Skip to contents

Stable Takes Episode 7: Nvidia, agentic shopping and tokenized deposits

15 min read

Stable Takes is a weekly mini-podcast from Tempo on stablecoins, tokenization, and agentic payments. Watch Episode 7 above or read the transcript below.

Tempo's Chief Business Officer, Dan Romero, and Head of Market Development, Simon Taylor, open on Nvidia's leadership news and what it means for the AI labs. From there: merchants blocking agents in agentic commerce, building a happy path for agents, identity, and agent spending limits on stablecoin balances. They close on the banks' joint stablecoin and why tokenized deposits need a utility stablecoin.

This conversation is lightly edited for clarity.


Dan Romero: How are you?

Simon Taylor: I'm good, man. Happy Friday. How are you doing?

Dan Romero: I'm great. I'm excited for Fintech NerdCon next week.

Simon Taylor: Thank you, my friend. Just a few more days to go. It's six days, I think, until we kick everything off. Very excited for it. fintechnerdcon.com for anybody interested. I'll see you there.

Dan Romero: I'll be there. So I think the big news this week that we'd have to talk about, even if it's not directly stablecoin-related, although I'm going to make a case that there is an impact on stablecoins, is the big Nvidia news: Jensen is out as CEO.

Simon Taylor: Yeah. I don't know what happens next there.

Dan Romero: But here's my pitch on why this has an impact on stablecoins. It's so big within the AI world that anything where AI was going to overlap with stablecoins is probably affected one way or another. Now, how it plays out, I'm not exactly sure.

I'll give you one example. In a world where Nvidia is at the leading edge of everything, if you're a frontier lab, OpenAI or Anthropic, you're thinking, "OK, it's going to be really hard to compete with Nvidia. Maybe for some of the lower-intelligence or simpler workflows, we can use our RSI-level models to help us design chips." They have hardware teams. Both companies have hardware teams. But it's still going to be hard to get to the true frontier in the same way that they are at the frontier of LLMs. In a world where he's not leading the company, there are a lot of talented people at Nvidia, but I do think founder-led is a different category.

So I do think it maybe shifts some attention. If they were spending 5 to 10% of their time before thinking about silicon, maybe that share of attention increases, because they think there's an opportunity for competitive differentiation. Now, it's a little fantastical, because there's such a real-world component to this. Just because you have AGI doesn't mean you can magically go get TSMC fab space.

But I do wonder if it has an impact, at least on the agentic side of things, on the marginal effort the frontier labs might have been putting into commerce. We already saw this with ChatGPT: when Claude started to move pretty quickly on coding, they killed or reprioritized a lot of the consumer initiatives, and we didn't really hear much about agentic commerce. I do think it gives an opportunity to the Instincts and Muses of the world. Granted, Meta has its own semiconductor team, but they can stay relentlessly focused on this consumer use case, whereas the frontier labs have this multidimensional resource allocation problem: how do I maintain the true frontier on intelligence? They have to think about the whole stack: data centers, silicon and, obviously, the models themselves.

So maybe that's a long-winded way of saying that I think it probably benefits the existing leading personal agents, and it makes the frontier labs prioritize those a bit less. I'd be curious what your take is.

Simon Taylor: Yeah, as you said, it's been sort of whack-a-mole with where their attention is and what's getting cycle time. Now it's all coding, and now it's not. It's entirely plausible. Claude Duet just launched, and ChatGPT Dots has launched, but they're not very heavy on the commerce side post-Muse. Everybody's reacting to that. It's like kids' soccer: somebody does something, gets a lot of attention, and then all the others crowd in to try to compete with that thing. It was probably all Grok's fault originally.

What I find so interesting is that Meta has got a frontier-class model, but it has been tuned and trained for that consumer use case. They're directly going after that thing, versus these broader frontier-class models that are trying to be many things to many people, or pushing the boundaries of math and science and wherever else.

