themselves

52:56 20VC: 70% of Neolabs Will Die | There Will be a $100BN US Open-Source Model | Data is a Trillion $ Market | Governments Cannot Regulate Models: It is Too Late | The Cyber Attacks to Come Will be Insane with Anastasios Angelopoulos @ Arena

Harry Stebbings · Mar 3, 2025 · hedged

Horizontal, homogeneous use cases like translation, transcription and customer service are the natural path for model providers to move into applications

40:24 20VC: Anthropic CPO Mike Krieger: Where Will Value Be Created in a World of AI | Have Foundation Models Commoditized | When Do Model Providers Become Application Providers | What Anthropic Learned from Deepseek

Severin Hacker · May 19, 2025 · hedged

ChatGPT building a good language tutor and pulling away casual users is somewhat of a concern

Scope: only 'somewhat' a concern; he has counterarguments

24:41 20VC: Duolingo Co-Founder on Why $3M is Harder than $100M to Raise | Why You Should Always Take Tier 1 VCs Even at Worse Terms | Why Europe Can't Win Unless the US Screws Up | How AI Impacts the Future of Work and Education with Severin Hacker

Harry Stebbings · Nov 15, 2023

Many investors dismiss AI application companies as thin layers on top of LLMs with little value accrual, since almost none build their own model

Scope: characterising a widespread investor view rather than fully endorsing it; poses it to Des as a challenge, asking if it's a 'bullshit VC-ism'

9:01 20VC: How to Survive and Thrive in a World of OpenAI, Are LLMs Being Commoditised, Where Does the Value Lie; Infrastructure or Application Layer, How Apple Could Win in a World of AI, How Amazon Could Threaten OpenAI and Why Google Struggle with Des Trayn

Harry Stebbings · Sep 8, 2025

20VC passed on ElevenLabs partly because it seemed implausible that two very young founders in London would not simply be beaten by OpenAI

the check was small and the company was extremely early, so the 'why wouldn't OpenAI do this?' objection dominated

Scope: recounting his and Kieran's reasoning at the time

13:41 20VC: ElevenLabs Hits $200M ARR: The Untold Story of Europe's Fastest Growing AI Startup | The Real Cost of AI from Talent to Data Centres | How US VCs are in a Different League to Europeans | The Future of Foundation Models with Mati Staniszewski

Harry Stebbings · Nov 15, 2023 · hedged

OpenAI's recent Dev Day releases likely killed off overnight most of the application companies that had been innovating over the past year, including in niches like prosumer wealth management.

OpenAI's horizontal capabilities now cover verticals nobody expected them to build for.

12:04 20VC: How to Survive and Thrive in a World of OpenAI, Are LLMs Being Commoditised, Where Does the Value Lie; Infrastructure or Application Layer, How Apple Could Win in a World of AI, How Amazon Could Threaten OpenAI and Why Google Struggle with Des Trayn

Harry Stebbings · Feb 10, 2025

OpenAI intends to steamroll startups

23:38 20VC: Why Model Providers Will Kill Many Startups Moving into the Application Layer | Why Deepseek is not a Threat to OpenAI & Why OpenAI Beats Anthropic | Apps vs Models vs Infrastructure: Where is Value in AI with Sridhar Ramaswamy, Snowflake CEO

Amjad Masad · Apr 25, 2026

Knowing how to evaluate and extract performance from foundation models is the core competency and IP of an agent lab

It is tacit knowledge — engineers act like psychologists probing a new model's limits — combined with proprietary benchmarks and A/B testing

14:30 20Product: Replit CEO on Why Coding Models Are Plateauing | Why the SaaS Apocalypse is Justified: Will Incumbents Be Replaced? | Why IDEs Are Dead and Do PMs Survive the Next 3-5 Years with Amjad Masad

Josh Browder · May 18, 2026

Custom evals will become a commonplace job function within five years because the future of AI is organization-specific rather than foundation-model-centric

Value will come from evaluating models against a company's own proprietary data, as DoNotPay does

61:54 20VC: Turning Peter Thiel's $100K into $10M Angel Portfolio | The One Man Accelerator at The Four Seasons | Why VCs Can Be Sharks and What Founders Need to Know | Why Stocks and Cash are BS and You Should Invest in Land with Josh Browder

