WEBVTT

1
00:00:00.080 --> 00:00:04.879
<v Speaker 1>AI daily breathing, I'm ed sharp, thanks for joining me today.

2
00:00:05.040 --> 00:00:08.080
<v Speaker 1>The inference chip arms race. Why seven hundred million dollars

3
00:00:08.080 --> 00:00:12.560
<v Speaker 1>flows to Etched, not models? Seven hundred million dollars. That's

4
00:00:12.599 --> 00:00:15.519
<v Speaker 1>what just landed on a chip company most people outside

5
00:00:15.519 --> 00:00:19.160
<v Speaker 1>Silicon Valley haven't heard of yet. Etched closed its Series

6
00:00:19.239 --> 00:00:22.519
<v Speaker 1>D at a twenty one billion dollar valuation with Jane

7
00:00:22.600 --> 00:00:25.839
<v Speaker 1>Street and Sequoia among the backers, and shipped its first

8
00:00:25.839 --> 00:00:29.519
<v Speaker 1>production rack directly to Jane Street. That's not a demo,

9
00:00:29.839 --> 00:00:32.840
<v Speaker 1>that's revenue, and it tells you exactly where the smart

10
00:00:32.840 --> 00:00:36.119
<v Speaker 1>money thinks the AI bottleneck lives right now. It's not

11
00:00:36.200 --> 00:00:39.679
<v Speaker 1>the models, it's the hardware running them. The thesis behind

12
00:00:39.679 --> 00:00:44.000
<v Speaker 1>the Etched round is straightforward. As AI inference scales, throughput

13
00:00:44.079 --> 00:00:47.920
<v Speaker 1>and memory architecture become the real competitive constraint. Building a

14
00:00:47.920 --> 00:00:51.640
<v Speaker 1>bigger model is yetting cheaper. Running it fast enough at

15
00:00:51.640 --> 00:00:55.280
<v Speaker 1>the right cost per query at scale is the harder problem.

16
00:00:55.759 --> 00:00:58.560
<v Speaker 1>Etched is betting. The inference chip is where margin and

17
00:00:58.600 --> 00:01:02.439
<v Speaker 1>mote get built. James Street taking the first wreck isn't

18
00:01:02.479 --> 00:01:05.239
<v Speaker 1>just a vote of confidence. It's a signal that a

19
00:01:05.239 --> 00:01:09.159
<v Speaker 1>sophisticated trading firm sees speed and cost per compute as

20
00:01:09.200 --> 00:01:14.480
<v Speaker 1>strategically important enough to deploy capital twice over. Etch isn't alone.

21
00:01:14.959 --> 00:01:17.840
<v Speaker 1>The Laura AI just raised one hundred and ten million

22
00:01:17.879 --> 00:01:21.640
<v Speaker 1>dollars at evaluation above one billion to attack the power

23
00:01:21.680 --> 00:01:26.159
<v Speaker 1>consumption side of the same problem, different layer, same thesis.

24
00:01:26.599 --> 00:01:30.480
<v Speaker 1>Compute per watt is becoming as valuable as compute per dollar.

25
00:01:30.840 --> 00:01:34.000
<v Speaker 1>The important distinction is that investors aren't just choosing between

26
00:01:34.079 --> 00:01:37.359
<v Speaker 1>chip start ups. They're converging on a shared view that

27
00:01:37.439 --> 00:01:41.359
<v Speaker 1>the AI economy's real infrastructure isn't the model there, it's

28
00:01:41.439 --> 00:01:45.519
<v Speaker 1>everything underneath it. Stripe made the same bet differently its

29
00:01:45.519 --> 00:01:49.319
<v Speaker 1>acquisition of open Router, the model routing platform spanning over

30
00:01:49.359 --> 00:01:52.959
<v Speaker 1>four hundred models, reportedly landed somewhere between seven and eight

31
00:01:53.000 --> 00:01:56.799
<v Speaker 1>billion dollars, depending on which figure you use. The strategic

