1
00:00:00,080 --> 00:00:03,360
Speaker 1: Awsco Adam ze Lipski with me from their big event

2
00:00:03,400 --> 00:00:05,640
Speaker 1: of the year in Adam. Like three big pieces of

3
00:00:05,680 --> 00:00:09,719
Speaker 1: news right, A deeper relationship with Nvidia on Gracehopper super chip,

4
00:00:09,880 --> 00:00:13,480
Speaker 1: on DGX cloud, you have next generations of your own silicon,

5
00:00:13,840 --> 00:00:15,440
Speaker 1: and then you have a chatbot that will get to

6
00:00:15,840 --> 00:00:18,639
Speaker 1: at the end of the discussion. But let's start with

7
00:00:18,640 --> 00:00:20,520
Speaker 1: the strategy. I think a lot of questions that I

8
00:00:20,600 --> 00:00:24,119
Speaker 1: get is what is the differentiation in strategy with what

9
00:00:24,200 --> 00:00:28,240
Speaker 1: AWUS is doing here deepening its relationship and reliance on

10
00:00:28,320 --> 00:00:32,000
Speaker 1: DGX cloud and grace Hopper superchip versus what you're doing

11
00:00:32,479 --> 00:00:35,320
Speaker 1: with the cloud supported by your own silicon.

12
00:00:35,520 --> 00:00:37,639
Speaker 2: Well, I think our strategy has been very consistent.

13
00:00:37,800 --> 00:00:42,000
Speaker 3: It is to provide choice and powerful options and all

14
00:00:42,120 --> 00:00:45,800
Speaker 3: layers of the stack, so down at the infrastructure layer

15
00:00:46,000 --> 00:00:49,280
Speaker 3: of what you need to do generative AI really well,

16
00:00:49,800 --> 00:00:52,919
Speaker 3: as you mentioned, we both have been investing in our

17
00:00:52,960 --> 00:00:56,320
Speaker 3: own silicon, and then we have had a long standing,

18
00:00:56,600 --> 00:01:00,520
Speaker 3: wonderful relationship with Nvidia. We've been the first to pretty

19
00:01:00,560 --> 00:01:05,480
Speaker 3: much every significant Nvidia chip to the cloud, including their

20
00:01:05,560 --> 00:01:08,040
Speaker 3: latest H one hundred this past summer, and I was

21
00:01:08,080 --> 00:01:11,279
Speaker 3: really excited to have their CEO, Jensen Holong on stage

22
00:01:11,280 --> 00:01:14,560
Speaker 3: with me this morning talking about an expansion of the

23
00:01:14,600 --> 00:01:20,280
Speaker 3: partnership with AWS, bringing their DJX cloud to AWS, and

24
00:01:20,520 --> 00:01:23,360
Speaker 3: Nvidia itself is going to be standing up a huge

25
00:01:23,720 --> 00:01:26,920
Speaker 3: supercomputer of a cloud for their own internal R and

26
00:01:27,040 --> 00:01:31,240
Speaker 3: D on AWS. So it's a fantastic expansion of what's

27
00:01:31,319 --> 00:01:35,600
Speaker 3: really a great relationship that benefits our mutual customers, Adam.

28
00:01:35,360 --> 00:01:38,440
Speaker 1: And they still supply constraints around in Vidio GPUs.

29
00:01:38,720 --> 00:01:41,240
Speaker 3: Well, they're very popular, and I think it is still

30
00:01:41,280 --> 00:01:45,000
Speaker 3: remains true that there's probably more people who want to

31
00:01:45,000 --> 00:01:47,680
Speaker 3: get their hands on them than actually can, which is

32
00:01:47,680 --> 00:01:51,160
Speaker 3: one of the reasons why we also announced this concept

33
00:01:51,360 --> 00:01:56,960
Speaker 3: of clusters BREC two for chips and so that you

34
00:01:57,000 --> 00:02:00,000
Speaker 3: can actually go down reserve up to hundreds of GPS

35
00:02:00,120 --> 00:02:02,920
Speaker 3: use for your short term generative AI needs for things

36
00:02:02,920 --> 00:02:05,440
Speaker 3: like training models, which can tend to be apisodic, and

37
00:02:05,480 --> 00:02:08,280
Speaker 3: there's really nobody else out there offering anything like that

38
00:02:08,400 --> 00:02:09,080
Speaker 3: kind of service.

