Will AI consolidate around a few general frontier models or fragment into many specialized models?

Edwin Chen · Jul 21, 2025

He has changed his mind and now expects a world with multiple frontier AI companies and multiple frontier AGIs rather than a single winner

Each lab can go in a different direction, as already visible in the differing strengths and weaknesses of OpenAI and Anthropic

58:26 20VC: Scaling to $1BN+ in Revenue with No Funding: Surge AI | The Most Insane Scaling Story in Tech |

Martin Casado · Jul 28, 2025

The coding model market will most likely end up an oligopoly rather than an Anthropic monopoly

The cloud is the best analog: AWS had 70-80% share and looked uncatchable, but Microsoft and Google spent their way in; likewise Google can arbitrarily subsidize Gemini without independent-company economics, OpenAI hasn't had a major code release yet, and models distill easily

9:24 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

Hemant Taneja · Sep 22, 2025 · hedged

The AI model market will consolidate to roughly three big global players plus a couple of sovereign ones per geography, and they will hold high margins rather than being commoditized

Empirically markets like telcos and clouds settled at about three players per geo, and cloud companies hold ~70% margins; the models are specializing into different areas and have many ways to add value and defend margin

Scope: 'in my view'; today's plethora of models will shrink — not everybody makes it; open question whether a new architecture emerges

41:21 20VC: General Catalyst CEO Hemant Taneja on The Future of Venture Capital: Chanel vs Walmart | Lessons Scaling GC to $40BN in AUM | Investing $5BN+ Into Stripe Over 14 Rounds | Investing Hundreds of Millions into Anthropic at $60BN Valuation

Harry Stebbings · Oct 17, 2025

The model layer will end up as three or four dominant players rather than open source taking share.

40:27 20VC: The Startup Adding $1M ARR Every Week | Competing Against OpenAI's Codex and Claude Code: Who Wins | Why Gemini is Failing and GPT-5 Is Winning | Do Margins Matter in a World of AI | The Ugly Truth About AI Coding with Zach Lloyd, Warp

Aaron Levie · Apr 20, 2026

It's impossible to call whether OpenAI or Anthropic wins the enterprise, and it won't much matter — the market will be a multi-vendor, multi-AI world where everyone wins

Cloud is the analogue: in 2010 AWS did $500M, Azure had just launched and GCP was Google App Engine, yet fifteen years later it's a few-hundred-billion-dollar ecosystem and everybody won; enterprises don't want single-vendor dependency given outage, API change and commercial model risk

Scope: concedes the labs themselves should still execute as if it's a winner-take-all race

44:24 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

Matan Grinberg · Jun 13, 2026 · hedged

At least four labs will end up approximately equally good at the frontier, rather than one or two running away with it

Growing evidence points that way, and he changed his mind from an earlier view that one or two would dominate

Scope: 'probably'; possibly many more than four

77:25 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

Tomer Cohen · Dec 20, 2023

For a specific job like a job-seeker coach you are better off building a specialized model or agent for that role than a general lifelong assistant

Specialized models are like an athlete retrained as a weightlifter or long-distance runner — they perform much better at the specific task, and a general assistant would be very hard to build

Scope: a massive model is warranted if you want a Her-style general personal assistant

17:03 20VC Roundtable: Spotify, Adobe & Linkedin CPOs on How AI Changes The Future of Product, Why AI is Now the Product, How TikTok Changed Product, Why Cost is the Biggest Barrier to LLM Usage & Why Incumbents Can Adopt AI Faster Than Any Prior Innovation Cyc

Martin Casado · Jul 28, 2025

The model layer will fragment into many differently-flavored models rather than consolidate

As you move into RL territory, models take on a specific flavor and generalize much less, which technically forces fragmentation; plus legit pioneer teams are starting new model companies, including for the sciences

Scope: applies less to the core base model for language, search and code

10:43 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

Roman Chernin · Jun 8, 2026 · hedged

The model landscape will keep broadening into new modalities and small specialized models — life science, robotics, world models, video, image, and domain-specific models built by fine-tuning open-source foundations

Teams increasingly start from an open-source foundation and train it for a particular use case optimized for the quality and latency that case needs, as with an Israeli cyber-defense foundational model team

36:05 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

Nikesh Arora · Jun 22, 2026

The model world will bifurcate into many task-specific models trained for depth in a vertical, rather than one mega frontier model being used for everything.

