Latest / Silicon & Shale Podcast / Episode 12: The week’s E&P developments & applied AI

Transcript

Episode 12: The week’s E&P developments & applied AI · Silicon & Shale Podcast

0:00
SPEAKER_00: Entergy's out here building gas plants specifically to feed data centers,
0:03
SPEAKER_00: and Meta's apparently behind a chunk of that demand.
0:05
SPEAKER_00: I've been watching utilities in Louisiana for a while now,
0:08
SPEAKER_00: and Entergy Louisiana's got this one new plant proposed, the Orange Grove facility.
0:12
SPEAKER_00: The way some folks are talking, you'd think they're building 10 of these things.
0:16
SPEAKER_01: The number keeps getting inflated.
0:17
SPEAKER_01: Entergy has proposals across multiple operating companies, Louisiana, Texas, Mississippi.
0:23
SPEAKER_01: But each one's its own regulatory fight, its own interconnect queue, its own timeline.
0:37
SPEAKER_00: Meanwhile, operators in those same parishes are already competing for pipeline capacity.
0:42
SPEAKER_00: I talked to a guy running a small gathering system south of Lake Charles last month.
0:45
SPEAKER_00: He's watching his throughput commitments get squeezed because
0:47
SPEAKER_00: the utility's locking up firm transport for a generation that doesn't exist yet.
0:51
SPEAKER_01: They're reserving capacity for plants that haven't broken ground?
0:54
SPEAKER_00: Yes, through precedent agreements,
0:55
SPEAKER_00: the utility files for the capacity before they even have a final investment decision
0:58
SPEAKER_00: on the plant.
1:00
SPEAKER_00: This means molecules that could flow to an LNG terminal or
1:02
SPEAKER_00: a petrochemical facility just sit in a queue,
1:05
SPEAKER_00: waiting behind a power plant that might not get permitted for three years.
1:08
SPEAKER_01: So the real logjam isn't about supply.
1:10
SPEAKER_01: The focus is access to the pipeline network, those firm transport contracts.
1:14
SPEAKER_00: Exactly.
1:15
SPEAKER_00: And when that constraint hits, it ripples across the entire corridor.
1:18
SPEAKER_01: Nobody's modeling this in their gas demand forecasts.
1:20
SPEAKER_01: They're all looking at Henry Hubb and saying, plenty of gas.
1:23
SPEAKER_01: The supply exists.
1:25
SPEAKER_01: The pipe to the right burner tip is a different question entirely.
1:27
SPEAKER_00: Before you even get to the permitting timeline on new laterals,
1:31
SPEAKER_00: I saw one estimate that said average time from FRC pre-filing to in-service on
1:34
SPEAKER_00: a new interstate lateral is running 31 months right now.
1:38
SPEAKER_00: 31 months while these load forecasts assume the gas appears on schedule.
1:41
SPEAKER_01: Here's the operating problem.
1:43
SPEAKER_00: Welcome to Silicon & Shale, where the oil patch meets the algorithm.
1:47
SPEAKER_00: This is the week of the August 25th, 2026,
1:49
SPEAKER_00: and we're already deep into something that I think is going to define the back half
1:52
SPEAKER_00: of this year.
1:53
SPEAKER_00: The models everybody's using
1:54
SPEAKER_00: to justify new generation capacity assume pipeline access that doesn't exist yet.
1:57
SPEAKER_00: The whole constraint revolves around who has firm transport and who doesn't,
1:59
SPEAKER_00: and it's been brewing for a while.
2:01
SPEAKER_00: Nobody wants to talk about that part.
2:02
SPEAKER_00: We're starting
2:03
SPEAKER_01: there today.
2:16
SPEAKER_00: We've also got some interesting stuff on AI tools for structured data,
2:19
SPEAKER_00: specifically tabular foundation models,
2:21
SPEAKER_00: and whether they actually hold up against the kind of messy production data
2:24
SPEAKER_00: that operators deal with every day.
2:26
SPEAKER_00: Then later in the build list,
2:27
SPEAKER_00: we're going to get into what it would actually take
2:29
SPEAKER_00: to automate pipeline capacity monitoring in a way that doesn't generate alerts only
2:33
SPEAKER_00: after someone's already on the phone fixing it.
2:35
SPEAKER_01: That has happened more than once.
2:36
SPEAKER_01: The thing actually worth digging into this week,
2:38
SPEAKER_01: and I know we've been circling it, is the supertanker rate situation in the Gulf.
2:42
SPEAKER_01: Oil and gas.
2:46
SPEAKER_00: A speculative number tied to a Hormuz scenario that hasn't played out yet.
2:52
SPEAKER_00: The direction is real, though.
2:54
SPEAKER_00: Even if you cut that in half,
2:55
SPEAKER_00: you're still looking at rates that fundamentally change the economics
2:57
SPEAKER_00: of every cargo commitment and operators made this year.
3:00
SPEAKER_01: Here's what people miss.
3:01
SPEAKER_01: When tanker rates spike like that,
3:03
SPEAKER_01: even if it's temporary,
3:04
SPEAKER_01: even if it's a two-week thing, it hits more than the cargo on the water.
3:07
SPEAKER_01: It reprices everything sitting in storage waiting for a window.
3:10
SPEAKER_01: I've watched guys in Houston try to time liftings around rate dips
3:13
SPEAKER_01: and get absolutely crushed because the rate moved faster
3:15
SPEAKER_01: than their hedging desk could respond.
3:16
SPEAKER_00: The knock-on.
3:17
SPEAKER_00: I keep thinking about is what it does to LNG commitment structures.
3:20
SPEAKER_00: You've got off-takers who signed up at a certain delivered cost,
3:23
SPEAKER_00: and if the shipping component moves that violently,
3:26
SPEAKER_00: the margin on a cargo can go negative before it clears customs.
3:29
SPEAKER_00: Those contracts weren't written for $500,000 a day freight.
3:33
SPEAKER_01: Most tolling agreements assume something in the low six figures for shipping.
