This summer, we wrote about how proposals targeting AI model development—or, how AI models are built—would impose steep compliance burdens on startups. As we outlined, imposing complex and costly compliance regimes on Little Tech can hinder US competitiveness, lead to a more concentrated AI market, and result in fewer choices for consumers.
Lawmakers are largely sympathetic to this concern. They understand the need for the United States to compete and win in AI and generally support small businesses and entrepreneurship. Yet, numerous state AI proposals—while intended to put safeguards in place for the biggest players—still risk sweeping in the startups at the forefront of AI innovation. The tools lawmakers reach for to carve out Little Tech, including compute and training-cost thresholds, aren’t built for the realities of how AI is made today.
In this conversation, Guido Appenzeller, investing partner, and Matt Perault, head of AI policy at a16z, discuss why thresholds based on either compute power and training costs fail to separate Little Tech from larger developers, and why revenue may be a more effective criteria for establishing what counts as an AI startup.
Key takeaways from their conversation:
Compute thresholds age quickly.
“If you pass a law today with any number, it’s just a time bomb.”
Processor speed has accelerated across the entire history of computing. The Intel 404, the very first CPU, could process about 100,000 operations per second. Today, Nvidia’s H100, one of the most widely used chips in AI development, can do about four quadrillion operations per second. Just 10 years ago, that was approximately 4 trillion operations per second. That’s a 10-million-fold leap since the 1970s, and a 1,000x jump in just the last decade.
Training costs don’t track size
“…you start with an open source model that gets trained for a certain number of operations. [Then] the question is for this derived model, how many training operations went into that? Is that just the operations that I invested? Is that the operations that the original trainer invested? Is it the sum of both operations?...there’s no clear lines you can draw here about what’s inside one company versus another company.”
AI development changes rapidly, and today, it is increasingly modular and remixable. Many developers start training their models using open-source models, fine tune it with additional data sets, or combine components built elsewhere. That makes “total training cost” nearly impossible to measure and a poor proxy for who is big or small and the associated risks of these models.
Revenue reflects use and market size
“Don’t regulate the scientists and researchers. Let them innovate...When there’s actually a product that hits the market and you can talk about use, that’s usually where regulation makes the most sense.”
Unlike compute power or training cost, revenue measures when a company’s product actually hits the market and gains traction. It’s a marker of real-world use and impact, not technical aspiration or achievement. Once a company hits a certain level of success in the market, they likely have the resources to build a legal function or hire outside counsel to help navigate compliance.
There are some exceptions. Guido and Matt point out that we have seen a company reach $500 million in annual run rate after approximately 15 months and with a small number of employees. According to California’s new disclosure law, this would put them in the same compliance category with the likes of Google, Meta, Microsoft, and OpenAI, which are magnitudes larger. This means that some rapidly-growing startups could bear the same compliance burdens as the largest tech companies in the world, which will saddle them with costs that make it harder for them to compete. That’s an undesirable outcome if we want AI markets to be as innovative and competitive as possible.
Regulate harmful use, not size
“If we’re blocking US companies from offering these models, I still can easily obtain them internationally. Bad guys don’t tend to follow rules.”
Imposing safeguards on developing large AI models might not actually result in the intended outcomes of keeping people safe. Regulating how models are built rather than how they are used could backfire, slowing down US developers while doing little to reduce risk. Powerful open-source models are already readily available abroad, including from China, making it easy for developers to train or fine-tune models outside US jurisdiction. Moreover, any kind of threshold is made to be gamed, with companies potentially finding novel workarounds like outsourcing their training costs to minimize exposure.
By contrast, a public policy approach that focuses on regulating the uses of AI holds all companies, regardless of size, accountable under the law and can be used to target and deter harmful conduct.
This transcript has been edited lightly for readability.
Guido Appenzeller (00:00)
So if you pass a law today with any it’s just a time bomb.
Matt Perault (00:04)
Instead of focusing on thresholds that are oriented around inputs, like training cost and compute, we should focus instead on thresholds that are oriented around impacts and what your presence is like in the market.
Guido Appenzeller (00:13)
Don’t regulate the scientists and researchers, right? Let them innovate, let the innovation thrive. When there’s actually a product that hits the market and you can talk about use, that’s usually where regulation makes the most sense.
Matt Perault (00:30)
All right, so I’m looking forward to this chat, Guido. We wrote this piece back in late May about the kinds of thresholds that could be used to delineate between startups and larger companies. And maybe I can just go into a little bit of the origin story of this piece that we wrote.