Sarah Tavel · May 6, 2024

Startups that are thin wrappers around an LLM face a nearly unwinnable battle, because roughly 90% of the value they deliver comes from OpenAI and an incumbent like Notion can add that feature contextualised by its own application value

If only ~10% of the delivered value is the startup's own workflow and focus, the incumbent can replicate the rest trivially and win on distribution

Scope: describes the first wave of AI startups

18:27 20VC: Benchmark's Sarah Tavel on Are Foundation Models Commoditising | Why Frontier Models Will Be Closed Source | Why the Value is in the Application Layer | The Future of AI is "Selling the Work" Not the Tools

Eléonore Crespo · May 9, 2025 · hedged

AI will kill thin-layer vertical SaaS apps but not products built on fundamental infrastructure — deep databases, compute, and multi-user workflow UX — which will win by embedding AI correctly.

Complex enterprise systems serving hundreds or thousands of users need more than an AI layer, so AI off the shelf can't replace them.

Scope: depends on the type of vertical app; she concedes there is some truth to the bear case

61:10 20VC: Four Traits of the Most Successful Founders | How to Hunt and Close Talent Like a Pro and Where All Founders Go Wrong | Lessons Raising $397M From the Best Investors in the World with Eléonore Crespo @ Pigment

Harry Stebbings · Mar 16, 2026 · hedged

Many verticalized voice/support agent startups look like thin layers on top of OpenAI and ElevenLabs, though some may become defensible via verticalized data and deep workflow embedding.

They amount to a plaster over the wound — the underlying capability comes from the foundation model and voice providers, not the startup.

Scope: Harry voices the counterargument that they build verticalized data, fine-tune their own models and embed in workflows, but does not adopt it

24:52 20VC: The 8 Moats of Enduring Software Companies: How to Analyse for Durability and Defensibility in a World of AI | Why Dropouts are "AI Maxing" the World & Remote Early-Stage Companies are Dying with Gokul Rajaram

Brendan Foody · Jun 1, 2026

Building defensibility in the software layer on top of foundation models will be incredibly difficult because the model itself is the product

Over the last two years it has become increasingly clear that the model, not the application wrapper, is what delivers the value

Scope: about the application/software layer built on top of models

0:00 20VC: Mercor CEO on Why Application Layer Companies Have No Defensibility, The Model is the Product | Token Spend Will Exceed Headcount Spend in 5 Years | The True Cost of Hiring AI Researchers in the Valley Today with Brendan Foody

Aravind Srinivas · Jun 5, 2024

The model is not where most of the value sits in search products — Google's AI Overviews was bad and ChatGPT browsing is still worse than Perplexity despite Google and OpenAI having the best index and models

Value comes from coordinating hard engineering feats beyond the LLM plus a large human element, which the labs have not assembled

Scope: specific to the search use case

37:07 20VC: Perplexity's Aravind Srinivas on Will Foundation Models Commoditise, Diminishing Returns in Model Performance, OpenAI vs Anthropic: Who Wins & Why the Next Breakthrough in Model Performance will be in Reasoning

Fernando Fanton · Jul 27, 2025

The decisive question for every AI application company is whether it adds enough value on its layer to justify existing above what the models do on their own — and fine-tuning no longer counts as that secret sauce

If it isn't user experience or learning, and fine-tuning doesn't seem to be a thing anymore, then there has to be something explaining why users prefer Lovable over Cursor or the next entrant

42:36 20Product: Should We Kill the PM Role Entirely | How Does Product Design Change Most in a World of AI | How Do Monzo Build Product Today: What Works, What Does Not | Why Most Product Sucks and What Makes Truly Great Product

Max Altman · Nov 21, 2025

Investors should avoid startups building adjacent to OpenAI and Anthropic, because those labs are trying to become multi-trillion-dollar public companies and will therefore try everything

A company aiming at that scale has to attack every adjacent surface — Amazon even tried making a phone — so assuming OpenAI won't build what you're building is naive