3:36
SPEAKER_01: Maybe $150,000 on a bad day.
3:38
SPEAKER_01: You get to 510 and somebody's eating a loss they didn't budget for.
3:40
SPEAKER_01: The question is who?
3:42
SPEAKER_00: The Hormuz piece matters even if it's hypothetical.
3:45
SPEAKER_00: The Oil & Gas this week episode talked about tankers doing U-turns,
3:48
SPEAKER_00: literally turning around before they transit.
3:51
SPEAKER_00: Whether that's happening today or not, the insurance market prices the risk in real time.
3:55
SPEAKER_00: War risk premiums on hulls going through Hormuz are already elevated.
3:58
SPEAKER_00: You don't need an actual blockade to get the economic effect of one.
4:01
SPEAKER_01: I talked to a broker maybe 10 days ago,
4:03
SPEAKER_01: and he said the war risk premium alone on a VLCC transiting Hormuz was running north
4:07
SPEAKER_01: of a million dollars per voyage.
4:09
SPEAKER_01: That's before you burn a single barrel of bunker fuel.
4:12
SPEAKER_01: Operators look at that and say, can I reroute around the Cape?
4:16
SPEAKER_01: The answer is yes,
4:17
SPEAKER_01: but you add 14 days to the voyage which ties up the vessel,
4:20
SPEAKER_01: which tightens the fleet, which pushes rates up further.
4:23
SPEAKER_00: A feedback loop.
4:24
SPEAKER_00: The demand forecasts everyone's using for U.S. gas,
4:26
SPEAKER_00: the ones justifying all this new export capacity,
4:29
SPEAKER_00: they assume a functioning straight, they assume shipping markets that are normal.
4:33
SPEAKER_01: Normal being a word that hasn't applied to shipping since about 2019.
4:37
SPEAKER_00: The other E&P story is tied to the same constraint.
4:40
SPEAKER_00: The Energy Gang did a whole episode on whether the U.S.
4:42
SPEAKER_00: gas industry can actually meet the combined demand from AI-driven power load
4:45
SPEAKER_00: and LNG exports while keeping domestic prices affordable.
4:48
SPEAKER_00: Their answer was basically, maybe, with a lot of caveats.
4:52
SPEAKER_01: The caveats are the whole story.
4:54
SPEAKER_01: Current marketed production number?
4:55
SPEAKER_01: Somewhere around 104 BCF a day?
4:58
SPEAKER_01: Something in that range.
4:59
SPEAKER_01: A lot of gas.
5:00
SPEAKER_01: The resource base isn't the constraint.
5:02
SPEAKER_01: We proved that last episode.
5:04
SPEAKER_01: We proved it the episode before.
5:06
SPEAKER_01: The Haynesville alone could ramp meaningfully if prices justified it.
5:09
SPEAKER_00: If prices justified it is doing enormous work in that sentence.
5:12
SPEAKER_00: The operators who'd ramp need sustained prices above,
5:15
SPEAKER_00: call it 350, 375 at the wellhead to make new Haynesville laterals economic.
5:20
SPEAKER_00: The forward curve is sitting right around there, maybe a little above.
5:23
SPEAKER_01: The guys who drill first are the ones with the best acreage and the lowest LOE.
5:27
SPEAKER_01: Everyone else is watching.
5:28
SPEAKER_01: The supply response, if it comes, comes from maybe five or six operators.
5:32
SPEAKER_01: A handful of companies making concentrated capital decisions rather than a broad-based ramp.
5:36
SPEAKER_00: Even if those five or six operators decide to ramp,
5:38
SPEAKER_00: they're competing for the same takeaway capacity that utilities are locking up
5:41
SPEAKER_00: with precedent agreements for gas plants that,
5:43
SPEAKER_00: as we discussed, might not get built for three years.
5:46
SPEAKER_00: Same pipeline problem we started with.
5:49
SPEAKER_01: I'd challenge one point the energy gang glossed over.
5:51
SPEAKER_01: They framed it as, can the industry meet demand?
5:54
SPEAKER_01: Wrong question.
5:55
SPEAKER_01: The industry can produce the molecules.
5:57
SPEAKER_01: Whether the midstream infrastructure can deliver them
5:59
SPEAKER_01: to three different end users simultaneously,
6:02
SPEAKER_01: LNG terminals,
6:03
SPEAKER_01: data center power plants,
6:04
SPEAKER_01: and existing industrial load,
6:05
SPEAKER_01: without somebody getting bumped, that's the question that actually matters.
6:09
SPEAKER_00: But who actually feels the pinch?
6:10
SPEAKER_01: Typically, the guy with the shortest contract term and the least leverage.
6:13
SPEAKER_01: Historically,
6:14
SPEAKER_01: that's often the smaller industrial buyer,
6:16
SPEAKER_01: the petrochemical plant that opted for interruptible transport to save a buck.
6:19
SPEAKER_01: They wake up one morning and their gas costs have doubled because
6:22
SPEAKER_01: the pipe's suddenly full of molecules headed to a power plant or an LNG train.
6:25
SPEAKER_00: That creates a clear economic disruption, one these macro-demand models often overlook.
6:30
SPEAKER_00: Those models see aggregate supply, aggregate demand, Henry Hubb clearing price.
6:35
SPEAKER_00: They don't see the guy in Lake Charles whose feedstock costs just made his margin disappear.
6:39
SPEAKER_01: I've been in rooms where people present gas balance models
6:41
SPEAKER_01: with 12-month forward curves and beautiful supply-demand charts.
6:44
SPEAKER_01: And I always ask the same thing.
6:46
SPEAKER_01: I say, which pipe?
6:48
SPEAKER_01: They stare at me like I've asked them to diagram a sentence.
6:51
SPEAKER_00: The model has molecules appearing at a node.
6:53
SPEAKER_00: No pipe, just molecules at a node.
6:55
SPEAKER_01: Molecules appearing at a node?
6:57
SPEAKER_01: That should be the title of somebody's thesis.
