Models Are Not a Moat: Silicon Valley's Second AI Wave, Where the Money Starts Moving to Hardware
Silicon Valley has reached a consensus: the model itself is neither a product nor a moat — what's valuable is the use case built on top. Meanwhile, the GPU shortage is pushing investment from software toward hardware, and big-company strategic investors are entering earlier than pure VCs.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
The model is not a product, and even less a moat
Kevin says a consensus has gradually formed in Silicon Valley: the model itself is not a product, you can't sell the model directly; and the model, big or small, is not a moat. The reason is that there are more and more open-source models — maybe not the best, but not much worse than the best — and open-source software usually iterates faster. So the companies that can survive long term are the ones that treat the model as foundational software and build the use case on top into a product, rather than just making a feature. Xiaojun adds that the discussion in China has changed over the past few months too: a few months ago investors were still asking, "the underlying layer is a general large model, so if I build a product on top, what's my moat?" Now the discussion has all become about products and applications.
— KevinThe GPU shortage pushes investment from software to hardware
One shocking thing in the first half was the GPU shortage on the hardware side, with NVIDIA alone holding the most core GPUs. Kevin observed that many investment directions started shifting from software to hardware, and many hardware-related startups are working on this. He judges that this shortage will eventually be resolved, maybe in a year or two, but the hardware bottleneck affects both startups and big companies developing new products internally. The finer mechanism: for every GPU or set of GPUs you buy, you have to satisfy three audiences at once — the people doing research and models, the users you need to satisfy when releasing a beta product, and the internal teams building applications — so internally adjusting GPU resource allocation is itself a big problem, before you even get to whether you can buy them.
— KevinCopilot is a marketing position, not a replacement for people
Sophia says the Copilot concept isn't actually new — there was robotic process automation (RPA) before — but RPA never reached its expected potential, while Copilot is another level, and maybe RPA will finally work through Copilot. She points out that the Copilot concept is the autopilot on a plane — a person can take control and modify at any time, with AI helping alongside. Honestly, this is also a marketing positioning: because if AI is too strong, many industries in the US have already started worrying about whether it will replace their jobs, most notably Hollywood writers and actors starting to strike, and journalists at some news organizations are doing the same. So "I'm not trying to replace people, I'm just helping people do their work more efficiently" is a very good position.
— SophiaTwo acquisitions: one offense, one defense
Several acquisitions in the first half revealed different signals. Mosaic ML was bought by Databricks for about 1.2 billion, and Kevin's analogy is Facebook buying Instagram back then — also around a billion, with very few people: Mosaic maybe fifty or sixty people, while Instagram at the time was no more than twenty. The other is CaseText in the legal industry being bought by Thomson Reuters for 650 million in cash; Sophia says this company worked hard for ten years and ultimately, because of the generative AI wave, was acquired for cash. Kevin contrasts the two: Databricks is already a high-tech company, thinking about how to quickly open up AI possibilities — that's offense; Thomson Reuters is more about seeing its own industry changing — that's defense, thinking about how to let the business continue to survive.
— KevinMeituan bought Light Year Beyond, and no one in Silicon Valley discussed it
Xiaojun asks whether this acquisition is popular in Silicon Valley, and Kevin answers, "What I saw was also in Chinese, but no one — no one really talks about it." He personally follows China's tech industry, so he knows about it, but Silicon Valley itself already has enough M&A news, and not many people pay attention to China's moves. Xiaojun says Light Year Beyond was one of the large-model companies Chinese investors were most optimistic about in the first half, not because the founder had a technical background, but because he was a very successful entrepreneur who co-founded Meituan, a company with a market cap in the hundreds of billions — very rare in China to have successfully started a company once and then enter large models so high-profile. Sophia thinks this Meituan acquisition has a bit of an emotional flavor — after all, they're old comrades in entrepreneurship, and when you hit difficulties you need to draw a finish line and give an explanation — whereas the AI-related acquisitions and future IPOs seen in the US are relatively pure business logic, without much emotional flavor.
— SophiaPaper authors are lobbied out by VCs to start companies
Kevin describes a very common pattern: behind every large model there is a paper, and people who like reading papers will look at who exactly wrote it; a month or two later, several of the people who wrote the paper get poached to do a startup, or some VC funds lobby them out and invest in them. The vast majority of these people still have fairly technical, research-leaning backgrounds. Kevin says his personal view is a bit pessimistic; he hopes they can quickly start pragmatically thinking about how to productize, how to bring out the value of large models with real real-life use cases, rather than just using VC money or big-company money to do a research project — otherwise it's still a bit of a pity.