32
00:01:56.840 --> 00:02:00.359
<v Speaker 1>logic is clean. If AI applications are routing query across

33
00:02:00.400 --> 00:02:04.000
<v Speaker 1>dozens of models based on cost, latency, and capability, then

34
00:02:04.079 --> 00:02:08.000
<v Speaker 1>controlling the routing layer is controlling the billing layer. Stripe

35
00:02:08.000 --> 00:02:12.599
<v Speaker 1>already owns payments infrastructure. Now it's extending that position into

36
00:02:12.599 --> 00:02:16.879
<v Speaker 1>token economics. That's not a bolt on, that's a platform play.

37
00:02:17.280 --> 00:02:20.439
<v Speaker 1>The most urgent story in this cycle doesn't involve funding.

38
00:02:20.759 --> 00:02:24.439
<v Speaker 1>Greg Brockman's public warning about zipu AI's GLM dash five

39
00:02:24.479 --> 00:02:28.520
<v Speaker 1>point three raises a genuinely difficult problem. The model reportedly

40
00:02:28.560 --> 00:02:32.199
<v Speaker 1>scored eighty four point five percent on cybergym exploit benchmarks.

41
00:02:32.719 --> 00:02:37.680
<v Speaker 1>Zipu delayed release after discovering unexpected cybersecurity capabilities. The catch

42
00:02:37.759 --> 00:02:40.439
<v Speaker 1>is that delay is in recall once in open weight

43
00:02:40.479 --> 00:02:45.039
<v Speaker 1>nodel ships, it ships everywhere. Here's the thing open AI's framing.

44
00:02:45.039 --> 00:02:49.000
<v Speaker 1>Here is the defender's window. Formal verification tools that could

45
00:02:49.039 --> 00:02:52.240
<v Speaker 1>authenticate AI behavior at the code level are years away

46
00:02:52.240 --> 00:02:56.199
<v Speaker 1>from production scale hacking capable open weight models are weeks

47
00:02:56.199 --> 00:02:59.560
<v Speaker 1>away from public release. That gap is the real exposure.

48
00:03:00.080 --> 00:03:03.240
<v Speaker 1>The benchmark figure itself is self report. The benchmark figure

49
00:03:03.280 --> 00:03:07.439
<v Speaker 1>itself is self reported and not independently verified, so some

50
00:03:07.520 --> 00:03:10.719
<v Speaker 1>caution is warranted, but the structural problem doesn't depend on

51
00:03:10.759 --> 00:03:15.280
<v Speaker 1>the exact number. The asymmetry between offensive accessibility and defensive

52
00:03:15.319 --> 00:03:19.639
<v Speaker 1>readiness is real and widening. August second mark the point

53
00:03:19.639 --> 00:03:23.919
<v Speaker 1>where EUAI Act compliance stopped being a planning exercise. Role

54
00:03:23.960 --> 00:03:29.360
<v Speaker 1>based accountability, high risk classification requirements across employment, education and

55
00:03:29.400 --> 00:03:33.800
<v Speaker 1>safety systems, and contractual allocation of liability are now active.

56
00:03:34.439 --> 00:03:38.319
<v Speaker 1>Organizations that treated compliance as future facing policy work are

57
00:03:38.360 --> 00:03:42.120
<v Speaker 1>finding out it's a procurement and contracting issue. Today, fines

58
00:03:42.199 --> 00:03:44.840
<v Speaker 1>run up to fifteen million euros or three per cent

59
00:03:44.879 --> 00:03:49.120
<v Speaker 1>of global turnover for general purpose AI violations. The timing

60
00:03:49.199 --> 00:03:52.439
<v Speaker 1>is uncomfortable for one more reason. New research shows that

61
00:03:52.479 --> 00:03:56.360
<v Speaker 1>the compliance detectors widely deployed across the guardrails market don't

62
00:03:56.360 --> 00:03:59.199
<v Speaker 1>actually apply the rules they're trained on. They pattern match