Scope: explicitly says this is not based on inside information

54:51 20VC: Max Altman on The New Seed War: Can Anyone Compete with Sequoia and a16z | Leaving $2BN on the Table with Reddit | Lessons from Backing Rippling at $25M Post | Why Climate Tech is a Mirage and Disaster

Max Altman · Nov 21, 2025

ElevenLabs is the exception rather than the rule — most companies whipping up a thin layer over an API do get killed by the labs, which is why his firm targets 'second order effect' AI companies further from what OpenAI will look like at $3 trillion

People remember the survivors and forget the many weekend-built API wrappers that OpenAI simply absorbed

56:20 20VC: Max Altman on The New Seed War: Can Anyone Compete with Sequoia and a16z | Leaving $2BN on the Table with Reddit | Lessons from Backing Rippling at $25M Post | Why Climate Tech is a Mirage and Disaster

Max Altman · Nov 21, 2025 · hedged

The best test of whether a company is safe from the labs is asking what world-class engineers in downtown San Francisco won't want to touch — unglamorous operational domains like trucking logistics are defensible even in categories like AI voice

Lab engineers want to stay in their ivory castle and won't want to deal with a late shipment of oranges from Miami to New York

Scope: says it is an art, not a science

56:43 20VC: Max Altman on The New Seed War: Can Anyone Compete with Sequoia and a16z | Leaving $2BN on the Table with Reddit | Lessons from Backing Rippling at $25M Post | Why Climate Tech is a Mirage and Disaster

Andrew Macdonald · Aug 17, 2026

Uber is relatively insulated from being cannibalized by frontier labs because its business has a physical-world component that doesn't extend from a model company's playbook

A lab would have to launch ride-sharing across tens of thousands of cities, get locally licensed, and put boots on the ground to run a physical-world service

Scope: he says he hasn't seen the cannibalization risk materialize yet, not that it can't

43:58 20VC: Uber President on The Untold Uber Stories: Travis, China and Self-Driving | Why Autonomy Is Existential | How to Beat DoorDash to #1 in Food with Andrew MacDonald

Harry Stebbings · Jul 29, 2024

Vertical SaaS like ServiceTitan cannot be replicated in a weekend with AI

These products embody intense workflows, deep functionality, and accumulated brand credibility, trust and customer love

54:54 20VC: Twitter's Most Controversial VC Delian Asparouhov on Inside the Walls of Founders Fund: What the World Does Not See | Why Western Europe Will Be Like the Third World | Why SaaS as an Industry Might Be Dead

Jake Saper · Mar 10, 2025

Lessons from prior SaaS eras still apply to AI-era software, above all that workflow is sticky and software people live in daily is very hard to rip out.

Even if AI changes the product, the majority of professional users will still spend their day inside a piece of software in the medium term, which entrenches it.

Scope: medium term

44:37 20VC: Lessons from Investing $2BN and Returning $8BN in Cash | Why Most Venture Partnerships are Broken | We Sold Salesforce Early and Lost Out on Billions | Are The Best Deals Always Expensive and Competitive with Jake Saper @ Emergence Capital

Jeff Seibert · Nov 22, 2023

AI should be treated as a technology component like MySQL, not as a product in itself; in five years a company built as a form on top of an LLM will look as unfundable as a form on top of MySQL

MySQL was cool and popular twenty years ago but building a thin form on top of it is not a fundable business today

26:14 Thin wrapper apps on llms will become as unfundable as a form on top of mysql

Jeff Seibert · Nov 22, 2023

Startups are more vulnerable than incumbents, and any startup solving a horizontal problem OpenAI will need to solve within five years to scale is a bad investment and a bad use of a founder's time

Those companies are filling incremental gaps that are already on OpenAI's roadmap, so they get 'Sherlocked' the way Apple killed Watson with Sherlock in 2002

27:23 Horizontal problems on openais roadmap will sherlock thin startups

Raman Malik · Nov 15, 2024

Companies at the application layer that go very specific on high-TAM areas can make a lot of money

Evidence has emerged over the last eight or nine months from those who are crushing it at the application layer

54:32 Narrow specific high tam application layer companies can be very profitable

Nikesh Arora · Jun 22, 2026

Frontier labs do not threaten incumbent security vendors at the perimeter — there is no OpenAI or Claude endpoint agent — the hard AI problem is post-intrusion: building the enterprise context and intelligence to find and remove an attacker already inside

Perimeter gatekeeping requires deployed sensors that labs don't have, while the real challenge is that breaches always get through and detecting them requires context and intelligence about what an intrusion means

31:54 Labs lack deployed sensors and post intrusion context

Clay Bavor · Jul 4, 2026 · speculative

The open question is whether general-purpose frontier tools themselves become the security solution rather than narrowly focused cybersecurity products.