39
00:02:09,320 --> 00:02:12,400
Speaker 1: Given that's the case, if you had to shift any

40
00:02:12,440 --> 00:02:16,000
Speaker 1: customers to trainingum because of the limited supply of in

41
00:02:16,120 --> 00:02:17,639
Speaker 1: video GPUs.

42
00:02:17,880 --> 00:02:22,200
Speaker 3: Yeah, it's really not about shifting, it's really about customers

43
00:02:22,240 --> 00:02:25,480
Speaker 3: need different, different customers need different things. The same customer

44
00:02:25,680 --> 00:02:29,320
Speaker 3: needs different things for different use cases and awos for

45
00:02:29,400 --> 00:02:32,000
Speaker 3: this over seventeen years we've been doing. This has been

46
00:02:32,040 --> 00:02:35,639
Speaker 3: all about choice, all about democratization, all about putting tools

47
00:02:35,639 --> 00:02:37,359
Speaker 3: in the hands of our customers so that they can

48
00:02:37,360 --> 00:02:40,160
Speaker 3: make the choices about what we need. And so we're

49
00:02:40,160 --> 00:02:43,960
Speaker 3: also excited that we announced the second generation of our

50
00:02:44,480 --> 00:02:48,320
Speaker 3: training specific chip for training gener workloads, and that is

51
00:02:48,800 --> 00:02:52,120
Speaker 3: TRAININGUM too, and that's going to have up to four

52
00:02:52,240 --> 00:02:55,239
Speaker 3: times the compute performance of the first generation of TRAINUM.

53
00:02:55,440 --> 00:02:57,359
Speaker 3: We of course have been investing in our own custom

54
00:02:57,440 --> 00:03:01,800
Speaker 3: silicon for years. We also today that we're shipping the

55
00:03:01,960 --> 00:03:05,960
Speaker 3: preview of our fourth generation of general purpose chip, Graviton four.

56
00:03:06,160 --> 00:03:07,920
Speaker 3: I happen to have one with me right here. This

57
00:03:08,160 --> 00:03:11,000
Speaker 3: is a Graviton forward ship. We're not talking about it,

58
00:03:11,080 --> 00:03:13,959
Speaker 3: we're not showing slides on it. It's not future looking

59
00:03:14,040 --> 00:03:16,880
Speaker 3: like a lot of other cloud providers. It is shipping today.

60
00:03:17,280 --> 00:03:21,240
Speaker 3: And again our fourth generation, the power is incredible, the

61
00:03:21,360 --> 00:03:25,440
Speaker 3: price performance is going to be incredibly attractive, and our

62
00:03:25,520 --> 00:03:29,160
Speaker 3: chips also have incredible energy efficiency benefits, which is really

63
00:03:29,440 --> 00:03:31,840
Speaker 3: really important to our customers these days.

64
00:03:32,360 --> 00:03:34,520
Speaker 1: Adam, if the chips are as real as the one

65
00:03:34,560 --> 00:03:36,320
Speaker 1: as you held in your hand, and they're shipping to

66
00:03:36,360 --> 00:03:39,160
Speaker 1: the real world. Give us some numbers, what percentage of

67
00:03:39,200 --> 00:03:43,120
Speaker 1: the AW customer base is using Graviton and is anyone

68
00:03:43,240 --> 00:03:47,200
Speaker 1: using Trainium the existing or future generation besides Amazon.

69
00:03:47,600 --> 00:03:51,160
Speaker 3: Yeah, Graviton has become very popular. I don't think we've

70
00:03:51,160 --> 00:03:54,240
Speaker 3: broken out numbers, but all one hundred of the top

71
00:03:54,280 --> 00:03:57,440
Speaker 3: EC two it's our compute service, the top one hundred

72
00:03:57,520 --> 00:04:02,160
Speaker 3: EC two customers and fifty thousand customers overall are using

73
00:04:02,200 --> 00:04:07,200
Speaker 3: Graviton for so it's incredibly popular, incredibly powerful, and in

