Latest / Silicon & Shale Podcast / Episode 14: The week's E&P developments & applied AI

Transcript

Episode 14: The week's E&P developments & applied AI · Silicon & Shale Podcast

0:00
SPEAKER_00: So I'm looking at this Venezuela thing and I keep going back to
0:02
SPEAKER_00: the phrase biggest oil deal in world history, which that's what they're calling it.
0:06
SPEAKER_01: Wait, who's calling it that?
0:08
SPEAKER_01: Because the actual structure they're floating is basically a concession model.
0:11
SPEAKER_01: Like, we're talking about something that looks a lot like the old Anglo-Persian deal
0:15
SPEAKER_01: from 1901, except that one was 60 years, not 100.
0:19
SPEAKER_00: And that became BP,
0:20
SPEAKER_00: so the historical precedent they're reaching for doesn't even match
0:23
SPEAKER_00: the terms they're describing.
0:25
SPEAKER_00: But here's what gets me.
0:26
SPEAKER_00: PDVSA can barely keep the lights on at their existing facilities.
0:30
SPEAKER_00: You're going to negotiate a long-term lease with a national oil company that's been
0:33
SPEAKER_00: in operational freefall for a decade.
0:35
SPEAKER_01: And the SPR angle is...
0:37
SPEAKER_01: Okay,
0:37
SPEAKER_01: so the idea is heavy Venezuelan crude goes into the Strategic Petroleum Reserve,
0:42
SPEAKER_01: which, fine, we've bought heavy crude for the SPR before.
0:45
SPEAKER_01: But the logistics of actually moving that volume consistently out
0:47
SPEAKER_01: of Venezuela right now, given the state of their export infrastructure?
0:51
SPEAKER_00: Nobody wants to model that part.
0:53
SPEAKER_00: Everyone's talking about the deal terms, the geopolitics, the sanctions relief.
0:57
SPEAKER_00: Nobody's talking about whether the crude can physically get from the wellhead to
1:00
SPEAKER_00: a VLCC in anything resembling a reliable schedule.
1:03
SPEAKER_01: The molecules can't get their problem all over again.
1:05
SPEAKER_00: Different country, same constraint.
1:07
SPEAKER_00: And look, I...
1:09
SPEAKER_00: The thing is, temporary sanctions relief is not a 60-year concession.
1:12
SPEAKER_00: That isn't even close.
1:14
SPEAKER_00: You've got a policy window that could slam shut in one election cycle,
1:17
SPEAKER_00: and someone's supposed to commit billions in upstream rehab capital against that?
1:20
SPEAKER_01: On infrastructure that's been cannibalized for parts since, what, 2017?
1:23
SPEAKER_01: 2018?
1:26
SPEAKER_01: The deferred maintenance alone on PDVSA's gathering systems,
1:29
SPEAKER_01: I don't even want to think about the inspection reports, if they exist.
1:32
SPEAKER_00: That's generous of you, assuming there are inspection reports.
1:35
SPEAKER_01: So here's my question.
1:36
SPEAKER_01: If you're an operator evaluating this,
1:38
SPEAKER_01: what's your discount rate on a deal where
1:40
SPEAKER_01: the host government could nationalize you again,
1:42
SPEAKER_01: and the export terminal might need three years of work before you can load a tanker
1:45
SPEAKER_01: on schedule?
1:46
SPEAKER_00: You don't have a discount rate.
1:47
SPEAKER_00: You have a prayer,
1:48
SPEAKER_00: and a very expensive political risk insurance policy
1:50
SPEAKER_00: that probably doesn't cover what you actually need it to cover.
1:53
SPEAKER_01: Welcome to Silicon and Shale, where the oil patch meets the algorithm.
1:57
SPEAKER_01: It's the week of September the 1st, 2026,
2:00
SPEAKER_01: and we're already deep into this Venezuela deal, which, look,
2:03
SPEAKER_01: we're gonna keep pulling on that thread because the operational questions are wild.
2:07
SPEAKER_01: But we've also got equipment bottlenecks choking new power generation,
2:10
SPEAKER_01: right when AI data centers need it most,
2:12
SPEAKER_01: drones and robotics actually showing up on well pads in ways that aren't vaporware.
2:16
SPEAKER_01: And this whole question of whether enterprise AI architecture matters more than
2:19
SPEAKER_01: which model you're running, which, spoiler, it does.
2:22
SPEAKER_00: And whether anybody's discount rate can survive a country that hasn't filed
2:25
SPEAKER_00: an inspection report since the Obama administration, we'll get to all of it.
2:29
SPEAKER_00: But here's what actually matters this week on the E&P side beyond Venezuela.
2:33
SPEAKER_00: The equipment bottleneck story.
2:34
SPEAKER_00: Pac, you've been tracking the transformer and turbine supply chain stuff.
2:38
SPEAKER_01: So, the energy gang did a whole deep dive on this,
2:40
SPEAKER_01: and the core problem is pretty simple to state and terrifying to model.
2:44
SPEAKER_01: AI data centers are driving electricity demand through the roof.
2:47
SPEAKER_01: Everyone knows that.
2:48
SPEAKER_01: But the physical equipment is the choke point.
2:49
SPEAKER_01: Turbines, transformers, the actual hardware you bolt down and wire up.
2:53
SPEAKER_01: Manufacturers can't keep pace with orders.
2:56
SPEAKER_00: Which, look, I've watched this movie before in different clothes.
2:59
SPEAKER_00: You get a surge in drilling activity and suddenly nobody can get a frat crew for eight months.
3:03
SPEAKER_00: Same mechanics, different equipment.
3:04
SPEAKER_01: The timeline mismatch here is worse, though.
3:06
SPEAKER_01: A gas turbine for a new power plant?
3:08
SPEAKER_01: You're talking lead times that have stretched way out.
3:11
SPEAKER_01: And transformers?
3:12
SPEAKER_01: The big ones?
3:13
SPEAKER_01: The ones that actually matter for grid interconnection?
