Should AI investors favor the infrastructure layer or the application layer?

Harry Stebbings · Sep 18, 2023

Value in AI accrues more to the infrastructure layer than the application layer for the time being

Democratization of fine-tuning, security, and data acquisition for startups are the live opportunities

Scope: for the time being

63:35 20VC: Benchmark General Partner, Miles Grimshaw on The Five Pillars of Venture Capital, Why Data Can Be a Trap When Early-Stage Investing, Investing Lessons from Missing Figma and Plaid & The New Business Model for AI & Why Co-Pilot is an Incumbent Strate

Miles Grimshaw · Sep 18, 2023

New tooling and frameworks for AI development are needed now, but over the arc of time the applications will capture profound value and pull better infrastructure through with them

Every developer will build AI-native or AI-enabled applications; historically application winners like Facebook became major contributors to infrastructure like MySQL

Scope: infrastructure is happening now; applications may not look like applications as we know them; conditional on this being a shift as large as the internet

64:03 20VC: Benchmark General Partner, Miles Grimshaw on The Five Pillars of Venture Capital, Why Data Can Be a Trap When Early-Stage Investing, Investing Lessons from Missing Figma and Plaid & The New Business Model for AI & Why Co-Pilot is an Incumbent Strate

Matan Grinberg · Jun 13, 2026

Value accrual among model, application, and infrastructure companies is time-dependent rather than a steady state in which one player captures all the value — pricing power shifts from one group to another over successive periods — and every layer is trying to commoditize the others

Everyone in the stack has an incentive to argue the other layers are irrelevant; app companies push model providers to compete on price and speed, model companies push to make apps trivial to build, and infra players have their own spin

12:36 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 Founder with Matan Grinberg, Founder @ Factory

Mike Mignano · Jul 6, 2026

Value will continue to accrue at the infrastructure layer, but there will also be massive value creation at the application layer.

The build-out isn't finished, but there are now enough new technological 'toys' for applications to create large value.

Scope: does not deny continued infrastructure value creation

10:51 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

Julien Bek · Aug 24, 2026

You should invest in both the infrastructure and application layers rather than choosing, because opposing theses can both be correct with truths materializing at different times

Infra like Fireworks is ripping now at the start of the AI revolution, while application companies like Rillet are signing customers whose finance teams will compound into sticky businesses over the next few years

54:18 20VC: Inside Sequoia's Investment Committee: Lessons from Don Valentine, Doug Leone and Alfred Lin | How the SpaceX and Citadel Deals Went Down | What Sequoia Specifically Looks for in Founders with Julien Bek

Harry Stebbings · Mar 16, 2026

Hardware companies are extremely popular with investors right now because investors are terrified that Anthropic will eat the lunch of software companies.

Anthropic moving into security has already coincided with security stocks plunging.

Scope: based on what he sees in Project Europe's pipeline

10:01 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

The next twelve months will be dramatically better for infrastructure companies upstream of the foundation labs than for application layer companies downstream of them

Application layer businesses sit close to the labs' own businesses — Claude Cowork can extend from software engineering into medicine, law and finance — so software-layer defensibility is very hard, while infrastructure players compound network effects, data moats and long compute R&D cycles into high-margin sustainable businesses

Scope: next twelve months

33:29 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 15, 2026

With unlimited money the highest-value thing to build is data centers on land, because physical infrastructure buildout is the return of the industrial age

Like the industrial revolution's pipelines, bridges and factories, AI now needs people thinking hard about scaling physical infrastructure cost-efficiently; and Perplexity is already good at converting infra into valuable output tokens

Scope: would start with land on earth rather than space, lacking that expertise; infra only matters if you can utilize it to produce value for users

69:24 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

Jerry Murdock · Aug 22, 2026

Fireworks' speed to scale is justified because it is the infrastructure layer riding on the frontier labs and open source, whereas comparable app-layer growth stories are not credible

They ride on the shoulders of the model builders as the next layer that needs to be developed, so they can scale right behind them