— KevinThe first-tier products are all priced by mature companies
Xiaojun asks what the first-tier products were in the first half, and Sophia counts them off: Github Copilot set a price, competitor GitLab's similar copilot for assisting programmers also set a price, Salesforce's product set a price, Microsoft's Office Suite set a price, Adobe set a price, and Notion, the productivity tool, set a price; AWS's own copilot called Code Whisper also counts as having set a price. Her criterion is: if you don't dare to price it and haven't sold it separately, it doesn't count as a separate product. So this list is all mature companies, and no completely new company has come out yet. She also says, "We are still at the infancy" — it's still very, very early, and checking in again this time next year should give a different list.
— SophiaMeta open-sources to erode others' moats
Kevin discusses open source and closed source on two levels. Among the big players: companies currently doing large models are still mostly closed source, just as the database industry started with mostly closed-source solutions, like Oracle. Open source is actually slowly lowering the moat of closed-source tech in competition. The best example is why Meta works hard to open-source its own Llama — it is not an enterprise services company itself, unlike Google and Microsoft which have their own cloud platforms and can directly package models into APIs and sell them; but it knows that if it only uses others' models while its own model lags, that's bad for its future competitive positioning, so it open-sources all its models — actually to lower the moat of other competing big players' models themselves, and at the same time it can iterate its own model quickly, because after open-sourcing anyone can download it and build new models on top, and it can also absorb the iterative progress brought by programmers.
— KevinThis time Silicon Valley goes to Washington voluntarily; before it hid
Kevin observes one thing different about this AI wave from before: Silicon Valley companies, big or small, are participating in regulation much earlier. Sam Altman has already done many hearings, Hugging Face's co-founder has testified before Congress and participated in White House AI summits. He says that before, there always seemed to be some hostility — the farther from DC the better, best not to bother me, until you absolutely have to deal with them — rather than actively accepting Congress's invitation to hearings to express views. This may show that the whole of Silicon Valley is more mature, or that its relationship with DC is more productive; it may also be that everyone has many concerns about AI's future safety and regulatory issues, and hopes to communicate with Congress first to reduce the risk that the government actually puts out some fairly dumb policies that affect the whole industry's development.
— KevinIn their own words · checked verbatim
I think there's a consensus that the model itself, as everyone knows, is not a product — you can't sell the model directly — and at the same time the model, big or small, is actually not a moat.
我觉得有一个共识 就是模型本身 大家知道不是一个产品 你不能直接卖模型 同时模型 不管是大还是小 其实不是一个护城河
Kevin7:07
Actually a lot of the investment direction is also shifting from software toward hardware as a focus.
其实很多的 这个投资方向吧 也是从软件开始 有往硬件转 的这么一个focus
Kevin14:11
Copilot is also a very good marketing position — that is, I'm not trying to replace people, I'm just helping people do this work more efficiently and better.
copilot也是一个很好的marketing position 就是说我不是要取代人 我只是帮助人把这个工作做的效率更高 做的更好
Sophia22:13
If you're a fairly mature industry, if you're not thinking about AI, you should think about AI.
如果你是一个 比较成熟的industry If you're not thinking about AI You should think about AI
Kevin32:27
If you're a startup and you're raising money now, if you're not talking about AI, what else do you talk about?
你是startup 现在要融资的话 If you're not talking about AI What else do you talk about
Sophia37:29
Personally I think if you don't dare to price it and haven't sold it separately, it doesn't count as a separate product.
我个人觉得 如果你不敢定价 没有单独卖出去的话 不算是一个单独的产品吧
Sophia43:37
The farther from DC the better, best not to bother me, until I absolutely have to deal with you — rather than actively accepting Congress's invitation to hearings to express your views.
离DC越远越好最好不要烦我 直到我就是不能不和你接触为止 而不是很主动的就是会去应国会的邀请去 听证会去表达自己的看法
Kevin58:59
Figures
| Mosaic ML team size | about fifty or sixty people | 28:21 |
| Instagram team size when acquired | no more than twenty people | 28:21 |
| Amount Meituan acquired Light Year Beyond for | 2.1 billion | 33:43 |
| Amount Microsoft acquired Github for | 7.5 billion USD | 2:00 |
| Amount GM acquired Cruise for | 1 billion USD | 53:52 |
Glossary
- Copilot
- An AI assistance tool model where AI helps alongside a person, who can take control and modify at any time, rather than replacing the person.
- RPA
- Robotic process automation, a technology that automates repetitive office processes, which had previously never met expectations.
- magnificent seven
- The seven US companies: Apple, Microsoft, NVIDIA, Tesla, Meta, Google, Amazon.
- red team
- A team inside big companies that tests their own models to ensure they don't produce harmful content.
- self regulation
- Currently the most reliable form of AI regulation in the US, where companies manage themselves well rather than following government rules.
How to listen
Chinese founders and investors watching AI startups and investment directions, and anyone who wants to understand models, applications, compute, and the M&A landscape in Silicon Valley's second wave.
From 47:25 to 48:46 they talk about Silicon Valley housing prices and Kevin moving, which is low on information — skippable.