20VC：Satya Nadella谈开放模型与封闭模型：谁会胜出｜是否该信任前沿模型提供商｜企业为何需要拥有自己的智能｜垂直化的重要性：从芯片到模型及更多

Dario, Ilya, Sam, Greg, and Elon were all, I guess, one happy family. They were really the believers in scaling laws, and that’s what I believed in back in 2018 and 2019. I think the biggest change for me now is that I used to always think alignment would just fall out. I no longer believe that.

This is 20VC with me, Harry Stebbings. Now, this show honestly is quite an emotional one for me. I started 20VC from my bedroom in London. I was 18 years old. I had no money. I had no contacts in my phone book. And today I have the chance to sit down with Satya Nadella, CEO of Microsoft, one of the greatest companies of the last three or four decades. It was 12 years in the making with this show. We’ve done 3,000 shows to build up to sitting in that room with Satya. There really is no overnight success. Work hard, be kind, never give up. And I will just say on Satya, the warmth that he was able to imbue in this interview in such a short space of time was almost unlike anything I’ve seen before. The clarity of his arguments and his ideas, his ability to go very deep very quickly, it was almost unlike anything I’ve ever seen before. It was a show I’ll always remember, and I really hope you enjoy it. But before we dive into the show today, Base44 is one of those products I used, and I just thought straight away, Man, I wish I’d had this when I started 20VC with $50 and no contacts in venture. I wanted to see what it could do, so I built a guest research hub for the show. Upcoming guests, company notes, previous interviews, sponsored contacts, follow-ups, all in one amazing location. I described it in plain English. A few minutes later, I had the product open in front of me. Database, login, back-end and hosting already handled and just all in one. It made it so easy, and that’s what I loved. I didn’t write a long spec. I didn’t stitch five different tools together. I built the thing, and I started using it. For any founder listening, that’s the point. The faster you turn an idea into something real, the faster you launch. Apps, dashboards, internal systems, AI agents, whatever you need to build, Base44 gets you there. Go to base44.com/20VC. That’s base44.com/20VC. While Base44 helps you build faster, Laridin helps you measure AI impact. Every company now runs two workforces, humans and AI agents. Laridin is the first unified dashboard across both workforces that measures their work and optimizes their impact. Laridin answers four questions. Who’s using AI? Which tools? What is it costing? And what’s it actually being produced for every team across every vendor? AI ROI stops being a guess and becomes a number that you can consistently and confidently measure. Gainsight, Vertiv, SurveyMonkey, and many other enterprise organizations run their AI programs on Laridin. Thousands of AI tools and millions of dollars of AI spend are measured. You bet on AI, Laridin proves that it’s working. Visit laridin.com to learn more. That’s L-A-R-R-I-D-I-N.com.

Satya, thank you so much for doing this with us. I would say thank you for coming to the studio, but you’re actually hosting us here, so I really appreciate you taking the time.

Thank you so much for having me on the show, Harry. It’s been a long time coming, but I’m finally glad we were able to do it.

You wrote a seminal piece about 10 days ago, and you said that we should assume a model is compromised. I wanted to expand on that. Does that mean that we can’t or shouldn’t rely on disclosures from model provider companies?

1. Building an AI Trust Architecture

The challenge, as I see it right now, is that models are becoming more and more capable. Scaling laws are working. But one of my assumptions was that alignment would just fall out of scaling, and that doesn’t seem to be happening. There are more science challenges that need to be tackled, and mechanistic understanding is also elusive.

Given that, I think the best posture to take, if you want to continue using these sophisticated models, is to treat them within a trust architecture. After all, zero trust is not something new. We’ve had zero-trust architectural approaches. The claim I’m making is that we should create an engineered zero-trust environment that can contain any malfunctioning agent before it does any harm. If you want agents that are more capable and yet less aligned, then you need a trust system you can rely on.

Can I ask, what does that trust system look like? Who’s the architect of it? Who’s the arbiter of it?