The same models that provide offensive capability may end up serving as the defensive layer.

55:06 General frontier tools may subsume security point products

Anastasios Angelopoulos · Aug 3, 2026

Frontier model providers will keep moving up into the application layer, and this is a serious risk to application companies

Inference is going to commoditize, so labs must own more of the application stack to stay close to the value they deliver to end customers and avoid being commoditized themselves

50:47 Inference commoditization pushes labs into the app layer

Nikhil Basu Trivedi · Sep 6, 2023

There are few good AI-first or AI-enabled venture opportunities for Footwork to invest in at this phase

The hype cycle has made valuations too expensive and created huge competitive noise, and in the LLM world the models themselves will serve many use cases rather than an application layer on top

48:48 Hype driven valuations plus model absorption leave few good app layer ai opportunities right now

Richard Socher · Aug 18, 2023

Foundational models have lowered the barrier for startups: a small team can now reach an 80% solution quickly and then layer task-specific data on top, which was impossible before large language models

Before foundation models it would have been unthinkable for a small company to build a search engine understanding many languages; now general natural-language understanding comes off the shelf

12:12 Foundation models lower startup barriers to an 80 percent solution though incumbents still win the last mile

Martin Casado · Jul 28, 2025

Believing AGI arrives soon is not in conflict with investing in enterprise SaaS, because AGI does not imply unlimited capability that makes existing software markets disappear

Humans are already general intelligence and enterprise SaaS still gets built and bought; people wrongly equate AGI with omnipotence

57:44 Agi does not imply omnipotence so saas persists

Aidan Gomez · Aug 19, 2024

Agentic products are best built by the model builders themselves; a pure consumer of someone else's model is structurally disadvantaged

The model is the reasoner behind the agent, so agent quality depends entirely on model quality and you need to be able to intervene at the model level to make it better at what you care about

40:33 Model level control is required for competitive agentic products so pure app layer consumers are disadvantaged

Aidan Gomez · Aug 19, 2024

The application/product layer remains extremely attractive for investors, and new AI products will transform social media

People love talking to these models and the usage time is insane

45:06 App layer remains highly attractive and ai will transform social media products

Harry Stebbings · May 19, 2025

Language learning is one of the four core application areas OpenAI will own, alongside chat, customer support and coding, because it is unusually well aligned with what LLMs are best at

LLM is 'kind of in the name' — the use case maps directly onto language models, so it's not a stretch for them

25:53 Language learning is core lab territory well aligned with llm strengths

Eran Zinman · Mar 2, 2026 · hedged

Proprietary accumulated training data is not essential to most of the value in AI application companies — you can probably get the vast majority of the value without that specific training

Over time those companies accumulate data, but the bulk of the product value does not depend on it

11:03 Proprietary training data is not where app layer value sits

Arvind Jain · Jul 11, 2026 · hedged

Microsoft is currently a more formidable competitor than the frontier labs

In prospecting, Glean hears 'we already get Copilot as a Microsoft customer' far more often than it hears prospects saying they've standardized on a lab product instead

20:19 Incumbent bundlers not labs are the real threat

Aaron Levie · May 22, 2024

OpenAI publicly stating which areas it will steamroll is good for the ecosystem, and founders should pay close attention to what that guidance implies for what they build

It is very useful when the lead platform player tells you the types of things it will go after versus leave to the ecosystem

26:07 Labs publicly signaling their roadmap helps founders avoid doomed areas

Brendan Foody · Jun 1, 2026

Network effects are the litmus test for which software companies survive: those with them will gain dramatic value from 10x product velocity, those without will become worthless

Companies like Salesforce, Slack Connect and Carta have marketplace and cross-company network effects that let them iterate ten times faster while compounding customer value, whereas pure software without network effects has no defensible moat once agents can recreate it

