Mercado Libre Rebuilds Search with AI at Massive Scale

1

00:00:10,427 --> 00:00:11,344

Hey,

2

00:00:11,344 --> 00:00:14,347

welcome to CoreWeave's booth here in our podcast studio

3

00:00:14,472 --> 00:00:19,102

on the second Story of CoreWeave’s booth at NVIDIA GTC 2026.

4

00:00:19,310 --> 00:00:20,854

I'm your host, Lisa Martin,

5

00:00:20,854 --> 00:00:23,440

and we're going to walk you through a great customer story.

6

00:00:23,440 --> 00:00:25,316

Sebastian Yunge is here with us.

7

00:00:25,316 --> 00:00:28,028

He is the senior staff software engineer at

8

00:00:28,028 --> 00:00:30,572

CoreWeave customer Mercado Libre.

9

00:00:30,572 --> 00:00:32,198

We're going to tell you a story

10

00:00:32,198 --> 00:00:35,368

through Sebastian's perspective of one of Latin America's

11

00:00:35,618 --> 00:00:40,165

most sophisticated and largest consumer technology companies.

12

00:00:40,498 --> 00:00:43,209

Sebastian is going to talk to us about how they evaluated,

13

00:00:43,209 --> 00:00:46,546

they piloted, and then moved into production with CoreWeave

14

00:00:46,755 --> 00:00:50,008

to really deliver this purpose built AI infrastructure

15

00:00:50,216 --> 00:00:53,928

for demanding scale and a huge volume of transactions.

16

00:00:54,012 --> 00:00:56,598

Sebastian, thank you so much for taking some time with us today.

17

00:00:56,598 --> 00:00:57,766

Thank you.

18

00:00:57,766 --> 00:01:00,894

So we did some research on Mercado Libre, but I want you to share with the audience

19

00:01:01,102 --> 00:01:04,105

a little bit about this organization is huge in Latin America.

20

00:01:04,272 --> 00:01:04,856

Yeah, yeah.

21

00:01:04,856 --> 00:01:07,442

I mean, if you could, if I could condense it

22

00:01:07,442 --> 00:01:12,113

and if you were to feel like Amazon people rolled into one, you know,

23

00:01:12,405 --> 00:01:15,366

because we do have Mercado Libre in, which is, as you know,

24

00:01:15,366 --> 00:01:18,787

akin to Amazon. Then we have Mercado Pago, which is akin to PayPal.

25

00:01:18,912 --> 00:01:20,830

Oh, okay. Fintech services.

26

00:01:20,830 --> 00:01:22,624

Okay. It's a pretty big company.

27

00:01:22,624 --> 00:01:26,044

We're already up to 120,000,

28

00:01:27,003 --> 00:01:28,880

I don't know, and 25.

29

00:01:28,880 --> 00:01:31,883

Wow. Huge. Across 18 countries.

30

00:01:31,925 --> 00:01:33,343

Yeah, we're we're really big in that.

31

00:01:33,343 --> 00:01:38,640

I learned and I was also reading that this is the sixth largest

32

00:01:39,390 --> 00:01:43,686

e-commerce platform in the world by gross merchandise volume, massive scale.

33

00:01:43,686 --> 00:01:46,689

We're also the biggest company Latin America by GMV.

34

00:01:47,232 --> 00:01:50,235

And you said e-commerce, payments, logistics,

35

00:01:50,610 --> 00:01:52,862

credit, logistical network in Latin America.

36

00:01:52,862 --> 00:01:57,909

That gives us a huge advantage over anyone that wants to enter into the regions.

37

00:01:58,493 --> 00:01:59,369

Excellent.

38

00:01:59,369 --> 00:02:01,121

So tell us about Mercado Libre.

39

00:02:01,121 --> 00:02:04,124

As are use case. What does that look like?

40

00:02:04,833 --> 00:02:09,754

So Duda doesn't fly all across the board, for instance, for prevention

41

00:02:10,296 --> 00:02:13,299

use a lot of public provision.

42

00:02:13,508 --> 00:02:16,511

Where I can speak both about is the search engine.

43

00:02:17,137 --> 00:02:19,806

Right now, we're in the process of rebuilding our search engine,

44

00:02:19,806 --> 00:02:22,976

and we're leveraging AI for different capabilities.