So I think it came from a conversation that we were having where I was saying to you, we’re often sitting across the table from lawmakers and lawmakers ask us, they’ll tell us the bill that we just introduced is not intended to cover small companies. And we will say, we’re really concerned about the impact on Little Tech And then the policymaker will say, well, that’s not what our bill was intended to cover at all. Tell us how we can ensure that Little Tech companies are carved out of the legislation.
So then I went to you and I said, okay, so how can we ensure that little tech companies are actually carved out of the legislation? And you had a sort of rubric for thinking about what are the kinds of thresholds that policymakers might include that are more problematic, like from our standpoint, aren’t really going to successfully separate out smaller companies and large companies versus thresholds that might be a little more productive. So what was your starting point? What’s your anchor for thinking about how to achieve that separation?
Guido Appenzeller (01:33)
Yeah, I think it’s a great topic. I think it’s very easy to accidentally craft legislation which is aimed at a large company and so the setup that a large company can deal with it and still build a good product, but it kills a startup because they can’t, you know, don’t have the lawyers, don’t have the financial resources, you know, just don’t have the capabilities to respond to these things.
So I think it’s a super important topic. I’m on the investment side, so invest in these small startups. Then when we’re sitting in the board meetings, we see this day to day, how they’re struggling with some of this regulation. Europe has shown us how not to do it. And some good lessons there to be learned.
Matt Perault (02:08)
So can we just pause on that actually for a second before going into the specifics of thresholds? Cause I feel like we often gloss over that point or particularly in my work, I feel like I make the point and gloss over it, but you’re like literally sitting at the board table with lots of these companies. Like what is the experience that they’re having when they’re looking at these various different regulatory proposals being introduced in past like disclosure mandates or impact assessment mandates or audit mandates or things that might increase their liability on day one, even before they’ve put a product out in the market? What is the sensibility right now at the companies that you’re working with?
Guido Appenzeller (02:42)
Yeah, first of all, I think it’s important to wrap your head around what’s currently happening. These startups, specifically in the AI space at the moment, they’re often very small. So you might have seven people. There’s the entire company sitting around a table. Most of them have PhDs in computer science. None of them has any background in politics or law or anything like that. They know how to train models. They train very good models. And they often take a good chunk of the money that we invest in them and invest it in training capacity. The amount of compute infrastructure they have is actually substantial, but they have very little understanding of how to deal with these regulations. So for them, for example, writing an impact report, that’s a major thing that would basically substantially slow them down.
Does this mean they’re not set up to do this? They have to hire people for they have to work with an external firm potentially, right? So it adds a lot of complexity and really slows them down against other competitors. And then often the question is, what do you do? Can you stay out of certain markets if they’re regulated that way? Do you just not go into a market? Do you take the hit and slow down your development? These are tough conversations.
Matt Perault (03:47)
And so then a lawmaker would say, as some lawmakers have, like we’ve seen these proposals, don’t worry about it. This just applies to frontier models. So this just applies to four or five, six companies that are building at the frontier.
Guido Appenzeller (03:57)
Yeah, I think what we’ve seen is that defining a frontier model in theory sounds easy. You have this billion dollar company that has a couple of hundred people, right? So clearly they train a frontier model. So this must be easy to describe. It’s surprisingly hard to do this distinction. And I can try to sketch out a little bit why this is hard if this is how.
There’s two aspects to it. The first one is some people suggested things like, for example, can we say a total number of training operations that go into a model? Now, that sounds logical. If I want to train a larger model, I need a lot more training operations. So training operations is basically how many cycles your little chip has to spin and multiply by number of chips and so on. So you say, OK, let’s take some really large number, like 10 to the 23 or something like that…a really large number. Anything above that counts as a frontier model. Anything below that, right? That’s more of a smaller company or hobbyist or so, right?
Matt Perault (04:55)
And to be clear, this isn’t hypothetical. What you just described is the threshold in the EU AI Act. The Biden administration executive order proposes a similar threshold. like this is an operative concept.
Guido Appenzeller (05:01)
That’s right, yes. They had a draft in California. At some point, they had the same threshold. And so you look at that number, you’re like, this looks really, really big. So now in practice, that totally doesn’t work. And there’s two reasons for that. And the first one is that if you look back at the history of technology, the speed of processors has continued to grow the entire time.