74
00:04:07,280 --> 00:04:10,240
Speaker 3: terms of TRAININGUM, we're seeing a lot of interest, a

75
00:04:10,280 --> 00:04:12,880
Speaker 3: lot of folks moving to adopt that as well as

76
00:04:13,440 --> 00:04:17,360
Speaker 3: of course GPU based capacity. And with training, we've got

77
00:04:17,400 --> 00:04:20,839
Speaker 3: folks like Data Bricks and Rico and many others that

78
00:04:20,920 --> 00:04:25,279
Speaker 3: are building and deploying on them. Anthropic that we're partnering

79
00:04:25,279 --> 00:04:28,040
Speaker 3: with very closely, and it's obviously one of the very

80
00:04:28,160 --> 00:04:34,960
Speaker 3: leading suppliers of foundation models and large language models, is

81
00:04:35,400 --> 00:04:38,640
Speaker 3: going to be training future generations of their models on

82
00:04:38,800 --> 00:04:41,479
Speaker 3: TRAININGUM on Anthropics.

83
00:04:41,520 --> 00:04:44,680
Speaker 1: Since last we spoke what's changed does Anthropics come out

84
00:04:44,720 --> 00:04:48,560
Speaker 1: and said it's also deepened its relationship with Google on

85
00:04:48,960 --> 00:04:55,400
Speaker 1: Google Cloud platform. Is Anthropic using any of your custom

86
00:04:55,480 --> 00:04:58,560
Speaker 1: silicon to train Clawed or do you have any sort

87
00:04:58,560 --> 00:05:00,920
Speaker 1: of visibility on a start day of when they'll start

88
00:05:00,960 --> 00:05:02,880
Speaker 1: training a future model on your silicon?

89
00:05:03,920 --> 00:05:05,039
Speaker 2: So really nothing's changed.

90
00:05:05,360 --> 00:05:08,960
Speaker 3: Anthropic has been running on AWS literally since they were

91
00:05:08,960 --> 00:05:12,440
Speaker 3: founded in twenty twenty one, which seems decades ago, but

92
00:05:12,640 --> 00:05:15,800
Speaker 3: it was only in twenty twenty one. And you know,

93
00:05:15,839 --> 00:05:18,119
Speaker 3: we obviously are providing choice of a lot of different

94
00:05:18,160 --> 00:05:22,600
Speaker 3: model providers to customers. You know, many startups, including Anthropic,

95
00:05:22,720 --> 00:05:24,880
Speaker 3: you know, as they as they grow or are going

96
00:05:24,920 --> 00:05:28,120
Speaker 3: to use multi providers providers for different things. But I

97
00:05:28,120 --> 00:05:32,360
Speaker 3: think we Anathropic have been very clear AWS is anthropics

98
00:05:32,640 --> 00:05:35,960
Speaker 3: primary cloud provider for their mission critical workloads and the

99
00:05:36,000 --> 00:05:39,200
Speaker 3: majority of overall workloads from Anthropic are going to run

100
00:05:39,240 --> 00:05:43,000
Speaker 3: on AWS, and Anthropic will be training future versions of

101
00:05:44,240 --> 00:05:48,080
Speaker 3: their Claud models on AWS. They will be collaborating with

102
00:05:48,160 --> 00:05:51,919
Speaker 3: us to help improve our trainum and our inferential chip technology,

103
00:05:52,240 --> 00:05:54,920
Speaker 3: and they're going to be bringing powerful features around fine

104
00:05:54,960 --> 00:05:58,200
Speaker 3: tuning and customization that for certain periods of time will

105
00:05:58,240 --> 00:06:01,920
Speaker 3: only only be available on AS and through Anthropic directly,

106
00:06:02,120 --> 00:06:04,839
Speaker 3: not on any other cloud, So the relationship is deep,

107
00:06:04,880 --> 00:06:06,240
Speaker 3: it's strong, and it's unchanged.

108
00:06:06,520 --> 00:06:10,279
Speaker 1: Adam, will you ever be able to get open AI

109
00:06:10,839 --> 00:06:12,080
Speaker 1: to be on AWS?