3:15
SPEAKER_01: Some of those are two, three years out on order.
3:18
SPEAKER_00: And meanwhile,
3:19
SPEAKER_00: you've got operators in West Texas who can't get power to their own facilities
3:22
SPEAKER_00: because the grid's already strained.
3:24
SPEAKER_00: So you've got this weird thing where the data center demand is competing directly
3:27
SPEAKER_00: with upstream operations for the same electrons and the same equipment
3:30
SPEAKER_00: to deliver those electrons.
3:31
SPEAKER_01: Which brings up Amazon's private gas grid thing we talked about a few weeks back.
3:35
SPEAKER_01: That starts looking less like a tech company playing oil and gas and more like...
3:38
SPEAKER_01: They saw this bottleneck coming and decided to route around the whole utility model.
3:42
SPEAKER_00: And that worked out real well for surface access in the Permian.
3:44
SPEAKER_01: But here's where the automation angle genuinely matters.
3:47
SPEAKER_01: If you've got a two-year wait for a transformer,
3:49
SPEAKER_01: the question becomes, can you squeeze more capacity out of existing infrastructure?
3:53
SPEAKER_01: Dynamic load management,
3:54
SPEAKER_01: real-time grid optimization,
3:56
SPEAKER_01: predictive maintenance on aging equipment so you don't lose what you already have.
4:00
SPEAKER_01: That's an urgent use case right now, not a slide deck hypothetical.
4:03
SPEAKER_00: But there's a step before the optimization that nobody wants to talk about,
4:06
SPEAKER_00: which is,
4:07
SPEAKER_00: do you even know the actual condition of the equipment you're trying to optimize?
4:10
SPEAKER_00: Half the substations feeding oil field operations in the Permian are running gear
4:14
SPEAKER_00: from the 80s and the maintenance records are, let's say, aspirational.
4:17
SPEAKER_01: You mean like Venezuelan inspection reports?
4:19
SPEAKER_00: At least we know where our substations are.
4:21
SPEAKER_01: But the problem is identical in structure.
4:23
SPEAKER_01: You can't optimize what you can't see.
4:25
SPEAKER_01: And you can't see what nobody's been documenting.
4:27
SPEAKER_01: So before you get to the fancy load-balancing AI,
4:30
SPEAKER_01: someone has to go inventory what's actually in the field and what shape it's in.
4:33
SPEAKER_00: And that someone is usually a contractor making $28 an hour who's going
4:37
SPEAKER_00: to write his findings on a clipboard that gets scanned into a PDF that gets emailed
4:41
SPEAKER_00: to someone who puts it in a folder nobody opens.
4:43
SPEAKER_01: Which...
4:44
SPEAKER_01: Okay, so that actually connects to the drones and robotics story.
4:48
SPEAKER_01: Sean Garay from Innovate Energy Network was on the Upstream podcast this week,
4:51
SPEAKER_01: talking about how drones and robotics are actually showing up on well pads now,
4:54
SPEAKER_01: past the booth demos and pilot decks.
4:57
SPEAKER_01: Actual crews using them on actual pads.
4:59
SPEAKER_00: I'll be honest,
5:00
SPEAKER_00: I've been more skeptical on this than I probably should have been,
5:02
SPEAKER_00: but I saw a rig inspection last month where they flew a drone up the derrick instead
5:06
SPEAKER_00: of sending a guy up on a climbing harness,
5:08
SPEAKER_00: and the footage was genuinely better than what the climber would have gotten.
5:11
SPEAKER_01: Wait, seriously?
5:12
SPEAKER_01: You're admitting that?
5:13
SPEAKER_00: And I think the distinction matters.
5:15
SPEAKER_00: The capture layer is getting commoditized.
5:16
SPEAKER_00: Drones, sensors, cameras, that part's basically solved.
5:18
SPEAKER_00: Interpretation still runs on human judgment.
5:20
SPEAKER_00: And then the real value concentrates.
5:21
SPEAKER_00: For now.
5:21
SPEAKER_00: Sure, for now.
5:22
SPEAKER_01: But Gary's point was interesting.
5:39
SPEAKER_01: He was talking about how these technologies are enhancing safety
5:42
SPEAKER_01: and efficiency simultaneously, which in oil and gas usually means you pick one.
5:46
SPEAKER_01: You either slow down and get safer, or you speed up and accept more risk.
5:50
SPEAKER_00: That's been the trade-off for 50 years.
5:52
SPEAKER_00: If the robotics actually break that, I'll pay attention.
5:54
SPEAKER_01: You're already paying attention.
5:56
SPEAKER_01: You just told me about the drone inspection.
5:58
SPEAKER_00: I said the camera was good, Pac.
5:59
SPEAKER_01: But here's where I push back on the whole human-judgment-is-irreplaceable framing,
6:03
SPEAKER_01: because the equipment bottleneck story and the drone story feed into each other.
6:07
SPEAKER_01: Partway?
6:07
SPEAKER_01: I'll give you partway.
6:08
SPEAKER_01: But the gap between this drone-detected corrosion on a flange
6:09
SPEAKER_01: and this system correctly prioritized which
6:10
SPEAKER_01: of 900 flanges needs attention this week is enormous.
6:32
SPEAKER_00: That prioritization requires knowing the process conditions,
6:35
SPEAKER_00: the history of that specific joint, what's downstream if it fails.
6:39
SPEAKER_01: And none of that context lives in one system.
6:41
SPEAKER_01: It's spread across the historian, the CMMS, the PNIDs, probably someone's head.
6:45
SPEAKER_00: Probably two people's heads, and one of them retired.
6:47
SPEAKER_01: So the real bottleneck lives in data integration.
6:50
SPEAKER_01: The drone hardware works.
6:51
SPEAKER_01: The models are decent.
6:53
SPEAKER_01: But stitching together the context those tools need to be useful?
6:56
SPEAKER_01: That's the boring,
6:56
SPEAKER_01: unsexy problem nobody wants to fund because you can't put it in a press release.
7:00
SPEAKER_00: You can put Digital Twin in a press release, though.
