Are users loyal to specific AI models, or is the model layer an interchangeable commodity?

Shyam Sankar · Jan 17, 2024 · hedged

LLMs will be commoditized, with most models becoming good enough even if some remain better than others

Scope: some models will still be better than others

53:03 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

Harry Stebbings · Jan 20, 2025

Model providers have been commoditized very quickly

47:58 20VC: Why Large Seed Rounds Increase the Chances of Success | When to Sell in Venture | Why Multi-Stage Firms Do Not Do The Work | Is Europe Totally F****** and Why AI Means London Can Compete with the US with Hussein Kanji

George Sivulka · Jan 22, 2025

Models are far less defensible than clouds because switching models is an API key change while switching clouds costs a mid-sized startup $10-20M and is almost never worth it

Switching costs determine stickiness; an entire industry will emerge just to fail models over from OpenAI to Anthropic when OpenAI goes down

42:42 20VC: Why All AI Companies Are Under-Valued | The Future of Foundation Models: Scaling Laws, Generalised vs Specialised, Commoditised? | From Unable to Afford Rent to Raising $130M From Index and Peter Thiel with George Sivulka @ Hebbia

Richard Socher · Apr 18, 2025

Switching costs are very low both between AI coding tools and between consumer LLMs; real switching costs only appear where there is company-internal or unique proprietary data

None of the consumer LLMs are that amazing yet and none do enough personalization

Scope: 'yet' — could change with better personalization

43:57 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

Harry Stebbings · Nov 24, 2025

The common analogy between the LLM market and the cloud market breaks down, because you cannot switch cloud providers with a line of code multiple times a month the way you can switch models

The cloud comparison assumes commoditization with a couple of core winners capturing a mega market, but that rests on switching friction that does not exist for LLMs

25:59 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?

Harry Stebbings · Nov 24, 2025 · hedged

Frontier labs are delusional about their moat — they don't have what they think they have

64:45 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?

Harry Stebbings · Jan 12, 2026

Customer promiscuity has never been higher — buyers of AI products like Cursor and the foundation models are customers, not hostages, and can switch easily

26:51 20VC: a16z's $15BN Fundraise with Alex Rampell | The Best Companies Have Hostages Not Customers | The Best Founders Materialise Capital, Customers and Labour | Mid-Sized Funds with Die and The Future of Venture Capital

Brendan Foody · Jun 1, 2026

The model API layer will be commoditized, even though stickiness can still be built in workflow products layered on top of the APIs

Switching costs at the API layer are zero and a new frontier model arrives every two months, so buyers with an eval for a workflow can compare models one-for-one and hot swap — which is close to the definition of a commodity

Scope: applies to pure API-based products, not workflow products built on top

45:07 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

Arvind Jain · Jul 11, 2026

The model layer is commoditized for enterprise use cases, allowing Glean to control costs by routing each task to the appropriate model, including open source models when they will yield a high-quality answer

Because many models are interchangeable for most enterprise tasks, Glean can route to the cheapest adequate model as a cost-control value add

Scope: scoped to enterprise use cases

14:11 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

David Frankel · Aug 8, 2026

Suno's value lies in the consumer experience and interface, not the underlying model, which is obfuscated from and irrelevant to the user

Suno's own CTO said without hesitation he would swap in an outside LLM if it were better than the model he built — the user only cares about getting there

74:46 20VC: The AI Boom Will Create Enormous Roadkill: Who Wins & Loses | Why Founders Should Never Take Multi-Stage Money at Seed | Why Triple, Triple, Double, Double is Good Enough

Harry Stebbings · Aug 10, 2026

The model layer is becoming a commoditized utility layer that users have no brand loyalty to

Using Arena, he submits a prompt and picks the best of several outputs without caring whether it came from Kimi, Muse, Claude or Sonnet, and it surfaces models he would never otherwise have tried

Scope: based on his own usage pattern

44:31 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 · hedged

Model personality and response style will not fully commoditize, so users of software will stay wired into a particular model for a given use case rather than having responses served interchangeably from any model

Interchangeable responses across Gemini, GPT-4 and Claude for the same user interaction is impractical

Scope: customers may still choose different models for different use cases

7:43 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?

Steeve Morin · Feb 24, 2025

Nobody is currently a real threat to OpenAI, because outside the industry bubble ChatGPT has the brand and consumers don't even recognize rivals like Gemini

Ordinary consumers know ChatGPT but not other models; OpenAI has a strong brand and strong product

Scope: balance between product and models matters

60:51 20VC: Why Google Will Win the AI Arms Race & OpenAI Will Not | NVIDIA vs AMD: Who Wins and Why | The Future of Inference vs Training | The Economics of Compute & Why To Win You Must Have Product, Data & Compute with Steeve Morin @ ZML

Jake Saper · Mar 10, 2025

Sustaining value in AI accrues to products people use every single day such that usage becomes muscle memory, which is part of why OpenAI is powerful.

