How far down (or up) its stack should a company vertically integrate, and when?

Matan Bar · Feb 11, 2021 · hedged

Melio will most likely partner for credit rather than build it, because underwriting and holding funds on the balance sheet would be defocusing

The focus is on becoming the best B2B payments platform and user experience, which is already a lot of work; credit underwriting and balance sheet funds are complicated and distracting from product strategy

Scope: 'most likely' / 'probably' — decision not final

14:58 20VC: Scaling to a $1.3Bn Valuation While in Stealth, The Power of Different Network Effects Within Payment Platforms & How To Leverage Your Board and Investor Base for the Most Value with Matan Bar, Founder & CEO @ Melio

Mike Salguero · Apr 5, 2023

A new subscription company should partner with established operators rather than build its own distribution, fulfillment and supply stack, because a startup cannot out-execute a firm with a century of experience straight out of the gate.

It takes a special kind of entrepreneur to believe they can get into distribution and beat a company that has been doing it for 110 years; ButcherBox instead partners for farms, slaughterhouses, cutting, distribution, last-mile and customer service while holding partners accountable through quarterly business reviews.

Scope: contrasted with Blue Apron-style peers who raised hundreds of millions to own everything

14:20 20VC: The Memo: Scaling to $600M Revenues with No Venture Funding, The Most In Detail Breakdown of Consumer Subscription Unit Economics & Why D2C and Consumer Subscription is Not a VC Backable Model with Mike Salguero, Founder @ ButcherBox

Kevin Niparko · Sep 29, 2023

Product teams fall into the build trap and should put a premium on partnerships to expand product capabilities instead of building

Products that look simple on the surface hide 'dragons' a few levels deeper and turn out far harder than teams credit; there is more partnership opportunity than product teams typically see

Scope: especially in areas outside your core competency

41:14 20Product: Why Product Memes Are More Important Than a Product Roadmap, Why Writing is the Essential Skill for Product People, How AI Changes The Role of Product, Big Mistakes Founders Make When Hiring Product Teams with Kevin Niparko, VP Product @ Twilio

Harry Stebbings · Sep 29, 2023

Shopify was right not to build its own payments stack in-house and to focus on where its business is instead

Though Shopify could build a Stripe internally, it isn't the right focus for their business

Scope: offered as an illustrative example of build-vs-partner

41:59 20Product: Why Product Memes Are More Important Than a Product Roadmap, Why Writing is the Essential Skill for Product People, How AI Changes The Role of Product, Big Mistakes Founders Make When Hiring Product Teams with Kevin Niparko, VP Product @ Twilio

Roman Chernin · Jun 8, 2026

The higher up the stack an infrastructure provider moves, the more value it creates and the larger the addressable customer population — roughly a dozen customers at bare metal, hundreds at managed infrastructure, thousands at inference, tens of thousands at the agentic layer

Each higher layer removes more complexity, so it serves a broader class of less specialised builders

19:08 20VC: Nebius Co-Founder on AI Infrastructure Bubbles | The Real Impact of Open Source on OpenAI & Anthropic | How Price Elastic is Demand for Compute | Could Nebius Sell 10x More Compute If They Had It & more with Roman Chernin

Roman Chernin · Jun 8, 2026 · hedged

Building the full software stack is necessary for long-term protection of the business, though in a world of infinite demand a pure bare-metal contract business could be sustained for the medium or even long term

As competition on the demand side grows you can be selective and work with customers who value the platform, rather than being locked into concentrated capacity supply relationships

Scope: acknowledges nobody knows where the world ends up; infinite-demand scenario is the exception

22:27 20VC: Nebius Co-Founder on AI Infrastructure Bubbles | The Real Impact of Open Source on OpenAI & Anthropic | How Price Elastic is Demand for Compute | Could Nebius Sell 10x More Compute If They Had It & more with Roman Chernin

Roman Chernin · Jun 8, 2026

Nebius's differentiation is full-stack integration both downward (data centers, racks, servers, platform) and upward (products serving customer segments), which lets it move faster, squeeze cost, and serve a broader customer base than raw-infrastructure buyers

Controlling the physical layer downstream enables speed and better economics, while integrating upstream removes the limit of serving only the small population who just need infrastructure