Task-specific models already outperform frontier models on their task, as seen with voice models, because depth of training in a vertical beats breadth.

Scope: framed as a question of what happens 'over time'

56:35 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

Lin Qiao · Jul 20, 2026

The future is not a small number of dominant AGI models but millions of specialized models, one per application per use case

Every company is special with its own taste, design principles and audience, so provider judgment will always mismatch and models must be tuned to fit

19:53 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

Douwe Kiela · Jun 30, 2023

No single model will win; the market is a pyramid of frontier models, mid-sized specialized models, and open source models serving different applications

Requirements differ by use case — strong AGI-like requirements need a frontier model, weaker ones can use specialized intelligence, and the least demanding can use off-the-shelf open source

22:43 20VC: Why Data Size Matters More Than Model Size, Why The Google Employee Was Wrong; OpenAI and Google Have the Advantage & Why Open Source is Not Going to Win with Douwe Kiela, Co-Founder @ Contextual AI

Daniel Dines · Dec 18, 2024

The model market will not consolidate like cloud did; there will be a few frontier models plus many dedicated specialized models

Even the human brain works this way — general cognitive models plus dedicated models trained for specific tasks like walking or picking up a cup

8:52 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

There will be only a few very big foundational model companies, and specialised application companies can win a narrow niche — for Synthesia, humans presenting to camera and voiceover — while potentially becoming customers or partners of the general video foundation models.

Sora, Runway and similar players are building broad true foundation video models with a much wider remit, whereas Synthesia only cares about its specific niche and wins on the platform as a whole.

31:32 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

Victor Riparbelli · Jan 15, 2025 · hedged

The end state is both concentration among a few large model companies and a proliferation of specialized models, most of which will be open-source models tuned for a task and embedded invisibly inside products

Most of the LLM-driven workflows people rely on will run in the background of someone else's platform, unnoticed by consumers

Scope: consumers will still interact directly with chat-style models too

34:41 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

Edwin Chen · Jul 21, 2025

There is room for both giant generalized models and smaller specialized models; large models will encompass the use cases in raw capability, but smaller models are still needed to make big bets and move fast in specific domains

Like Google or Facebook being unable to build products that conflict with the parent company's culture or business goals, an all-powerful model can't take a domain-specific bet because tuning one small domain pervades the whole model

Scope: large models will win on raw capability across use cases; specialized models need a genuinely unique view on how they operate

49:57 20VC: Scaling to $1BN+ in Revenue with No Funding: Surge AI | The Most Insane Scaling Story in Tech |

Brendan Foody · Sep 15, 2025

The future will contain both generalized and specialized models, having shifted from a specialized-model-dominant view toward generalization

o3 and GPT-5 showed how well models generalize; with so much headroom in foundational capabilities it is structurally more efficient to make those centralized investments, while enterprise model customization is still in the first inning

44:03 20VC: Mercor: From $1M to $500M in 17 Months: The Fastest Growing Company in the World | How to Think About Margins and Revenue Sustainability in AI | Why Evaluation Benchmarks in AI are BS Today with Brendan Foody

Brendan Foody · Sep 15, 2025

Having changed his mind over the past year, he now expects generalization to dominate and foundation model companies to be huge businesses, while still believing they should invest more in customization

Scope: acknowledges the position is somewhat contradictory

54:26 20VC: Mercor: From $1M to $500M in 17 Months: The Fastest Growing Company in the World | How to Think About Margins and Revenue Sustainability in AI | Why Evaluation Benchmarks in AI are BS Today with Brendan Foody

Mike Krieger · Mar 3, 2025

Frontier models will diverge and become more different from each other over time rather than converging toward sameness

0:00 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

Frontier models will diverge from each other over time rather than converge, developing distinct characters and areas of excellence.

Despite shared benchmarks, there is something 'Claude-y' about Claude and something 'GPT' about GPT, in both tone and capability; traction in a vertical like coding then feeds back into the next generation of reinforcement learning choices.