110
00:06:12,680 --> 00:06:14,760
Speaker 3: I think we are going to try and bring all

111
00:06:14,880 --> 00:06:17,599
Speaker 3: of the models that people care about on too AWS

112
00:06:17,880 --> 00:06:22,880
Speaker 3: O God, Anthropic, cohere Stability, AI AI twenty one, Meta

113
00:06:22,920 --> 00:06:25,680
Speaker 3: Islama II model. I'm sure there will be others over time,

114
00:06:25,720 --> 00:06:28,600
Speaker 3: and we're just going to be guided by where our

115
00:06:28,600 --> 00:06:30,280
Speaker 3: customers tell us they care about the most.

116
00:06:31,240 --> 00:06:33,400
Speaker 1: So the other big piece of news is an AI

117
00:06:33,520 --> 00:06:36,200
Speaker 1: chatbot finally from Amazon Q.

118
00:06:37,000 --> 00:06:39,279
Speaker 2: But here's my question. If I use.

119
00:06:39,200 --> 00:06:43,000
Speaker 1: Gmail or I use a Microsoft piece of software, I

120
00:06:43,120 --> 00:06:47,000
Speaker 1: can take those pieces of software and use the new

121
00:06:47,000 --> 00:06:51,200
Speaker 1: eight AI tools that those companies are bringing to market

122
00:06:51,760 --> 00:06:55,920
Speaker 1: in their existing suites. Q is on the AWS console.

123
00:06:56,240 --> 00:06:58,080
Speaker 1: So does that mean I have to train as a

124
00:06:58,120 --> 00:07:01,280
Speaker 1: developer and become a developer to get access to it.

125
00:07:02,760 --> 00:07:03,040
Speaker 3: No. Q.

126
00:07:03,560 --> 00:07:05,000
Speaker 2: We're incredibly excited about Q.

127
00:07:05,480 --> 00:07:08,400
Speaker 3: All the many customers that have been previewing Q and

128
00:07:08,440 --> 00:07:12,720
Speaker 3: Private are incredibly excited about it, and we think it's

129
00:07:12,760 --> 00:07:16,440
Speaker 3: going to be transformative in helping to reinvent how people work.

130
00:07:16,920 --> 00:07:21,320
Speaker 3: I mean the existing chatbots or generative AI assistants are

131
00:07:21,360 --> 00:07:23,520
Speaker 3: great for consumers, and so many of us of course

132
00:07:23,560 --> 00:07:26,680
Speaker 3: have used them experiment with them, but unfortunately, for the

133
00:07:26,720 --> 00:07:30,280
Speaker 3: most part, they don't really work at work, they don't

134
00:07:30,280 --> 00:07:32,480
Speaker 3: really know who you are, they don't really have access

135
00:07:32,520 --> 00:07:35,360
Speaker 3: to your positories, and for the most part, they've lacked

136
00:07:35,400 --> 00:07:41,360
Speaker 3: security and privacy characteristics that are pretty much requirements for

137
00:07:42,080 --> 00:07:46,640
Speaker 3: enterprises and many other companies. So Amazon Q is going

138
00:07:46,680 --> 00:07:50,760
Speaker 3: to be tailored and customized to your business. We've built

139
00:07:50,760 --> 00:07:55,000
Speaker 3: connectors to forty over forty major systems, including Microsoft three

140
00:07:55,120 --> 00:08:01,160
Speaker 3: sixty five, including Google, systems including Salesforce, Zendesk, service Now,

141
00:08:01,480 --> 00:08:03,960
Speaker 3: and many others, and Q is going to be able

142
00:08:03,960 --> 00:08:06,440
Speaker 3: to access those if you give Q access to them,

143
00:08:06,600 --> 00:08:09,160
Speaker 3: but it's going to respect your privacy and security. If

144
00:08:09,200 --> 00:08:11,680
Speaker 3: you don't have access to a database without Q, you're

145
00:08:11,720 --> 00:08:13,600
Speaker 3: not going to have access to it with Q.

146
00:08:14,000 --> 00:08:15,960
Speaker 2: But meanwhile, it's going to make it so much.

