Will a new wave of foundation model companies displace today's incumbent labs?

Larry Aschebrook · Jun 16, 2025

The way to play AI is to buy the established leaders, because nothing will rival OpenAI and Anthropic's scale within three years

Nothing is holding back the next generation of LLMs except time and capital, which the incumbents already have

71:09 20VC: How We Made $800M on Coursera | We Lost Money on Uber and Made Money on Lyft | We Did 3x on Postmates in 18 Months | DPI is King, MOIC is BS | We Dodged Theranos and I Still Lost Millions with Larry Aschebrook @ G Squared

Larry Aschebrook · Jun 16, 2025 · hedged

The LLM space is very hard for new entrants because catching up requires both time and enormous capital

Time and capital are the binding constraints, and incumbents already have both

73:18 20VC: How We Made $800M on Coursera | We Lost Money on Uber and Made Money on Lyft | We Did 3x on Postmates in 18 Months | DPI is King, MOIC is BS | We Dodged Theranos and I Still Lost Millions with Larry Aschebrook @ G Squared

Harry Stebbings · Jun 16, 2025

At the foundation model layer there is really only OpenAI and Anthropic, with little room for additional winners

74:50 20VC: How We Made $800M on Coursera | We Lost Money on Uber and Made Money on Lyft | We Did 3x on Postmates in 18 Months | DPI is King, MOIC is BS | We Dodged Theranos and I Still Lost Millions with Larry Aschebrook @ G Squared

Brendan Foody · Sep 15, 2025 · hedged

The largest model builders that will exist have already been created

The CapEx on data and compute, plus the cost of assembling research teams, has become prohibitively expensive for new entrants

Scope: not 100% sure; breakthroughs enabling further model progress could still come from startups

40:53 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 Mignano · Jul 6, 2026 · hedged

If recursive self-improvement happens, it is most likely achieved by one of the existing frontier labs rather than a newcomer

They already have vast compute advantages, frontier models, infrastructure and chips

Scope: 'if I were to bet'; a new architecture from labs like Thinking Machines or Safe Superintelligence could still leapfrog everything

17:02 20VC: Why Now is the Time for the Application Layer | Why OpenAI & Anthropic Won't Win the App Layer | Why Startups Should be TokenMaxxing | Why VCs Should Reduce Weighting on Price & Ownership in an Age of AI with Mike Mignano, USV

Douwe Kiela · Jun 30, 2023 · speculative

Today's leading labs like OpenAI and Anthropic represent a first-generation technology analogous to Lycos and AltaVista, while Contextual's approach is the PageRank-equivalent that could make it the Google of language models

Their technology is the right architecture at the right time, so with the right execution it could displace the incumbents as PageRank did early search engines

Scope: conditional on 'all the stars aligning' and right execution; stated as what he would hope for

38:42 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

Aravind Srinivas · Jun 15, 2026

No AI company can sit comfortably on a current win — if Anthropic treats Claude Code as already won, they won't even be around in six to twelve months

Even OpenAI's seemingly unassailable consumer position two years ago has become contestable, showing anyone can be pushed into fighting from behind

Scope: about the field generally, not a specific prediction of Anthropic's demise

7:49 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

David Frankel · Aug 8, 2026

OpenAI and Anthropic becoming seismic changes the landscape for the better, and they in turn will eventually be displaced

The ecosystem has repeatedly displaced apparently unassailable incumbents — Microsoft looked undisruptable and Google did it; Google looked undisruptable and search now starts elsewhere

57:24 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

Matt Murphy · Jul 27, 2026

Neolabs are overheated — round sizes are far too large for company stage and there is no way 60+ independent model companies survive alongside open source

There are 60+ neolabs, many generic 'get the band together' research plays; not all can have great acquihires, creating huge concentrated risk positions for some firms

Scope: he still likes the sector and Menlo is in seven; distinguishes focused labs like Chai and Axiom from generic ones

55:42 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

Anastasios Angelopoulos · Aug 3, 2026

Most neolabs are overrated, but Black Forest Labs is underrated

60:44 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

Alex Atallah · Aug 10, 2026 · hedged

Roughly 50% rather than 70% of neolabs will die or consolidate over the next three years

There aren't that many neolabs and 70% seems too high, though counting acquisition by a model lab as death raises the number

Scope: only if acquisition/consolidation counts as dying

52:46 20VC: Will OpenRouter Sell for $10BN to Stripe? | Why Chinese Open Models Are Beating America—and What Happens Next | Why Enterprises Are More Fearful of Anthropic and OpenAI Than China | Is the Routing Layer Becoming a Commodity with Alex Atallah

Emad Mostaque · May 17, 2023

There will only be five or six foundation model companies in the world in three to five years, and all of them already exist today (Stability, NVIDIA, Google, Microsoft/OpenAI, Meta, Apple).