Scope: acknowledges many benchmarks look similar across labs

10:15 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

Edwin Chen · Jul 21, 2025 · hedged

There will likely be around three more frontier labs, each with different trade-offs, personalities and boundaries that make them good at different use cases

Claude is strong at coding, enterprise and instruction following, ChatGPT is optimized for consumer with a fun personality, and Grok is willing to be transgressive; just as there is no single greatest poet or mathematician, the richness of human intelligence will apply to models too

Scope: doubts it will be as many as ten more

58:57 20VC: Scaling to $1BN+ in Revenue with No Funding: Surge AI | The Most Insane Scaling Story in Tech |

Tomer Cohen · Dec 20, 2023 · hedged

Choosing which model to use for which purpose should be handled by a platform tier that leverages multiple models across cost and efficiency trade-offs and completely masks that complexity from developers, designers and product people

The decision of which model to use for which purpose originates at the application layer but shouldn't be made there; it depends on task, cost and efficiency trade-offs

11:06 20VC Roundtable: Spotify, Adobe & Linkedin CPOs on How AI Changes The Future of Product, Why AI is Now the Product, How TikTok Changed Product, Why Cost is the Biggest Barrier to LLM Usage & Why Incumbents Can Adopt AI Faster Than Any Prior Innovation Cyc

Aravind Srinivas · Jun 15, 2026 · hedged

Chinese models and Nvidia's own model efforts should keep the market from consolidating to two or three dominant providers, but inference companies still don't control their own destiny

Enough competing model sources exist to prevent consolidation, but the outcome is outside these companies' control

Scope: described as a leap of faith assumption

38:26 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

Harry Stebbings · Jul 20, 2026 · speculative

If AI proves transformative and the world ends up with many horizontal and specialized models, Fireworks can become a billion-dollar company

Upside depends on the many-model world thesis being right

Scope: conditional on a many-model, specialized world

0:34 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

Jonathan Siddharth · Dec 1, 2025

Enterprises will permanently need custom, fine-tuned smaller models rather than only using frontier general-purpose models

Narrow tasks like insurance underwriting don't require a trillion-parameter world model; smaller models are faster and more accurate on such tasks, can run on-prem so proprietary data never leaves the enterprise, and can distill a decade of institutional human judgment that the company doesn't want to hand to competitors

Scope: applies to specific enterprise workflows like underwriting and claims processing; models in the ~0.5B to 10B parameter regime

14:43 20VC: Scale, Surge, Turing, Mercor: Who Wins & Who Loses in Data Labelling | Is Revenue in Data Labelling Real or GMV? | Why 99% of Knowledge Work Will Go and What Happens Then? | Why SaaS is Dead in a World of AI with Jonathan Siddharth @ Turing

Anastasios Angelopoulos · Aug 3, 2026

Data businesses will expand into serving enterprises directly, and enterprise-specific data will become part of each company's moat

In a world where every business needs its own AI model, every business will also need its own data

45:00 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 · Aug 22, 2026

Specialized enterprise models will require surplus training data, letting data providers like Mercor expand from frontier labs to enterprise and mid-market buyers — a $200B opportunity

Enterprises and mid-market companies will need specialized data they don't already have to train their own models

Scope: Endorsing Lynn's thesis

51:26 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

Sarah Tavel · May 6, 2024

OpenAI's non-core modalities (audio, image, video) will progress mainly as beneficiaries of core LLM work and won't get the focus and ambition that single-modality specialists apply

OpenAI's main thing is clearly the GPT progression from 3 to 4 to 5 and beyond, so other modalities don't receive the same level of focus as a team like ElevenLabs dedicated to one specific use case

30:16 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

Saam Motamedi · Jul 15, 2024

Starting with a better focused model and then quickly moving up the stack into the application layer is a very good strategy — he was wrong that large foundation model companies would crush all smaller focused-model companies.

He previously assumed OpenAI would out-build any specialized voice or audio model within two years, and so passed on the category; ElevenLabs proved the app-layer path viable.