Scope: applies to infrastructure, not app layer or legal

33:27 20VC: The AI Bubble Will Burst: Half the Neoclouds Will Die | China: Should We Ban Chip Exports & Be Fearful of Chinese Open-Source | Mag7: Who Dies and Who Thrives: Why Meta is Meh and Microsoft is Mega

Harry Stebbings · Aug 24, 2026 · hedged

Infrastructure investments are more durable and certain than application-layer investments, because whoever wins the app layer still has to use the infrastructure

He can't tell which application providers survive — the app layer feels transient — whereas infra like Fireworks, ClickHouse and MotherDuck gets used by whoever wins on top

Scope: framed as personal uncertainty / humility

53:39 20VC: Inside Sequoia's Investment Committee: Lessons from Don Valentine, Doug Leone and Alfred Lin | How the SpaceX and Citadel Deals Went Down | What Sequoia Specifically Looks for in Founders with Julien Bek

Vince Hankes · May 3, 2023 · hedged

The speaker believes most AI value will accrue to the application layer rather than the infrastructure layer, though infrastructure providers can still clip a coupon as toll roads

In software, the value of AWS, Azure and Google Cloud versus the software built on top is roughly an order of magnitude to one, and value gets captured where it is created for end customers

Scope: infrastructure providers still act as toll roads and may be capitalized at high multiples; applications may compete more aggressively with each other

27:29 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

Shyam Sankar · Jan 17, 2024

Value in AI biases towards incumbents who own applications at the app layer

Models alone don't deliver the value — business outcomes happen in an application, and owning the pixels lets you apply the model faster; also most problems are only about a third AI, a third deterministic code, a third human thought, and that integration can only be done at the app layer

52:28 20VC: Palantir CTO on The Broken Incentive Structure of How Governments Buy Defence, The Danger of Defence Spending at Historic Lows, How Elections and Wars Change Government Defence Buying & Why Budgets are Anti-Creative with Shyam Sankar

Sarah Tavel · May 6, 2024

The application layer, not the infrastructure or model layer, will capture most of the value in AI

Whoever owns the end user can provide more and more value to them over time and capture that value

Scope: acknowledges intense competition and unresolved market structure at the model layer

16:39 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

Michael Eisenberg · Jun 19, 2024

The key strategic question in AI is where users start their day — the entry-point agent layer — rather than which foundation models get acquired by cloud providers

As agents proliferate, people will ask their AI screen to fetch content rather than visiting sites directly, making the starting point the valuable position, with an API-accessed model layer underneath serviced by hyperscalers

Scope: Anthropic appears ahead on the underlying API layer right now

8:51 20VC: Foundation Models are the Fastest Depreciating Asset in History, Lina Kahn is a Threat to American Capitalism, PE is Not Coming to Save the M&A Market & How China Could Overtake the US in the AI Race with Michael Eisenberg

Tomasz Tunguz · Apr 21, 2023

Startups' odds of success in AI are significantly higher at the application layer than at the foundational model layer

The diversity of customer needs at the application layer is far greater, leaving room for many winners

0:00 20VC: Who Wins in AI; Startup vs Incumbent, Infrastructure vs Application Layer, Bundled vs Unbundled Providers | From 150 LP Meetings to Closing $230M for Fund I; The Fundraising Process, What Worked, What Didn't and Lessons Learned with Tomasz Tunguz

Tomasz Tunguz · Apr 21, 2023

Investor odds of success are significantly higher at the AI application layer than at the infrastructure layer

In web two the top three clouds and the top 100 public cloud application companies each total roughly $2.1T of market cap — the same value spread across 3 companies versus 100 — because the diversity of needs at the application layer is greater

Scope: if the web two analogy holds

27:29 20VC: Who Wins in AI; Startup vs Incumbent, Infrastructure vs Application Layer, Bundled vs Unbundled Providers | From 150 LP Meetings to Closing $230M for Fund I; The Fundraising Process, What Worked, What Didn't and Lessons Learned with Tomasz Tunguz