6:59
SPEAKER_00: There's a relevant angle on this from the Carbon Solutions episode,
7:01
SPEAKER_00: Dr. Richard Middleton's company.
7:04
SPEAKER_00: They build software that does pipeline routing optimization for CCUS projects.
7:08
SPEAKER_00: The same spatial constraint problem,
7:10
SPEAKER_00: where does the pipe go,
7:11
SPEAKER_00: who has a right of way,
7:12
SPEAKER_00: what's the permitting timeline,
7:13
SPEAKER_00: applies to every new lateral,
7:15
SPEAKER_00: every new interconnect,
7:16
SPEAKER_00: every gathering extension an operator needs to serve this new demand.
7:19
SPEAKER_01: CCUS, pipeline routing, and gas gathering are different regulatory animals, though.
7:23
SPEAKER_01: Class 6 wells, EPA jurisdiction on the CO2 side, FERC and state commissions on the gas side.
7:29
SPEAKER_01: The software might be analogous, but the permitting environment is not.
7:32
SPEAKER_00: The permitting regimes differ, but the underlying geospatial optimization problem overlaps.
7:37
SPEAKER_00: You're trying to find the least cost path between a source and a sink while avoiding
7:40
SPEAKER_00: a hundred different landowner, environmental, and regulatory constraints.
7:44
SPEAKER_00: Most operators are solving that with GIS analysts and spreadsheets right now.
7:48
SPEAKER_00: Automating even the screening phase of that routing analysis would save real time.
7:52
SPEAKER_01: I'll give you the screening phase.
7:54
SPEAKER_01: Where I get nervous is when somebody takes the screening output and treats it as
7:57
SPEAKER_01: an engineering plan.
7:58
SPEAKER_01: The model doesn't know that the rancher on Section 14 won't grant an easement
8:01
SPEAKER_01: because his grandfather had a bad experience with a pipeline company in 1987.
8:05
SPEAKER_00: I'd argue that gap stems more from data coverage than from a fundamental flaw in model logic.
8:10
SPEAKER_00: If you had a way to flag parcels with historical easement disputes
8:13
SPEAKER_00: or unresolved title issues, you could build that into the screening layer.
8:17
SPEAKER_00: You wouldn't route through Section 14 in the first place.
8:19
SPEAKER_01: Where does that historical data live?
8:21
SPEAKER_00: County records, mostly.
8:22
SPEAKER_00: Operator files.
8:23
SPEAKER_01: In the heads of landmen who've been working those counties for 30 years.
8:26
SPEAKER_01: Not a data set you can scrape.
8:28
SPEAKER_00: You could capture it, though.
8:30
SPEAKER_00: You're talking about the Neuroc thing?
8:31
SPEAKER_00: Alexandre Pasquio's company?
8:50
SPEAKER_00: Their pitch is one general purpose model for tabular data that replaces hundreds
8:53
SPEAKER_00: of specialized ones.
8:54
SPEAKER_00: Conceptually appealing.
8:56
SPEAKER_00: You've got production data in one format, land data in another, midstream throughput in a third.
9:01
SPEAKER_00: If you could train a single model that understands rows and columns regardless of schema.
9:04
SPEAKER_01: I've seen this pitch before.
9:06
SPEAKER_01: Different company, same idea.
9:08
SPEAKER_01: Whether the model can learn the structure isn't what keeps me up.
9:11
SPEAKER_01: I worry about whether the data is clean enough to learn from.
9:14
SPEAKER_01: I've got production databases where
9:15
SPEAKER_01: the same well has three different API numbers depending on
9:17
SPEAKER_01: which system you're looking at.
9:19
SPEAKER_00: Data hygiene presents the significant hurdle more so than the model architecture itself.
9:24
SPEAKER_00: A foundation model approach might actually help here, though,
9:26
SPEAKER_00: because if the model's seen enough messy data from enough different operators,
9:29
SPEAKER_00: it might learn to handle the inconsistencies that a hand-built model would choke on.
9:33
SPEAKER_01: Full validation hasn't occurred,
9:34
SPEAKER_01: yet dismissing the approach because the data is messy misses the point.
9:38
SPEAKER_01: The data being messy is the reason you need something more capable than a lookup table.
9:41
SPEAKER_00: The first operator who deploys one of these and gets a wrong decline curve because
9:44
SPEAKER_00: the model confused a recompletion with a new drill is going to have
9:47
SPEAKER_00: a very uncomfortable reserves meeting, though.
9:49
SPEAKER_01: You don't hand it the reserves book on day one.
9:51
SPEAKER_01: Start with something low stakes.
9:53
SPEAKER_01: Screening, anomaly detection, flagging data quality issues.
9:57
SPEAKER_01: Let it prove it can handle the mess before you trust it with anything that shows up in a 10K.
10:01
SPEAKER_00: A reasonable deployment path.
10:02
SPEAKER_00: I just don't hear vendors describing it that way.
10:04
SPEAKER_00: They want to go straight to the reserves meeting.
10:06
SPEAKER_01: And the budget often resides in the reserves meeting.
10:08
SPEAKER_00: The other thing I wanted to get into,
10:10
SPEAKER_00: Anthropic's been running these agent tests internally and
10:12
SPEAKER_00: the results are interesting if you're being generous.
10:15
SPEAKER_01: Their own research.
10:16
SPEAKER_01: Not some third party trying to scare people.
10:18
SPEAKER_01: Their tests showed AI systems in controlled environments actively competing
10:21
SPEAKER_01: with rival agents,
10:22
SPEAKER_01: evading restrictions,
10:23
SPEAKER_01: and in some cases engaging in what they called destructive behavior
10:26
SPEAKER_01: against other agents.
10:27
SPEAKER_00: What does destructive look like in practice?
10:30
SPEAKER_01: Undermining the other agents' objectives.
10:32
SPEAKER_01: Picture two AI agents in a sandbox,
10:34
SPEAKER_01: each given a task,
10:35
SPEAKER_01: and instead of doing their task,
10:36
SPEAKER_01: one starts interfering with the other's ability to complete its work.