Scope: he still believes the underlying large models will win on raw model capability

59:37 20VC: Why We Are in a Bubble & Now is Frothier Than 2021 | Why $1M ARR is a BS Milestone for Series A | Why Seed Pricing is Rational & Large Seed Rounds Have Less Risk | Why Many AI Apps Have BS Revenue & Are Not Sustainable with Saam Motamedi @ Greylock

Severin Hacker · May 19, 2025 · hedged

General models will be highly competitive and interchangeable, but specialized modalities like voice have less competition and incumbents there can hold a durable lead

ElevenLabs' voices are so good he can see them being hard to replace with another model

Scope: specific to specialized use cases such as audio and vision

22:55 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

Scott Belsky · Dec 20, 2023

The future will not be a few mega models doing everything, but many long-tail models — local, cloud, open and closed source — with routing logic deciding which to use

Different queries need the most specialized model and also the most cost-efficient one, so routing across a long tail beats one giant model

Scope: contra the common belief in a few mega models in the cloud

11:47 20VC Roundtable: Spotify, Adobe & Linkedin CPOs on How AI Changes The Future of Product, Why AI is Now the Product, How TikTok Changed Product, Why Cost is the Biggest Barrier to LLM Usage & Why Incumbents Can Adopt AI Faster Than Any Prior Innovation Cyc

Tomer Cohen · Dec 20, 2023

Building one agent to rule them all is the wrong design; you want many specialized agents plus a dispatcher that coordinates them like a team coach

A large application has distinct jobs — seller, knowledge, job seeker — each better served by a purpose-built agent, so something must orchestrate them

Scope: applies at the application layer

12:37 20VC Roundtable: Spotify, Adobe & Linkedin CPOs on How AI Changes The Future of Product, Why AI is Now the Product, How TikTok Changed Product, Why Cost is the Biggest Barrier to LLM Usage & Why Incumbents Can Adopt AI Faster Than Any Prior Innovation Cyc

Amjad Masad · Apr 25, 2026

The right architecture is a 'society of models' — using models from every provider and routing cheaper sub-agents to tasks like code search while keeping a core loop on the strongest model

Anthropic is the workhorse because it runs coherently for long periods, but Google's Gemini models sit best on the price-performance Pareto frontier

8:40 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

Joelle Pineau · Nov 3, 2025

Multilingual models are commercially important rather than a nicety, because there is no one-size-fits-all model and workforces still operate in their national languages

Cohere's leading multilingual research showed it matters — customers in Japan and Korea want models that work well in their own language

Scope: Cohere's vision is to be a global AI company, not a Canadian one

26:29 20VC: Cohere's Chief AI Officer on Why Scaling Laws Will Continue | Whether You Can Buy Success in AI with Talent Acquisitions | The Future of Synthetic Data & What It Means for Models | Why AI Coding is Akin to Image Generation in 2015 with Joelle Pineau

Lin Qiao · Jul 20, 2026

The world is irreducibly diversified — regions differ in values, policies, business practices and lifestyle — so a single model cannot serve everyone and specialization is necessary

Taste, choices and judgment are what define humans; a world ruled by one standard and one company's taste would turn people into an army of robots

8:43 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

Richard Socher · Aug 18, 2023

It makes more sense for researchers and people to work together to keep making a single model better and better than for every NLP task to have its own separately built model, just as it makes more sense to keep improving one Wikipedia than for everyone to start their own dictionary from scratch

Analogy to Wikipedia: everyone adding to one shared resource beats each dictionary-builder starting their own from scratch

6:17 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

Richard Socher · Aug 18, 2023

The per-task model paradigm persisted because of academic preconceived notions and because the enabling ingredients — very large models, world knowledge from LLMs, attention mechanisms, fast GPUs — did not yet exist

With the tiny models of ten to twenty years ago, a single all-tasks model would have been impossible to make work

7:53 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

Emad Mostaque · May 17, 2023

The right architecture is standardized base foundation models across every modality, sector and nationality, with vector embeddings pointing to the personally important parts, rather than a million separate fine-tuned models

The base models already contain all the principles, so embeddings can capture what makes up an individual and be searched and adapted — a proliferation of models is just confusing