63
00:03:59.240 --> 00:04:02.759
<v Speaker 1>regulatory length, which then ignore it. Only chain of thought

64
00:04:02.800 --> 00:04:06.759
<v Speaker 1>reasoning approaches appear to escape that failum mode. The signal

65
00:04:06.800 --> 00:04:09.319
<v Speaker 1>here is that a two point one billion dollar guard

66
00:04:09.400 --> 00:04:13.400
<v Speaker 1>rails market may be operating on architecture that can't satisfy

67
00:04:13.479 --> 00:04:16.920
<v Speaker 1>enforcement grades. Screw to me. The execution gap isn't limited

68
00:04:16.959 --> 00:04:20.319
<v Speaker 1>to Europe. In the uae AI spending rows one hundred

69
00:04:20.319 --> 00:04:23.160
<v Speaker 1>and five percent by one measure, but only seven percent

70
00:04:23.199 --> 00:04:27.160
<v Speaker 1>of enterprises have deployed autonomous workflows. Seventy seven percent cite

71
00:04:27.240 --> 00:04:31.360
<v Speaker 1>data readiness as the primary barrier. That pattern large budget

72
00:04:31.639 --> 00:04:35.879
<v Speaker 1>limited operationalization isn't unique to the region. It's the dominant

73
00:04:36.000 --> 00:04:40.160
<v Speaker 1>enterprise AI story globally. Right now, consider this on a

74
00:04:40.199 --> 00:04:43.879
<v Speaker 1>smaller scale, discovered materials closed a nine million dollar seed

75
00:04:43.959 --> 00:04:48.800
<v Speaker 1>round to deploy AI agent swarms identifying firmly efficient chip materials.

76
00:04:49.279 --> 00:04:51.879
<v Speaker 1>It's a narrow bit, but it addresses a constrain that

77
00:04:51.920 --> 00:04:55.759
<v Speaker 1>pure compute scaling can't solve. Keat and energy density are

78
00:04:55.759 --> 00:05:00.639
<v Speaker 1>physical limits. Finding better materials is a discovery problem. AI

79
00:05:00.800 --> 00:05:04.600
<v Speaker 1>Accelerated Discovery is now a funded category. The through line

80
00:05:04.600 --> 00:05:07.600
<v Speaker 1>across today's developments is that the competitive advantage in AI

81
00:05:07.759 --> 00:05:10.120
<v Speaker 1>is shifting from who has the best model to who

82
00:05:10.120 --> 00:05:13.240
<v Speaker 1>controls the infrastructure underneath it, and who can demonstrate that

83
00:05:13.319 --> 00:05:17.040
<v Speaker 1>infrastructure is actually safe and lawful etched. The Laura and

84
00:05:17.079 --> 00:05:20.519
<v Speaker 1>Stripe are building three different layers of that stack. The

85
00:05:20.560 --> 00:05:23.360
<v Speaker 1>cybersecurity window and the guard rail research are both telling

86
00:05:23.360 --> 00:05:26.360
<v Speaker 1>you that governance hasn't kept pace. The near term metrics

87
00:05:26.360 --> 00:05:29.519
<v Speaker 1>to watch are James Street's inference performance data from the

88
00:05:29.600 --> 00:05:32.959
<v Speaker 1>Edge Deployment ZIPU, AI's release decision on g L and

89
00:05:33.040 --> 00:05:35.560
<v Speaker 1>DASH five point three, and the first wave of EU

90
00:05:35.639 --> 00:05:38.879
<v Speaker 1>enforcement actions. Those three data points will tell you more

91
00:05:38.879 --> 00:05:42.360
<v Speaker 1>about where this industry is actually headed than any benchmark announcement.

92
00:05:43.079 --> 00:05:46.720
<v Speaker 1>Thanks for listening. This podcast was built using AI technology

93
00:05:47.160 --> 00:05:48.399
<v Speaker 1>AI s WE production