45

00:02:24,644 --> 00:02:25,520

Actually, one of the use

46

00:02:25,520 --> 00:02:30,191

cases that drove us to for we was actually one of the use cases

47

00:02:30,191 --> 00:02:34,237

we were right now building within search search 2.0, if you will.

48

00:02:34,237 --> 00:02:37,240

Okay. So,

49

00:02:37,907 --> 00:02:40,410

right now what we're trying to do is, well, we're rebuilding.

50

00:02:40,410 --> 00:02:41,870

But of course, you know,

51

00:02:41,870 --> 00:02:44,873

when you're doing the AI, you also want to start small. Yes.

52

00:02:46,207 --> 00:02:49,919

See what once was a little land and build something responsibly.

53

00:02:49,919 --> 00:02:52,922

See if there's any revenue.

54

00:02:53,089 --> 00:02:53,923

Is into this.

55

00:02:53,923 --> 00:02:55,633

You see all that?

56

00:02:55,633 --> 00:02:58,636

So one of the experiments that we wanted to go forward with

57

00:02:58,720 --> 00:03:01,639

was actually mechanism of query expansion,

58

00:03:01,639 --> 00:03:05,143

where you take the query that your user

59

00:03:05,143 --> 00:03:08,104

makes and enrich, and with all the context that you have put out, the.

60

00:03:09,189 --> 00:03:12,192

So after that, you can you create new rules

61

00:03:12,442 --> 00:03:15,445

and then merge them all together to get a more precise result

62

00:03:15,528 --> 00:03:18,323

on what they might be actually looking for.

63

00:03:18,323 --> 00:03:21,743

Talk us through what you were looking for in an AI

64

00:03:21,743 --> 00:03:25,496

cloud provider and what stood out with query.

65

00:03:25,830 --> 00:03:26,289

Sure.

66

00:03:26,289 --> 00:03:30,043

Well, first of all, we were looking for someone that had capacity, you know,

67

00:03:30,043 --> 00:03:34,422

because right now you only know looking for GPUs is a slugfest.

68

00:03:34,464 --> 00:03:39,427

Yes. Capacity, especially in the Virginia area

69

00:03:39,427 --> 00:03:42,305

where all the data centers are specialty scores. Gaffney.

70

00:03:43,890 --> 00:03:46,392

But core we actually had capacity there.

71

00:03:46,392 --> 00:03:50,021

And in fact, when we were doing the policy, they even told us.

72

00:03:50,021 --> 00:03:53,691

So you can use like, our 6200,

73

00:03:53,983 --> 00:03:56,945

which are very powerful GPUs in our class.

74

00:03:57,654 --> 00:04:00,573

And that was a huge surprise for us for

75

00:04:00,573 --> 00:04:03,576

having all the capacity of the platform at a below sea level.

76

00:04:03,868 --> 00:04:07,622

So that was huge for us, you know, to test out what Corwin was all about.

77

00:04:07,830 --> 00:04:11,376

Sounds like that what Corey was offering was kind of a tipping point.

78

00:04:11,376 --> 00:04:12,919

And why you said yes.

79

00:04:12,919 --> 00:04:14,295

Let's sign up for a PFC.

80

00:04:14,295 --> 00:04:15,046

Know for sure.

81

00:04:15,046 --> 00:04:18,424

And all the capabilities that we also heard from,

82

00:04:18,758 --> 00:04:22,303

you know, Slurm, just got south on the platform.

83

00:04:22,929 --> 00:04:24,764

The fact that it is run on bare metal.

84

00:04:24,764 --> 00:04:28,101

So you all do anything, you know, and performance wise,

85

00:04:29,894 --> 00:04:32,772

and also,

86

00:04:32,772 --> 00:04:34,232

a clear pricing structure.

87

00:04:34,232 --> 00:04:36,609

That was that was a huge runo as well.

88

00:04:36,609 --> 00:04:39,487

Transparency. Yes. Yes, yes.

89

00:04:39,487 --> 00:04:42,490

So you you made the decision to do the PLC.

90

00:04:42,740 --> 00:04:43,992

What were you mentioned?

91

00:04:43,992 --> 00:04:47,620

What were some of the core technologies, the key technologies of core waves

92

00:04:47,745 --> 00:04:48,997

that you put to the test?

93

00:04:48,997 --> 00:04:52,250

And how did they match up to your expectations?

94

00:04:52,625 --> 00:04:55,628

Well, obviously we tried, you know,

95

00:04:56,045 --> 00:04:59,716

the supernatants with the platform and that's where we are today, actually

96

00:04:59,716 --> 00:05:04,721

running all of our inference for the query expression that I mentioned before.