Commerce has already been too early once. It might be too early now. Who knows? My experience of commerce with mobile was Google Wallet in 2010 and 2011. PayPal tried to launch a wallet. NFC and tap to pay all worked four years before Apple Pay came out. But it wasn't until the market was ready, the security stuff was ready, the merchants were ready and the issuing banks were ready that any of that took off. So if you're leaning into it now and it's "50% of the time, it works every time," then it might be smart for them to go focus on chips for a little while if there's a gap in the market, and it might change the nature of the market as well. I know OpenAI has its own chips with Cerebras, I think, or something that begins with a C.

Dan Romero: Yep.

Simon Taylor: Maybe that's the thing. Maybe they just go vertically integrated. Elon's doing it.

Dan Romero: Yeah. So I think it obviously shakes up the whole market from an AI standpoint, and it's going to take a little time to settle out. Who knows, Jensen joins SpaceX and all of a sudden Terafab is being run by the Nvidia co-founder. I don't know how it plays out, other than that I think focus and attention are the scarcest resources. Literally, for LLMs, attention is all you need. So in a world where the opportunity around silicon has increased for those who can go there, I do think it gives those focused on agentic commerce the opportunity to double down on that while the eye is pointed somewhere else.

Simon Taylor: Well, speaking of agentic commerce, did you see the news that United and eBay have joined the list of merchants blocking agents? United said on the 11th of November it'll ban unauthorized AI agents like Meta's Muse from booking its flights on its app, on its website and through travel agencies, which I'm curious to see how they do. And a day earlier, The Verge found that Muse can no longer browse eBay's listings. They're joining a long list: Amazon, JPMorgan, Schwab, Kohl's, Neiman Marcus. What do you think of all this?

Dan Romero: I sound like a broken record at this point. It's an opportunity for agent-native businesses to pop up. I am interested in how the details play out with something like United. If I'm Booking.com or Expedia and I'm agent-friendly, or whatever meta-aggregator, how do you stop that? Here's another example: Google Flights. I can get pretty far, where all I have to do is click through from there, and then I'm at United, and OK, fine, I have to log in and click it. But you can get pretty far in terms of how you navigate stuff with agents.

So again, I think it's a losing battle. For sure they can block things in the near term, but two things. One, I think it's just an opportunity for agent-native experiences. Granted, airlines are a little different. It's a little hard to go start a new at-scale carrier airline. But it was a little surprising, because I listened to the Scott Kirby podcast with John Collison on Cheeky Pint, and I was really impressed with all the stuff he was talking about, being so tech-forward. They made this big investment during COVID. I would almost have expected Delta to be the one to do this.

But it's one thing to do it this year, and then next year it's, "OK, we've got the systems more robust or agent-friendly," or, "Hey, you need to use a United agent that talks to your other agent," so at least they control the consumer experience. If I were to make a bold prediction, 12 months might be too soon, but if two years from now there isn't an official channel through the major agent platforms to book a flight on United, I would eat my hat. How about that?

Simon Taylor: See, I'm kind of with you. I don't know about the time horizon, but the big thing is building a happy path. Walmart built its own agent gateway, because the problem is you're not getting a ton of volume from agents yet. I think in the Uber earnings release, they said that in the past quarter, agents they could identify as agents had booked four Uber rides and about 500 Uber Eats orders. I don't know if that was all the Tempo office, but I'm pretty sure it's more than that.

Still, if those are the numbers these merchants are seeing, it doesn't look worth enabling. You're not getting a meaningful upside. The downside is there's a real risk of fraud. There was a study that came out today from Harvard MBA saying that you can convince agents to do a lot more fraud on your behalf quite easily. And you speak to any of the merchants and they're like, "Yeah, we don't know whether this was a stolen card or not. We don't know if it was you, and we're seeing refunds spike whenever we see agent activity." Their laundry list of complaints is legit. So your cost of turning it off is almost zero, except maybe a negative headline from the tech crowd, and your benefit is, "Well, now I've reduced my attack surface area and my fraud losses." But agents are going to come, so you have to build the happy path. You can't just block it and do nothing. You're going to have to block it and build the happy path, and I think we're going to have to build that better experience.

There was a good study from Visa a couple of days ago. I don't know if you saw this one. They found that even pre-checkout, agents are only succeeding 57% of the time at selecting the right item for the right price, following the user's intent. The stronger models do a little bit better, up toward 80%. So there's a real qualitative issue that turns into refunds and chargebacks if you're on the merchant side, and you have to deal with that.