37:59 Network effects separate survivors from worthless software

Sarah Tavel · May 6, 2024

Benchmark's AI application-layer investments will only get better as underlying models improve — taking on more work and earning better margins as human-in-the-loop QA becomes unnecessary

Their value depends on model capability improving, so a 100x better OpenAI is a tailwind rather than a threat

17:38 Better underlying models are a tailwind not a threat for app layer companies

Lin Qiao · Jul 20, 2026

Each company's uniqueness is baked into its product design, software, data and understanding of user intent, and is not learnable or shareable by an outside company

That uniqueness is the fundamental basis for why a company exists at all

10:34 Company specific taste and user intent cannot be learned by outsiders

Becca Lindquist · May 2, 2026

General LLMs can't replace go-to-market tooling because the hard part is not one rep running a prompt but scaling and iterating the same motion across a hundred reps

Getting a hundred humans to change and iterate in the same way is the bottleneck; people are not bots

61:01 Multi rep scaling and iteration is the defense

Matan Grinberg · Jun 13, 2026

The best outcome for customers requires the model layer to be separate from the application layer

A model provider selling the coding application is an API business incentivized to maximize token consumption rather than token efficiency; an independent application layer that lets enterprises arbitrate between providers forces models to compete on being best, cheapest or fastest and lets customers route by language or task

59:25 Token consumption incentives misalign labs from owning the app layer

Saam Motamedi · Jul 15, 2024

OpenAI will not need to own focused applications like copilots for lawyers or physicians, or software development tooling such as incident response, SRE workflow and debugging

Those are not core foundational capabilities OpenAI needs to own, so the platform-kills-you fear doesn't apply there

13:13 Opeanai wont own vertical copilots or specific dev tooling since theyre not core platform capabilities

Harry Stebbings · Mar 16, 2026

If Anthropic or OpenAI shipped an enterprise note-taking product connected to their broader product suite, it would heavily threaten the market size Granola can expand into in large enterprise

An incumbent model provider's integrated suite would cap how far a standalone notetaker can penetrate enterprise

36:44 Lab suite bundling caps standalone app enterprise tam

Matt Murphy · Jul 27, 2026

Anthropic and Lovable are not meaningfully competitive with each other — Anthropic comes at the technical user and Lovable at the lay user, and the market is big enough for both to do extremely well

Any overlap is in the middle; and Cursor was as directly in Anthropic's crosshairs as possible and still had a very good outcome

21:52 Lab and app serve different user segments so both win

Jesse Zhang · Sep 19, 2025

Dependence on third-party model performance is less of a risk in customer service than in a business like Windsurf, because customer service needs instruction following rather than frontier reasoning and today's models are already good enough at it

Most meaty support inquiries require following rules and executing workflows correctly, not reasoning; so the remaining challenge is orchestration — picking the right model per task and optimizing latency and consistency

25:36 Instruction following tasks face less frontier dependency risk than reasoning heavy tasks

Harry Stebbings · Mar 30, 2026

The market reaction to Anthropic's security announcement — CrowdStrike and Cloudflare dropping 8-9% — was ridiculous, since the announcement obviously does not threaten those businesses today

53:36 Markets overreact lab announcements are no near term threat to incumbents

Harry Stebbings · Feb 10, 2025 · hedged

Models appear to be getting commoditized quickly (as DeepSeek recently showed) and much application-layer work built atop commoditized models looks light, making it unclear where sustainable value will be generated

DeepSeek's emergence in recent weeks demonstrated how fast model advantage erodes

19:44 Rapid model commoditization makes durable app layer value unclear

Sridhar Ramaswamy · Feb 10, 2025

New value creation in AI is murky because the line between infrastructure and application providers is blurry — model providers like OpenAI, Anthropic, Microsoft or Google can quickly build any application category that takes off — so durable value sits with companies that already have customer relationships and adopt AI fast enough to avoid disruption

Half a billion loyal users and the surrounding feature set give staying power; users won't leave for a free model wrapper, and OpenAI would happily host DeepSeek underneath ChatGPT if useful

20:23 Existing customer relationships plus fast ai adoption defend against lab encroachment