Scope: Roman declines to compare directly to competitors

29:51 20VC: Nebius Co-Founder on AI Infrastructure Bubbles | The Real Impact of Open Source on OpenAI & Anthropic | How Price Elastic is Demand for Compute | Could Nebius Sell 10x More Compute If They Had It & more with Roman Chernin

Harry Stebbings · Jun 13, 2026

Nebius's ambitions to go full stack don't make sense and would put it in competition with application-layer companies, whereas CoreWeave stays out of their way

Full-stack expansion eats into the plans of application companies

68:20 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

Lin Qiao · Jul 20, 2026

Fireworks will not move up into the application layer, though moving down the stack into data centers remains on the table

Scope: data center expansion described as 'always on the table', not planned

0:00 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

Lin Qiao · Jul 20, 2026

Fireworks will not move up into the application layer, though moving down into building data centers remains on the table as a question of timing

Scope: data center move is a timing question, not a commitment

49:15 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

David Luan · Jun 24, 2024

There will be strong vertical integration pressure between model builders and chipmakers, with each moving into the other's layer

AI forces the question of which offerings get bundled or integrated and which get unbundled

24:58 20VC: Why Foundation Model Performance is Not Diminishing But Models Are Commoditising, Why Nvidia Will Enter the Model Space and Models Will Enter the Chip Space & The Right Business Model for AI Software with David Luan, Co-Founder @ Adept

David Cahn · Aug 5, 2024

Vertical integration between the model layer and the data centre will matter increasingly — you cannot have a separate team running the data centre and a separate team building the model.

As models get bigger, data centre operation and model building must be deeply coupled; Musk and Zuckerberg already control their own data centres while OpenAI/Microsoft and Anthropic/Amazon are split across separate companies.

Scope: a view he changed his mind on over the last few months/year

23:53 20VC: Sequoia's David Cahn on AI's $600BN Question | Why the Data Centre is the Most Important Asset | Servers, Steel and Power: The Core Pillars Powering the Future of AI

Aravind Srinivas · Jun 15, 2026

CoreWeave has been more successful at building data centers than OpenAI because the work is operationally intensive and demands total focus on permits, power, supply chain bottlenecks and total cost of operations

It's hard-to-replicate operational work — securing permits, planning ahead, testing systems, handling random physical issues

35:04 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

Lin Qiao · Jul 20, 2026

Data centers are not commoditized — construction, power, fiber, and liquid cooling for newer chips require deep expertise that cannot be acquired quickly

Every stage from construction to power to cooling to part replacement demands specialized expertise; she says she herself could not become a data center operator tomorrow

51:29 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

Shyam Sankar · Jan 17, 2024

Model companies should push into the app layer to own more of the value they create rather than just selling an API endpoint, contrary to standard venture advice to build something simple that scales via API

Scope: framed as what he would do if he ran a model company

53:50 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

Eiso Kant · Oct 7, 2024

Value in AI will not accumulate only at the model layer but all the way to the end user, so deep vertical integration — building end to end — is how a foundation model company avoids becoming merely an IP source for others

BYD started as a battery company and became the largest-volume seller of electric cars in the world, showing the payoff of deep vertical integration

Scope: still expects more value to be built on top of Poolside than Poolside can unlock itself

55:06 20VC: Raising $500M To Compete in the Race for AGI | Will Scaling Laws Continue: Is Access to Compute Everything | Will Nvidia Continue To Dominate | The Biggest Bottlenecks in the Race for AGI with Eiso Kant, CTO @ Poolside

David Cahn · Oct 27, 2025

OpenAI and Anthropic are becoming vertically integrated — effectively becoming server and power companies — and the big labs will keep moving down the supply chain

They are developing their own chips and procuring their own gigawatts of power; this has been one of the biggest trends of the last twelve months

12:59 20VC: Sequoia's David Cahn on The Winners and Losers in AI | The $0-$100M Revenue Club: Is Triple, Triple, Double, Double Dead? | The Future of Defence: Who Wins and Who Loses | How to Analyse Margins and Growth Rates in a World of AI

David Cahn · Oct 27, 2025

Competitive pressure will push all model providers toward vertical integration into compute and power, and this trend will be durable

Competitive pressures force every provider to spend time on and staff teams for this