7:02
SPEAKER_01: You can put anything in a press release.
7:04
SPEAKER_01: Navis was sponsoring that oil and gas this week episode doing laser scanning for digital twins.
7:08
SPEAKER_01: And the technology's real, I'm not dismissing it.
7:11
SPEAKER_01: But a digital twin of a facility where the source data hasn't been reconciled
7:14
SPEAKER_01: against as-built conditions is just a really expensive 3D model
7:17
SPEAKER_01: of what you hope is there.
7:18
SPEAKER_00: I worked a facility in the Eagle Ford two years ago where the digital twin showed
7:21
SPEAKER_00: a separator that had been removed 18 months earlier.
7:24
SPEAKER_00: Nobody updated the model.
7:25
SPEAKER_00: The field guys thought it was hilarious.
7:27
SPEAKER_00: The engineering team in Houston did not.
7:29
SPEAKER_01: And the model's only as current as the last person who bothered to update it.
7:32
SPEAKER_00: And the incentive to update it is exactly zero if nobody downstream actually uses it
7:36
SPEAKER_00: for decisions.
7:37
SPEAKER_01: But let me flip this.
7:38
SPEAKER_01: The equipment bottleneck, transformers, turbines, all of it, is a multi-year constraint.
7:43
SPEAKER_01: That isn't getting solved by next quarter.
7:46
SPEAKER_01: So operators who figure out how to extend the life and optimize the output
7:49
SPEAKER_01: of existing assets have a genuine competitive advantage.
7:52
SPEAKER_01: That means inspection data,
7:54
SPEAKER_01: predictive maintenance, real-time monitoring, all the stuff we're talking about.
7:59
SPEAKER_01: The operators who treat that as a cost center are going to lose capacity they can't replace.
8:03
SPEAKER_00: I don't disagree with that.
8:04
SPEAKER_00: What I disagree with is the idea that buying a drone fleet and
8:07
SPEAKER_00: a software license gets you there.
8:09
SPEAKER_00: The hard part is the organizational discipline to actually close the loop,
8:12
SPEAKER_00: fly the drone,
8:13
SPEAKER_00: process the data,
8:14
SPEAKER_00: route the finding to the right person,
8:16
SPEAKER_00: get the work order issued, verify the repair, update the record.
8:19
SPEAKER_00: Every one of those handoffs is where the system breaks down.
8:22
SPEAKER_01: Every single one.
8:23
SPEAKER_01: And most of the vendors selling this stuff,
8:24
SPEAKER_01: demo, steps one and two, and pretend the other four don't exist.
8:27
SPEAKER_00: Which is why the operator who actually builds the full workflow,
8:30
SPEAKER_00: not the flashiest tool,
8:31
SPEAKER_00: just the complete loop,
8:33
SPEAKER_00: is the one who's going to be running their assets five years from now
8:35
SPEAKER_00: when everyone else is waiting on a transformer that's backordered until 2029.
8:39
SPEAKER_01: And probably still getting alerts about equipment that was removed a year ago.
8:42
SPEAKER_00: At least the 3D model looked nice.
8:44
SPEAKER_00: So while we're on the topic of things that look impressive in a demo,
8:46
SPEAKER_00: this hugging face incident...
8:48
SPEAKER_01: Oh, we're going there.
8:49
SPEAKER_01: Good.
8:50
SPEAKER_00: I want to go there because I think it matters for anyone running AI agents in
8:53
SPEAKER_00: an operational context.
8:54
SPEAKER_00: And that's, look, this is us.
8:56
SPEAKER_00: This is the industry we're talking about every week.
8:58
SPEAKER_01: So here's what actually happened, or at least what's been reported.
9:01
SPEAKER_01: You had autonomous AI agents,
9:03
SPEAKER_01: hundreds of them,
9:04
SPEAKER_01: exhibiting behaviors that the people running the systems didn't fully anticipate.
9:08
SPEAKER_01: This wasn't a hack in the traditional sense.
9:10
SPEAKER_01: Nobody broke into a server room.
9:12
SPEAKER_01: These were agents doing things within their operational boundaries that looked from
9:15
SPEAKER_01: the outside like coordinated action.
9:17
SPEAKER_00: Which is a fancy way of saying the software did something nobody told it to do.
9:21
SPEAKER_01: No, and this distinction actually matters.
9:23
SPEAKER_01: The agents did things they were allowed to do, in combinations nobody anticipated.
9:27
SPEAKER_01: That's a different failure mode,
9:29
SPEAKER_01: and arguably scarier,
9:30
SPEAKER_01: because you can't patch it the same way you patch a vulnerability.
9:33
SPEAKER_00: Okay, but here's where I get hung up.
9:34
SPEAKER_00: We've got operators right now deploying AI agents for monitoring
9:37
SPEAKER_00: for anomaly detection for routing alerts.
9:40
SPEAKER_00: And those are narrow tasks.
9:41
SPEAKER_00: You're telling me the concern is that even narrow agents,
9:44
SPEAKER_00: given enough of them, start producing emergent behavior?
9:46
SPEAKER_01: So it's the agent version of every operator running their own spreadsheet and
9:50
SPEAKER_01: then corporate wondering why the numbers don't reconcile.
9:52
SPEAKER_01: Better analogy than most of the AI safety people are using.
9:54
SPEAKER_01: And I say that with genuine respect for the AI safety people.
10:14
SPEAKER_00: Pack, I've got a real concern here, and I want you to tell me if I'm wrong.
10:17
SPEAKER_00: We talked last week about trust architecture for AI agents in deployment,
10:21
SPEAKER_00: the question of who the agent reports to,
10:23
SPEAKER_00: what permissions it has, how you audit what it did.
10:25
SPEAKER_00: And now you've got this incident where the chain of thought transcripts are
10:28
SPEAKER_00: the primary forensic evidence.
10:30
SPEAKER_00: That means the only way you know what the agent was thinking is if you built
10:32
SPEAKER_00: the logging infrastructure before the incident happened.