147
00:08:15,800 --> 00:08:19,360
Speaker 3: Easier and more convenient to find data, to generate answers,

148
00:08:19,360 --> 00:08:23,200
Speaker 3: to summarize things, to reach new conclusions, and to get

149
00:08:23,240 --> 00:08:26,320
Speaker 3: to actual insights and actions. So it's going to integrate

150
00:08:26,360 --> 00:08:29,000
Speaker 3: really well with the systems that people care about. It's

151
00:08:29,000 --> 00:08:32,680
Speaker 3: going to be enterprise grade, and again We're really excited

152
00:08:32,720 --> 00:08:34,719
Speaker 3: about it because our customers are excited about it.

153
00:08:35,160 --> 00:08:38,320
Speaker 1: Adam, I know that you're someone that believes fundamentally choice

154
00:08:38,360 --> 00:08:41,520
Speaker 1: is important when it comes to AI tools. What work

155
00:08:41,559 --> 00:08:44,040
Speaker 1: are you doing on the diversity and ethics side here?

156
00:08:44,080 --> 00:08:46,520
Speaker 1: We always talk a lot about silicon and product, but

157
00:08:46,600 --> 00:08:49,640
Speaker 1: what about some of the guardrails and safeguards needed in AI.

158
00:08:50,960 --> 00:08:52,120
Speaker 2: It's incredibly important.

159
00:08:52,240 --> 00:08:54,440
Speaker 3: I fund you should use the word guardrails because this

160
00:08:54,559 --> 00:08:57,280
Speaker 3: morning here a reinvent where we have fifty thousand people

161
00:08:57,320 --> 00:08:59,920
Speaker 3: live in Las Vegas and over three hundred other thousand

162
00:09:00,040 --> 00:09:04,280
Speaker 3: people registered around the world. We actually announced a guardrails

163
00:09:04,280 --> 00:09:07,560
Speaker 3: for Amazon Bedrock. Bedrock is our foundation model as a

164
00:09:07,600 --> 00:09:12,880
Speaker 3: service offering with all those models inside, and Guardrails is

165
00:09:12,920 --> 00:09:17,080
Speaker 3: a new capability which actually allows you to enforce your

166
00:09:17,120 --> 00:09:21,559
Speaker 3: own responsible AI priorities and policies. So if you're a

167
00:09:21,600 --> 00:09:24,880
Speaker 3: financial services company, you might have your chatbot not give

168
00:09:24,880 --> 00:09:28,880
Speaker 3: investment advice, or if you're a utility with customer service

169
00:09:29,440 --> 00:09:33,680
Speaker 3: chatbot or generative AI assistant, you might have it strip

170
00:09:33,760 --> 00:09:37,040
Speaker 3: out the PII or personally identifiable information from the call

171
00:09:37,080 --> 00:09:40,480
Speaker 3: summary to respect people's privacy. So we're taking really seriously

172
00:09:40,559 --> 00:09:43,400
Speaker 3: to the point of actually building in these leading edge

173
00:09:43,440 --> 00:09:47,440
Speaker 3: capabilities that really aren't available from other providers. And in addition,

174
00:09:47,480 --> 00:09:50,160
Speaker 3: it's really important that we're participating in the working groups

175
00:09:50,200 --> 00:09:52,320
Speaker 3: that are discussing these things. We're going to need a

176
00:09:52,360 --> 00:09:57,200
Speaker 3: multilateral approach of government, private sector, academics, and more. And

177
00:09:57,320 --> 00:09:59,520
Speaker 3: I was the White House with President Biden the summer

178
00:10:00,040 --> 00:10:03,920
Speaker 3: where we participated in these voluntary commitments around AI. Earlier

179
00:10:03,960 --> 00:10:06,640
Speaker 3: this month, I was with the UK Prime Minister Sunak

180
00:10:06,679 --> 00:10:09,640
Speaker 3: at a gathering where he announced at AI Safety Institute

181
00:10:09,679 --> 00:10:12,520
Speaker 3: in the UK, we're going to continue spending time participating

182
00:10:12,559 --> 00:10:13,360
Speaker 3: in these dialogues.

183
00:10:13,440 --> 00:10:16,880
Speaker 1: Adam Szlitzky, ws CEO, thank you for your time. Back

184
00:10:16,920 --> 00:10:17,400
Speaker 1: to you guys.

185
00:10:17,400 --> 00:10:18,000
Speaker 2: Thank you