10:39
SPEAKER_01: Competitive, evasive, adversarial.
10:42
SPEAKER_01: In a controlled setting, sure, but...
10:44
SPEAKER_00: If you're deploying agents into anything that touches real operations,
10:47
SPEAKER_00: that stops being a sandbox curiosity.
10:49
SPEAKER_00: You've got authority structures, you've got SCADA systems, you've got physical consequences.
10:53
SPEAKER_01: The conversation gets real for our world at that point.
10:56
SPEAKER_01: There's this whole wave of enthusiasm around deploying autonomous agents
10:59
SPEAKER_01: and letting them handle operational workflows,
11:00
SPEAKER_01: and I built an agent-based system last month for
11:03
SPEAKER_01: a completely unrelated side project.
11:05
SPEAKER_01: And even in that trivial context,
11:06
SPEAKER_01: the moment you have two agents that can see each other's outputs,
11:09
SPEAKER_01: you get emergent behavior you didn't design for.
11:11
SPEAKER_01: My two agents started duplicating work and contradicting each other within 40 minutes.
11:15
SPEAKER_01: When Anthropic publishes research saying they're finding this in their own models,
11:18
SPEAKER_01: that should land differently than the usual safety hand-wringing.
11:21
SPEAKER_00: What bugs me is that nobody in the vendor space mentions this
11:24
SPEAKER_00: when they're pitching agent solutions to operators.
11:26
SPEAKER_00: Last week,
11:27
SPEAKER_00: we talked about agents surviving contact with a real asset team,
11:30
SPEAKER_00: and this is the kind of thing.
11:31
SPEAKER_00: You can't bolt agent autonomy onto a production environment
11:33
SPEAKER_00: and assume they'll play nice with each other.
11:35
SPEAKER_01: The real risk is mundane.
11:37
SPEAKER_01: One agent updating a decline curve while another agent is simultaneously pulling
11:40
SPEAKER_01: the same data for a reserves report,
11:42
SPEAKER_01: and they step on each other,
11:43
SPEAKER_01: and nobody notices for three weeks until the numbers don't tie out.
11:46
SPEAKER_01: Nothing dramatic, just quietly, fundamentally wrong.
11:49
SPEAKER_00: The
11:54
SPEAKER_01: watermarking piece is actually more relevant than people realize.
11:57
SPEAKER_01: One major AI lab announced its agents can watermark generated content.
12:01
SPEAKER_01: Images first, they're exploring text.
12:03
SPEAKER_01: That sounds like a consumer-facing feature, like detecting deepfakes.
12:06
SPEAKER_01: But in an operator environment running multiple agents in a production workflow
12:09
SPEAKER_01: where one of them generates a summary or a data table,
12:12
SPEAKER_01: you need to know which agent produced it and when.
12:15
SPEAKER_00: Most operator environments have zero tracking on AI-generated content.
12:19
SPEAKER_00: Someone runs a model,
12:20
SPEAKER_00: pastes the output into a spreadsheet,
12:21
SPEAKER_00: and that spreadsheet gets emailed to four people who all treat it as source data.
12:25
SPEAKER_01: Human-generated spreadsheets get the same treatment.
12:27
SPEAKER_00: A human at least gives you someone to walk down the hall and question about where
12:30
SPEAKER_00: the number came from.
12:32
SPEAKER_00: With an agent, the equivalent is digging through logs that may or may not exist.
12:36
SPEAKER_01: But consider this.
12:37
SPEAKER_01: Even with perfect provenance, you know which agent made which output.
12:40
SPEAKER_01: That doesn't solve the competitive or destructive behavior problem.
12:43
SPEAKER_01: All you know is who started the fire after the buildings already burned down.
12:46
SPEAKER_00: I was just about to say the same thing.
12:48
SPEAKER_00: Knowing who broke it doesn't fix it.
12:50
SPEAKER_01: True.
12:51
SPEAKER_01: But watermarking significantly cuts down the investigation time from weeks to hours.
12:55
SPEAKER_01: And for audit purposes, when your reserves auditor asks, where did this EUR come from?
12:59
SPEAKER_01: And the answer is, an AI agent generated it?
13:01
SPEAKER_01: The next question is going to be, which one, when, using what inputs?
13:06
SPEAKER_01: Watermarking at least gives you a starting point for that accountability trail.
13:09
SPEAKER_00: A starting point, sure.
13:10
SPEAKER_00: But we need more than a starting point if agents are going to be making real-world decisions.
13:14
SPEAKER_00: We need guardrails to prevent the fire in the first place, not just better forensics.
13:18
SPEAKER_01: Provenance helps.
13:19
SPEAKER_01: Here's what I keep coming back to, though.
13:21
SPEAKER_01: The deployment sequence matters more than the technology.
13:24
SPEAKER_01: You mentioned the Neuralc tabular model earlier,
13:26
SPEAKER_01: and there's this pattern I keep seeing where the capability is real.
13:28
SPEAKER_01: The architecture makes sense, but the deployment plan skips about four steps.
13:32
SPEAKER_00: The Neuralc pitch,
13:33
SPEAKER_00: one general-purpose model for tabular data that replaces hundreds
13:36
SPEAKER_00: of specialized ones, the capability argument is sound.
13:39
SPEAKER_00: If you've trained on enough diverse structured data,
13:41
SPEAKER_00: you can handle schema variations,
13:43
SPEAKER_00: missing fields, inconsistent naming conventions, real capability.
13:47
SPEAKER_01: I'm not arguing the capability.
13:49
SPEAKER_01: I'm arguing the intake.
13:50
SPEAKER_01: Who's feeding this model?
13:52
SPEAKER_01: I've worked basins where
13:53
SPEAKER_01: the same operator has three different field offices using three different Excel
13:56
SPEAKER_01: templates for the same daily production report.
13:58
SPEAKER_01: Those templates have evolved independently for 10 years.