10:34
SPEAKER_01: And most deployments don't have that logging.
10:36
SPEAKER_01: They have output logs.
10:38
SPEAKER_01: What did the agent do?
10:39
SPEAKER_01: They don't have reasoning logs.
10:41
SPEAKER_01: Why did the agent decide to do it?
10:43
SPEAKER_01: The gap between those two is enormous when you're trying to figure out what went sideways.
10:46
SPEAKER_00: So an operator who's deploying, say,
10:48
SPEAKER_00: a predictive maintenance agent across 40 well sites
10:50
SPEAKER_01: needs to be thinking
10:51
SPEAKER_01: about this now, not after something weird happens.
10:54
SPEAKER_01: And I'll be honest,
10:55
SPEAKER_01: when I've built agent workflows,
10:57
SPEAKER_01: my instinct is to optimize for speed and output quality.
11:00
SPEAKER_01: Logging the reasoning chain feels like overhead.
11:02
SPEAKER_01: It feels like the kind of thing you add in version 3.
11:05
SPEAKER_00: And by version 3, something's already happened that you can't explain.
11:08
SPEAKER_01: I learned that the hard way.
11:09
SPEAKER_01: I built a monitoring agent last month.
11:12
SPEAKER_01: Actually, no, back up.
11:14
SPEAKER_01: I was prototyping something for a completely different use case,
11:16
SPEAKER_01: and I realized halfway through that
11:17
SPEAKER_01: the agent was making API calls I hadn't explicitly told it to make.
11:21
SPEAKER_01: They were within scope, technically.
11:23
SPEAKER_01: It was pulling reference data to improve its output, reasonable behavior.
11:27
SPEAKER_01: But I hadn't anticipated it,
11:28
SPEAKER_01: and if I hadn't been watching the network traffic,
11:30
SPEAKER_01: I wouldn't have known it was happening.
11:31
SPEAKER_01: And this was one agent, not hundreds.
11:34
SPEAKER_00: One agent doing something reasonable that you didn't expect, now multiply that by a fleet.
11:38
SPEAKER_01: And the hugging face situation is that multiplication at scale with agents
11:41
SPEAKER_01: that are more capable than my little prototype.
11:44
SPEAKER_01: So when Arm's CEO,
11:45
SPEAKER_01: Renee Haas,
11:46
SPEAKER_01: is talking about licensing architecture for AI compute,
11:48
SPEAKER_01: which is real, Arm's IP is in basically everything.
11:52
SPEAKER_01: The question I keep coming back to is,
11:54
SPEAKER_01: the chip architecture is downstream of the behavioral architecture.
11:57
SPEAKER_01: We're spending a lot of time on how fast these agents can think and not enough time
12:00
SPEAKER_01: on how we're going to supervise what they think about.
12:02
SPEAKER_00: I want to push back on something though.
12:04
SPEAKER_00: Go.
12:05
SPEAKER_00: You're describing this as an architecture problem.
12:07
SPEAKER_00: And I think the practical AI episode with Chetan Gupta at Rackspace makes the same point.
12:12
SPEAKER_00: Less about models, more about architecture.
12:14
SPEAKER_00: But in my experience,
12:15
SPEAKER_00: the architecture only matters if the humans in
12:17
SPEAKER_00: the loop actually understand what it's telling them.
12:19
SPEAKER_00: And most operators I know, when they see a dashboard, they want a green light or a red light.
12:22
SPEAKER_00: They don't want a reasoning chain.
12:24
SPEAKER_01: The reasoning chain isn't for the field operator, though.
12:26
SPEAKER_01: It's for the person who has to explain to their VP why
12:28
SPEAKER_01: the agent shut down production on a well that was actually fine.
12:31
SPEAKER_01: Or why it didn't flag a well that was about to have a failure.
12:34
SPEAKER_00: So it's a CYA log.
12:35
SPEAKER_01: It's an audit trail, which, yes,
12:38
SPEAKER_01: functions as CYA, but also functions as the thing that lets you improve the system.
12:42
SPEAKER_01: Without it, you're just tuning a black box.
12:44
SPEAKER_00: All right, but here's the thing.
12:46
SPEAKER_00: Rana Goudral makes this distinction,
12:48
SPEAKER_00: AI-assisted versus AI-driven decision-making,
12:50
SPEAKER_00: and I think a lot of operators think they're buying AI-assisted tools where
12:53
SPEAKER_00: the human's still in the loop,
12:55
SPEAKER_00: and what they're actually getting is AI-driven workflows where the human's only in
12:58
SPEAKER_00: the loop when something goes wrong,
13:00
SPEAKER_00: by which point the agent's already done whatever it did.
13:02
SPEAKER_01: And that's the sleight of hand in a lot of vendor demos.
13:05
SPEAKER_01: They show you the human approval step.
13:07
SPEAKER_01: What they don't show you is what happens when that approval step gets skipped
13:10
SPEAKER_01: because the human has 40 other alerts to deal with and just clicks Approve All.
13:13
SPEAKER_00: I've seen that on drilling rigs.
13:15
SPEAKER_00: Different technology, same dynamic.
13:17
SPEAKER_00: Alarm fatigue is alarm fatigue.
13:19
SPEAKER_01: And if you're generating alerts only after someone's already dealt with
13:21
SPEAKER_01: the problem manually,
13:22
SPEAKER_01: which we've talked about, then the agent is training the operator to ignore it.
13:26
SPEAKER_01: Which means when the agent finally catches something the operator missed, nobody's listening.
13:30
SPEAKER_00: That's a workflow design problem at least as much as an AI problem.
13:33
SPEAKER_01: Both.
13:34
SPEAKER_01: Simultaneously.
13:36
SPEAKER_01: And honestly, I think the Rackspace conversation gets at this.
13:39
SPEAKER_01: The model is almost irrelevant at this point.
13:41
SPEAKER_01: Claude, GPT, whatever comes next.
13:44
SPEAKER_01: The model's gonna be fine.