Harry Stebbings · May 17, 2023

Infrastructure and application layers have similar total TAM (~$2T each), but infrastructure concentrates value into about three companies versus roughly 50 in the application layer, so average enterprise value per company is far higher in infrastructure

Tunguz analyzed the two layers and found equivalent TAM but very different company counts

42:31 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 @

Aaron Levie · May 22, 2024

Far fewer juggernaut companies will be created at the foundation model layer than at the application layer, with room for perhaps only one to three independent horizontal LLM players at scale

Players like Zuckerberg are willing to spend billions commoditizing the model layer, so differentiation is fragile and a single large training run from OpenAI, Google or Meta can take a model company out

Scope: excludes hyperscalers; niche, industry-specific or domain-specific model companies remain viable

6:18 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?

Martin Casado · Jul 28, 2025

The only investing sin in this AI cycle has been zero-sum thinking; every layer of the stack has captured value and produced winners

Markets are so large and growing so fast that things people assumed were silly are now profitable businesses; the worry about defensibility and margin at each layer has been answered 'yes' nearly unilaterally

Scope: observational, based on what has happened so far

4:50 20VC: a16z's Martin Casado on Anthropic vs OpenAI: Where Value Accrues | Cursor vs Replit vs Lovable: Who Wins and Who Loses | The One Sin in AI Investing | Why Open Source is a National Security Risk with China

Maor Shlomo · Nov 24, 2025

The key open question for the AI market is whether model improvements benefit the whole ecosystem or only the model companies, whose margins will erode as they cut prices to compete

If value from each model generation accrues only to the lab that ships it, model companies fight and margins worsen; if it spreads, the market keeps compounding

52:06 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?

KR Sridhar · Jun 29, 2026

Value in the AI buildout will accrue broadly across every layer of the stack rather than concentrating in a couple of winners

AI will become ubiquitous, and ubiquity requires an entire sustaining ecosystem — consumers, edge data centers, service providers and suppliers all have to benefit or the ecosystem doesn't hold

31:22 20VC: Leo Aschenbrenner's Largest Holding: Inside the $90BN Bloom Energy | Why Electricity, Not AI Models, Will Decide the Winners of the AI Race | Why We Are Not in an AI Capex Bubble | Energy Sovereignty and The Future of Power with KR Sridhar

Emad Mostaque · May 17, 2023

Incumbents win the next three to five years, but many startups will still reach billion-dollar outcomes, including thin-layer ones

Value and moats are not necessarily innovation-first — ITA sold for $700M while Kayak, a layer on top of ITA, sold for $2B

42:17 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 @

Sanjit Biswas · Dec 8, 2023

In AI, large-scale infrastructure favors the best-capitalized incumbents while the application layer favors startups, so it is not an either/or but a question of where you play in the stack

The hyperscaler wars were won by companies able to invest billions ahead of time in data centers, whereas rapid feedback loops and creativity at the application layer suit startups

Scope: distinguishes infrastructure layer from application layer

44:13 20VC: $18BN Market Cap and $1BN in ARR in 8 Years; Samsara | How to Find Product Market Fit Reliably | How to Create a Multi-Product Company | The Pros and Cons of Serial Entrepreneurship with Sanjit Biswas, Founder & CEO @ Samsara

Eran Zinman · Mar 2, 2026

Nobody can currently tell who wins the AI race — narratives on Microsoft, Google, OpenAI and Anthropic flip every three months, and we are probably missing 90% of how it plays out

We're at the beginning of an exponential; it echoes 1998 when people assumed Yahoo and Netscape would capture all the value of the internet revolution

Scope: calls for humility rather than a directional call

59:02 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

Harry Stebbings · Jun 22, 2026

Venture investors genuinely do not know where value will accrue across the AI stack of infra, models and apps