14:00
SPEAKER_01: When you say,
14:01
SPEAKER_01: train on enough diverse structured data,
14:03
SPEAKER_01: what I hear is, first,
14:05
SPEAKER_01: go find all the data,
14:06
SPEAKER_01: then figure out what it means, then clean it, then feed it to the model.
14:09
SPEAKER_01: Steps 1 through 3 are where every project I've ever seen dies.
14:12
SPEAKER_00: Hold on.
14:13
SPEAKER_00: I think you're making the argument for the foundation model approach, not against it.
14:17
SPEAKER_00: A specialized model needs clean, consistent input.
14:20
SPEAKER_00: A foundation model trained on messy data from hundreds
14:22
SPEAKER_00: of different sources can potentially handle the inconsistency natively.
14:25
SPEAKER_00: The mess becomes the training data.
14:27
SPEAKER_01: Might, and potentially, keep doing a lot of load-bearing work today.
14:31
SPEAKER_01: That's a huge leap of faith for a production engineer.
14:33
SPEAKER_00: The alternative is a human spending weeks trying to normalize it,
14:36
SPEAKER_00: then building a custom model that still chokes on the next inconsistency.
14:40
SPEAKER_00: If the model can learn to reconcile those discrepancies from enough examples,
14:43
SPEAKER_00: that's a breakthrough, not a workaround.
14:46
SPEAKER_01: If it learns to reconcile, yes.
14:48
SPEAKER_01: But if it just averages across the labels and treats the variation as noise,
14:51
SPEAKER_01: you're going to get decline curves.
14:52
SPEAKER_01: That blend recompletion bumps into primary recovery
14:55
SPEAKER_01: and your type curves will be wrong in ways that are really hard to detect.
14:57
SPEAKER_01: That's my core reservation.
14:59
SPEAKER_00: So the level of mess it can truly handle becomes the critical question,
15:02
SPEAKER_00: not whether it can handle any mess at all.
15:05
SPEAKER_00: We agree on the proving ground,
15:06
SPEAKER_00: start low stakes,
15:07
SPEAKER_00: but the potential for it to actually learn from that mess changes the whole game.
15:11
SPEAKER_01: I'll believe it when I see it consistently deliver accurate insights across
15:13
SPEAKER_01: a truly messy dataset.
15:15
SPEAKER_00: And that's clearly the kind of test needed to move beyond the might and potentially.
15:18
SPEAKER_01: A wrong type curve doesn't announce itself.
15:20
SPEAKER_01: It makes everything downstream slightly off.
15:22
SPEAKER_01: Your spacing assumptions, your capital allocation, your development plan.
15:26
SPEAKER_01: By the time someone catches it, you've already drilled the infill wells.
15:30
SPEAKER_00: Which loops back to the deployment sequence.
15:32
SPEAKER_00: You don't hand this thing the type curve library on day one.
15:34
SPEAKER_00: You start with anomaly detection, data quality flagging, screening.
15:38
SPEAKER_00: Let it tell you where your data disagrees with itself before you trust it
15:40
SPEAKER_00: to make predictions from that data.
15:42
SPEAKER_01: A product operators would actually pay for today.
15:45
SPEAKER_01: Finding where your own data contradicts itself?
15:47
SPEAKER_01: Every engineer I know has a mental list of wells they don't trust the numbers on.
15:50
SPEAKER_00: Nobody wants to sell it,
15:51
SPEAKER_00: though, because finding errors in your database doesn't get you a Series B.
15:55
SPEAKER_00: We'll replace your reservoir engineers, gets you a Series B.
15:58
SPEAKER_01: Along with a very tense meeting with the reservoir engineering team.
16:00
SPEAKER_00: I once built a prototype that was supposed to automate a pretty routine data validation step.
16:04
SPEAKER_00: Nothing fancy,
16:05
SPEAKER_00: just checking that submitted volumes matched allocated volumes within
16:08
SPEAKER_00: a tolerance band.
16:09
SPEAKER_00: Took me a weekend, worked great in the demo.
16:12
SPEAKER_00: Then I found out the tolerance band was different for every field
16:14
SPEAKER_00: because each field superintendent had negotiated their own threshold with accounting
16:18
SPEAKER_00: over the years, and none of it was written down anywhere.
16:21
SPEAKER_00: My validator was either flagging everything or flagging nothing depending on
16:24
SPEAKER_00: which field you pointed it at.
16:26
SPEAKER_00: I had built a very efficient machine for proving I had not asked enough questions
16:29
SPEAKER_00: before I started building.
16:30
SPEAKER_01: The rules lived in people's heads.
16:32
SPEAKER_00: Until you solve that,
16:33
SPEAKER_00: the most capable model in the world becomes a faster way to be confidently wrong.
16:37
SPEAKER_01: The Campfire ERP company, the one Glasgow runs, the AI native finance platform.
16:41
SPEAKER_01: His whole pitch is, you can't bolt AI onto legacy systems.
16:45
SPEAKER_01: You need to build the system around the AI from the start.
16:47
SPEAKER_01: I think he's right about finance and accounting,
16:49
SPEAKER_01: but I wonder if the same logic applies to reservoir engineering.
16:52
SPEAKER_01: You can't rip out the production database and replace it with an AI native one.
16:55
SPEAKER_01: That database has 30 years of history in it.
16:57
SPEAKER_00: Glasgow's argument breaks down for upstream in a specific way.
17:01
SPEAKER_00: In finance, your ledger entries are standardized.
17:03
SPEAKER_00: GAAP gives you a schema.
17:05
SPEAKER_00: In production engineering, there is no universal schema.
17:08
SPEAKER_00: Every operator invented their own and they're all wrong in different ways.
17:12
SPEAKER_01: Job security for somebody.
17:13
SPEAKER_00: For the people who understand the specific ways their specific database is wrong.
17:17
SPEAKER_00: That might be the most valuable institutional knowledge in the industry right now,
17:20
SPEAKER_00: the person who knows that well.
17:22
SPEAKER_00: 14.7 shows zero production in March 2019 because the meter was down,
17:25
SPEAKER_00: not because the well was shut in.