97

00:05:04,804 --> 00:05:07,807

I also got personally to, to be able

98

00:05:07,807 --> 00:05:10,810

to test the same capability.

99

00:05:10,935 --> 00:05:14,355

That becomes especially useful when you actually

100

00:05:14,397 --> 00:05:17,400

to times like deep and actually

101

00:05:18,026 --> 00:05:20,111

start prioritizing the to use it.

102

00:05:20,111 --> 00:05:22,113

Once you start to have a lot of compute there,

103

00:05:22,113 --> 00:05:25,241

you actually need to keep the Q fields because otherwise you're losing money.

104

00:05:25,491 --> 00:05:27,994

But you actually need to know the inventory of

105

00:05:27,994 --> 00:05:30,246

who wants to win one on one time frame.

106

00:05:30,246 --> 00:05:31,956

So having that,

107

00:05:33,374 --> 00:05:36,044

and as recently as sunk is,

108

00:05:36,044 --> 00:05:40,840

is really a great feature to have because everyone will know something about them.

109

00:05:40,882 --> 00:05:44,177

Customer on your own is it's it's not a simple task.

110

00:05:44,344 --> 00:05:46,512

It can take a while, but you have the visibility now.

111

00:05:46,512 --> 00:05:48,639

Yeah, yeah, it's one click. It's done.

112

00:05:48,639 --> 00:05:52,685

You have a Slurm cluster running right beside your Kubernetes infra.

113

00:05:52,935 --> 00:05:54,729

It's really, really, really easy.

114

00:05:54,729 --> 00:05:56,981

So how long was the POC and what were some of the

115

00:05:56,981 --> 00:05:59,734

the benefits that you immediately saw.

116

00:05:59,734 --> 00:06:04,322

So if I recall correctly, the PLC ran around three weeks.

117

00:06:04,489 --> 00:06:05,281

Okay.

118

00:06:05,281 --> 00:06:06,866

And we were surprised that

119

00:06:06,866 --> 00:06:09,869

they even left us the infrastructure over that period of time.

120

00:06:09,994 --> 00:06:12,288

It's a hot commodity.

121

00:06:12,288 --> 00:06:13,039

Things that we saw.

122

00:06:13,039 --> 00:06:15,833

Well, yes, of of durability, you know, of operations.

123

00:06:18,044 --> 00:06:20,880

There was no feature created whatsoever.

124

00:06:20,880 --> 00:06:25,426

So from the get go, we were able to see how that meshed with our operations.

125

00:06:26,260 --> 00:06:27,428

Because sometimes you get

126

00:06:27,428 --> 00:06:30,515

they give you this platform demos and you get to give your tires around for a bit.

127

00:06:31,307 --> 00:06:33,643

And, you know, the performance is there,

128

00:06:33,643 --> 00:06:36,062

but then how does that mesh on a day to day basis?

129

00:06:36,062 --> 00:06:36,771

Right.

130

00:06:36,771 --> 00:06:39,899

You kind of do, the way that the PLC was handled.

131

00:06:40,274 --> 00:06:43,236

We were from day zero, be able to see, okay,

132

00:06:43,236 --> 00:06:46,239

this is how we would interact from scratch.

133

00:06:47,031 --> 00:06:50,201

It was the giveaways to see that Ames was really well,

134

00:06:50,201 --> 00:06:53,204

and we would like to move forward with this contract.

135

00:06:53,538 --> 00:06:57,834

You said from scratch, and I've heard that a number of times this week from Corps.

136

00:06:57,834 --> 00:07:00,837

We've engineers, leaders that they built

137

00:07:00,920 --> 00:07:04,006

the platform from scratch to be purpose built for AI.

138

00:07:04,340 --> 00:07:07,760

We've heard that from partners as well, like the Aston Martin Formula

139

00:07:07,760 --> 00:07:10,179

One team that they that built that team from scratch.

140

00:07:10,179 --> 00:07:11,889

Core is a part of that.

141

00:07:11,889 --> 00:07:14,183

You're in early production, I understand. Yeah.

142

00:07:14,183 --> 00:07:15,476

Talk about how that's going

143

00:07:15,476 --> 00:07:18,813

and what are some of the things that you are gleaning or learning so far?

144

00:07:19,564 --> 00:07:23,401

the team on for week has a ton of expertise in a lot of areas.