If I look at the very first CPU that came out, the Intel 404, that thing could do about 100,000 operations per second. That’s a lot, 100,000 math operations per second. But the H100, which is probably the most widely used AI accelerator today, like a single chip, can do about four quadrillion operations per second. So how much is that more? Like a factor of 10 million more or something like that. So it’s a massive, massive difference in performance. [The] Intel chip was a long time ago. But at the moment, we’re actually growing faster than historically.
So if you look specifically at AI accelerators, 10 years ago, was maybe at 4 trillion operations per second. Now we’re at 4 quadrillion operations per second. So over 10 years, we went up by a factor of 1,000 in speed. So if you pass a law today with any number, it’s just a time bomb. Eventually, we’ll hit a point where a single chip can probably do this. So basically, our hobbyist at home would be able to train a model of that time.
And that doesn’t make sense. If somebody can train a model on their laptop, you don’t want to regulate that and have a proper regulation around that. That would just mean people either [ignore] the law, or it happens abroad. But it would not in practice change anything about how these models are trained.
Matt Perault (06:46)
And then one concept that we are repeatedly raised in our conversations with lawmakers that I think is, that I want to hear you talk about from the investing side is, our belief is that startups should be able to compete at the frontier. So even if you had lawmakers like repeatedly updating definitions, so it’s keeping pace with the technological advancement that you’re describing.
[I think] most people understand it will be impossible for lawmaking to keep pace, lawmaking is just a complicated process. It’s always going to lag behind technology. But just for the sake of argument, let’s assume that it’s keeping pace. Still, from our point of view, the goal is to have the frontier be a place where it’s not just a small handful of companies that are operating. And you’re living this in what you’re trying to do on the investment side of our firm.
So can you describe that a little bit? Why is it so important to us that the frontier is not left to us?
Guido Appenzeller (07:34)
Yeah, totally. At the end of the day, innovation is driven by startups. If you look at today, what are the top most valuable US companies, right? That’s companies like Google or Tesla or Meta or Microsoft or Amazon. 20 years ago, I think all but Microsoft would have qualified as a startup still or we’re not founded yet.
So what starts as small companies that become the industry leaders that drive the US economy and sort of create this leadership position in the world for the US economy. And if we want to hold on to that, we need to enable these companies to innovate. Otherwise, we’ll have a Chinese company instead taking this top spot.
So model companies, can today, like one of the beauty of the effective US capital markets is that these small companies can raise money very efficiently. So if you have a really good idea, you have the best possible team, you can grow very quickly and raise a lot of money, and then you can very quickly get a lot of compute to train these models. That doesn’t mean that you necessarily have a very large organization that can deal with a lot of complexity. But like a small, a couple of PhD students today, in some cases, they raise north of $100 million. And for that, you can train model that gets us into the order of magnitude where some people would argue is frontier.
Matt Perault (08:55)
So I think the counter we would hear in conversations with policy folks is some version of the Spider-Man, I think it’s the Spider-Man line, with great power comes great responsibility. So you’re using great power at the frontier. Therefore, why wouldn’t it make sense to treat that company as a larger entity, as a compute-based threshold might do, right? A compute-based threshold would basically be saying, if you’re using X amount of power, then for the purposes of this regulatory initiative we are treating you as a large frontier subject to heightened regulatory restrictions. Why is that not like a sensible approach?
Guido Appenzeller (09:30)
I mean, there’s two issues. The first one is, what does it mean to be treated as a frontier model? If this, the end of the day, requires a large amount of complexity, a large amount of process, some people were suggesting things like, have to hire an external auditing firm and work with them on an impact assessment and so on.
If I’m a couple of PhD students that are training a model, that slows me down massively. That’s just not something I’m set up for. If I’m Meta or Google, they can hire an army of lawyers and deal with this. And for them, it’s very easy. But for a small company, you just don’t have the resources. You don’t have the depth to do that.
There’s actually a second problem which I think amplifies that, which is that many people think of training a model as a one-time thing. It’s like building a house, and then I’m done, and now I can value [the house] or look at how much construction materials went in, and can determine the size. The problem is that’s not quite how it works in practice. Today, specifically, if I have open source models, and today many startups use open source models as part of their model training, you can basically take a trained model and from that derive a new model or on top of that train a better model. Does that make sense? I can sort of, you take a model that was trained by somebody else and then I invest more in it to make it even better.