Everyone is looking at the same stack and none of them can tell

Scope: speaking for venture investors generally

26:33 20VC: Nikesh Arora on the Frontier Model Problem: Breadth vs Depth | The Future of Token Costs | Memory Becoming the Moat | Where Value Accrues: Infra, Models, or Apps? | Why Enterprise AI is Not Ready & Systems of Record vs Systems of Intelligence

Miles Grimshaw · Sep 18, 2023

The framework and tooling layer for building AI apps will become a major opportunity, analogous to Rails/New Relic, Next.js/Vercel and Docker in prior paradigms

Each new development paradigm produced a dominant framework plus tooling and monitoring layer that empowered tens of millions of developers; the same will happen for AI apps

77:04 Framework and tooling layer for ai apps repeats past platform shift patterns

Arthur Mensch · Apr 29, 2024

Two opposing forces act on where value accrues: improving models and better tooling thin out the application layer, while falling price per unit of intelligence and competitive pressure thin out the model layer

Models get better, making verticalized apps easier to build; but models also get cheaper through compression and efficiency gains plus competition, driving down dollar per intelligence unit

15:40 Opposing forces model improvement thins app layer while falling intelligence price thins model layer

Arthur Mensch · Apr 29, 2024

Widely available open source LLMs push value up from the model itself to the platform and customization layer, and accelerate that shift

24:54 Open source availability accelerates value shift from model to platform and customization layer

Mamoon Hamid · Oct 21, 2024 · hedged

The middleware layer between foundation models and applications is overinvested and much of the value created there is fleeting

When the rate of change is this high in early innings, a lot of investing crowds into that layer, but the technologies (e.g. vector databases, fine-tuning weights) may not hold value

15:09 Middleware layer is overinvested given capital abundance

Nikesh Arora · Jun 22, 2026 · hedged

Value will be shared between frontier models and the context accumulated in enterprise applications, and frontier labs recognize context/memory is where the gap is

The more context a model holds about a user, the easier it is to answer them well in future, which creates switching costs; storing 30-90 days of personalized interaction requires significant storage and personalization

27:33 Value splits between frontier models and application context

Zach Lloyd · Oct 17, 2025

If model providers stay competitive with each other, the app layer will capture a lot of the value; a single model provider running away with the market would be horrible for app-layer companies.

There is real margin somewhere in the stack; who captures it depends on whether the model layer is competitive or monopolized

39:26 App layer value capture depends on model market competitiveness

David Cahn · Aug 5, 2024

The AI supply chain is where the most interesting investment opportunities are right now, because it is less picked over than the application layer

It's less picked over than more obvious AI categories

34:57 Ai supply chain is less picked over than app layer and offers the best opportunities

Alexandr Wang · Jun 12, 2024

In AI, where value accrues in the stack is constantly shifting, mirroring Intel's repeated migrations described in High Output Management

The field is so new and nascent that the point of value capture keeps moving, just as Intel had to repeatedly migrate to different parts of the stack

24:18 Value capture point keeps migrating across the stack like intels history

Harry Stebbings · Aug 3, 2026 · hedged

The physical and mechanical infrastructure for compute — cooling systems, steel, data center buildout — is relatively underhyped

58:49 Physical datacenter buildout is the underhyped layer

Mike Mignano · Jul 6, 2026

The AI infrastructure build-out phase is essentially complete and we are now entering the application layer phase, mirroring the fiber and broadband build-out that preceded internet applications.

The labs consumed vast capital to build magical technology; once that infrastructure exists, an application layer emerges to take advantage of the new technology, as happened after fiber/broadband.