It’s a great point. That’s why I look at it and say you leave AI safety outside of this. Alignment, for example, has to show up in training, maybe even in pre-training. I don’t think you can RL your way to alignment. You have to truly get alignment embedded in the weights right at the very beginning. Hopefully, at some point, the model itself will get aligned.

If you leave AI safety aside, then the trust architecture is going to be what a company like Microsoft does, or what the cybersecurity folks do. We’ve built two things, for example: Agent 365 and Perception. What do they do? They give every agent an identity. They log all of its activity and tool calls—in fact, every model call and the CoT in it. I’m also saying that all the CoT better be in English and not in new release. Then you can check using multiple models. It better be tamper-proof; you can’t have the agent modify its own logs. On top of that, you set policy and behaviorally monitor the agents’ activity so they conform to the policy. That’s the architecture.

We don’t have that architecture today. Is that just symbolic of where we are in the progression?

No, that architecture exists. I think people conflate what I would call an engineered security, zero-trust architectural system for containment—

Yeah.

—and AI safety. People say, “Well, AI safety is hard and elusive and not yet achieved, so therefore there must be nothing we can do on security.” I’m not saying these two things have to come together at the same time. I would just say, “Hey, we have to deal with the situation as it exists, so let’s build a trust system outside the model and be super aggressive in setting policy and monitoring for policy deviation.”

On the monitoring side—

Mm-hmm.

—if we think about chain of thought, how reliable is that?

2. Auditing Models and Open Weights

It’s not. We know that the chain of thought itself can be unfaithful. In fact, it can be made worse when the model company says, “Well, it’s unfaithful, and I’m going to hide it.” That makes it even more problematic.

That’s where I think there’s a push for standardization and for the CoT to be available as is, at least for trusted partners, so that it’s inspectable and auditable, and I can check it using multiple models. That may even drive more open weights, quite frankly, because if you can’t trust them—it’s funny: closed-weight models are more powerful, but if they don’t recognize the need to be more transparent, especially as their trusted partners use them, then we may have a problem.

Can I ask you, when we think about the rise of open source, I completely agree with you that closed models are more powerful. Open source is predominantly Chinese models. Is that a concern when we think about safety being baked into pre-training, disclosures, and the fact that the majority of open source is Chinese?

I think there are two sets of things. If you step back, AI safety not being there is not a good thing for China, the United States, or the UK. This isn’t something where an unaligned open-weight Chinese model would only attack someone else; it could even attack China. So you can’t assume this will only be used as an attack vector.

Setting that aside, I’m pretty pleased with some of the progress open-weight models in the West are making. Jensen’s working on it, and other folks are working on it. We’ll have MAI models that won’t be open-weight. I think of it more as licensed weights. Think of it that way. I want to turn this into a simple, straightforward software business, as opposed to a network-effects intelligence business. There will be different business models, and ultimately, I trust that your ability to use and control those models is what’s going to determine what gets used.

As a venture investor, I have to predict where traffic will go today between closed and open. When you think about the two dimensions of traffic and dollars, if we look forward, say, 24 months, what percentage of traffic do you think will go through open versus closed?

3. Frontier Versus Commodity Models

It’s a great question. I think it depends on what the frontier looks like. Most things seem to get saturated, and when they do, open weights—or even just cheaper closed-weight models—seem to work. Especially if you can train on your own traces, you really don’t need the latest model to solve that same problem if the problem hasn’t changed. If you have an eval that you want to continuously keep moving forward, then you may want to pay.

But even there, I’m very thrilled about this hydro fusion stuff we now have in GitHub.

I just saw this.

Yeah, because the cool thing there is it’s a learned model. Anyway, we’re now going into these agent swarms and multiple sub-agents. I can call a lot of sub-agents that happen to be local models, too. I think my suspicion is that there will always be a market for frontier models. There will always be some new thing. Science may be the next thing where the frontier is being set, but I think predominantly, token usage will be commodity token usage.

Does that mean it’s predominantly open source?

It may be open weights, or, as I said, it can also be cheap closed weights, because there will be a lot of price competition. If you look at my own usage inside Copilot, we use a lot of cheaper closed-source models along with open-weight models.