Training frontier models requires capital at a scale only a handful of players can sustain — Google alone spends ~$20B a year on AI and DeepMind's salary budget is ~$1.2B.

Scope: three to five year horizon

42:51 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 @

Aravind Srinivas · Jun 5, 2024

Competing as a foundation model trainer against OpenAI is one of the worst arenas to be in, with only about five players left standing.

The game is extremely hard to play given the capital and talent required.

Scope: counts Google, Anthropic, Meta, Mistral and possibly xAI post-funding

21:17 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

Aravind Srinivas · Jun 5, 2024

Second-tier models — not cutting edge but cheap enough to build a business on — will get commoditized, while a small set of frontier models stay differentiated and are contested by only three or four players.

23:26 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

Matt Clifford · Jul 1, 2024

It is not too late to build an AGI company — the most ambitious founders today should seriously consider whether that's what they can do.

Scope: reversal of his view from twelve months earlier that the ship had sailed and there would only be three or four such companies

55:35 20VC: LLMs Are Reaching a Stage of Diminishing Returns: What is the Next S Curve | The Bull & Bear Case for China's Ability to Challenge the US' AI Capabilities | How AI Changes the Future of War & How Agents Will Reshape Society with Matt Clifford @ EF

Edwin Chen · Jul 21, 2025

We are only a few percent of the way toward true AGI, so there is enormous headroom for unexpected breakthroughs and new entrants despite the capital intensity

Real AGI means systems that cure cancer, send rockets to Mars and design new philosophical systems, not just automating an L3 or L4 engineer; assuming today's leaders are final is like assuming ten years ago that Google would be the last search engine, and future breakthroughs may come from AIs themselves working with humans

Scope: conditional on a demanding definition of AGI

60:28 20VC: Scaling to $1BN+ in Revenue with No Funding: Surge AI | The Most Insane Scaling Story in Tech |

Anastasios Angelopoulos · Aug 3, 2026

A neolab valued at $10B today must generate roughly $4B of revenue within two to three years to justify a 10x outcome, and failing that it will hemorrhage talent

At a 25-30x revenue multiple, becoming a $100B business requires ~$4B in revenue; without it the business bleeds out its people

Scope: assumes a 25-30x revenue multiple

36:59 20VC: 70% of Neolabs Will Die | There Will be a $100BN US Open-Source Model | Data is a Trillion $ Market | Governments Cannot Regulate Models: It is Too Late | The Cyber Attacks to Come Will be Insane with Anastasios Angelopoulos @ Arena

Anastasios Angelopoulos · Aug 3, 2026

Mistral will do great and Eleven Labs will become a public company

These are not zero-revenue valuations; unlike the labs he is criticising, they have substantial revenue behind their valuations

39:00 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

David Frankel · Aug 8, 2026

Anthropic and OpenAI will unequivocally be disrupted, and there is an excellent chance the disruption comes from China

Platforms never stay dominant forever — the entire history of venture is disruption of incumbents like Microsoft and Google

67:33 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 8, 2026

Innovation cycles have compressed so far that the AI incumbents are being attacked by Chinese open source models before they have even established their incumbency

The speed of innovation cycles has accelerated

67:53 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

Carles Reina · Apr 11, 2026

A new wave of foundational model companies is coming, and the incumbents (OpenAI, Anthropic, Google, ElevenLabs) will end up acquiring them for a few billion each and folding in their research

The new wave unlocks a brand new world of opportunities and incumbents will buy the research teams

79:25 Incumbents acquire the next wave research teams

Arthur Mensch · Apr 29, 2024 · hedged

He would not recommend starting a new foundational model company today, but it would be arrogant to claim no new entrant could arise and beat Mistral

Sam Altman advised against it a year ago and Mistral did it anyway and changed things, which shows such advice can be wrong

42:50 Starting a new foundation model company today is ill advised but not impossible to succeed

Hemant Taneja · Sep 22, 2025 · hedged

Mistral will build a compelling business and has caught back up on model quality after falling behind