145

00:07:23,901 --> 00:07:26,571

So sometimes, I mean, we have our own experts as well,

146

00:07:26,571 --> 00:07:29,907

but sometimes we may be drawing a blank or seeking out an opinion.

147

00:07:29,907 --> 00:07:33,244

So for we've is obviously a cloud provider.

148

00:07:33,244 --> 00:07:36,205

So they can get information for other use cases

149

00:07:36,205 --> 00:07:38,082

and actually see what sticks and whatnot.

150

00:07:38,082 --> 00:07:41,878

So having that support there is it's a blessing.

151

00:07:42,295 --> 00:07:45,798

What our day to day operations like right now in early production.

152

00:07:46,924 --> 00:07:49,510

So it's it's not I can't say a lot to that.

153

00:07:49,510 --> 00:07:51,387

Basically we just

154

00:07:51,387 --> 00:07:55,099

throw our inferencing, you know workloads and everything works well.

155

00:07:55,266 --> 00:07:56,559

So far so good.

156

00:07:56,559 --> 00:07:58,519

We haven't had an issue yet. That's nice.

157

00:07:58,519 --> 00:08:02,106

So from an efficiency a productivity perspective it sounds like

158

00:08:02,106 --> 00:08:04,734

it's made a pretty big impact already.

159

00:08:04,734 --> 00:08:06,110

Yeah, it's been transparent.

160

00:08:06,110 --> 00:08:06,694

Basically

161

00:08:06,694 --> 00:08:09,864

you're working with other cloud providers who are running these workloads.

162

00:08:10,281 --> 00:08:11,282

Another stacks.

163

00:08:11,282 --> 00:08:14,285

And we just, you know, make the click and

164

00:08:14,827 --> 00:08:17,830

on running, you know, like we were never switching in the first place.

165

00:08:18,080 --> 00:08:20,958

Or most places like in your past experience, kind of like

166

00:08:20,958 --> 00:08:23,961

features are locked down used to give us access to the GPUs.

167

00:08:24,086 --> 00:08:27,924

Every feature was available and that's not normal or common.

168

00:08:28,049 --> 00:08:29,842

I wouldn't know if it's coming or not,

169

00:08:29,842 --> 00:08:32,845

but I've run into a few instances where features were gated.

170

00:08:33,221 --> 00:08:38,392

Features did not work the same way in the field, and then the final version,

171

00:08:39,477 --> 00:08:40,561

stuff like that happens.

172

00:08:40,561 --> 00:08:44,106

I, I'd like to say it's not common, but I've seen it time and time again.

173

00:08:44,148 --> 00:08:47,693

Yeah, well, maybe a refreshing experience that you've had with Huawei.

174

00:08:47,860 --> 00:08:49,737

You know, we were talking yesterday,

175

00:08:49,737 --> 00:08:53,324

with the Aston Martin F1 team and three things.

176

00:08:53,449 --> 00:08:56,452

The CMO Jean English as performance about core performance.

177

00:08:56,827 --> 00:08:59,705

Piece and partnership.

178

00:08:59,705 --> 00:09:03,834

And it sounds like that's what you're getting from this that partnership angle,

179

00:09:04,126 --> 00:09:07,922

the word transparency I think that's one of the words I,

180

00:09:08,548 --> 00:09:11,175

I think of when I think of core, we've and how they operate,

181

00:09:11,175 --> 00:09:14,011

because a lot of companies have trouble doing that.

182

00:09:14,011 --> 00:09:15,054

They don't.

183

00:09:15,054 --> 00:09:19,267

How is the the performance, the pace, the partnership really helped define

184

00:09:19,809 --> 00:09:23,062

your decision to choose core to your leaders.

185

00:09:23,938 --> 00:09:25,731

I think it's important that,

186

00:09:27,483 --> 00:09:28,693

or we miss our organization

187

00:09:28,693 --> 00:09:32,071

show from the get go any time we've had any issue,

188

00:09:32,572 --> 00:09:35,908

I don't think more than a couple of hours fast and feel we get a response.

189

00:09:37,493 --> 00:09:40,871

Whenever there was an issue, they did not try to hide what the issue was.

190

00:09:40,871 --> 00:09:44,000

They told us, hey, like this, this, that happened here, you know, like,

191

00:09:44,667 --> 00:09:48,379

a document detailing what actually happened, how they solved it.