Matt Perault (10:45)
So this is a critical concept that I think we should walk through extremely slowly. I think the model that people still have of model development is based on the sense of how open AI developed. A group of super smart people, super talented people take enormous troves of data, throw enormous compute resources on it, keep it all inside this one house, and out of that magic box, a model emerges.
One of the things that I learned in the course of the conversations that I had with you is like, that has already in a very short period of time become, the, I don’t know if it’s the dominant model still, but certainly not the sole model and not the only model that regulators should take into account when they’re thinking about what regulation would look like. Can you say a little bit more about like, that’s model A, what is model B?
Guido Appenzeller (11:32)
So what you described is true for the large frontier labs. It’s not like if you look at the average startup in our portfolio that develops an internal model, that’s not what they do. What they instead do is they take a model and build on top of that. And I think the best analogy to understand that is maybe if you look at open source software. Today, if somebody asked me on my server, what was the budget that went into the open source software that went into my server?
My answer is like, look, I have no idea, but it’s impossible to calculate. And the reason is that, for example, the server runs Linux. So how much, what’s the budget for Linux? Well, Linux was developed by thousands, 10,000s of companies. I don’t even know. They all contributed bits and pieces. And it took various meandering paths. Things were contributed back. So sort of this group effort that gets you to an end result over decades in the case of Linux.
Now, [AI] models are still younger, so it’s usually a little bit easier to trace their history. But what happens very often is that you start with an open source model that gets trained for a certain number of operations. And then a startup takes that and further fine-tunes it. So fine-tuning is basically you take a model and you train it for a specific domain area. So if I say I want to create a model for lawyers, I might start with an open source model. And then I basically adapt it to be particularly good in understanding legal texts.
So basically, then the question is, well, this derived model, how many training operations went into that? Is that just the operations that I invested? Is that the operations that the original trainer invested? Is it the sum of both operations? If I’m starting with a good open source model, which we want companies to do, then probably the combined thing is above many thresholds for frontier models. So again, we’re running into this thing. It’s like there’s no clear lines you can draw here about what’s inside one company versus another company, just like in the open source case.
Matt Perault (13:24)
So we’ve talked about compute power as one type of threshold. You’re also sort of alluding here to a second type of threshold that we’ve seen. There’s currently one significant bill pending in New York that uses a training cost threshold. So it’s sort of a complicated mixture of things, but basically there’s a per model training cost threshold, which I think is $5 million. And then there’s an aggregate threshold. So models are captured if they’ve spent $100 million in aggregate training the model.
Can you talk a little bit about why these two components of it–one is why a training cost threshold also in your view doesn’t separate out little and big, and then also like when you hear a number like a hundred million dollars for training costs, I think most people think those are [large companies] who are spending a hundred million dollars in aggregate training costs–and so I’m also interested in your assessment of that specific number.
Guido Appenzeller (14:16)
Yeah $100 million is a lot of money, I think by any metric. That said, we have startups with fairly small teams that raise more. But I think the real problem is this thing that, let’s assume I want to, say, build a particularly good model to analyze legal data. I might start with an open source model. Does this open source model tell me how much it was trained for? No, doesn’t, right? I can get the weights, but this may come from a different country or a different state. There’s no information about how much they train for. How do I even estimate my aggregate training cost in this case? If I have an open source component and then I know what my training cost is, but I don’t know what the baseline is. And whoever trained this may also not know what the baseline is because they may have used a base model themselves, they may have used training data collections that were generated by others, today we’re seeing things such as fine-tuning reinforcement learning, where basically I can use a model to generate training data and then use that training data to train a new model.
So how would you account for this? Would you only take what it takes to generate the training data or not, right? And again, the training data may be open source, so I may just grab it from somewhere. So I ended with this thing where for me as a startup, it’s very hard to manage this risk. I have a lot of unknowns here. Like somebody gives me this number, I can’t calculate that number. So what do I do? Do I say, I’m not going to offer it in the state, but maybe the state will pursue me wherever I am. Do I say, I’m just not going to train a model at all? But then I have a competitive disadvantage.
Guido Appenzeller (15:52)
And what makes all of this crazy is we’re now seeing, in some areas, actually the best open source models coming from China. There, we see very rapid innovation. In the United States, innovation in open source has slowed down in part because companies are starting to be worried about liability. If I’m releasing an open source model here, can somebody sue me for it? What are the possible outcomes? And it’s a very unhealthy dynamic, I think, for innovation.