They fell behind because they were compute- and capital-constrained, but Arthur has grown from scientist to CEO, learned to aggregate capital, and shifted to a commercial customer relationship rather than build-it-and-they-will-come

45:15 Capital constrained challengers can catch back up by shifting to commercial focus

Aravind Srinivas · Jun 15, 2026

Frontier model providers stay relevant only by continuously shipping new capabilities; anyone winning today can lose tomorrow

The prize is enormous — Anthropic reached roughly Meta's valuation in six years versus Meta's twenty — so no one can relax and a six-month gap without new capability is damaging

66:24 Leadership is perpetually contestable shipping cadence decides

Matt Clifford · Jul 1, 2024 · hedged

A new path to building a frontier AI company is now open: a genuinely different idea plus a compelling demo can unlock the capital to compete, rather than the old route of having trained a large model inside a lab and then raising on that.

Like DeepMind ten years ago, a demo that makes people sit up can attract the capital needed to go up against the incumbents, because ideas now carry more value.

9:58 A differentiated idea and compelling demo can unlock capital to compete with incumbent labs

Anastasios Angelopoulos · Aug 3, 2026 · hedged

Top researchers have largely concentrated at the frontier labs, but a growing cohort is leaving because those labs now feel like big companies where they can't have impact — and this exodus will become more extreme once the labs go public

Scale reduces individual impact, and IPOs will sharpen that dynamic

36:08 Talent exodus from scaled labs accelerates after ipos

Christian Kleinerman · Sep 22, 2023 · hedged

The next wave of AI innovation will come from a hybrid of existing model incumbents and new startups rather than only one camp

Incumbents are certainly chasing these innovations, and the current moment has generated such a spur of creativity that most new startups want to chase some aspect of the generative revolution

18:38 Innovation comes from both incumbents and new startups not one camp

Christian Kleinerman · Sep 22, 2023

There is innovation everywhere in the core AI technology layer, not just in applications of existing models, so many new startups will create new things alongside continued research at OpenAI and Anthropic

At a recent dinner, separate startups described blending fine-tuning with prompting, addressing transformer model limitations, and improving computer vision

19:21 Innovation is happening broadly across many startups in the core model layer not just applications

Shyam Sankar · Jan 17, 2024 · hedged

Some core LLM winners already exist — OpenAI unquestionably, with Meta having impressively reinvented itself and Google still to be judged — but more winners are still to come

54:59 Openai and meta already established as winners google still undetermined

Emad Mostaque · May 17, 2023

Competing with OpenAI in proprietary models is extremely difficult, and generic 'OpenAI for Europe' plays will fail; the only viable edges are open standardization, full multi-modality, emerging markets, or deep verticals like defense, government or healthcare.

OpenAI is executing incredibly well and there's no reason to choose a regional clone over GPT-4 or Palm 2; verticals let you understand a domain deeply and become sticky, as Scale is doing by going into defense.

47:16 Vertical focus multimodality or emerging markets not generic cloning let new entrants compete with openai

Julien Bek · Aug 24, 2026

The only way to back a genuinely novel AI lab is to back an N-of-one founder pursuing a fundamentally different architecture rather than the same approach done better

If a differentiated architecture works the outcome is massive; being incrementally better at the same thing is not a path to success now

12:17 Only an n of one founder with a different architecture can win

Alex Lebrun · Jun 19, 2023

Incumbents can use AI to stay competitive in their existing markets but will not invent the disruptive products that would kill their own business

They mostly sprinkle AI dust on existing products as enhanced features rather than rethinking the paradigm, and like Kodak with the digital camera they won't release what cannibalises their main revenue

18:51 Capital talent and focus can still launch competitive new foundation model companies like mistral

Eiso Kant · Oct 7, 2024

Consolidation of smaller AI labs is largely finished — there are very few companies left that are far enough along to be acquisition targets

Only a handful of sufficiently advanced independent players remain (Cohere, Reka, Mistral), and xAI is very unlikely to be acquired

39:48 Consolidation of smaller ai labs is largely complete few viable targets remain

Aravind Srinivas · Jun 5, 2024 · hedged

Whether the frontier ends with one winner or three or four depends on who cracks bootstrapped reasoning first — if they then put all their capital into scaling it, it ends as one player; if they hedge, it doesn't.

Cracking self-improving reasoning gives a lead that can be extended indefinitely by concentrating capital on scaling it.

23:46 Whether the frontier consolidates to one winner or several depends on whether the reasoning breakthrough leader concentrates capital on scaling it