Is investing in AI foundation model companies a viable strategy for venture investors?

Tom Hulme · May 8, 2024

Investing in foundation models is like putting a few hundred million into a power station that everyone else is building next door with the same GPUs and little edge — and you have to depreciate the asset over weeks or months

Training time is the build, inference is the output, but competitors build near-identical models on the same hardware with only marginal improvements, so the asset's value decays almost immediately

43:28 20VC: GV's Tom Hulme on Why Investing in Foundation Models is like Investing in "Power Stations", The Conventional Wisdom in VC that is BS & Lessons from a 24x Angel Track Record, 255x on Robinhood and Making Billions on Uber

Tom Blomfield · May 13, 2024 · hedged

Foundation models will likely end up as five or six roughly equivalent providers attached to big tech, with commodity pricing — great for startups and humanity, bad for the investors in those companies

Only Google, Microsoft, Facebook and Apple have the funds to power them; they will beat each other down on price like GCP vs AWS vs Azure, leaving models interchangeable and swappable

Scope: opens with 'I don't know' on whether there's money to be made; framed as the best case / ten years out

41:18 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

Aravind Srinivas · Jun 5, 2024

Base model pretraining is a losing business game — not because the science is hard, but because the ROI doesn't work.

Every large training run burns huge money, the model is destroyed by the next competitor update, and you can't recover it through APIs because nobody uses your API if someone else offers a better model cheaper and faster.

21:49 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

Hussein Kanji · Jan 20, 2025 · hedged

Investing in AI foundational models is somewhat binary — it either works or it doesn't

47:34 Foundation model bets are binary either works or doesnt

Hussein Kanji · Jan 20, 2025

Bespoke domain-specific foundational models, such as an AI that designs next-generation materials, are a fundamentally different bet from generic models and can work

A Microsoft Research paper from two years ago showed this kind of approach can work

48:02 Bespoke domain specific foundation models are a different viable bet

Hussein Kanji · Jan 20, 2025

You can make money across the whole spectrum of AI, including foundational model deals, even as a small seed fund

He had ruled foundational models out as too capital intensive for a small seed fund, but a foundational model company for material science came in and they used reserves for one large investment owning 11%, which works even if the company later raises $100-200M

67:49 Small seed funds can profit across the ai spectrum including foundation models via concentrated reserves

Nikhil Basu Trivedi · Sep 6, 2023

Multistage funds are burning a lot of money on AI seed companies, and the foundation-model seed rounds raising hundreds of millions of dollars do not make sense even for the big funds

The money is funding CapEx and going straight out the door to Nvidia H100s

15:56 Massive foundation model seed rounds dont make sense even for large funds since money just funds capex

Brad Lightcap · Apr 15, 2024

OpenAI was a fundamentally different investment from other deep-tech projects because its systems improved continuously with scale rather than carrying binary technical risk

Fusion reactors, quantum computers, self-driving cars and satellites were all-or-nothing bets, whereas OpenAI's systems just kept getting better over time — first unpredictably, then predictably

5:53 Continuous scaling improvement unlike binary deep tech risk

Reid Hoffman · Jun 10, 2024

There will be viable standalone businesses purely in foundation models, including frontier models

It follows the classic venture pattern where early operating margins look terrible (as with Airbnb's first thousands of stays) until scale arrives; software ultimately has good operating margins because it is cheap to execute, and the expensive development curve gets amortized over time

13:39 Foundation model companies follow classic venture margin curve to profitability at scale

Tom Hulme · May 8, 2024

A foundation model investment only makes sense if the company has a uniquely defensible approach such as cracked memory or the ability to take agency; otherwise it is very difficult to see a return

If the plan is just throwing huge data and hundreds of millions of dollars of H100 compute at the problem like everyone else, there is no defensibility

48:39 Foundation model investing only viable with a uniquely defensible technical edge

Sarah Tavel · May 6, 2024

It is still too early to know whether the large multi-hundred-million-dollar investments into model-training companies will work out

Benchmark hasn't engaged in those situations because they fall outside its model, and the outcomes remain unresolved

47:57 Outcome of massive foundation model bets remains too early to judge