Matt Perault (16:16)
One thing that we’ve seen in training cost thresholds is that some of them don’t specify the number of years that the training would have to occur over. So it’s one thing to say, if you spend a hundred million dollars next year, you’re liable. It’s a different thing to say if you in aggregate spend a hundred million dollars training a model, you will be liable because at some point that means every startup will be liable in some point in their history, the question is like, are they liable in a year, or five or 10 or 20 or 50 or 100? I’m interested in your take of what that timeline looks like. At this point when you’re advising companies, if you were sort of saying like, here’s, and let’s just assume they’re doing a lot of, they’re spending a lot of the cost themselves. They’re not able to utilize underlying models to address a lot of their training costs. Like when are they hitting $100 million?
Guido Appenzeller (16:47)
We certainly had startups that hit that in their first 12 months of existence.
Matt Perault (17:10)
And if you’re aiming to build a competitive model, you’re going to be hitting that in the course of a number of years.
Guido Appenzeller (17:14)
Yeah, it depends a bit what kind of model you want to build, right? For example, if you want to build a new competitive large language model, 100%, yes, you’ll be there. If you want to say, a music model that will be the other end of the spectrum, yeah, it’s a lot more relaxed, right? There, $100 million in training cost is a lot.
Matt Perault (17:32)
But this where I’m really interested in your assessment as an investor because you want to be able to invest in a whole range of competitive markets, right? Like you want the frontier to be competitive. You want more niche applications to be competitive. You’re not looking for a single vertical for investment. You’re looking for the most number possible. I assume you’re looking at the most number possible. And so the fact that in some verticals, companies might not be likely to hit a hundred million dollars for several years, I would assume is relatively immaterial to your overall investing [picture].
Guido Appenzeller (18:09)
[There’s also] a bad correlation here where the most competitive, most interesting areas are the ones where you probably need the largest investment. There’s one other thing about these sorts of total amounts for the company, [which is there’s lots of questions] we haven’t seen getting resolved yet. And I’m not sure that lawmakers thought about, like, for example, what happens in the case of a merger or acquisition, right? I have two companies that both spend 550.
What was your threshold, a hundred million? So they both spent 55 million. Can they now merge until their models are certified or how does it work?
Matt Perault (18:37)
Yeah, presumably once they’ve merged they’re subject to whatever the obligations are.
Guido Appenzeller (18:47)
Exactly, right. So these things, so you’re just introducing a lot of complexity for these companies to manage ultimately for I think, in many cases, fairly unclear payback.
Matt Perault (18:59)
Let’s talk about that specifically, because again, if I can just try to channel a policymaker perspective, I think what they would say is like, yes, we’re talking about obligations, but we’re talking about the obligations in the interest of safety. So yes, we may capture some small companies. Yes, we might make it somewhat harder for them to compete. But what we’re trying to do is come up with a proxy for what is a set of companies that have some heightened safety and security risk in the ecosystem generally, and then for companies that have that risk, they should bear more obligations.
What’s your sense in terms of how these thresholds correlate to safety in the area of model?
Guido Appenzeller (19:34)
It’s a great question. [When] people talk about safety, there’s sort of, unfortunately, no clear threat model. Different people have very different opinions of what is dangerous. So let’s walk through a couple of ones I’ve heard.
So some people are like, well, AI will take over the world and kill us all. If you’re using these models today, this is actually kind of funny, right? I mean, you have to work incredibly hard to get them to run. It’s very difficult to build the infrastructure. It’s very difficult to get them to function and practice. They make mistakes all the time. They rat-hole all the time. You have to dig them back out. They’re very, very high maintenance, right? It’s this very complex thing that you constantly have to babysit and constantly have to tweak to just get something working, right? So this is, we’re light years away from this, know, being able to do something on itself, right? Currently they can maybe operate for a couple of minutes or a couple of hours on very simple tasks, but any more complex tasks, they still fall flat on their face, right? So that’s [a threat] I don’t quite understand. It seems to be a little bit drummed up on that one.
[Some people] are worried about a model potentially helping people to do bad things, right? Can you use this to design a bio-weapon or something like that? And I think the important thing to understand is these models can’t really innovate. What a model does is, think Andrej Karpathy recently called it, they’re more like ghosts. They’re echoes of the past. We’re basically taking lots of data from humans, and these models are very effective in playing us back that data. So if I have a question about something they saw in their training corpus, then they can answer that question based on the data that they’ve seen.