192

00:09:48,379 --> 00:09:51,382

And you know, what to look out for in the future.

193

00:09:51,716 --> 00:09:53,843

So, right from the get go, everything is transparent.

194

00:09:53,843 --> 00:09:56,178

There's no like, oh, let me talk to my team,

195

00:09:56,178 --> 00:09:58,264

and they'll like, I'm in the corner and see what's what.

196

00:09:58,264 --> 00:09:58,973

No, no, no.

197

00:09:58,973 --> 00:10:01,976

We have a live one on one chat with everyone on slack.

198

00:10:02,310 --> 00:10:03,394

They tell us everything.

199

00:10:03,394 --> 00:10:05,688

24, seven and so far.

200

00:10:05,688 --> 00:10:07,773

And that's the way that we've been operating.

201

00:10:08,941 --> 00:10:09,775

So you have access

202

00:10:09,775 --> 00:10:12,778

to their direct to expert engineering team.

203

00:10:12,862 --> 00:10:14,530

You said slack, for example.

204

00:10:14,530 --> 00:10:19,577

Really getting, getting collaborating with you on the channel or with or with,

205

00:10:20,911 --> 00:10:22,330

you know, on organization.

206

00:10:22,330 --> 00:10:23,497

So whenever we need to chat

207

00:10:23,497 --> 00:10:27,627

with someone, you know, an expert on X, Y, and Z technology, they're there.

208

00:10:27,627 --> 00:10:30,171

You have I mean, there is a time difference, I'm sure.

209

00:10:30,171 --> 00:10:30,963

There we go. Yep.

210

00:10:30,963 --> 00:10:33,758

But they answer as soon as possible.

211

00:10:33,758 --> 00:10:36,385

From a user, a customer perspective.

212

00:10:36,385 --> 00:10:38,679

You know, we think of Mercado Libre, we think of Amazon.

213

00:10:38,679 --> 00:10:43,142

And and as consumers, we want to find whatever we want right away.

214

00:10:43,434 --> 00:10:48,022

We want them to show us relevant products and services.

215

00:10:48,272 --> 00:10:52,360

And, and the consumer is just becoming more and more demanding

216

00:10:52,860 --> 00:10:55,404

that that experience is seamless.

217

00:10:55,404 --> 00:11:00,159

So how what are some of the business outcomes in that context

218

00:11:00,576 --> 00:11:04,330

that you expect Mercado Libre to be able to deliver to those

219

00:11:04,330 --> 00:11:07,375

demanding customers because of the partnership with what we've.

220

00:11:08,417 --> 00:11:09,835

That's a great question.

221

00:11:09,835 --> 00:11:12,838

The results that we're trying to achieve is,

222

00:11:13,381 --> 00:11:16,550

as you mentioned, for the experience, the buying experience of our customers

223

00:11:16,550 --> 00:11:20,721

to actually be better, to actually get to what they're trying to buy.

224

00:11:20,930 --> 00:11:23,891

Sometimes not by asking the question, but by us

225

00:11:23,891 --> 00:11:26,852

knowing our customer can see all this talk about

226

00:11:27,228 --> 00:11:31,440

AI is capturing the context and truly knowing that customers

227

00:11:31,440 --> 00:11:33,401

to get deep down into what they actually mean.

228

00:11:33,401 --> 00:11:36,904

Yes, when they search for something and that's what we're trying to get it.

229

00:11:37,530 --> 00:11:41,075

It's something really easily measurable, you know,

230

00:11:41,409 --> 00:11:44,620

basically circumcision rate whenever we do something.

231

00:11:45,037 --> 00:11:48,290

And if they buy something, hey, back to what they were looking for, right?

232

00:11:49,208 --> 00:11:53,671

So in the next like 6 to 12 months, AI is moving so incredibly quickly.

233

00:11:53,963 --> 00:11:56,966

GTC, as you can see behind us this

234

00:11:57,007 --> 00:12:00,010

amazing time square, but the vibe is still great since day three.

235

00:12:00,344 --> 00:12:01,595

What are some of the things you're hoping

236

00:12:01,595 --> 00:12:05,099

in the next 6 to 12 months from an energetic perspective?

237

00:12:05,391 --> 00:12:10,521

To really help continue this business transformation, what's on your wish list?

238

00:12:11,689 --> 00:12:14,692

Honestly, I want to continue to see the

239

00:12:16,485 --> 00:12:17,194

the performance.