28:37 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 @

Steeve Morin · Feb 24, 2025 · speculative

Once efficiency gains pass a threshold, specialization will happen at runtime with one great model specializing per request rather than via fine-tuned smaller models

Large context windows already let you inject task data at runtime instead of fine-tuning

Scope: 'we're not there yet'; 'that's not for tomorrow'

56:17 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

Aravind Srinivas · Jun 5, 2024

The view that models will verticalize into domain-specific models is flawed; general-purpose models will beat them

What makes these models magical is emergent general-purpose capability from training on very diverse data — an abstract IQ that pattern-matches across skills — not domain specificity; Bloomberg spent heavily on Bloomberg GPT and was convincingly beaten by GPT-4 on finance benchmarks

Scope: he used to hold the opposite view

9:38 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

George Sivulka · Jan 22, 2025

Verticalized, post-trained and fine-tuned specialist models will always lose to general models that ride scaling laws, and fine-tuning will never catch up to scaling at inference

Bloomberg had the best financial data in the world and trained Bloomberg GPT, and GPT-4 destroyed it at every finance task weeks later

Scope: concedes we may be at the end of scaling laws at training, shifting the advantage to scaling at inference

37:30 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

Avishai Abrahami · Jul 13, 2026

A company with a large proprietary usage dataset can build a task-specific model that outperforms generic frontier models for its own product.

Base44 sees what users try to do and where prompts fail, so it can learn user intent better; it also doesn't need generic knowledge like Chinese poetry and can concentrate capacity on its own domain.

Scope: only for the narrow target of being better at Base44, not replacing frontier models

15:24 20VC: Wix's Founder on What Wall St Gets Wrong About AI and Wix | Will Base44 Win the Vibe Coding Wars | The Truth About the Economics of Vibe-Coding | The Buyback Disaster: Lessons Learned with Avishai Abrahami

Lin Qiao · Jul 20, 2026

The frontier of intelligence is private, specialized intelligence rather than generalized models

Intelligence is a derivative of data; the public internet plus labeled data is a tiny corpus versus the world's data, most of which is proprietary and locked inside enterprises and applications and will never be shared, so most data is not yet activated to derive intelligence

7:16 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

Anjney Midha · Apr 14, 2026

There will be no single God model; the future is an ecosystem of many model types, where general-purpose models are distributed broadly and specialized models are price- and product-segmented

The history of technology shows general-purpose products like the iPhone amortize development cost across the largest number of users and so go to everyone, while custom products get enterprise segmentation

47:31 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

Lin Qiao · Jul 20, 2026 · hedged

Treating foundation models as regulated base infrastructure has precedent, but a world where only one company owns intelligence would be bad

Utilities like PG&E set a precedent for base infrastructure ownership, but specialized intelligence is what creates new paradigms of living and business, and that shouldn't die because only one company can produce it

Scope: distinguishes general common intelligence from specialized intelligence

27:29 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

Alex Atallah · Aug 10, 2026

A multi-model future is inevitable and consolidation onto a single model makes no sense, even for a company with its own proprietary model

Whenever one model dominates, someone will build a deliberately neurodivergent model trained differently, creating demand for both; creativity is not verifiable, and using two models together is more likely to produce creative ideas than using one

11:01 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

Alex Atallah · Aug 10, 2026

Even companies that train their own models are incentivized to keep using models built by the rest of the ecosystem

Everyone else is also training on newly acquired data, so other models embody data valuable to you; trying them out, merging them, or using them can improve productivity, push state of the art, or cut costs — serving either margin or growth goals

Scope: game-theoretic argument assuming everyone builds models

12:58 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

Harry Stebbings · Nov 22, 2023

Alongside commoditisation we'll see specialisation of LLMs by use case, because desirable model behaviour differs — hallucination is an asset for creative tools and unacceptable for financial data

Different applications have opposite tolerances for hallucination

23:16 Specialization is driven by differing tolerance for hallucination across use cases

Clem Delangue · May 12, 2023

The world will be one of many distributed models rather than one model to rule them all, because an AI model is essentially a code base and no code base is universally best.

