Dwarkesh Podcast · Aug 25, 03:32 PM UTC

Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028

xyz:NVDAHOLDMEDIUM confidenceUpdated Aug 26, 07:47 AM UTC

Historical operation · Retained

“now is this: Before, if they served a model — GPT-4 being served on Nvidia Hopper GPUs — it was”

Quoted source · Dylan Patel · 02:38–01:14:41 · 123 source segments

[1] 02:38–02:47
[2] 02:47–02:54
[3] 02:55–03:03
[4] 03:03–03:08
[5] 03:08–03:15
[6] 03:15–03:19
[7] 03:19–03:24
[8] 26:16–26:22
[9] 26:22–26:28
[10] 26:28–26:34
[11] 26:34–26:36
[12] 26:36–26:39
[13] 26:40–26:43
[14] 26:43–26:48
[15] 26:48–26:52
[16] 26:52–26:55
[17] 26:55–27:00
[18] 27:02–27:06
[19] 27:06–27:11
[20] 27:11–27:15
[21] 27:15–27:19
[22] 27:19–27:22
[23] 34:15–34:21
[24] 34:22–34:26
[25] 34:26–34:32
[26] 34:32–34:34
[27] 34:37–34:45
[28] 34:45–34:48
[29] 34:49–34:52
[30] 34:52–34:58
[31] 34:58–35:02
[32] 35:02–35:09
[33] 35:09–35:13
[34] 35:13–35:18
[35] 35:18–35:25
[36] 35:25–35:31
[37] 35:31–35:37
[38] 35:37–35:41
[39] 35:41–35:43
[40] 35:43–35:46
[41] 35:46–35:51
[42] 35:51–35:59
[43] 35:59–36:04
[44] 36:04–36:10
[45] 36:10–36:15
[46] 36:15–36:21
[47] 36:21–36:27
[48] 36:27–36:32
[49] 36:32–36:36
[50] 36:36–36:39
[51] 36:39–36:45
[52] 36:45–36:50
[53] 36:50–36:54
[54] 36:54–37:01
[55] 37:01–37:08
[56] 37:08–37:13
[57] 37:13–37:18
[58] 37:19–37:23
[59] 37:23–37:32
[60] 37:32–37:36
[61] 37:36–37:42
[62] 37:42–37:48
[63] 37:48–37:54
[64] 37:56–38:00
[65] 38:00–38:03
[66] 38:03–38:10
[67] 38:10–38:16
[68] 38:16–38:25
[69] 38:25–38:28
[70] 38:28–38:30
[71] 38:30–38:34
[72] 38:34–38:36
[73] 38:37–38:42
[74] 45:32–45:36
[75] 45:36–45:40
[76] 45:40–45:46
[77] 45:46–45:48
[78] 45:48–45:55
[79] 45:55–46:00
[80] 46:00–46:04
[81] 46:04–46:08
[82] 46:08–46:12
[83] 46:12–46:17
[84] 46:17–46:21
[85] 46:21–46:25
[86] 46:25–46:31
[87] 46:31–46:36
[88] 46:36–46:42
[89] 46:42–46:45
[90] 46:46–46:51
[91] 46:51–46:56
[92] 46:56–47:00
[93] 47:00–47:03
[94] 47:03–47:10
[95] 47:10–47:13
[96] 47:13–47:17
[97] 47:17–47:21
[98] 47:21–47:26
[99] 01:12:51–01:12:58
[100] 01:12:58–01:13:00
[101] 01:13:00–01:13:07
[102] 01:13:07–01:13:09
[103] 01:13:09–01:13:12
[104] 01:13:12–01:13:15
[105] 01:13:15–01:13:18
[106] 01:13:18–01:13:22
[107] 01:13:22–01:13:28
[108] 01:13:28–01:13:32
[109] 01:13:32–01:13:34
[110] 01:13:34–01:13:46
[111] 01:13:46–01:13:48
[112] 01:13:48–01:13:52
[113] 01:13:52–01:13:55
[114] 01:13:55–01:13:58
[115] 01:13:58–01:14:00
[116] 01:14:00–01:14:04
[117] 01:14:04–01:14:09
[118] 01:14:09–01:14:15
[119] 01:14:15–01:14:19
[120] 01:14:19–01:14:24
[121] 01:14:24–01:14:30
[122] 01:14:32–01:14:37
[123] 01:14:37–01:14:41

Model assessment

The transcript discusses Nvidia's role in the AI compute supply chain, noting that AI labs are generating high revenue per megawatt and that Nvidia has pricing power, but it does not provide a direct, material catalyst for Nvidia's near-term fundamentals beyond the existing BUY signal.

123 exact transcript citations

Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028

Current publication unavailable.