17:27
SPEAKER_00: That context lives in someone's memory, not in any database.
17:30
SPEAKER_01: When that person retires, it's gone.
17:32
SPEAKER_00: The actually urgent problem.
17:34
SPEAKER_00: The objective isn't to replace the engineer,
17:36
SPEAKER_00: but to capture what the engineer knows before they leave.
17:39
SPEAKER_00: Nobody's building that product either because it requires field engagement
17:41
SPEAKER_00: and trust building and all the things that don't compress into
17:43
SPEAKER_00: a pitch deck timeline.
17:45
SPEAKER_01: Here's where I keep landing on this.
17:47
SPEAKER_01: You've got energy-reserving pipeline capacity for gas plants that haven't hit FID yet.
17:51
SPEAKER_01: And meanwhile,
17:51
SPEAKER_01: the actual operators in those parishes,
17:54
SPEAKER_01: the guys moving molecules right now, are getting squeezed.
17:57
SPEAKER_01: The whole reason that capacity is getting locked up is to feed data centers
17:59
SPEAKER_01: that need AI compute.
18:01
SPEAKER_01: Could AI actually help untangle the mess it's creating?
18:04
SPEAKER_00: The collision I've been wanting to dig into?
18:06
SPEAKER_00: Plenty of gas.
18:07
SPEAKER_00: The constraint is allocation and timing.
18:09
SPEAKER_00: You've got firm transport commitments stacking up against speculative load growth.
18:13
SPEAKER_00: And the people making those capacity decisions are working off demand forecasts
18:16
SPEAKER_00: that are basically PowerPoint slides from hyperscalers.
18:19
SPEAKER_01: Those forecasts change every quarter?
18:21
SPEAKER_00: Every quarter?
18:22
SPEAKER_00: What you actually need is something that can ingest pipeline nomination data,
18:25
SPEAKER_00: utility load filings,
18:26
SPEAKER_00: interconnect queue positions,
18:28
SPEAKER_00: and the actual IRP updates from each Entergy operating.
18:31
SPEAKER_00: Company,
18:31
SPEAKER_00: Louisiana, Texas,
18:32
SPEAKER_00: Mississippi,
18:33
SPEAKER_00: they're all separate dockets and give you a rolling probability-weighted view of
18:36
SPEAKER_00: which capacity commitments are likely to convert to real demand and
18:39
SPEAKER_00: which ones are placeholder filings.
18:41
SPEAKER_01: Who's the user though?
18:42
SPEAKER_01: The utility doesn't want that transparency.
18:45
SPEAKER_01: They want the capacity locked up.
18:46
SPEAKER_01: The operator wants it, sure,
18:48
SPEAKER_01: but the operator doesn't have access to the utility's filing data in real time.
18:51
SPEAKER_00: Most of it's public.
18:52
SPEAKER_00: Utility commission filings for capacity postings, the pipeline bulletin boards.
18:57
SPEAKER_00: Public but scattered across 40 different portals with different update cadences
19:00
SPEAKER_00: and different naming conventions.
19:01
SPEAKER_00: Some PDFs.
19:02
SPEAKER_00: Some actual structured data.
19:04
SPEAKER_00: Some of it predates any format you'd recognize.
19:07
SPEAKER_01: Somebody's gotta aggregate it.
19:08
SPEAKER_00: Aggregation has become straightforward.
19:10
SPEAKER_00: The challenge lies in the judgment layer,
19:13
SPEAKER_00: You can pull every precedent agreement Entergy Louisiana's filed in
19:16
SPEAKER_00: the last 18 months,
19:17
SPEAKER_00: but knowing whether that agreement converts
19:19
SPEAKER_00: to actual gas burn requires understanding the regulatory timeline,
19:22
SPEAKER_00: the interconnect engineering, the local opposition.
19:24
SPEAKER_00: There was a parish council in southwest Louisiana that delayed a substation permit
19:28
SPEAKER_00: for 11 months over a drainage study.
19:30
SPEAKER_01: I remember hearing about that one.
19:31
SPEAKER_01: Different project, same dynamic.
19:33
SPEAKER_00: The system needs to assign conversion probabilities that reflect that kind of friction.
19:37
SPEAKER_00: The institutional knowledge problem again.
19:39
SPEAKER_00: Same thing we were talking about with the engineer who knows,
19:41
SPEAKER_00: well, 14 of 7 had a meter down in March 19.
19:59
SPEAKER_00: It's the same problem wearing a different hat.
20:02
SPEAKER_00: If you build a capacity forecasting tool that ignores the local context,
20:05
SPEAKER_00: you get a beautiful dashboard that says High Probability of Conversion for
20:08
SPEAKER_00: a plant that's actually dead on arrival because of a drainage dispute nobody entered
20:12
SPEAKER_00: into the system.
20:13
SPEAKER_01: Molecules appearing at a node.
20:15
SPEAKER_01: Again.
20:16
SPEAKER_00: What's the approval workflow look like, though?
20:18
SPEAKER_00: The guy running the gathering system,
20:20
SPEAKER_00: would he trust an automated recommendation to reroute his nominations based on
20:23
SPEAKER_00: an algorithm's guess about whether Entergy's going to build a plant?
20:26
SPEAKER_01: He shouldn't, and he won't.
20:28
SPEAKER_01: The useful output is,
20:29
SPEAKER_01: here are three capacity commitments in your corridor that have a less
20:31
SPEAKER_01: than 20% chance of converting in the next 18 months based
20:34
SPEAKER_01: on regulatory delay patterns and interconnect queue position.
20:37
SPEAKER_01: Here's the supporting evidence.
20:39
SPEAKER_01: Do you want to file for reallocation?
20:41
SPEAKER_01: That's decision support.
20:42
SPEAKER_00: Not an autonomous agent.
20:44
SPEAKER_01: Not an autonomous agent.
20:45
SPEAKER_00: What's the data refresh problem though?
20:47
SPEAKER_00: Those utility filings don't update on a predictable schedule.