But there was another thing between the lines. Scott Kirby said, "I don't want to create another middleman." In other words, I don't necessarily want to be a commodity here. I'm like, isn't that what Google is? You're getting aggregated already. You can't push back against that. I can see the worry about being commoditized, but I really think this is about real fraud issues post-Project Glasswing and very little volume. This is a low-cost move to make at the moment.

Dan Romero: Yeah. And on the volume side of things, I think it's just too difficult to set up. I always put myself in my parents' shoes. If they're using ChatGPT, what if it just popped up somehow? Let's say it identified the email address, and then Uber, and it says, "Hey, looks like you use Uber or Uber Eats or DoorDash. Link that account, and next time you want to order, you can just do it right from ChatGPT." And it's an official first-party integration. DoorDash or Uber is excited about it because of whatever shared economics, or because it increases the amount of ordering. The companies figure it out bilaterally, and I think you're going to see a lot of that.

I think Instinct has actually been at the forefront of this. I think they just released a partnership with Zocdoc. They did OpenTable. So I think you're going to see a lot of bilateral moves, and it'll be Instinct and Meta, going back to my point that the other labs are a little distracted, whereas the personal agents are all in on trying to find the companies that want to lean into this but want to do it with some level of control.

And I think once you get a sufficient number of companies that have done bilateral deals, you probably get some standard that emerges, probably built off of MCP. But even with MPP, for example, we've added identity. Brendan this week actually did an extension to the identity primitive that's extremely flexible. We don't want one enshrined version of identity with MPP. It's whatever the server wants from an identity standpoint. So whether you're using X41 or something else, like, "Hey, I just want an email address verification," I think it can be extensible. So I think the market's just going to sort this out over the next 12 to 18 months, and we'll start to clearly see how companies that want to maintain that customer relationship and experience are willing to work with agents. To think that they're not going to do it at all, I just think it's not realistic.

Simon Taylor: When Apple Pay launched in 2014, it was a tiny fraction of commerce, and now mobile payments are somewhere like 33% of commerce. So certainly a significant portion. You can imagine agentic being as high as that, if not higher, over a similar time horizon. Over a decade, this is inevitable. The interesting question with AI is whether that gets accelerated so much because of the accelerated timelines of everything, when building stuff takes less time.

I think the big blocker is going to be network tokenization. Until the payments work and there's a liability framework, the merchants are sitting there going, "I'm on the hook for a refund here, and it's not clear to me what evidence I've got to refund this or not." I think that's going to be the big blocker, and that's commercial and human.

But one thing I liked: I don't know if you saw this week, Rain launched agent vaults, which is essentially a prepaid balance done as a stablecoin wallet, with a card attached so that you can give your agent a budget. It's a little bit like a virtual card, but not, because it's a stablecoin balance. Virtual cards and prepaid cards break a lot. Merchants don't like them so much, in that prepaid cards are used for a lot of fraud. They're expensive to run. They're quite painful. This is a much lighter-weight, neater model of doing it, very similar to what Stripe did with Link and balances.

So you can just prefund it, give it, I don't know, $700, and it'll go buy your lunches for the week. And it can do the thing where if tickets go on sale at 2 a.m. for that thing you wanted to see, it can just go buy them. Now you have agentic commerce that is actually better than regular commerce, because most of the time, if you just give it your card details, it's going to come back to you and say, "Hey, can you just authorize this, please?" And I'm like, "Well, I should have just done that myself." So I think there are all of these little speed bumps of experience that we're going to have to iron out over the next couple of years.

Dan Romero: Yeah. I think it leans into the programmability of stablecoins, which is always the forgotten component that really makes it powerful: the ability to say, "Hey, instead of having to issue a new virtual card for every purchase or vendor, I can have one card but pull from different balances onchain." So you can change the abstraction of where that is, and in that case it's still self-custodial. And if you're benefiting from some economic upside of a stablecoin balance, under whatever reward-sharing scheme exists, I think that over time, even if the consumer doesn't realize it's a stablecoin, it just becomes a natural way to make this stuff seamless.

