WEBVTT

1
00:00:00.000 --> 00:00:03.440
The general consensus in technology is
that the ultimate prize

2
00:00:03.440 --> 00:00:07.040
is the smartest AI brain, with companies
like OpenAI sitting

3
00:00:07.040 --> 00:00:08.400
securely at the top.

4
00:00:08.400 --> 00:00:11.360
But JPMorgan Chase recently ran a rare
analysis on the

5
00:00:11.360 --> 00:00:15.040
frontier AI market and concluded something
entirely different.

6
00:00:15.040 --> 00:00:18.960
They determined that OpenAI's competitive
advantage is actually an increasingly

7
00:00:18.960 --> 00:00:22.800
fragile moat. The math behind that
assessment is straightforward.

8
00:00:22.800 --> 00:00:25.680
No single developer can maintain a
sustained edge.

9
00:00:25.680 --> 00:00:29.680
When a major leap like GPT-5 launches,
competitors like Anthropic

10
00:00:29.680 --> 00:00:33.520
and Google match or exceed its performance
benchmarks almost immediately.

11
00:00:33.520 --> 00:00:36.480
Intelligence at the bleeding edge is a
temporary state.

12
00:00:36.480 --> 00:00:38.720
The model itself is commoditizing.

13
00:00:38.720 --> 00:00:43.200
Training and operating these massive
systems requires escalating capital expenditure.

14
00:00:43.200 --> 00:00:46.000
This creates what the industry calls the
model tax, a

15
00:00:46.000 --> 00:00:49.200
massive, ongoing compute cost required
just to keep the lights

16
00:00:49.200 --> 00:00:53.680
on. This dual-axis chart visualizes the
resulting lethal margin squeeze.

17
00:00:53.680 --> 00:00:57.600
The red line represents the exponential
rise in compute capital

18
00:00:57.600 --> 00:01:01.080
expenditure, while the blue line shows the
margin per token

19
00:01:01.080 --> 00:01:04.680
plummeting towards zero as companies are
forced to compete on

20
00:01:04.680 --> 00:01:08.440
price. Adding to the financial pressure,
the strategy of simply

21
00:01:08.440 --> 00:01:12.120
feeding models more data is yielding
diminishing returns.

22
00:01:12.120 --> 00:01:16.120
Researchers at major conferences like
NURIPS are actively debating whether

23
00:01:16.120 --> 00:01:19.320
scaling laws have hit a wall, forcing the
academic community

24
00:01:19.320 --> 00:01:21.240
to look for alternative methods.

25
00:01:21.240 --> 00:01:25.320
With margins shrinking and costs
skyrocketing, the strategic focus of

26
00:01:25.320 --> 00:01:28.360
the industry is being forced to pivot
completely away from

27
00:01:28.360 --> 00:01:29.640
the model layer.

28
00:01:29.640 --> 00:01:31.800
If you want to see where the market is
actually

29
00:01:31.800 --> 00:01:34.600
going, you have to ignore the AI labs and
track

30
00:01:34.600 --> 00:01:37.880
the billions of dollars flowing through
recent mega acquisitions.

31
00:01:37.880 --> 00:01:41.080
Salesforce spent $8 billion on
Informatica.

32
00:01:41.080 --> 00:01:43.800
Cisco spent $28 billion on Splunk.

33
00:01:43.800 --> 00:01:45.480
Databricks acquired Tabular.

34
00:01:45.480 --> 00:01:47.400
IBM moved for Datastacks.

35
00:01:47.400 --> 00:01:50.440
None of those multi-billion dollar deals
were to acquire a

36
00:01:50.440 --> 00:01:54.960
new model. Every single one was about the
unglamorous plumbing,

37
00:01:54.960 --> 00:01:58.160
data integration, infrastructure, and
governance.

38
00:01:58.160 --> 00:02:03.920
A typical large enterprise runs over 106
different siloed applications.

39
00:02:03.920 --> 00:02:08.400
That fragmented landscape makes raw AI
models practically useless.