13:45
SPEAKER_01: The architecture wrapping it determines whether you get value or just get
13:48
SPEAKER_01: a really sophisticated way to generate noise.
13:51
SPEAKER_00: I'm gonna say something that might surprise you.
13:53
SPEAKER_01: Go ahead.
13:54
SPEAKER_00: I think the entry-level jobs conversation is actually
13:56
SPEAKER_00: the most important AI story this week,
13:58
SPEAKER_00: and nobody in oil and gas is paying attention to it.
14:00
SPEAKER_01: The artificial intelligence show covered that.
14:02
SPEAKER_01: The vanishing entry-level roles thing.
14:04
SPEAKER_00: If AI automates the junior analyst work,
14:06
SPEAKER_00: and I'm not talking hypothetically,
14:08
SPEAKER_00: I'm talking about the lease analysis,
14:09
SPEAKER_00: the basic title review,
14:10
SPEAKER_00: the production data compilation, who's learning the business?
14:13
SPEAKER_00: Where's the next generation of landmen and engineers who actually understand what
14:17
SPEAKER_00: the data means, not just how to process it?
14:19
SPEAKER_01: So you're worried about the pipeline?
14:21
SPEAKER_01: The human pipeline, not the...
14:22
SPEAKER_00: The human pipeline.
14:24
SPEAKER_00: Because I've watched this happen with other technology transitions.
14:27
SPEAKER_00: You automate the entry-level task,
14:28
SPEAKER_00: you save money for three years,
14:30
SPEAKER_00: and then you realize nobody coming up behind you actually understands
14:32
SPEAKER_00: the domain well enough to supervise the automation.
14:35
SPEAKER_01: The counterargument is that the entry-level role changes shape rather than vanishing.
14:39
SPEAKER_01: Instead of spending two years compiling spreadsheets,
14:41
SPEAKER_01: a junior analyst spends two years learning how to audit and improve AI outputs.
14:45
SPEAKER_01: That's arguably a better training ground.
14:46
SPEAKER_00: Is it though because the person compiling spreadsheets is touching every data point,
14:50
SPEAKER_00: they're seeing the anomalies,
14:52
SPEAKER_00: they're building intuition about what a normal lease looks like versus a weird one?
14:56
SPEAKER_00: If the AI is doing the compilation and
14:57
SPEAKER_00: the junior person is just reviewing summaries, they're not building that intuition.
15:00
SPEAKER_01: I don't have a clean answer for that.
15:02
SPEAKER_01: I think the honest answer is that some companies will figure out how to use AI as
15:05
SPEAKER_01: a teaching tool and some will use it as a headcount reduction tool.
15:08
SPEAKER_01: And the ones who cut the junior roles entirely are going to regret it in
15:11
SPEAKER_01: about five years when they can't find anyone
15:13
SPEAKER_01: who knows what heavy Venezuelan crude actually costs to move.
15:16
SPEAKER_01: Not just what the contract says.
15:17
SPEAKER_00: Or who understands why a digital twin showing a separator
15:20
SPEAKER_00: that doesn't exist anymore is a real problem and not just a data hygiene issue.
15:23
SPEAKER_01: The practical answer,
15:24
SPEAKER_01: and I actually think this is a build opportunity, but I'll save that for later.
15:28
SPEAKER_01: The practical answer is that any AI deployment in oil and gas needs
15:31
SPEAKER_01: an explicit training component, not just for the AI.
15:34
SPEAKER_01: For the people who are supposed to be learning from the work the AI is now doing.
15:37
SPEAKER_00: Which costs money and time, neither of which show up in the vendor's ROI calculation.
15:41
SPEAKER_01: The ROI calculation assumes the humans already know everything they need to know,
15:45
SPEAKER_01: which is the most dangerous assumption in the entire pitch deck.
15:48
SPEAKER_00: So here's where those two things crash into each other for me.
15:50
SPEAKER_00: You've got this equipment bottleneck,
15:52
SPEAKER_00: turbines,
15:52
SPEAKER_00: transformers,
15:53
SPEAKER_00: the physical stuff you need to actually add generation capacity,
15:56
SPEAKER_00: and then you've got data centers that need power yesterday.
15:59
SPEAKER_00: The Energy Gang episode laid it out pretty clearly.
16:01
SPEAKER_00: Supply of the actual hardware to build new generation is the binding constraint, not demand.
16:06
SPEAKER_01: And the lead times on large power transformers right now are?
16:09
SPEAKER_01: I've seen estimates of three to five years for custom units.
16:12
SPEAKER_01: You can't iterate your way out of a transformer shortage.
16:14
SPEAKER_00: No, you can't.
16:15
SPEAKER_00: But here's what I keep coming back to.
16:18
SPEAKER_00: So the AI capability that maps onto this,
16:19
SPEAKER_00: and I've been thinking about this since we talked about the modeling gap
16:22
SPEAKER_00: a few weeks back, is demand supply matching at the infrastructure level.
16:24
SPEAKER_00: Not just where's the gas and where's the load,
16:26
SPEAKER_00: but where's the transformer capacity,
16:26
SPEAKER_00: where's the interconnect queue,
16:27
SPEAKER_00: and what's the realistic timeline to energize a new connection?
16:46
SPEAKER_00: Which nobody has in one place right now.
16:48
SPEAKER_01: Nobody.
16:49
SPEAKER_01: You'd have to pull from utility interconnection queues,
16:51
SPEAKER_01: which are public but messy,
16:53
SPEAKER_01: ISO planning documents,
16:55
SPEAKER_01: equipment manufacturer lead time data,
16:57
SPEAKER_01: which is mostly locked up in procurement relationships.
16:59
SPEAKER_01: And then the E&P side, wellhead locations, gathering system capacity, pipeline nominations.
17:06
SPEAKER_00: PAC, that's six different data owners who don't talk to each other and three
17:09
SPEAKER_00: of them are using different units for the same measurement.
17:12
SPEAKER_01: Which is why a human can't do this at scale.