Habitual daily use means users open your app instead of the default alternative, as people now do with ChatGPT instead of Google.

45:17 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

Harry Stebbings · Jun 20, 2025

There is real lock-in in consumer AI assistants; he has become unwaveringly committed to ChatGPT after previously believing the products were interchangeable

Memory features and brand affinity created stickiness he didn't expect

Scope: personal experience

71:55 20Product: How Duolingo Build Product 10x Faster with AI | Duolingo's Biggest Lessons on Paywalls, Push Notifications and In-App Purchases | Why Small Teams are the Future of Product | Why PMs Will Become Extinct with Cem Kansu, CPO @ Duolingo

Kieran Flanagan · Jul 11, 2025

Memory will make ChatGPT's paid product extremely strong and will make analytics and optimization far harder than it was on Google

Results become personalized to each user's history, so two users asking the same question get different recommendations and there is no way to collate that into decipherable analysis the way Google allowed

31:48 20Growth: The Death of Growth Teams? | How Hubspot Use AI to Triple Email Conversion | The Future of AI SEO | Why Prompt Engineering is the New Coding | What Every CMO Needs to Know About AI in 2025

Harry Stebbings · Jul 11, 2025

LLM providers are not commoditized because memory is an incredibly powerful retention mechanism

Switching to DeepSeek or Claude means losing all accumulated memory, making it a completely different product

32:38 20Growth: The Death of Growth Teams? | How Hubspot Use AI to Triple Email Conversion | The Future of AI SEO | Why Prompt Engineering is the New Coding | What Every CMO Needs to Know About AI in 2025

Kieran Flanagan · Jul 11, 2025

Memory is the most defensible part of an LLM

As a power user of all the apps he finds it very hard to leave ChatGPT because memory has fine-tuned it to who he is and what he wants to do

32:55 20Growth: The Death of Growth Teams? | How Hubspot Use AI to Triple Email Conversion | The Future of AI SEO | Why Prompt Engineering is the New Coding | What Every CMO Needs to Know About AI in 2025

Nick Frosst · Sep 1, 2025

Cohere would tolerate large increases in Cursor pricing because Cursor's product quality is worth it, and would only switch to plugging its own models into an editor if costs rose on the order of a thousandfold

Cursor has built genuinely good UX and product; Cohere's own models would not currently match the output quality, though Cohere makes its own models and could plug them into a VS Code extension

Scope: as a business rather than personally; own-model quality parity not there today

61:26 20VC: Cohere Founder on How Cohere Compete with OpenAI and Anthropic $BNs | Why Counties Should Fund Their Own Models & the Need for Model Sovereignty | How Sam Altman Has Done a Disservice to AI with Nick Frosst

Alex Atallah · Aug 10, 2026

Real developer loyalty to specific models exists even when switching costs are near zero and better models are available

Churn data shows developers sticking with models for three reasons: their app already works and they don't want to break their optimizations and guardrails; newer models aren't cheaper; and they trust the outputs, often judged by idiosyncratic personal evals

Scope: 'some' loyalty; based on OpenRouter's own retention and churn data

35:49 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

Arthur Mensch · Apr 29, 2024

Brand is a critical determinant of model adoption because trust matters and developers cannot afford to evaluate everything themselves

People use models known to be good; community vouching substitutes for exhaustive evaluation, and open source generates trust

Scope: described as something Mistral learned along the way

19:34 20VC: Mistral's Arthur Mensch: Are Foundation Models Commoditising | How Do We Solve the Problem of Compute | Is There Value in the Application Layer | Open vs Closed: Who Wins and Mistral's Position

Harry Stebbings · Mar 3, 2025 · hedged

Brand is the most important thing in AI, because users already identify as a Claude person or a ChatGPT person rather than switching daily on capability

People aren't switching every day; they identify with their model of choice

28:27 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 · Mar 10, 2025

Brand is the overlooked Trojan horse in AI, since everyone now has a consumer front end but few of them get used.

Anthropic has a consumer front end, yet hardly anyone reaches for Claude.

45:33 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

Vince Hankes · May 3, 2023

Commoditization of raw model output will not be the dimension on which builders choose a model; the ecosystem around the model will decide winners

Many companies give search results similar to Google, but the ones that scale do so because their ecosystem grows; OpenAI's plugin launch and ChatGPT's record user growth show ecosystem and multi-modal convenience are what builders optimize for, plus access to hundreds of millions of users

Scope: not dismissive that model accuracy and size matter; framed as how people will look back in five years

21:42 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

Daniel Dines · Dec 18, 2024

The product experience around a model matters more than the model itself, because models are interchangeable components chosen on cost, speed and accuracy

Qwen is only the best model for that particular job today and may be swapped for Llama 3.3; the model is unusable without the surrounding experience of tagging documents and retraining on the fly

Scope: in the context of document understanding at UiPath

7:58 20VC: UiPath's Daniel Dines on Why Agents Do Not Mean RPA is F***** | Why We Have Reached the Upper End of Scaling Laws | The Future of Work in an Agent World and What Everyone Misunderstands About Enterprise AI

Victor Riparbelli · Jan 15, 2025

The text generation layer is being commoditized, and from here competition will be decided by distribution and product rather than raw model capability.