240

00:12:17,194 --> 00:12:22,533

For what better chance to introduce this whole new product for inferencing,

241

00:12:22,867 --> 00:12:25,870

where we can get a lot more tokens, more bang for our buck account?

242

00:12:25,953 --> 00:12:30,791

Yeah, because we are seeing some barriers, on the trading front.

243

00:12:30,875 --> 00:12:34,712

Okay, so once we get better inferencing performance,

244

00:12:35,463 --> 00:12:38,591

then we can really start doing better work on.

245

00:12:40,718 --> 00:12:43,387

On production for inferencing is real time.

246

00:12:43,387 --> 00:12:46,348

It's when the user is actually using your platform.

247

00:12:46,348 --> 00:12:50,102

And having better performance there allows us to think

248

00:12:50,102 --> 00:12:53,105

for more time to write more on what their needs are.

249

00:12:53,272 --> 00:12:56,275

So I'd really like to see that trend to continue to go further.

250

00:12:56,859 --> 00:12:59,695

It's exciting if there are any AI pioneers out there,

251

00:12:59,695 --> 00:13:03,407

and we know there are, and they're maybe embarking on a similar journey

252

00:13:03,991 --> 00:13:07,119

that you've done in the last year or so, what would you

253

00:13:07,411 --> 00:13:13,417

what's your three key takeaways be to advise them you got this. So

254

00:13:14,376 --> 00:13:14,919

I think the

255

00:13:14,919 --> 00:13:18,172

first one would be to try out something

256

00:13:18,172 --> 00:13:22,968

that is very near to your best to a non use case

257

00:13:23,636 --> 00:13:27,640

like try to work with something where you truly have a ton of metrics

258

00:13:28,015 --> 00:13:30,893

so that when you actually do something with the AI,

259

00:13:30,893 --> 00:13:33,896

you know, in a few seconds, if there's a good return, right?

260

00:13:34,271 --> 00:13:34,730

Yeah.

261

00:13:34,730 --> 00:13:36,982

Because otherwise you're just burning money, right?

262

00:13:36,982 --> 00:13:39,485

So having that use case really well established.

263

00:13:39,485 --> 00:13:42,154

And when you talk to your business,

264

00:13:42,154 --> 00:13:45,574

you can actually start seeing right away if there's any return there.

265

00:13:46,033 --> 00:13:49,370

So that's one of the things that we got started to actually send out

266

00:13:50,287 --> 00:13:52,540

where we,

267

00:13:52,540 --> 00:13:54,333

another one.

268

00:13:54,333 --> 00:13:57,503

there's a lot of hype around the AI and so much not unfounded,

269

00:13:57,670 --> 00:14:00,673

but really try to take a breath.

270

00:14:00,756 --> 00:14:04,885

There's a lot to learn and you can get started with simpler mechanisms.

271

00:14:04,885 --> 00:14:08,889

You do not need a full blown, you know, frontier model

272

00:14:08,889 --> 00:14:12,476

training with skill training on the fly with real time data.

273

00:14:12,685 --> 00:14:14,645

You do not need all of that from the get go.

274

00:14:14,645 --> 00:14:18,232

Know, patience and truly getting to know what you actually need.

275

00:14:18,649 --> 00:14:21,277

I think might be the best advice I can give at this point.

276

00:14:21,277 --> 00:14:25,030

I think that's fantastic advice because, you know, the speed with

277

00:14:25,030 --> 00:14:28,868

which air is accelerating is just can't even quantify it.

278

00:14:29,243 --> 00:14:32,246

And there's so many companies that are under pressure,

279

00:14:32,329 --> 00:14:35,499

as I mentioned, to move so fast.

280

00:14:36,000 --> 00:14:39,670

But you bring you bring up a great point like start small, identify

281

00:14:39,962 --> 00:14:42,840

the business processes where you're going to have

282

00:14:42,840 --> 00:14:46,302

metrics, measurable metrics where you can see the needle moving.

283

00:14:46,594 --> 00:14:49,430

And if it's not moving, start from scratch.

284

00:14:49,430 --> 00:14:50,306

Go somewhere else.

285

00:14:50,306 --> 00:14:51,932

Very start up minded in that way.

286

00:14:51,932 --> 00:14:53,642

Yeah. Failure.

287

00:14:53,642 --> 00:14:55,269

Yes. You're going to fail. Failure.