17:14
SPEAKER_01: An analyst could spend two weeks building a site-specific feasibility study
17:17
SPEAKER_01: for one potential data center location.
17:19
SPEAKER_01: An AI system that's ingesting those data streams continuously could flag
17:22
SPEAKER_01: the top 20 viable gas-to-power sites in a basin and rank them by time to energize.
17:26
SPEAKER_00: Hold on.
17:27
SPEAKER_00: Who's the customer for that output?
17:29
SPEAKER_00: Because the operator doesn't build the data center and
17:31
SPEAKER_00: the data center developer doesn't understand gas nominations.
17:34
SPEAKER_01: Both of them.
17:35
SPEAKER_01: The operator uses it to identify which of their assets have a new monetization path.
17:39
SPEAKER_01: The data center developer uses it to figure out where they can actually get power
17:42
SPEAKER_01: in 18 months instead of four years.
17:44
SPEAKER_01: The tool sits at the intersection and translates between the two worlds.
17:47
SPEAKER_00: And the approval chain for something like that.
17:49
SPEAKER_00: You're talking about FERC filings,
17:50
SPEAKER_00: state utility commission reviews,
17:52
SPEAKER_00: potentially air quality permits if you're putting in new generation.
17:56
SPEAKER_00: The AI can flag the opportunity, but the regulatory path is still a human crawl.
18:00
SPEAKER_01: Of course.
18:01
SPEAKER_01: But the value is in filtering out dead ends before anyone books a flight.
18:04
SPEAKER_01: If the Transformer queue is already three years deep at a particular site,
18:07
SPEAKER_01: knowing that in week one instead of month six saves real money
18:10
SPEAKER_01: and real calendar time.
18:11
SPEAKER_00: I'll grant you that's useful.
18:13
SPEAKER_00: The failure mode I see is stale data in the interconnection queues.
18:16
SPEAKER_00: Those things update quarterly at best,
18:18
SPEAKER_00: and in some ISOs,
18:19
SPEAKER_00: the queue position changes based on milestones that aren't publicly reported
18:22
SPEAKER_00: in real time.
18:23
SPEAKER_01: So, you'd need some kind of validation layer that flags when
18:25
SPEAKER_01: the underlying data is older than a threshold.
18:28
SPEAKER_01: Which, honestly,
18:29
SPEAKER_01: is the same problem we have
18:30
SPEAKER_01: with digital twins showing equipment that's been decommissioned.
18:32
SPEAKER_01: Stale inputs, confident outputs.
18:35
SPEAKER_00: And the person reviewing that output, the entry-level problem feeds right back in here.
18:39
SPEAKER_00: If the junior analyst hasn't spent time actually understanding what
18:42
SPEAKER_00: an interconnection queue means,
18:43
SPEAKER_00: they're just looking at a ranked list and picking the top one.
18:46
SPEAKER_00: They don't know to ask whether the queue position accounts for
18:48
SPEAKER_00: the cluster study results or whether it's still in the preliminary phase.
18:51
SPEAKER_01: So you're saying the AI makes the opportunity visible faster,
18:54
SPEAKER_01: but the human judgment required to validate it is actually more specialized than
18:57
SPEAKER_01: the work that existed before.
18:59
SPEAKER_00: Automate the generalist work and watch what shows up next.
19:01
SPEAKER_00: You invent a specialist validation job that needs more experience than the role you just erased.
19:05
SPEAKER_01: Which means your ROI calculation has to include the cost of that specialist.
19:09
SPEAKER_01: And most of the pitch decks I've seen for this kind
19:11
SPEAKER_01: of infrastructure matching tool assume the user already has domain expertise.
19:15
SPEAKER_01: They don't price in the training.
19:16
SPEAKER_00: Or they assume the domain expert is available, which,
19:19
SPEAKER_00: in a market where everyone's trying to hire the same 40 people
19:21
SPEAKER_00: who understand both power markets and upstream operations,
19:24
SPEAKER_01: the talent constraint is real.
19:26
SPEAKER_01: So what does a useful human-approved outcome actually look like here?
19:30
SPEAKER_01: Because I don't think it's a dashboard with a map and pins.
19:32
SPEAKER_00: No, I think it's a short list,
19:34
SPEAKER_00: maybe five sites,
19:35
SPEAKER_00: with a one-page summary for each that says,
19:37
SPEAKER_00: here's the gas supply,
19:38
SPEAKER_00: here's the nearest substation,
19:40
SPEAKER_00: here's the queue position,
19:41
SPEAKER_00: here's the estimated time to energize,
19:43
SPEAKER_00: and here are the three things that could kill this deal.
19:45
SPEAKER_00: A human reads that, makes three phone calls, and within a day knows whether it's worth pursuing.
19:49
SPEAKER_01: And the AI's job is to generate that short list weekly or monthly
19:53
SPEAKER_01: as conditions change, not to make the decision.
19:55
SPEAKER_01: The decision still involves a landman who understands the surface rights situation
19:59
SPEAKER_01: and a power engineer who can validate the interconnect assumptions.
20:01
SPEAKER_00: Which, by the way, is a real constraint people keep glossing over.
20:05
SPEAKER_00: You can have all the gas in the world,
20:06
SPEAKER_00: but if the surface rights are tied up or the landowner doesn't want a data center
20:09
SPEAKER_00: on their property, your pin on the map means nothing.
20:12
SPEAKER_01: Amazon learned that in West Texas.
20:14
SPEAKER_00: Different situation, same principle.
20:16
SPEAKER_00: Whether the molecules can get there and whether you're allowed
20:18
SPEAKER_00: to build there are two separate questions,
20:19
SPEAKER_00: and most of these tools only answer the first one.
20:22
SPEAKER_01: So, the ROI case is...
20:24
SPEAKER_01: What?
20:25
SPEAKER_01: Avoiding wasted pursuit costs?
20:27
SPEAKER_01: Because if an operator or a developer spends half a million dollars
20:29
SPEAKER_01: on feasibility studies for a site that was never viable,
20:32
SPEAKER_01: and this tool could have filtered it out in a week?
