Latest / Silicon & Shale Podcast / Episode 13: The week’s E&P developments & applied AI

Transcript

Episode 13: The week’s E&P developments & applied AI · Silicon & Shale Podcast

0:00
SPEAKER_01: So diesel cracks hit $102 a barrel.
0:02
SPEAKER_00: Wait, per barrel?
0:03
SPEAKER_01: Not per- Per barrel, 102.
0:06
SPEAKER_01: And the Treasury Secretary apparently couldn't figure out why oil spiked.
0:09
SPEAKER_01: Which, look,
0:10
SPEAKER_01: I'm not going to editorialize on that too much,
0:12
SPEAKER_01: but if you're sitting on a refining margin that wide and the person in charge of
0:15
SPEAKER_01: the nation's finances is publicly confused about it-
0:17
SPEAKER_00: That's not a crack spread.
0:18
SPEAKER_00: That's a chasm.
0:19
SPEAKER_00: You're telling me the margin between crude input
0:21
SPEAKER_00: and refined diesel output is sitting at $102 and the official line is, huh, weird?
0:26
SPEAKER_01: That's what I'm telling you.
0:28
SPEAKER_01: And this is happening while Red Sea Transit is still a mess.
0:30
SPEAKER_01: Not Hormuz.
0:31
SPEAKER_01: I know everyone keeps saying Hormuz, but it's Bab el-Mandeb.
0:34
SPEAKER_01: It's the southern chokepoint.
0:35
SPEAKER_01: Ships are still rerouting around the Cape.
0:37
SPEAKER_01: That's been going on for, what, six months now?
0:40
SPEAKER_00: More than that.
0:41
SPEAKER_00: And that rerouting adds 10 to 14 days on a tanker voyage.
0:44
SPEAKER_00: So you're compressing refining capacity on one end and stretching the supply chain on the other.
0:48
SPEAKER_00: The fact that the crack spread is triple digits isn't mysterious.
0:50
SPEAKER_00: It's arithmetic.
0:51
SPEAKER_01: And here's the part that got me.
0:53
SPEAKER_01: Last week we were talking about pipeline constraints keeping gas from reaching power plants.
0:57
SPEAKER_01: Molecules exist but can't get where they need to go.
1:00
SPEAKER_01: This is the same story playing out on water instead of steel.
1:03
SPEAKER_01: Diesel's there, the refining margin says the world desperately wants it.
1:06
SPEAKER_01: But the logistics layer between barrel and buyer is broken in a way
1:09
SPEAKER_01: that keeps widening the spread.
1:11
SPEAKER_00: So, what does a $102 crack actually mean for an operator in,
1:14
SPEAKER_00: say, the Permian who's looking at their condensate split?
1:17
SPEAKER_01: It means your light suite is suddenly a lot more interesting to
1:19
SPEAKER_01: a refiner who's short feedstock because their usual cargo is taking
1:22
SPEAKER_01: the scenic route around Africa.
1:24
SPEAKER_01: You might actually see some of those condensate splitters
1:26
SPEAKER_01: that never quite penciled out start getting dusted off.
1:29
SPEAKER_01: Not tomorrow.
1:30
SPEAKER_01: But people are running the numbers again.
1:32
SPEAKER_00: Which is funny because six months ago the line was condensates a stranded molecule.
1:36
SPEAKER_00: Now it's the molecule everyone's chasing.
1:38
SPEAKER_01: In this business every stranded molecule is about 18 months
1:40
SPEAKER_01: from being someone's strategic priority.
1:43
SPEAKER_01: Give or take a shipping lane.
1:45
SPEAKER_01: Anyway, welcome to Silicon and Shale where the oil patch meets the algorithm.
1:49
SPEAKER_01: Week 36, 2026.
1:51
SPEAKER_01: And I think we've got a good one today because
1:52
SPEAKER_01: that crack spread conversation actually feeds right into something I've been wanting
1:56
SPEAKER_01: to dig into,
1:56
SPEAKER_01: which is what happens when the refining layer gets this dislocated and
1:59
SPEAKER_01: the people building supply models downstream just don't account for it.
2:03
SPEAKER_00: We're going to get into that.
2:04
SPEAKER_00: The modeling gap,
2:05
SPEAKER_00: specifically how forecasting tools handle logistics disruptions versus price
2:08
SPEAKER_00: signals.
2:09
SPEAKER_00: Because those are two very different inputs and most systems treat them like they're
2:12
SPEAKER_00: the same thing.
2:13
SPEAKER_00: And then there's this whole thread on AI agent security that I think operators need
2:16
SPEAKER_00: to hear about,
2:17
SPEAKER_00: because the trust architecture for these systems is basically non-existent
2:20
SPEAKER_00: in most deployments right now.
2:22
SPEAKER_00: The permissions layer alone is a mess.
2:24
SPEAKER_01: And we'll hit the data center ban question because
2:26
SPEAKER_01: a couple states are now actively pushing back on new builds,
2:28
SPEAKER_01: and that has implications for where compute lands relative to where
2:31
SPEAKER_01: the energy actually is.
2:32