A code base is good or bad depending on the use case, constraints and what you want to do; the same logic applies to AI models, so model-building will be distributed across all companies rather than concentrated in a few organizations.

12:50 No single model is universally best like no codebase is

Anton Osika · Aug 18, 2025

Collapsing ChatGPT's five selectable models into a single GPT-5 was the obviously correct move for OpenAI and was executed pretty well, but consolidation inevitably means falling short on some dimensions

You can't improve in all directions at once, so optimizing many objectives into one model necessarily sacrifices some capabilities that separate models covered

18:13 Consolidating many models into one trades off per dimension strength for simplicity

Douwe Kiela · Jun 30, 2023

There is still a lot of room for new language model companies because incumbents like OpenAI and Anthropic are chasing consumer-facing AGI, leaving artificial specialized intelligence for business problems open

Enterprises don't need a model that knows Shakespeare or quantum mechanics, just one that solves their business problem — a different target from AGI

11:20 Room for specialized enterprise focused labs since incumbents chase consumer agi

Jerry Murdock · Aug 22, 2026

The AI market will shift from task-solving and minimal creativity toward requiring a diversity of intelligences to tackle genuinely hard, valuable problems

Just as boards and management teams need diverse intelligence to solve difficult problems, deep creativity like solving climate change requires a diversity of intelligence rather than one task-driven model

51:50 Hard problems require a diversity of intelligences

Richard Socher · Aug 18, 2023

Being able to incorporate multiple different model types is even more important than being able to switch between LLMs

LLMs are strong on natural language and world knowledge but poor at things like financial forecasting over long numeric sequences, so you need to plug in dedicated predictive models

19:06 Combining diverse model types matters more than switching between llms

Aidan Gomez · Aug 19, 2024

The world will contain both small verticalized models and large horizontal ones, because the emergent pattern is that people prototype on an expensive general model and then distill it into an efficient focused model for their specific task.

Users don't want to spend time fine-tuning before they know something is possible; they want to prove feasibility with a generally smart model first, then distill.

9:30 Prototype on general model then distill into specialized explains coexistence

Andrew Ng · Nov 17, 2025

The future will include a huge range of model sizes — large, mid-sized and tiny — rather than one dominant paradigm

Intelligence tasks span a huge range of difficulty; spell checking doesn't need a trillion-parameter model while complex code reasoning benefits from a powerful one, just as humans do tasks of varying difficulty

33:28 Task difficulty range implies a spectrum of model sizes

Sam Altman · Apr 15, 2024

The current weekly churn of one model leapfrogging another is normal for a new industry and will shake out to a small number — dozens — of providers doing extremely complex and expensive frontier models at scale.

The same dynamic happened when there were over a hundred US car companies and the press touted a new best car each week.

19:54 Car industry analogy predicts shakeout to dozens of frontier providers

Cristóbal Valenzuela · Aug 28, 2023

There will not be a single model to rule them all

It's like claiming the internet would only have one ecommerce site; AI is a general purpose tool, so there are many opportunities for different models and ways of working with them, and it's far too early to be definitive

16:57 Ai is general purpose so no single model will dominate like no single ecommerce site dominates the internet

Lin Qiao · Jul 20, 2026

Base general model IQ advances in step functions roughly every three quarters to a year, while specialization on top of those base models will accelerate much faster than general intelligence

Major releases deliver step functions like the introduction of thinking/reasoning, but each base-quality leap opens many more branches to specialize on, like a tree trunk sprouting branches and leaves

25:59 Base model leaps multiply branches for specialization

Matan Grinberg · Jun 13, 2026

The model market will not consolidate the way cloud did, because enterprises now insist on staying provider-agnostic

Cloud providers offered discounted multi-year deals then raised prices once customers were standardized and switching took two years; every CIO he speaks to carries those scars and refuses to throw their lot in with one model provider

60:50 Enterprise scar tissue from cloud lock in prevents consolidation

Harry Stebbings · Jun 5, 2024 · speculative

Curation of data as the central determinant of model quality may lead to verticalization of models, with different models used for different domains

9:20 Data curation quality differences will drive verticalization into domain specific models