13:38 20VC: Sequoia's David Cahn on The Winners and Losers in AI | The $0-$100M Revenue Club: Is Triple, Triple, Double, Double Dead? | The Future of Defence: Who Wins and Who Loses | How to Analyse Margins and Growth Rates in a World of AI

Lin Qiao · Jul 20, 2026

Companies should not try to own the full stack early; you earn the right to build a layer only once your own traffic is giant enough that building saves multiples of cost

Agility and focus matter most early; Meta didn't build everything when it was young but it made sense once it was big, and Cursor grew 1000x in two years precisely by focusing on product and research and outsourcing the platform layer

Scope: depends on company philosophy; changes once the business is very large

38:52 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

Matt Murphy · Jul 27, 2026

Building your own chips is a worthwhile bet for very large AI companies because specific workloads can be materially improved by custom silicon and it brings down cost structure

At scale you look at your bill and see you're overpaying; some model-specific behaviors (memory cache etc.) would be far better on purpose-built hardware, so it's worth the swing at ~$100B revenue scale

Scope: only at very large scale; the chip business is hard and requires a special team to compete with Nvidia; applies to some percentage of training or inference workload, not all

26:24 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

Alex Bouaziz · Oct 22, 2025

Founders face a choice between being an acquisition outcome with one to three great products and owning the rails as close to the metal as possible, and Deel chose to own the rails even though building payroll for 100+ countries looks crazy

Single-country payroll companies like PayFit and Gusto became unicorns/decacorns, so building global payroll rails from the ground up is a bold long-term play only justified by ambition about company size and long-term customer value

37:18 Choose between acquisition scale few products or owning the rails end to end

Alex Bouaziz · Oct 22, 2025

Building and owning Deel's own local entity infrastructure for employer of record, rather than using third-party partners, was the best decision the company ever made

Partners ran outdated technology and did not share Deel's standard of customer service; owning the entities let Deel deliver modern service, and the EOR product now generates a very large monthly revenue

60:45 Own critical compliance infrastructure rather than rely on partners

Andrew Feldman · Oct 6, 2025

Full vertical integration from chip to data centre to software is not required to win in AI

The two most successful AI companies to date are not vertically integrated — OpenAI ran on Azure infrastructure exclusively for years and Anthropic uses a mix of AWS and Google — proving other working models exist

38:30 Full vertical integration is not required to win openai and anthropic succeeded without it

Jerry Murdock · Aug 22, 2026

Owning the chip layer is the wrong direction for model companies long term, even though it makes short-term sense for large players optimizing chipsets for their models

The real opportunity sits above the silicon, in the complexity between model, agent and human-in-the-loop

28:59 Value sits above silicon so do not own the chip layer

Sebastian Siemiatkowski · Feb 16, 2026

Owning its own payment rails gives Klarna a structural data advantage over other fintechs, because it sees SKU-level digital receipts rather than just transaction amounts, which is what makes genuinely useful financial advice possible

To advise a customer on their everyday spending you need to know what they actually bought, not just where and how much — e.g. that the contact lenses they bought are overpriced

27:34 Owning the rails yields granular data rivals cannot see

Aidan Gomez · Aug 19, 2024

Companies will keep verticalizing into building their own chips, and the current lack of choice in the chip market will change faster than most people expect

Chips are exceptionally high margin right now and there is very little choice in the market

17:54 Chip market lack of choice will change fast as more companies verticalize into building own chips

Aidan Gomez · Aug 19, 2024

Building its own data centers is not currently an attractive path for Cohere because partner pricing beats doing it in-house

They have run the numbers and the prices from providers make self-build uneconomical; they would switch only if it became cheaper or if a compellingly cost-effective chip could not be procured by any provider

19:29 Building own data centers is unattractive when partner pricing beats in house cost

David Luan · Jun 24, 2024

Model builders must control their own chips, because a cost advantage on training compute compounds into a model quality advantage

If one company has, say, a 20% cost advantage via in-house chips like Google's TPU, it can train bigger models and invest more in post-training tricks, pressuring rivals to replicate the in-house effort

25:18 In house chip cost advantage compounds into model quality advantage

Mike Salguero · Apr 5, 2023

Owning bigger pieces of the stack does genuinely expand margins and improves EBITDA, since the amortization sits below that line.