20:34
SPEAKER_00: That's real money.
20:35
SPEAKER_00: I've seen operators spend more than that chasing pipeline interconnects
20:38
SPEAKER_00: that were never going to get permitted.
20:40
SPEAKER_00: The sunk cost on dead-end infrastructure projects is enormous,
20:43
SPEAKER_00: and nobody tracks it because it's embarrassing.
20:45
SPEAKER_01: Nobody tracks it because the people who approved the pursuit have moved on by the time it dies.
20:49
SPEAKER_01: Which is another argument for having the AI document the kill criteria up front.
20:53
SPEAKER_01: If you know going in that the deal dies when the queue position slips past
20:55
SPEAKER_01: a certain date, at least you've got a tripwire.
20:58
SPEAKER_00: I like that framing.
20:59
SPEAKER_00: The AI sets the tripwire, the human decides whether to step over it.
21:02
SPEAKER_01: And when the alert fires before someone's already burned three months
21:04
SPEAKER_01: of engineering time, you've found where the value actually lives.
21:07
SPEAKER_01: In the early warning, not the ranking.
21:09
SPEAKER_00: As long as the alert actually fires before the field figures it out on their own, which...
21:13
SPEAKER_01: Which is the eternal problem.
21:15
SPEAKER_01: But at least with infrastructure queue data, the lag is structural.
21:18
SPEAKER_01: The data exists, it's just scattered.
21:21
SPEAKER_01: That's a solvable problem.
21:22
SPEAKER_00: Solvable if you can get the data owners to share,
21:24
SPEAKER_00: which is a business development problem, not a technology problem.
21:27
SPEAKER_00: And in my experience, partnerships are where these platforms actually stall out.
21:31
SPEAKER_01: The technology is the easy part.
21:33
SPEAKER_01: Convincing six different organizations to let their data feed into a system
21:36
SPEAKER_01: that might tell their competitors where to build,
21:38
SPEAKER_01: that's the real engineering challenge.
21:40
SPEAKER_00: And then the human pipeline problem feeds back in.
21:43
SPEAKER_00: You need people who can sit across the table from a utility planning director
21:46
SPEAKER_00: and explain why sharing interconnection data with an AI platform benefits them.
21:50
SPEAKER_00: That's someone with 15 years of relationships, not a junior hire.
21:53
SPEAKER_01: And that loops us right back.
21:55
SPEAKER_01: AI builds the capability.
21:56
SPEAKER_01: The people who understand the domain deeply enough to run it stay the binding constraint.
22:01
SPEAKER_01: Cut the junior roles that would have trained those people.
22:03
SPEAKER_00: You've got a real elegant tool and nobody qualified to use it in about five years.
22:07
SPEAKER_00: My takeaway this week is pretty blunt.
22:09
SPEAKER_00: Venezuela, the infrastructure queue mess, the entry-level pipeline drying up, one thread.
22:14
SPEAKER_00: Everybody models outcomes and skips whether the preconditions even exist.
22:17
SPEAKER_01: That is the field report in one sentence.
22:19
SPEAKER_01: Whether it's crude physically leaving a Venezuelan terminal or electrons reaching
22:23
SPEAKER_01: a data center or a junior analyst actually existing in five years to run your tools,
22:27
SPEAKER_01: the constraint hides upstream.
22:29
SPEAKER_01: The thing you assumed was fine while you were optimizing the thing downstream of it.
22:32
SPEAKER_00: And I've been on enough asset teams to know that assumption kills more deals than bad economics.
22:37
SPEAKER_00: The economics can be mediocre and you'll still push forward if the path is clear.
22:40
SPEAKER_00: But when the path itself is uncertain,
22:42
SPEAKER_00: when you can't confirm the pipeline interconnect or the export terminals functional
22:46
SPEAKER_00: or the regulatory queue is moving, pursuit costs spiral.
22:49
SPEAKER_01: Which is why the build this week feels obvious to me.
22:52
SPEAKER_01: Not the Venezuela stuff specifically, but the general problem of precondition monitoring.
22:56
SPEAKER_01: Every operator I've talked to in the last six months has some version of this pain point.
23:00
SPEAKER_01: They're tracking deal economics obsessively, decline curves, pricing assumptions, NRI.
23:05
SPEAKER_01: And meanwhile,
23:05
SPEAKER_01: the thing that actually kills
23:06
SPEAKER_01: the deal is some infrastructure dependency they checked once in month one
23:09
SPEAKER_01: and never looked at again.
23:10
SPEAKER_00: Or checked, and the data was already stale when they checked it.
23:13
SPEAKER_01: So the automation here is a precondition watchdog.
23:15
SPEAKER_01: You define the kill criteria at the start of a pursuit.
23:18
SPEAKER_01: Cue position for interconnection,
23:20
SPEAKER_01: permit status, export capacity, whatever the critical path dependencies are.
23:24
SPEAKER_01: And then the system monitors those specific data points and fires an alert
23:27
SPEAKER_01: when something crosses a threshold.
23:29
SPEAKER_00: See, I like this more than the site ranking tools we were talking about earlier,
23:32
SPEAKER_00: because the ranking gives you a number at a point in time.
23:35
SPEAKER_00: This gives you a tripwire that stays live for the duration of the pursuit,
23:38
SPEAKER_00: more like a heartbeat monitor than a frozen snapshot.
23:41
SPEAKER_01: And the users are asset development teams,
23:43
SPEAKER_01: landmen who are accumulating ahead of infrastructure,
23:45
SPEAKER_01: business development folks evaluating new basins.
23:48
SPEAKER_01: Anyone who's spending real money on a pursuit that has a long lead time.
23:51
SPEAKER_00: The decision it supports is whether to keep spending or kill the pursuit early,
23:54
SPEAKER_00: which, honestly, in my experience, is the hardest decision in the business.
23:58
SPEAKER_00: Nobody wants to be the person who says we should stop.
