Does commoditization mean model and infrastructure companies cannot be great businesses?

Peter Singlehurst · Mar 19, 2025

Baillie Gifford has stayed out of the large language model companies because it cannot yet define what competitive advantage looks like at the LLM layer

They still don't know what durable competitive advantage will look like at the large language model level

0:00 20VC: The 10 Question Framework a $217BN Manager Uses to Make Investment Decisions | Lessons from Turning Down Stripe, Coinbase and Losing Money on Northvault | The Bull Case for Bytedance | How Anduril Could Be a $200BN Company with Peter Singlehurst

Peter Singlehurst · Mar 19, 2025

They have not invested in the large LLM companies because competitive advantage at the model layer remains undefined, even though it is definable at the infrastructure and distribution layers

Forces of commoditisation such as open source models and DeepSeek make it unclear what durable advantage looks like at the LLM level; the question is where value accrues over a ten-year holding period

Scope: not a judgement that the products or revenue bases are unimpressive; they do own AI-adjacent positions like Databricks and Tenstorrent

16:39 20VC: The 10 Question Framework a $217BN Manager Uses to Make Investment Decisions | Lessons from Turning Down Stripe, Coinbase and Losing Money on Northvault | The Bull Case for Bytedance | How Anduril Could Be a $200BN Company with Peter Singlehurst

Peter Singlehurst · Mar 19, 2025

He would buy none of OpenAI at $300BN, xAI at $50BN or Anthropic at $60BN

He does not know what creates enduring competitive advantage at the large language model layer, and won't buy without a strong thesis on that

Scope: at the stated valuations; position could change with a thesis on LLM moats

61:38 20VC: The 10 Question Framework a $217BN Manager Uses to Make Investment Decisions | Lessons from Turning Down Stripe, Coinbase and Losing Money on Northvault | The Bull Case for Bytedance | How Anduril Could Be a $200BN Company with Peter Singlehurst

Byron Deeter · Aug 25, 2025

Calling foundation models commodities does not imply price erosion or bad economics — the hyperscaler playbook shows commodity layers can be exceptional businesses.

AWS is the best business in the history of software and sits in what people call a commodity layer, so the same playbook can run in foundation models.

Scope: accepts foundation models may be 'commodities' only in the hyperscaler sense

6:30 20VC: Do Margins Matter in AI? | Is Defensibility Gone For Good? | Is Vertical SaaS Dead in a World of AI | What SaaS Rules Are BS and No Longer Apply in a World of AI | The Future of Venture: Why Chanel vs Walmart is BS with Byron Deeter

Harry Stebbings · Nov 10, 2025 · hedged

Commodity businesses can be among the best businesses in the world, contrary to the view that models won't make money because they are commoditised

AWS, Google Cloud and Azure are commodity businesses and are exceptional businesses

Scope: 'maybe'; stated as a revision of his prior view

27:52 20VC: Benchmark's Newest General Partner Ev Randle on Why Margins Matter Less in AI | Why Mega Funds Will Not Produce Good Returns | OpenAI vs Anthropic: What Happens and Who Wins Coding | Investing Lessons from Peter Thiel and Mamoon Hamid

Harry Stebbings · May 13, 2024

Big tech's foundation model deals are effectively acqui-hires serving the core cloud services cash cow

They realise cloud services is where the cash is, and the models are tools that help them acquire great customers

42:16 20VC: Behind the Scenes at Y Combinator: The Interview Process | What the Best & Worst Do in the Program | Do the Best All Raise Pre-Demo Day & YC's Fundraising Advice to Startups | Why the Value is in Application Layer AI with Tom Blomfield

Harry Stebbings · Jan 22, 2025

Cloud providers are happily using foundation models as a loss leader to build stronger moats in their cloud businesses

Visible in the Anthropic–Amazon and Microsoft–OpenAI arrangements

42:30 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

Saam Motamedi · Jul 15, 2024 · hedged

Greylock is less excited about investing in models-as-a-service / API cloud businesses because it's unclear whether leaders can persist a lead and command margin

Only one or two players may persist a lead, and if the separation between the top models and the next tier is small, those models can't command margin

Scope: there is a lot of value in API services this year and next; skepticism is about how it plays out longer term

15:23 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

Richard Socher · Apr 18, 2025

Companies confined to the API infrastructure layer face far more competitive pressure to keep building the best models than consumer-app owners do

Anthropic must keep leading on models because Claude's consumer market share is much smaller than ChatGPT's

9:40 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

Arthur Mensch · Apr 29, 2024

The foundational model layer's margin will not go to zero, because value accrues where the difficulty and defensibility are and that is where most innovation happens

Value tends to accrue where the hard part and defensibility lie; if the layer doing the most innovation captured nothing there would be a fairness problem

20:43 Value accrues to the model layer because that is where the hardest innovation and defensibility lie

Everett Randle · Nov 10, 2025

He was wrong to dismiss the AI cloud business model as commodity reselling; demand for AI inference is so astronomical that for the next few years it overwhelms business-quality concerns

He was negative when CoreWeave was raising at $3bn, and it is now a $60bn public company where investors have gotten liquidity

74:19 Demand overwhelms business quality concerns for now

Mamoon Hamid · Oct 21, 2024 · hedged

Selling tokens is not a good business today, but LLM providers will engineer gross margins that make them highly profitable over time

OpenAI's publicly visible financials show poor economics today, but the teams are smart enough to fix gross margin

17:28 Poor token economics today will improve as labs engineer margins

Douwe Kiela · Jun 30, 2023

The leaked Google memo claiming OpenAI and Google have no moat is wrong; both have a giant moat

It's all about data: OpenAI has a deep understanding of how people want to use language models that nobody else has, plus economies of scale that let them serve models very cheaply because of request volume

17:40 Openai and google have a real moat via data and economies of scale contra no moat memo

Aidan Gomez · Aug 19, 2024

A business that only sells models is currently a near-zero-margin commoditized business because of price dumping and free model releases

So many players are giving models away for free or dumping prices that margins are extremely tight, even though the market itself is large and growing fast

15:21 Pure model selling is near zero margin today due to price dumping and free releases

George Sivulka · Jan 22, 2025 · hedged

Model providers are unlikely to enjoy cloud-like economics because cloud has fewer, more entrenched players with OPEC-style pricing control and is technically more complex than training ever-larger models

Cloud's oligopoly structure and greater complexity give incumbents pricing power that model providers lack

42:13 Model providers lack clouds oligopoly pricing power because switching is easier

Richard Socher · Apr 18, 2025

The thin LLM infrastructure layer will resemble telcos: enormous CapEx that creates huge value in the world without capturing much of it

Just as you can't build Uber without ubiquitous internet, yet Vodafone and T-Mobile take no cut of Uber, LLM infrastructure enables applications without sharing their upside

8:13 Llm infrastructure resembles telcos huge capex creates value without capturing it

Richard Socher · Apr 18, 2025

LLM infrastructure is actually worse off than telcos because software has far less of a moat, and open source erodes it further

Unlike telcos, software providers have no retention lock-in, and open source models remove even more of it

9:13 Llm infra is worse than telcos lacking retention lock in and eroded by open source

Eiso Kant · Oct 7, 2024

In the extreme, as model prices fall, hardware/compute margin is what really matters to who wins — as already happened in cloud computing

Competitors are incentivized to cut margin down toward their true hardware cost, so the remaining differentiator is the cost of the compute itself (plus intelligence-layer advantages from distilling the best large model)

16:33 Compute cost basis not model quality decides winners as prices fall to marginal cost