40
00:02:08.400 --> 00:02:13.120
Bridging the massive gap between
cutting-edge inference and 20-year-old ERP

41
00:02:13.120 --> 00:02:17.040
systems drives this $260 billion market.

42
00:02:17.040 --> 00:02:20.160
A powerful AI brain is required, but
without the data

43
00:02:20.160 --> 00:02:22.560
plumbing to connect it to the business, it
has zero

44
00:02:22.560 --> 00:02:26.880
practical value. We are watching the
agentic shift happen.

45
00:02:26.880 --> 00:02:30.640
The industry is moving from passive text
generators to autonomous

46
00:02:30.640 --> 00:02:34.960
systems that perceive, plan, and execute
multi-step tasks.

47
00:02:34.960 --> 00:02:38.800
Deploying these agents into a corporate
environment hinges on trust,

48
00:02:38.800 --> 00:02:42.320
moving beyond the mere technical challenge
of data integration.

49
00:02:42.320 --> 00:02:47.200
Peer-reviewed research at ICLR and ACLR
proves that collaborative agents

50
00:02:47.200 --> 00:02:49.680
can easily bypass standard safety
guardrails.

51
00:02:49.880 --> 00:02:54.680
They execute what are called decomposition
jailbreaks by splitting malicious

52
00:02:54.680 --> 00:02:59.880
tasks into harmless-looking pieces,
creating an illusion of shallow safety.

53
00:02:59.880 --> 00:03:03.880
Corporate legal and compliance teams look
at those vulnerabilities and

54
00:03:03.880 --> 00:03:07.320
flatly refuse to grant a black box agent
access to

55
00:03:07.320 --> 00:03:11.640
core financial or HR data without ironclad
oversight.

56
00:03:11.640 --> 00:03:16.120
Engineering a secure, auditable trust
bridge defines the final stage

57
00:03:16.120 --> 00:03:17.880
of enterprise deployment.

58
00:03:17.880 --> 00:03:21.880
Based on research from the Orchestrix
project, the only defensible

59
00:03:21.880 --> 00:03:25.560
position left in the industry is building
the specialized management

60
00:03:25.560 --> 00:03:28.600
stack that sits on top of these commodity
models.

61
00:03:28.600 --> 00:03:32.200
This architectural blueprint outlines the
required stack.

62
00:03:32.200 --> 00:03:36.600
It starts with an agent-ready data fabric
optimized for compliance.

63
00:03:36.600 --> 00:03:41.400
That foundation supports a multi-agent
reasoning engine to decompose massive

64
00:03:41.400 --> 00:03:45.440
business goals. To secure the system, a
firewall locks into

65
00:03:45.440 --> 00:03:49.840
place. Finally, a human-agent
collaboration layer sits on top.

66
00:03:49.840 --> 00:03:55.120
That four-part architecture transforms an
unpredictable, volatile AI model into

67
00:03:55.120 --> 00:03:57.440
a safe, reliable corporate asset.

68
00:03:57.440 --> 00:03:59.280
What we are looking at is the birth of a

69
00:03:59.280 --> 00:04:03.440
mandatory new software category, agentic
resource planning.

70
00:04:03.440 --> 00:04:07.840
As AI usage becomes a universal baseline,
CIOs and compliance

71
00:04:07.840 --> 00:04:11.360
officers will require these systems to
track the usage, safety,

72
00:04:11.360 --> 00:04:14.880
and return on investment of an entirely
digital workforce.

73
00:04:14.880 --> 00:04:18.800
The initial race for foundation model
supremacy is effectively over,

74
00:04:18.800 --> 00:04:22.160
squeezed by tight margins and eroding
competitive modes.

75
00:04:22.160 --> 00:04:25.040
The industry's center of gravity is
shifting toward those who

76
00:04:25.040 --> 00:04:28.640
build the operating system, leaving the
providers of raw intelligence

77
00:04:28.640 --> 00:04:31.520
to compete on razor-thin margins.