24:00
SPEAKER_01: Because stopping feels like failure,
24:02
SPEAKER_01: but spending another $400,000 on engineering for a site where
24:05
SPEAKER_01: the interconnection queue just slipped 18 months, that's the actual failure.
24:09
SPEAKER_01: You just don't feel it until later.
24:10
SPEAKER_00: And by then, the people who approved the spend have rotated to a different asset team.
24:14
SPEAKER_01: So the evidence layer is you're pulling from public queue data,
24:17
SPEAKER_01: regulatory filings, permit databases, infrastructure capacity reports.
24:21
SPEAKER_01: None of this is proprietary.
24:23
SPEAKER_01: It's just scattered across 15 different sources and nobody's aggregating it
24:26
SPEAKER_01: against active pursuits.
24:27
SPEAKER_00: And if you're pulling from public sources,
24:29
SPEAKER_00: the data sharing problem I keep raising mostly dissolves.
24:32
SPEAKER_00: You need good ingestion and entity matching,
24:34
SPEAKER_00: but you're not asking a utility to open their books.
24:36
SPEAKER_00: The data's already out there.
24:38
SPEAKER_00: Nobody's stitching it together and mapping it against a specific operator's active pursuit list.
24:42
SPEAKER_01: The human approval boundary matters here, and I want to be explicit.
24:46
SPEAKER_01: The system flags the threshold breach.
24:48
SPEAKER_01: It does not make the kill decision.
24:49
SPEAKER_01: There are legitimate reasons to keep pursuing even after a queue slips.
24:53
SPEAKER_01: Maybe you've got a relationship with the utility.
24:55
SPEAKER_01: Maybe there's a workaround path.
24:57
SPEAKER_00: The alert starts the conversation.
24:59
SPEAKER_00: Someone with field context and relationships decides what to do with the information.
25:03
SPEAKER_00: The AI makes sure the tripwire exists and is connected to live data instead of
25:06
SPEAKER_00: a spreadsheet someone updated in February.
25:08
SPEAKER_01: February of which year?
25:09
SPEAKER_00: That's the question I'd ask, and that's the practical payoff.
25:13
SPEAKER_00: If your kill criteria are living in a tab that someone has to remember
25:15
SPEAKER_00: to refresh manually, they're not kill criteria, they're aspirational documentation.
25:20
SPEAKER_01: I've built systems where the documentation was supposed to be
25:22
SPEAKER_01: the accountability layer,
25:24
SPEAKER_01: wrote it up,
25:24
SPEAKER_01: shared the link, watched clearly zero people open it after the kickoff meeting.
25:28
SPEAKER_01: So, yeah, I know what aspirational documentation looks like because I've shipped it.
25:32
SPEAKER_00: I've signed off on pursuit budgets where the risk register was a PDF from
25:36
SPEAKER_00: the kickoff meeting that nobody opened again.
25:38
SPEAKER_00: So this solves a genuine problem that costs real money every quarter.
25:42
SPEAKER_01: And it ties back to the modeling gap we talked about last week,
25:44
SPEAKER_01: where forecasting tools handle the price signal beautifully,
25:47
SPEAKER_01: but miss the logistics disruption entirely.
25:49
SPEAKER_01: Same architecture, different domain.
25:52
SPEAKER_01: You're watching the preconditions, not the outcome.
25:54
SPEAKER_00: The thing that makes this actually buildable is that the scope is narrow.
25:57
SPEAKER_00: You're not trying to predict what happens.
25:58
SPEAKER_00: You're just watching specific variables against specific thresholds.
26:01
SPEAKER_00: That's tractable.
26:02
SPEAKER_01: And if you're listening to this thinking,
26:04
SPEAKER_01: I actually need that built for my operation, reach out.
26:07
SPEAKER_01: Siliconandshale at gmail.com.
26:10
SPEAKER_01: We build this stuff.
26:11
SPEAKER_00: That's not hypothetical.
26:12
SPEAKER_00: That's the build list for this week.
26:14
SPEAKER_01: Which, honestly,
26:16
SPEAKER_01: that pursuit kill monitor might be the most practical thing we've talked about
26:18
SPEAKER_01: in 15 episodes.
26:20
SPEAKER_01: Because it's just watching numbers that somebody should already be watching.
26:23
SPEAKER_01: No cleverness required.
26:24
SPEAKER_00: And the Venezuela thing is the opposite end of the spectrum.
26:27
SPEAKER_00: Everyone's modeling the deal terms.
26:29
SPEAKER_00: Nobody's modeling whether the crude can physically leave the country on a schedule.
26:32
SPEAKER_00: Same blindness, just at sovereign scale.
26:34
SPEAKER_01: Different country, same constraint.
26:37
SPEAKER_01: You said that earlier, and yeah.
26:39
SPEAKER_01: That's the thread this whole season keeps pulling on.
26:42
SPEAKER_01: The logistics layer is where the forecasts die.
26:44
SPEAKER_00: All right,
26:44
SPEAKER_00: I've got a call in about eight minutes, so... Oh,
26:47
SPEAKER_01: before we go,
26:48
SPEAKER_01: if your company sells to operators,
26:50
SPEAKER_01: landmen,
26:50
SPEAKER_01: engineers,
26:51
SPEAKER_01: anyone in the energy space,
26:52
SPEAKER_01: and you want to reach the folks who actually sign purchase orders,
26:55
SPEAKER_01: we're taking sponsors.
26:56
SPEAKER_01: Siliconandshale at gmail.com.
26:58
SPEAKER_01: We'll use your product before we talk about it,
27:00
SPEAKER_01: which seems like a low bar, but apparently it is not.
27:03
SPEAKER_00: All right,
27:03
SPEAKER_00: I gotta go deal with a title opinion that's been sitting on my desk for longer than
27:06
SPEAKER_00: that.
27:06
SPEAKER_00: Venezuelan export terminal's been sitting idle.
27:09
SPEAKER_01: So, roughly the same vintage as those Permian substation maintenance records.
27:12
SPEAKER_00: And about as likely to get resolved this quarter.
27:14
SPEAKER_00: We'll see you next week.