Simon Taylor: 100%. Did you see the other story this week about banks committing 270 million to Project Pharaoh, their joint stablecoin? This is Goldman, Citi, Deutsche Bank and several other firms. It's a new company that's going to directly issue a dollar stablecoin, according to Bloomberg. That money is to cover five years of running costs, and the launch is planned for the first half of 2027. There's some logic to this. Dan, your thoughts?

Dan Romero: Yeah. Maybe you could help me navigate this a bit, because the thing I struggle with is that these banks are obviously all interested in stablecoins, but as we've talked about before, from a capital ratio and banking regulatory standpoint, stablecoins are just not as amenable to the balance sheets of banks. Whereas something like a tokenized deposit is effectively a stablecoin but plays nicely with bank accounting. So I'm curious how you think about a consortium like this in a world where banks are also very interested in tokenized deposits.

Simon Taylor: A tokenized deposit is great if it's your tokenized deposit, but somebody else's might as well be a stablecoin. There's only so much fun you can have with your own blockchain, and there's only so much fun you can have with your own tokenized deposit. I've been saying this to bankers for multiple years, and I was at Blockworks DAS this week in London, and a very senior representative of a top-five bank in the world came up to me and said, "Simon, I'm finally with you. I finally get the need for a utility stablecoin." They had different challenges with this particular project, but the idea makes sense.

What you as the bank want to hold is a tokenized deposit. You want your own tokenized deposit. But if I receive, I don't know, BNY's or Citi's tokenized deposit, now I've got the same liquidity coverage ratio issue I had with a stablecoin in the first place. So I don't want to hold on to that, or the stablecoin. What I really want is some way to convert that into my own tokenized deposit, because against my own tokenized deposit I'll have a ratio of somewhere between 5 and 40%. The lower that ratio, the more capital I have available to lend, to buy other assets and securities, to play in capital markets and make a profit for myself. Stablecoins and another bank's tokenized deposit have a 100% liquidity coverage ratio. In other words, I can't do anything with them. They just sit there.

So the ideal scenario is that one bank sends its tokenized deposit, and in the middle it is converted to a utility stablecoin like OpenUSD or Pharaoh. That travels, and at the other end it's converted back into the other bank's tokenized deposit, with a market maker in the middle. This is extremely useful, especially in scenarios where you don't have an RTGS system. In the United States, you could use The Clearing House for this. Because they're the middleman, they'll deal with all of the tokenized deposits, and you never have to touch anybody else's tokenized deposit, so you're happy. But for cross-border, like Santander and Goldman, for instance, there is no equivalent, and the Swift ledger doesn't give you that either. What it does is synchronize the movement of tokenized deposits. It doesn't solve this issue from a balance sheet perspective. So you do need a utility stablecoin. It's a useful thing.

My questions about Pharaoh are: what's their use case going to be? Because if you look at these banks, they're all very different animals in the forest, with very different goals for that utility stablecoin. And what are the economics around that stablecoin going to look like? One thing I've heard a lot about is OpenUSD: no mint, no burn. What it's really trying to do is monetize velocity. It's trying to be the thing you get in and out of quickly, and it charges you for velocity, not for holding on to the thing. I think some of the banks are certainly coming round to the idea that there's an OpenUSD-shaped hole in the market, and perhaps this development even predated them knowing that OpenUSD existed. But it's interesting to see people who focus on balance sheets and capital ratios coming to the same conclusion all around the same time.

Dan Romero: Well, I don't know if I have anything else to add, so maybe that's it.

Simon Taylor: It's fun times with stablecoins if you're into TradFi right now. Maybe some of the hype has blown away, but I've never had more fun than at Blockworks DAS this week. I was like a kid in a candy store. So, really fun times if you like that sort of thing, as I do.

Dan Romero: We'll see you next week at Fintech NerdCon.

Simon Taylor: Yep. I'm a nerd after all. See you there.

Dan Romero: See you.

Stay updated

Get Tempo updates delivered to your inbox. No spam, unsubscribe anytime.