It is how other food companies achieve margin expansion, and depreciation treatment made asset ownership attractive; ButcherBox built two dry ice factories both as a margin play and to control a critical input.

16:05 Owning more of the stack expands margin via amortization benefits

Paul Erlanger · Jun 27, 2026

You should build anything user-facing in the core product yourself and only buy or acquire behind-the-scenes infrastructure

Core surfaces like a unified social trading experience across web and phone can't be bolted on from an acquired terminal, while things like bare metal servers or indexing would take too long to build expertise in

43:14 Build user facing surfaces buy backend infrastructure

Andrew Ng · Nov 17, 2025

Vertical integration wins while an industry is immature and API boundaries are unclear, but as standards emerge horizontal specialists can succeed — so AI stack ownership will shift from vertical to horizontal over time.

Early computing required integrated players like IBM to solve interoperability across keyboard, CPU and memory; standards like USB later let separate makers interoperate.

48:10 Integrate while standards are immature then horizontal specialists win

Jonathan Ross · Sep 29, 2025

Few companies will successfully move into the chip layer; building a competitive AI chip is extremely unlikely to succeed for any individual attempt

Google ran about three simultaneous chip efforts and only one (the TPU) beat GPUs, and industry efforts like Tesla's Dojo are getting cancelled; replicating NVIDIA's level of optimization is like deciding to replicate Google Search

10:11 Few companies can successfully build a competitive ai chip

Jonathan Ross · Sep 29, 2025

The real unique selling point of building your own chip is control over your own destiny and negotiating leverage, not deploying the chip in mass production

Google built 10,000 AMD servers it knew would be thrown away just to get a discount on Intel chips; similarly, when a hyperscaler threatens to build its own, NVIDIA suddenly finds the GPU allocation — and owning a chip means NVIDIA can't dictate your allocation

14:47 Chip building value is negotiating leverage not deployment scale

Nabeel Hyatt · Apr 4, 2025

At the frontier of innovation you want to be vertically integrated — maximising the connective layer between the model you're building and the problem the consumer is trying to solve

Moving really fast requires as much connectivity as possible between model and end user need

51:27 Maximize connective layer between model and customer at the frontier of innovation

Lin Qiao · Jul 20, 2026

Building your own chip only makes sense once your workload and business have stabilized, and AI workloads today are still too dynamic to warrant it

Once hardware is taped out it is extremely costly to change, so you need a durable, stable workload pattern before baking logic into silicon; the application space is still experimenting and we are early in the workload maturity funnel

59:22 Workload stability gates custom silicon

Mati Staniszewski · Sep 8, 2025

For a company training models continuously and moving large volumes of data, owning your own data centers beats renting because you break even on a two-year horizon and gain speed and control.

They ran the math: continuous training plus data transfer costs make ownership break even within two years, and it now lets them run more experiments faster.

43:38 Continuous training workloads justify owning data centers over renting

Andrew Feldman · May 26, 2026 · hedged

Google's full-stack ownership from land to tokens is a real cost advantage, but being able to sell TPUs only to itself historically limits volume and therefore the size of the opportunity

Volume has historically mattered a lot in hardware; a single internal customer constrains demand and the cost curve you can drive

17:05 Internal only volume caps the payoff from owning silicon

Andrew Feldman · May 26, 2026 · hedged

Investing in and acquiring companies in the application layer built on top of Cerebras is an opportunity newly available now that it is public, because public capital carries a different mandate than venture dollars

With venture dollars the question is whether your venture partners should invest instead; with public dollars the mandate and investor access differ

48:47 Public company capital enables moving up into the app layer

Andrew Macdonald · Aug 17, 2026

Divesting ATG was the right decision given the circumstances, and it is rose-colored glasses to assume Uber would have led autonomy had it stayed in

During COVID mobility lost 84% of top line in three weeks, the core was burning billions, and Uber was already trailing (at minimum trailing Waymo); the focus strategy turned the core into cash-flowing machines and every metric since is up and to the right

20:53 Divest a trailing layer and source it from partners

Harry Stebbings · Feb 17, 2025

Meta's $65B data center spend amounts to internalizing the margins it would otherwise pay to third-party compute providers, going full stack

40:13 Meta datacenter spend internalizes third party compute margins