288

00:14:55,269 --> 00:14:59,064

So that you spend a as little as possible and move on to the next item

289

00:14:59,064 --> 00:15:00,274

and just iterate.

290

00:15:00,274 --> 00:15:03,277

Do you think that the culture of Mercado Libre

291

00:15:03,611 --> 00:15:06,280

is was important to have that startup

292

00:15:06,280 --> 00:15:09,283

small, fail fast mentality too, right. Yes.

293

00:15:09,408 --> 00:15:12,036

You create a lot of startups

294

00:15:12,036 --> 00:15:15,039

working in conjunction, so that fail fast.

295

00:15:15,247 --> 00:15:17,333

I mean, we call it continues beta.

296

00:15:17,333 --> 00:15:19,460

I love that continuous manner. Yeah.

297

00:15:19,460 --> 00:15:20,127

Oh yeah.

298

00:15:20,127 --> 00:15:23,380

That's how we interact with the future and every single day of how we work.

299

00:15:23,797 --> 00:15:25,382

That's a great continuous.

300

00:15:25,382 --> 00:15:26,592

That's a great approach.

301

00:15:26,592 --> 00:15:29,553

Last question for you, Sebastian GTC

302

00:15:30,095 --> 00:15:32,348

2026 30,000 plus people.

303

00:15:32,348 --> 00:15:35,351

People are calling this the AI Super Bowl.

304

00:15:35,351 --> 00:15:37,811

A carnival, the heartbeat of AI.

305

00:15:37,811 --> 00:15:41,732

What are your takeaways when you board a plane to go home to Chile

306

00:15:41,899 --> 00:15:45,611

that you've learned this week from, from NVIDIA, from its partner ecosystem,

307

00:15:45,611 --> 00:15:46,862

from core? We've.

308

00:15:46,862 --> 00:15:49,698

Well, first of all, I learned that the ecosystem is huge.

309

00:15:49,698 --> 00:15:51,283

And it was bigger than I actually

310

00:15:52,534 --> 00:15:53,327

is to.

311

00:15:53,327 --> 00:15:56,288

Right. Yeah. And it's on the growing. Yeah it's yeah.

312

00:15:56,288 --> 00:15:57,790

And the thing that I take to heart the most

313

00:15:57,790 --> 00:16:01,210

is that I've seen a lot of talks that I've engaged with the speakers

314

00:16:01,585 --> 00:16:05,422

is that the open source movement is especially prevalent nowadays,

315

00:16:05,756 --> 00:16:09,426

like everybody is trying to open source so that we can continue to grow

316

00:16:09,426 --> 00:16:13,263

as a business and not hiding behind closed doors especially.

317

00:16:13,555 --> 00:16:17,935

I talked to Jensen, Gabe, about open sourcing, a lot of the wall stuff

318

00:16:18,227 --> 00:16:19,561

and working through the engage

319

00:16:19,561 --> 00:16:22,564

with companies to be able to push the new forward path.

320

00:16:22,606 --> 00:16:26,026

Yeah, that spirit of the ecosystem is is only going to get bigger.

321

00:16:26,026 --> 00:16:30,531

I can't even imagine what we'll see at GTC 2027.

322

00:16:30,698 --> 00:16:33,993

I mean, the I think the possibilities are limitless.

323

00:16:33,993 --> 00:16:34,952

Well, Sebastian, thank you for

324

00:16:34,952 --> 00:16:38,497

joining us on the podcast today and sharing the Mercado Libre story.

325

00:16:38,914 --> 00:16:42,418

Your air use case, why you went to Core Weave

326

00:16:42,418 --> 00:16:44,378

and what it's enabling you to do so far,

327

00:16:44,378 --> 00:16:48,632

and also for sharing that really wise advice for those out there

328

00:16:48,632 --> 00:16:51,635

who are struggling with how do I deal with the pressure and the speed?

329

00:16:51,760 --> 00:16:54,263

Start small, fail fast.

330

00:16:54,263 --> 00:16:56,390

Really be prescriptive in your processes.

331

00:16:56,390 --> 00:16:58,767

We thank you so much for your insights. Glad to be here.

332

00:16:58,767 --> 00:17:01,353

All right. For Sebastian Yunge, I am Lisa Martin.

333

00:17:01,353 --> 00:17:04,148

We are coming to you live GTC day three.

334

00:17:04,148 --> 00:17:06,442

CoreWeave booth. Thanks for watching guys.

‍