But this means that essentially there’s nothing new here. Any information that’s in what these models typically train on, all of them train on the internet. So if the model is able to give me a clear answer to something, it typically means that information is somewhere on the internet and I could find it with a Google search. So that’s also a threat model I have a hard time following.
Matt Perault (21:22)
I think you’re raising understandable skepticism about the threats which is one way to think of this issue. The second part that I’m curious about is like, let’s just take it that the allegations of the existence of safety and security risk are given. Let’s just assume that that’s the case. What do you think of compute and training cost thresholds as a way to essentially draw a circle around where we think the most acute safety and security risk will come from.
Guido Appenzeller (21:50)
I mean, it comes back to this, like a more capable model is necessary. I have a hard time making a correlation between large and safety, if that makes sense.
Matt Perault (22:02)
Where large means both either compute power or the number of dollars that I put into the training process.
Guido Appenzeller (22:05)
Yeah, either or. The reason is that basically we know today if I want to solve a specialized task, I can train a very small model on that specialized task. [You can have a] model that’s incredibly wide and knows everything about the world, and is very deep, then you sort of need a large model. [With] a smaller model, I can either go not particularly smart and wide, or I can go very narrow, but then I can go deeper, if that makes sense. So let’s say, I don’t know, I want to build a model that can help me, tell me something about rockets. I could basically take a fairly small model and just fine tune it on that particular topic with that particular training data set. And then I could actually get something that’s fairly good at that. So I think size is not necessarily a good correlation for risk here. Again, it’s a little bit hard to argue about this without a clear threat model. It’s not clear that you could come up with a particular threat where the large models will be more dangerous than a small specialized model.
Matt Perault (23:03)
I also think the point that you made previously about remixing model development throws a lot of this into disarray, right? Because you can imagine a world where a model is not all that powerful or is not all that risky, even though they have spent X number of dollars on it, because maybe that’s all spent in-house on a truly proprietary model. And if there are other circumstances, it sounds like, maybe the training costs are relatively small because they’ve used a remix approach where they’re building off a model that actually is already pretty powerful on its own, so the additional compute power that they’re using; the marginal compute, the marginal training costs are relatively small.
Guido Appenzeller (23:26)
That’s right. The other question is like, look, I can get some very large, very sophisticated models from other countries at this point, right? There’s some Chinese open source models that are currently very widely used just because they’re easy to fine tune, right? They’re readily available. We have some good models from Europe with Mistral. OK, so if we’re blocking US companies from offering these models, I still can easily obtain them internationally. Bad guys don’t tend to follow rules, so if something’s trivially available to me to use, then what’s really the benefit of making it difficult for somebody in the United States to develop that. It’s very hard to really understand the threat model here that you’re trying to guard against.
Matt Perault (24:16)
And then the framing that you’re talking about now, I think gets back to kind of what the core of our policy argument is here, which is focused on regulating harmful use. Don’t focus on regulating the underlying development. If you focus on development, you’ll do things like penalize US companies without necessarily actually even addressing potential harmful applications. Instead put the emphasis on targeting that harmful use. So that actually tracks, I think.
Guido Appenzeller (24:26)
That makes a lot of sense to me. [We] have a rich history of regulating the use of technology. And I think everybody agrees that that’s an important thing to do. And I think models are not any different than any other technology. And in fact, in many cases, we already have regulation on the books that allows us to regulate these models.
If my image model generates child pornography. It’s very clear that there’s laws that people will come after me with. So I can’t do this. And guess what? Pretty much every image generation site today has a filter after running their model to make sure that doesn’t happen. And there may be flaws, but I think everybody buys into the concept and everybody tries their best to make that happen.
I think the challenge is when you say people who develop a core technology that could potentially do something, you want to regulate them. In the past, we haven’t really done that. We’ve said, you can develop a good database, a good web server, and there’s no restrictions on what you can do there. If you then serve particular websites, well, there are certain restrictions on what you can do. But that’s the use of that technology. To some degree, just following the pattern we’ve used in the past here, honestly, to me, is the most pragmatic and I think the most straightforward approach.