9:33 Buildout phase is done so the application layer phase begins now

Shyam Sankar · Jan 17, 2024

Value will accrue in two directions — infrastructure and the app layer — with the model layer squeezed in between

Infrastructure players will capture a lot of the compute

53:34 Model layer gets squeezed value accrues to infrastructure and app layer on both sides

Des Traynor · Nov 15, 2023

Most AI value today is going to the infrastructure layer, but that layer will commoditize into AWS-like price-plus-product competition among three or four providers (GCP, AWS, OpenAI, Azure)

More competitors offering the same thing takes the market from monopoly to oligopoly to perfect competition, with price falling at each step; the sheer amount of investment in the space guarantees this

44:30 Infra layer currently captures most value but will commoditize into an aws like oligopoly

Roman Chernin · Jun 8, 2026

Building infrastructure is the comparatively easy part of AI; the people building end-user products are taking the real risk and driving most of the growth

In infrastructure you broadly know what is needed and customers tell you; in end-user products you risk building something people may not need

46:29 App builders carry the real risk and drive the growth

Andrew Ng · Nov 17, 2025

It is hard to deploy large amounts of capital at the application layer because trying an idea out is so cheap

If an idea only costs a million dollars to test, there is no obvious way to put $10bn to work the way you can with data centers, and much app-layer investment just flows through to OpenAI/Anthropic and ultimately NVIDIA

29:26 App layer cannot absorb large capital because experiments are cheap

Nabeel Hyatt · Apr 4, 2025

Investing in both foundation models and the application layer is not hedging — both layers can win and accrue sustainable value

47:46 Investing in both model and app layers is not hedging since both can independently win

Tom Hulme · Apr 10, 2025

In a world of commoditised foundation models, value accrues at the application layer and in hardware, where China leads and the West is playing catch-up

If models commoditise, the value moves to applications and to the devices that will be the conduit for commoditised AI, and China is far better at manufacturing and hardware value-add

63:13 Commoditized models push value to app layer and hardware where china has manufacturing edge

Tom Hulme · Apr 10, 2025

Over the next five years value in AI will aggregate to the application layer, where brand is decisive, and to hardware as a conduit for AI

Brand-driven consumer adoption is enormous and installed device bases can distribute AI, which is why they invested in Nothing with its 7 million devices

69:07 Value aggregates to app layer via brand and to hardware as ai distribution conduit

Bret Taylor · Oct 2, 2024 · hedged

The AI market will commercially play out much like the cloud market did, splitting into infrastructure, toolmakers, and a long tail of valuable software-as-a-service applications, because companies prefer buying working solutions over building and maintaining their own software.

Whatever happens at higher layers, the bottom layers capture value from everyone working above them

11:17 Ai market splits into infra tools and long tail saas like the cloud market did

Bret Taylor · Oct 2, 2024

The biggest misconception about the coming decade of AI is the focus on hardware and models rather than applications; many of the defining AI companies will be delivering consumer and business solutions that happen to be powered by AI, not the models themselves.

61:44 Defining ai companies will be application and solution companies not hardware or model companies

Varun Mohan · Jun 2, 2025

Model providers can differentiate at the API layer by specializing for particular workloads rather than only by moving into the app layer

E.g. optimizing latency for code output where much of the generated code resembles existing code makes a provider a better vendor for a coding company than for an image workload

51:13 Model providers differentiate via workload specific api specialization not app layer expansion

Matan Grinberg · Jun 13, 2026

Disagrees that the next twelve months will be most value-accruing for AI infrastructure while model and application layer companies are most at risk

The old software moat of exclusive capability is disappearing, but the value question becomes whether it is worth your time to build versus buy — which favors specialist application providers

11:10 Build versus buy economics favor specialist application providers

Richard Socher · Apr 18, 2025

Investing in pure LLM/infrastructure companies is unattractive unless the company owns the end user or owns a vertical

23:04 Pure infra plays unattractive unless owning end user or vertical

Matt Murphy · Jul 27, 2026

The AI infrastructure and developer tooling stack above the foundation model was a false negative and is now underinvested

Early bets 2-4 years ago failed because everyone focused on single models and didn't need surrounding infrastructure; now the ecosystem is bigger, spend needs managing, and observability, routing and abstraction layers are taking off

56:38 Middleware above the model is the underinvested layer now that the ecosystem is multi model