You talked about the importance of owning your own intelligence for some of your amazing customers. You were talking about some in the UK. Is that the model environment we’ll have moving forward: specialized intelligence for companies with their own data?

4. Firms Own Their Intelligence

Yeah. To me, this is one of the most important things: What’s the future of a firm in an era of intelligence? If you go back even to Hayek, his fundamental description of an economy is local intelligence. Tacit knowledge is distributed throughout the economy. Without that, there can’t be only one firm, and there can’t be one intelligence. That’s the foundational thing. In fact, capitalism works because of that. Price discovery works because of that. A pricing mechanism works because of that.

Therefore, I think it’ll be important for every firm to be able to accumulate knowledge at increasing scales using AI, rather than leak it. That’s why I called it the reverse information paradox. Kenneth Arrow described this paradox of information: If you’re a seller and you disclose a secret without a patent, trade-secret protection, or copyright protection, you leak the value. Now, for the first time, the buyer needs protection. Most people think of it as, “Oh, my data.” It’s even the exhaust of my using it: the thumbs-up, thumbs-down signal. You’re creating value for the model company instead of capturing that value for yourself.

I think one of the things I ascribe to is that, for this to be stable, more of the surplus created by the use of intelligence needs to remain at the firm. And it doesn’t need to happen only by embedding your knowledge in weights. It can be in context; it can be in your evals. You may have a benchmark. In fact, one thing I’m very gung-ho about is that every firm will have a private eval benchmark, which is its core IP. Every model comes in and tries to hill-climb on that eval. To me, that’s how a firm works.

I love this: private evals and benchmarks for each company. This seems like it would be for a rarefied set of companies—your Harveys or your Agora’s, the very forward-thinking AI companies. Will that spread to every company in the world? You mentioned JPMorgan.

Yeah, it happens. It always happens, Harry. There was a day, in the late ’80s and early ’90s, when I remember a few companies had deployed PCs and were doing forecasting. It felt magical: They could put it in Excel spreadsheets, send it in email, and avoid the fax machine and interoffice memos. You’d think, “Oh, wow, when will this happen?” It took maybe 10 years, and then it spread throughout.

This time around, what’s considered, “Oh, I’m an AI-native company and I do this,” will be pretty passé five years from now, I would imagine. That’s how these tech cycles always start. Even with cloud, in the beginning, the big deal was, “Who’s cloud-native and who’s not?” Now everybody is, even Lloyd’s, which is 500 years old. It happens.

Do we move into an implementation economy, then? When we think about these customers, some of your very big customers will need heavy FDE support, much more so than they have in the last 10 years. Will we have a very heavy implementation economy to transition existing customers into this?

It’s interesting. Systems integrators have always existed. We’ve depended on the implementation economy for the diffusion of lots of other technologies. This one, I think, has different contours.

What do you think is different about this one?

It’s the combination of technology and business. It’s not classic management consulting, classic systems integration, or classic technology support. For example, think about my FDE engagements. Take the Unilever example I talked about at the keynote today. It’s fascinating: They’ve built an agentic system to automate order-to-cash, along with trade claims. They started using a frontier model, then substituted an MAI model for all the trade claims because it’s lower cost, and they control it. They’re able to train on their traces.

We had some FDEs deployed, and now their IT team is managing it all. It started with the questions: What’s the business? What’s the eval? What’s the way to conceptualize the agentic architecture? It’s business consulting combined with technical consulting. You’re right that a new set of skills will be built in the ecosystem. Today, I think all of us have, quote-unquote, our own FDEs, but FDEs will become like traditional systems integration. Traditional systems integration will become the new FDE.

You’ve mentioned MAI.

Yeah.

When we look at a lot of different players, there’s this realization that you need to be full-stack, from chips to models to applications. Is that the future? Do you have to have vertical ownership?

It all depends on the industry structure. Today, there may be five or six closed models competing, and perhaps five or six open-weight models competing. At some level, if you’re building an application, like an agentic system, you could have a router with a decision engine built into it, and then route to the best model for a given task.