Matt Perault (25:47)
So this tracks with the distinction that we tried to make in this piece between thresholds that are a little bit better and thresholds that are a little bit worse. When you’re imposing a threshold to carve out certain companies, essentially what you’re saying is the way that you’re regulating [is not] targeting the entirety of the concept in the way that you might want to. So when we released this piece, I sent it around to people and I got some feedback on it. The most consistent piece of feedback I got, which I agreed with, I think was a fair criticism is why would you look to exempt companies from bad public policy? Why not get good public policy passed in the first place? And I think we agree, we want to see policies focused on regulating harmful use, not regulating development when for the most part, I think use-oriented public policy is not gonna require carve-outs for small companies. Small companies should be required to comply with the law and they can be enforced against when they do things that use AI to violate consumer protection law or other existing laws.
Instead of focusing on thresholds that are oriented around inputs, like training cost and compute, we should focus instead on thresholds that are oriented around impacts and what your presence is like in the market. So what does it look like when your product is in the market? And the thing that you and I started talking about as the more palatable, more desirable threshold that in our view would separate little and big is revenue. At some level of revenue, and we should talk a little bit about what the right level of revenue might be, but at some level of revenue a company has launched a product, it is succeeding in some form in the market, and it’s taking in money that it could divert to compliance. And so I think people would say at some point, whatever that is, your revenue levels would be so significant that you should hire a general counsel and you should hire a head of policy and you can hire a communications person and you can hire lobbying firms to support you and firms to conduct impact assessments on your behalf.
So Guido, from your perspective, can you give us a sense of how you see this evolution? Because you are dealing with companies that are maybe in some cases quite literally in the garage. They are at the very early stage, but we’re also companies in our portfolio that at some point they’ve crossed a revenue threshold where it’s appropriate for them to spend some money on compliance. And so what does that transition look like and can you in your own words talk about why you think revenue might be a better way to set a threshold in AI regulation?
Guido Appenzeller (28:09)
Yeah, first all, I completely agree with you, right? Don’t regulate the scientists and researchers, right? Let them innovate, let the innovation thrive. When there’s actually a product that hits the market and you can talk about use, that’s usually where regulation makes the most sense.
All that said, a revenue-based threshold is a lot better because if you’ve raised a bunch of money, a bunch of PhDs sitting on the table, literally, this is how our typical startup looks like. At that point, it’s very hard to deal with a complex compliance regime that can shut you down or just drastically delay what you’re doing.
If you’re at a point where you have half a billion dollars in revenue, it looks different. At that point, you need all the legal and accounting overhead. You probably have lawyers that negotiate your agreements. And you’ve created a little bit of an organization. There’s actually some exceptions. We’ve seen companies that basically get to half a billion in revenue just by accepting credit cards on their web page. And it’s still just a bunch of engineers. But in most cases, somewhere around that threshold, you really start seeing the company getting
more professional, [they’re a] large organization, and then it’s much easier for them to cope with these things. So it might still be a speed bump, but it’s at least not something that kills them anymore.
Matt Perault (29:17)
Yeah, I think that makes a lot of sense and you’re picking a number that’s not totally arbitrary, I think, because that is the revenue-based threshold in California’s new disclosure statute, SB 53.
Okay, so you’ve talked a little bit about hypothetically that you can have small teams building AI applications that will hit this number quickly. Are there any, without naming names, are there any examples that you can give that are more specific about when a company might hit a $500 million threshold?
Guido Appenzeller (29:44)
Yeah, we’ve had one company that they’ve publicly said that they reached the 500 million annual run rate, you know, after about 15 months, selling a product for 15 months. And, I don’t know the exact number, but there were tens of people at that point. So this is not a large corporation, this is still basically a bunch of engineers.
Matt Perault (30:01)
So still a tiny company. And then I think the important thing to recognize about how a $500 million threshold would work. That would mean that that company that has somewhere under a hundred people significantly under a hundred people is in the same compliance bucket as Google with over a hundred thousand employees, Microsoft, Meta, OpenAI, Anthropic.
I think that just to be fair, I think there are policymakers who would say, and that’s exactly right. That’s what we’re trying to achieve. You’ve hit that level of revenue, one of your tens of people should be included general counsel and your tens of people can afford with $500 million in revenue to hire outside firms to support your work and stuff. And I think that’s fair, but I think also it’s important to note that that company is really those kinds of companies we want to be able to compete with the larger companies and they’re likely to not have the same kind of capacity to manage a compliance framework that those much larger companies are going to have.
Guido Appenzeller (30:54)
It’s those fast growing companies that may be the [future market leaders for US tech leadership]. That’s really what we want to build for the next generation here.