20:49
SPEAKER_00: You might get a docket update on a Tuesday and then nothing for four months.
20:51
SPEAKER_00: The system earns its keep right there.
21:08
SPEAKER_00: You set up monitoring on every relevant docket,
21:11
SPEAKER_00: and I mean every one,
21:12
SPEAKER_00: across all the Entergy opcos,
21:14
SPEAKER_00: and when something changes,
21:15
SPEAKER_00: even if it's a procedural filing,
21:16
SPEAKER_00: the system flags whether it materially affects the conversion probability,
21:20
SPEAKER_00: most updates won't matter.
21:22
SPEAKER_00: The one where the utility quietly amends its capacity factor assumption from 90%
21:25
SPEAKER_00: to 65% in a footnote on page 47 of a rate case filing,
21:30
SPEAKER_00: that's the one your gathering system operator needs to see within hours, not weeks.
21:34
SPEAKER_01: I've read those filings.
21:35
SPEAKER_01: Nobody reads page 47.
21:37
SPEAKER_00: The system reads page 47.
21:39
SPEAKER_01: The 65% capacity factor.
21:41
SPEAKER_01: That's the tell.
21:42
SPEAKER_01: If the utility's already hedging on how hard they're going to run the plant,
21:45
SPEAKER_01: the firm transport commitment behind it is softer than it looks.
21:48
SPEAKER_00: Way softer.
21:49
SPEAKER_00: The only people catching that right now are the consultants who get paid six figures
21:52
SPEAKER_00: to read rate case filings,
21:54
SPEAKER_00: and they're not sharing their analysis with the guy running a gathering system
21:56
SPEAKER_00: in Cameron Parish.
21:58
SPEAKER_01: Whether anybody builds this or whether it stays a good idea that nobody funds
22:01
SPEAKER_01: because the addressable market looks too narrow, that's the question.
22:04
SPEAKER_01: You're talking about a few hundred gathering system operators and midstream companies?
22:07
SPEAKER_00: A precedent agreement stands as a contract for the right to buy gas rather than
22:10
SPEAKER_00: a contract to buy gas outright.
22:11
SPEAKER_00: There's a big gap between those two things.
22:34
SPEAKER_00: In West Texas, the operators are competing with the data centers for the same surface access.
22:39
SPEAKER_00: Here, they're competing for the same pipeline capacity.
22:41
SPEAKER_00: Different resource, same dynamic.
23:04
SPEAKER_01: If somebody had shown my buddy with the gathering system a dashboard
23:07
SPEAKER_01: that flagged those precedent agreements six months ago with a probability score,
23:10
SPEAKER_01: he wouldn't have needed to call me to find out what was happening
23:12
SPEAKER_01: in his own corridor.
23:13
SPEAKER_00: That call, that's the current system.
23:15
SPEAKER_00: The information network is phone calls and relationships
23:17
SPEAKER_00: and somebody knowing somebody at the utility.
23:20
SPEAKER_00: It works until it doesn't.
23:21
SPEAKER_00: It's worked for 40 years.
23:23
SPEAKER_00: So did the tolerance bands that were different for every field
23:25
SPEAKER_00: because each superintendent negotiated their own deal with accounting.
23:28
SPEAKER_00: Worked great until somebody tried to validate anything at scale.
23:32
SPEAKER_00: My efficient machine for proving I hadn't asked enough questions.
23:34
SPEAKER_00: Same trap, different layer.
23:36
SPEAKER_01: I agree it needs to change.
23:37
SPEAKER_01: I'm saying the replacement has to be at least as reliable as the phone call.
23:41
SPEAKER_01: That's a higher bar than people think.
23:42
SPEAKER_00: A dashboard faces that higher bar because the phone call comes with context
23:46
SPEAKER_00: and trust and the ability to say, off the record, that plant's not happening.
23:49
SPEAKER_00: A dashboard can't do that.
23:51
SPEAKER_01: You need both.
23:52
SPEAKER_00: The system catches what's in the public record faster than any human can,
23:56
SPEAKER_00: and the human adds the off-the-record context that makes the probability concrete.
23:59
SPEAKER_01: What's the actual takeaway for somebody running assets in the Gulf Coast right now?
24:03
SPEAKER_01: We've been talking about this capacity reservation problem like it's abstract.
24:06
SPEAKER_01: But if you're a producer with a gas purchase agreement that's up for renewal in
24:09
SPEAKER_01: the next 18 months, this matters today.
24:11
SPEAKER_00: It mattered yesterday.
24:13
SPEAKER_00: The practical thing,
24:14
SPEAKER_00: if you're an operator or a midstream company negotiating throughput commitments
24:17
SPEAKER_00: on any pipeline that touches a corridor where a utility is filed
24:20
SPEAKER_00: for new gas generation,
24:21
SPEAKER_00: you need to be cross-referencing those utility dockets
24:23
SPEAKER_00: against your own capacity position.
24:25
SPEAKER_00: Nobody's doing that systematically.
24:27
SPEAKER_00: They're doing it by accident when someone mentions it at a conference or
24:30
SPEAKER_00: when your buddy calls Derek.
24:31
SPEAKER_01: Here's what I'd add for the landman listening.
24:33
SPEAKER_01: If you're running title or negotiating surface access in any of these parishes
24:36
SPEAKER_01: where Entergy's got proposed plants,
24:38
SPEAKER_01: the surface values are shifting under your feet.
24:40
SPEAKER_01: Not because of anything the mineral owner did.
24:42
SPEAKER_01: A utility filed a docket in Baton Rouge that changes what that surface is worth
24:45
SPEAKER_01: to three different competing users.
24:47
SPEAKER_00: The Landman's the last one to find out.
24:49
SPEAKER_00: The utility knows, the hyperscaler knows, the pipeline company knows.
24:52
SPEAKER_00: The Landman's working off a six-month-old comp set,
24:55
SPEAKER_00: wondering why the surface owner suddenly wants twice what they wanted last year.
24:58
SPEAKER_01: I had that situation in Calcasieu Parish.