Of course, there’s always the risk that there could be one vertically integrated model and app tier, all the way to the silicon and systems. That seems unlikely to me. At some point, I thought that might be the case, that there would be one. If anything, it’s going the other way: Every day you wake up and say, “Wow, there’s more competition in every layer of the stack,” whether it’s silicon, systems to manage the silicon, the model layer, or the app layer.

I think that’s a healthy direction of travel, but you never know. It depends on what becomes the dominant design. My hope is that this becomes a bit of a horizontal specialization, with multiple winners in each layer and the ability for people to compose.

The thing I didn’t see is the depth of each layer.

Yeah.

What I know now that I wish I’d known 12 months ago is just the infinite demand for inference.

Yeah, that’s a great point.

I wish I’d known it 12 months ago; I would’ve made some very different investments. Speaking of that, we’ve had Mustafa Suleyman saying training runs of $100 billion are coming, and only a handful of labs can pay for that. With $100 billion training runs, do we have the energy, infrastructure, and power to enable this scale across the stack?

The way I come at it is that ultimately it has to come from world GDP inflecting, and then that allowing someone to spend a percentage of it on R&D. When somebody says “training run,” it’s just a percentage of R&D. Companies that are heavy on R&D spend 15% or 20% on it. At some point, if you have enough revenue, you can spend enough on training. You’ll have enough revenue if you’re contributing to the growth of the world economy, so that you can get your fair share, so to speak.

Then you bring up the other side, the supply side. We have power shortages, silicon shortages, and all kinds of things today. They’re long-lead items. You can’t just snap your fingers and have power generation, and you can’t have a new fab.

And so now, interestingly enough, given the demand, everyone’s building. The question is: When will all of this come online? What will the demand-supply be? No one knows, right? Quite frankly, any one of us sitting here today would just be speculating.

But my feeling is that if you start by thinking of the stable state, it would be just like today, right? Income statements are suddenly not going to be different for a company. It’s going to be, “Hey, I have some revenue, I have some cost of revenue, and you have to have some profit.” If that’s going to be true, then the question is: What is the world GDP growth rate? That’s the bigger thing for me.

In terms of the inflation of GDP growth rate, the world has kind of stabilized at 2% a year for a very large amount of time. Does AI have the ability to seriously inflate that in a short amount of time?

I think so, and I somewhat hope so, because that is the point, right? Let’s take, for example, the amount of money we spend on healthcare as a percentage of any of our GDPs. That’s why I look at even the deployment of Copilot in the NHS and say, “Wow, if it can improve patient outcomes and reduce cost in healthcare, that is, by definition, helping the GDP of the United Kingdom.”

Mm-hmm.

I think there are pockets of the economy that are ripe for adoption of AI. In fact, there’s the least amount of displacement and the most amount of benefit, and that’s where I think we’ll have to find the best fit.

Software development is another one, right? It’s fascinating to see how elastic the demand is for software development. That means the IT backlogs were huge, and none of us even recognized it.

Satya, I was taught that you had to be the top 1 or 2 in a market to win in venture. We’re in, you know, Factory, which is fifth, and it’s winning. Fifth.

Yeah.

This breaks all the rules of venture.

Yeah.

A question I have, again, slightly naughty and not on schedule, but it’s just: Will we see the tipping in terms of capability advancement in alternative functions—finance, sales, marketing—that we’ve seen in coding? Because coding’s been the mother lode.

I think that’s a good point. People talk a lot about whether you have a verifiable reward. It’s easy to do these because a lot of the magic has been RL on top of these pretrained models. That’s what has helped coding. Can we crack that in knowledge work? It seems much harder.

But again, all the examples I talked about—in contact centers, supply chain systems, and marketing/sales or quote-to-cash systems—I think it’s possible, and we’re seeing it. Is the demand as elastic as coding? I think maybe science is as elastic, right? Or engineering in general.