Matt Perault (31:05)
So one thing that you expressed concern about with a compute power training cost threshold is its durability. Like how quickly are companies going to get those numbers and then in 10 years, the numbers are going to seem...
Guido Appenzeller (31:13)
Yep, absolutely. Any threshold in 10 years will probably be meaningless.
Matt Perault (31:19)
Yes, it will seem out of date. So what do you think about a revenue threshold like, SB 53 in California has a $500 million revenue threshold. How does that look 10 years from now?
Guido Appenzeller (31:30)
Look, we all understand inflation. So over time, we have the same problem there. But it’s much, much slower. Before a dollar today is only worth 1,000 of its value it will take a lot longer than 10 years. So it’s not perfect, but it’s only a step in the right direction.
Matt Perault (31:48)
I’m curious about your conversations with founders on these kinds of issues. Like what are the ways that these topics are seeping into their head space and how do you talk to them about these issues and how to think about them as they’re building their companies?
Guido Appenzeller (32:04)
Yeah, it depends. One thing I’m hearing in some cases is that they’re saying, you know what, let’s not expand to Europe for now, it’s too complicated. So I think that’s just a good case study of what happens if you overregulate that. We now have open source models that are no longer available in Europe, right? It says the license can be used anywhere in the world. You can use this freely except in the European Union or something like that.
If you’re trying to build a startup over there, that really hurts. You suddenly have an important piece of technology that you no longer have access to. The other thing is, in some cases, if this is the main thing that you’re doing, you’re going to go for it. You’re going to take risks. You’re just going to try to muddle through somehow. And if you get slowed down, you’re going to need to take the hit. If there’s an adjacency where you want in one direction, could go into a second one, but the second area is tightly regulated. You may not do that and then give up on that part of the market. It just means things take longer, things take more money, you lose competitiveness. These startups are often a foot race where it’s about executing faster than other companies in other countries or other companies in the same country. And whoever’s the fastest wins. So I think this is where it comes in really, the competitive dynamics.
Matt Perault (33:17)
The other sort of bizarre thing about thresholds is sometimes, you’ve raised this point in our conversations that we’ve had about it, they’re almost made to be gamed. And so you gave a mergers and acquisitions example of like, you know, maybe companies hold off when they’re at, you know, if there’s a hundred million dollar training compute threshold, they’re at 49 million and they’re merging with another company with 49 million and they sort of hold off and investing additional training, money and training to try to avoid the threshold. Is that kind of gaming something that you’re starting to hear people talking about?
Guido Appenzeller (33:46)
I think the gaming will come from the large companies, not the small ones. There’s so many things you can do where you can say, some of the training job to, instead of doing it in-house, you give that to another company, maybe abroad, and let them train something and then take back the open source model. Now you can claim you don’t know how much training effort went into this thing. That’s probably true. You can potentially distill down models. There’s many, many ways to do this.
Look, the core, regulating this kind of technology is incredibly complex. If you would tell me, Guido, can you write a perfect regulation to regulate the R&D aspect of it, the honest answer would be probably, I can’t. I have absolute sympathy for any lawmaker that feels that this is difficult, because I think it really, really is. But I think, to some degree, that’s why in the past we focused on the use cases for the regulation, because those things are much easier when a particular model that does something with market applications, then we all understand how this thing should be behaving, what is good and what is bad. This is much harder than if I have a bunch of researchers in the data center staring at very large matrix multiplication and trying to regulate what they’re doing.
Matt Perault (34:54)
Guido, this was super fun. Thanks so much for doing this.
Guido Appenzeller (34:56)
Of course. Thanks, Matt.
This newsletter is provided for informational purposes only, and should not be relied upon as legal, business, investment, or tax advice. Furthermore, this content is not investment advice, nor is it intended for use by any investors or prospective investors in any a16z funds. This newsletter may link to other websites or contain other information obtained from third-party sources - a16z has not independently verified nor makes any representations about the current or enduring accuracy of such information. If this content includes third-party advertisements, a16z has not reviewed such advertisements and does not endorse any advertising content or related companies contained therein. Any investments or portfolio companies mentioned, referred to, or described are not representative of all investments in vehicles managed by a16z; visit https://a16z.com/investment-list/ for a full list of investments. Other important information can be found at a16z.com/disclosures. You’re receiving this newsletter since you opted in earlier; if you would like to opt out of future newsletters you may unsubscribe immediately.