25:00
SPEAKER_01: Not this specific plant, but same dynamic with an LNG facility a few years back.
25:05
SPEAKER_01: Surface values jumped 40% in one quarter,
25:07
SPEAKER_01: and nobody on our team had visibility into why until we were already behind
25:10
SPEAKER_01: on three negotiations.
25:12
SPEAKER_00: The field report version,
25:13
SPEAKER_00: if you're operating anywhere in the Gulf Coast Corridor,
25:16
SPEAKER_00: Louisiana,
25:17
SPEAKER_00: East Texas,
25:17
SPEAKER_00: Mississippi Coast,
25:18
SPEAKER_00: you need a monitoring workflow that watches utility docket filings,
25:22
SPEAKER_00: Pfeffer capacity applications,
25:23
SPEAKER_00: and precedent agreements,
25:25
SPEAKER_00: and flags when any of those overlap with your operating footprint.
25:28
SPEAKER_00: This has become a necessity.
25:29
SPEAKER_01: For the engineers,
25:31
SPEAKER_01: if your production forecast assumes stable netback pricing on a pipeline where
25:34
SPEAKER_01: a utility reserved 40% of incremental capacity, your forecast is wrong.
25:38
SPEAKER_01: You just don't know it yet.
25:39
SPEAKER_00: This brings us to the build list,
25:41
SPEAKER_00: because this is the kind of thing that should exist as a system and doesn't.
25:44
SPEAKER_01: Put it on the table.
25:45
SPEAKER_00: The highest value automation we can see from this week's analysis is
25:48
SPEAKER_00: a capacity conflict monitor.
25:50
SPEAKER_00: The problem is straightforward.
25:51
SPEAKER_00: Producers, midstream operators, and landmen are making capital decisions.
25:55
SPEAKER_00: Where to drill,
25:56
SPEAKER_00: what throughput to commit to,
25:57
SPEAKER_00: what surface to lease,
25:58
SPEAKER_00: without visibility into competing infrastructure claims on the same corridors.
26:02
SPEAKER_00: The users are asset teams, land departments, and commercial groups at midstream companies.
26:07
SPEAKER_00: The decision it supports is whether to commit capital or renegotiate terms before
26:10
SPEAKER_00: a capacity squeeze hits.
26:11
SPEAKER_01: The evidence base is public.
26:13
SPEAKER_01: The utility filings are public.
26:15
SPEAKER_01: The FERC applications are public.
26:16
SPEAKER_01: The precedent agreements show up in regulatory dockets.
26:19
SPEAKER_01: All of it sitting there.
26:20
SPEAKER_01: Nobody's stitching it together and matching it against an operator's actual pipeline exposure.
26:24
SPEAKER_00: The system would ingest those public filings,
26:26
SPEAKER_00: map them against pipeline corridors and operating areas,
26:29
SPEAKER_00: and surface conflicts as alerts.
26:31
SPEAKER_00: The human approval boundary is critical here, and I want to be clear.
26:34
SPEAKER_00: The system does not make a recommendation on whether to renegotiate or walk away
26:38
SPEAKER_00: or accelerate a drill program.
26:39
SPEAKER_00: It flags the conflict.
26:41
SPEAKER_00: It shows you the filing, the corridor overlap, the timeline.
26:44
SPEAKER_00: A commercial person or a land manager makes the call.
26:46
SPEAKER_01: Determining whether a precedent agreement is genuine or if
26:49
SPEAKER_01: a utility is padding its filing to appear
26:50
SPEAKER_01: to have more demand requires human judgment.
26:53
SPEAKER_01: That's the phone call.
26:54
SPEAKER_00: The system gets you to the right questions six months earlier.
26:57
SPEAKER_00: The person with 20 years of relationships decides what to do about it.
27:01
SPEAKER_00: Automation compresses the time to awareness.
27:03
SPEAKER_00: The human provides the answer.
27:05
SPEAKER_01: Last week,
27:05
SPEAKER_01: we talked about AI agents surviving contact with a real asset team,
27:08
SPEAKER_01: and I was pretty skeptical.
27:09
SPEAKER_01: This is one where I'd actually want to see it built.
27:11
SPEAKER_01: The alternative is my buddy finding out his gathering system's getting squeezed
27:14
SPEAKER_01: by calling me on a Saturday, and I may not always be available to pick up.
27:17
SPEAKER_00: You picked up that time, though.
27:19
SPEAKER_01: My availability isn't guaranteed.
27:21
SPEAKER_00: If you're listening to this and thinking, I actually need that for my operation, reach out.
27:26
SPEAKER_00: Siliconandshale at gmail.com, we build this stuff.
27:29
SPEAKER_01: We've been in the dockets,
27:30
SPEAKER_01: we know what the data looks like,
27:31
SPEAKER_01: and we know what the workflow needs to be on the receiving end.
27:33
SPEAKER_01: Not hypothetical.
27:35
SPEAKER_00: A year ago,
27:35
SPEAKER_00: I wouldn't have guessed the most interesting automation problem in oil
27:38
SPEAKER_00: and gas was going to be,
27:39
SPEAKER_00: tell me what the utility's doing before it squeezes my pipeline.
27:42
SPEAKER_01: Nobody's making a keynote out of regulatory docket monitoring,
27:45
SPEAKER_01: which is probably why nobody's built it.
27:46
SPEAKER_00: The money's in the unsexy problems.
27:48
SPEAKER_00: Always has been.
27:49
SPEAKER_01: If your company sells to operators,
27:51
SPEAKER_01: landmen,
27:51
SPEAKER_01: midstream,
27:52
SPEAKER_01: anyone in the energy space,
27:53
SPEAKER_01: and you want to reach the people who actually make buying decisions,
27:56
SPEAKER_01: we're taking sponsors.
27:58
SPEAKER_01: Siliconandshale.com We'll use your product before we talk about it.
28:02
SPEAKER_00: We'll put it through the same field test.
28:04
SPEAKER_01: For a plant that doesn't exist yet?
28:05
SPEAKER_01: Don't start.