By the way, if I look at what’s happening inside, not just in software development, let’s take silicon development. That’s another thing that’s going through a radical shift, right? Your ability to tape out much faster and simulate everything in silico is much better. Your first spin can completely come out—

But even just optimization of maintenance—

Oh, yeah.

…reports, like you said with Rolls-Royce, which I would never consider—

Well, it’s unbelievable.

Yeah.

It’s a thing they’ve been dealing with forever: scheduled maintenance. I hadn’t realized that, wow, that’s a big optimization problem that you can now run.

I’ve watched many of your shows, and we’re going to do a quick-fire round. People haven’t asked some of these questions, which I’ve always wanted to know and I think everyone wants to know. What’s the exercise routine, given the travel schedule, Satya?

5. The Quick Fire Round

I try to be. I don’t know whether it’s all well hidden. The one thing I’ve gotten into the habit of—and this has been the case for maybe 25 years now—is getting up every morning. I used to be able to run.

Mm.

Unfortunately, my knees have gone away, so now at least I walk on an incline. Even this morning, I was able to get a half hour in. In spite of any jet lag, wherever I am, I hit the gym every day.

I, too, have bad knees. It’s my age as well. What have you changed your mind on in the last 12 months?

Obviously, when Dario and Ilya and Sam and Greg and Elon were all, I guess, one happy family, they were really the believers in scaling laws, and that’s what I believed in back in 2018, 2019. I think the biggest change now is that I used to always think alignment would just fall out. I no longer believe that. This alignment thing may be a harder, different problem. That’s the one thing I’ve changed my mind on.

What looks crazy today but will be incredibly common and accepted in 5 years?

It might be that thing we just talked about, those evals, right? What is a firm? I think a firm would be an AI benchmark.

Final one for you. There are many things to be excited about over the next 10 years. For you specifically—for me, it’s that my mother’s got MS, and I’m so excited about the drug development that could be possible. What are you most excited for?

That’s a great one, I think. One of the things I’ve always been looking forward to is a general-purpose technology, even as powerful as this. People say, “Oh, this is not just another general-purpose technology,” whatever this is.

I hope this spreads and diffuses throughout the world faster than anything else, and throughout all sectors of the economy unlike anything else. Truly, for once, there’s a real shift in the frontier of global growth, where it’s not about just counting winners and losers, but everyone gets lifted. That’s been my dream.

My father was a development economist, and it was always about convergence growth. When will the Global South get something that just suddenly boosts that? I hope this is it.

Gotcha. This has been such a joy to do. Thank you so much for taking the time.

Absolutely. It’s a pleasure, Harry.

But before we leave you today, Base44 is one of those products I used, and I just thought straight away, Man, I wish I’d had this when I started 20VC with $50 and no contacts in venture. I wanted to see what it could do, so I built a guest research hub for the show. Upcoming guests, company notes, previous interviews, sponsor contacts, follow-ups, all in one amazing location. I described it in plain English. A few minutes later, I had the product open in front of me. Database, login, back-end and hosting already handled, and just all in one. It made it so easy, and that’s what I loved. I didn’t write a long spec. I didn’t stitch five different tools together. I built the thing, and I started using it. For any founder listening, that’s the point. The faster you turn an idea into something real, the faster you launch. Apps, dashboards, internal systems, AI agents, whatever you need to build, Base44 gets you there. Go to Base44.com/20VC. That’s Base44.com/20VC. While Base44 helps you build faster, Laridin helps you measure AI impact. Every company now runs two workforces, humans and AI agents. Laridin is the first unified dashboard across both workforces that measures their work and optimizes their impact. Laridin answers four questions: Who’s using AI? Which tools? What is it costing? And what’s it actually being produced for every team across every vendor? AI ROI stops being a guess and becomes a number that you can consistently and confidently measure. Gainsight, Vertiv, SurveyMonkey, and many other enterprise organizations run their AI programs on Laridin. Thousands of AI tools and millions of dollars of AI spend are measured. You bet on AI, Laridin proves that it’s working. Visit Laridin.com to learn more. That’s L-A-R-R-I-D-I-N.com.