For most use cases where LLMs will transform the world, current-generation models are already good enough, so value shifts to the product and scaffolding built around them; X and Anthropic have caught up quickly while OpenAI's edge is its distribution moat.

Scope: for most current use cases; base models still need improving

24:32 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

Reid Hoffman · Jun 10, 2024

Foundation models will not be commoditized like grain or copper — they will retain meaningfully different strengths and weaknesses

Comparative work on Gemini, ChatGPT, Copilot and Bing Chat already shows distinct behaviors and use-case fit, even between OpenAI models

8:32 20VC: Reid Hoffman on Foundation Models: Who Wins & How Do Incumbents Respond | The Inflection AI Deal: How it Went Down | Why Trump is a Threat to Democracy | The Future of TikTok | Lessons from Sam Altman, Brian Chesky and the OpenAI Board

Demis Hassabis · Apr 7, 2026 · hedged

Models are not commoditising — the gap between the three or four leading labs and everyone else is starting to pull away

AI tools like coding and math assistants help build the next generation, and it is getting harder to eke out the same gains from the same ideas, so labs able to invent genuinely new algorithmic ideas will gain a bigger advantage as the current set of ideas is exhausted

Scope: over the next few years

11:54 20VC: DeepMind's Demis Hassabis on Why AGI is Bigger than the Industrial Revolution | Why LLMs Will Not Commoditise & We Have Not Hit Scaling Laws | Bottlenecks in AI & The Energy Crisis Caused By AI | Whether AI Will Do More to Harm or Help Inequality

Arthur Mensch · Apr 29, 2024

AI developers primarily care about cost, customization, and portability of deployment — with portability mattering because it gives data control

Developers operate across clouds, on-prem and edge devices; LLMs become useful when connected to knowledge bases, so the model sees all the enterprise's sensitive data and enterprises need to secure that access

18:12 Developers value cost customization and deployment portability over raw capability

Harry Stebbings · Jul 1, 2024

The best companies will be defined by their ability to transition between underlying models

15:59 Ability to switch between underlying models defines the best companies

Victor Riparbelli · Jan 15, 2025 · hedged

For large-volume workloads like moderation, model choice is driven mainly by pricing and the surrounding scaffolding infrastructure rather than by raw capability differences

Some models are better at specific things, but for big workloads the economics and tooling around the model matter most

33:40 Pricing and scaffolding decide model choice for high volume workloads not raw capability

Mike Krieger · Mar 3, 2025 · hedged

Consumer AI users do attach to a provider based on personality, interface design and vibe, and the differentiators for AI products are likely model personality, the prescriptiveness of the product scaffolding, and vibes

It mirrors his Instagram-era formula that social networks are made of format, audience and vibes — Instagram stayed distinct from Snapchat and Facebook on vibes even when product features converged

28:42 Personality scaffolding and vibes drive consumer brand attachment

Brendan Foody · Jun 1, 2026

Every Fortune 500 will need a system of record for evaluating and specifying agent behavior across every workflow, and will use it to commoditize the model layer

Evals per workflow let a company identify the Pareto frontier of price performance and hot-swap providers, which enables perfect competition among models with zero switching costs

43:34 Enterprise eval systems of record force perfect model competition

Nabeel Hyatt · Apr 4, 2025

You should not evaluate any company by the models underneath it — not even model companies

63:50 Investors should evaluate companies independent of underlying model choice

Alex Atallah · Aug 10, 2026 · speculative

Fine-tuning may move from producing fully independent models to portable LoRAs or 'cartridges', such that swapping the base model layer could cost only a few dollars or a few dozen dollars

Many inference providers are already working on making these adapters portable between models

10:18 Portable adapters make base models cheap to swap

Varun Mohan · Jun 2, 2025 · hedged

The cloud-provider analogy does not currently hold for model providers, but it could become accurate later as models start holding persistent state about users and their codebases

Models could eventually hold stateful context — large codebases are billions of tokens — which would let providers raise switching costs

47:52 Persistent model state over codebase context could raise future switching costs

Eiso Kant · Oct 7, 2024

Speech recognition and image generation are commoditizing very quickly as AI use cases

47:36 Speech and image generation are commoditizing quickly