SPEAKER_00: Which ties back to everything we were talking about last week with pipeline constraints.
2:36
SPEAKER_00: The molecules can't get their problem doesn't stop at power plants.
2:39
SPEAKER_00: It extends to whether your AI infrastructure can even sit where the cheap electrons are.
2:42
SPEAKER_00: But we'll get there.
2:44
SPEAKER_01: So the Phillips 66 stuff,
2:45
SPEAKER_01: Sean Marr was on crude realities this week,
2:47
SPEAKER_01: and the thing that caught me wasn't the macro overview,
2:50
SPEAKER_01: it was how he framed refined products as the actual bottleneck.
2:53
SPEAKER_01: Not crude supply, not even demand.
2:55
SPEAKER_01: The refining layer.
2:56
SPEAKER_00: Which makes sense given what we just talked about with the crack spread.
2:59
SPEAKER_00: Phillips 66 is sitting on...
3:01
SPEAKER_00: What are they?
3:02
SPEAKER_00: The third largest refiner in the U.S. by capacity?
3:04
SPEAKER_01: In that range.
3:05
SPEAKER_01: And Meher's point the way I heard it was
3:07
SPEAKER_01: that American refining infrastructure is globally connected in ways
3:10
SPEAKER_01: that people outside the industry don't think about.
3:13
SPEAKER_01: When you lose transit through Bob El Mondeb,
3:15
SPEAKER_01: it doesn't just affect the tankers rerouting around the Cape.
3:17
SPEAKER_01: It reshuffles where every barrel of refined product lands.
3:21
SPEAKER_01: European buyers who were getting diesel
3:22
SPEAKER_01: from Middle Eastern refineries are now competing for Gulf Coast barrels.
3:25
SPEAKER_00: And that competition is what's inflating the crack spread to 102 because
3:29
SPEAKER_00: the demand signal hitting U.S. refiners isn't just domestic.
3:32
SPEAKER_00: It's everyone who lost their usual supply route suddenly bidding on the same molecules.
3:36
SPEAKER_01: Here's where it gets interesting for operators.
3:38
SPEAKER_01: If you're a Permian producer looking at this,
3:40
SPEAKER_01: your crude isn't just feedstock for Corpus Christi refineries anymore.
3:43
SPEAKER_01: The export economic shift when European and Asian buyers are scrambling.
3:47
SPEAKER_01: Your light sweet barrel that was maybe marginally economic
3:49
SPEAKER_01: to export six months ago is now, it's a different calculation entirely.
3:53
SPEAKER_00: So what does that actually look like for,
3:55
SPEAKER_00: say, a mid-cap operator running a couple hundred wells in the Delaware Basin?
3:58
SPEAKER_00: Are they seeing this in their netbacks yet?
4:00
SPEAKER_00: Or is there a lag?
4:01
SPEAKER_01: There's always a lag.
4:03
SPEAKER_01: The crack spread moves in real time, but your marketing contracts don't.
4:06
SPEAKER_01: Most of these guys are locked into 30-, 60-, 90-day pricing windows.
4:10
SPEAKER_01: So the guy who locked in July pricing is sitting there watching August blow out
4:13
SPEAKER_01: and he can't touch it.
4:15
SPEAKER_00: Which is the kind of thing you'd want a model to flag.
4:17
SPEAKER_00: If you had
4:17
SPEAKER_00: a system watching transit disruption data alongside your contract windows,
4:20
SPEAKER_00: you could at least...
4:21
SPEAKER_01: Hold on,
4:21
SPEAKER_01: I see where you're headed,
4:22
SPEAKER_01: and I'm not against it in principle, but the inputs you'd need are insane.
4:26
SPEAKER_01: You're talking about real-time shipping route data,
4:29
SPEAKER_01: refinery utilization rates by facility,
4:31
SPEAKER_01: contract expiration schedules across your entire marketing book,
4:34
SPEAKER_01: and then you need someone to tell the model what counts as
4:36
SPEAKER_01: a disruption versus normal variance in transit times.
4:39
SPEAKER_00: The complexity is real.
4:40
SPEAKER_00: I'm not saying you build that tomorrow.
4:42
SPEAKER_00: But the fact that the lag exists...
4:44
SPEAKER_00: The fact that an operator can watch a $102 crack spread on a screen and not be able
4:48
SPEAKER_00: to act on it because their contract structure doesn't allow it?
4:51
SPEAKER_00: That's a process problem, not just a market problem.
4:53
SPEAKER_01: It's a relationship problem.
4:55
SPEAKER_01: Your marketer isn't going to let you renegotiate a pricing window because an AI told you to.
4:59
SPEAKER_01: They've got their own book to manage.
5:00
SPEAKER_00: But you could at least shorten the information gap.
5:03
SPEAKER_00: Even if you can't act on it instantly,
5:04
SPEAKER_00: knowing three days earlier that the spread is widening gives you leverage in
5:07
SPEAKER_00: the next conversation with your marketer.
5:09
SPEAKER_00: That's not replacing the relationship.
5:11
SPEAKER_00: It's arming the person in the relationship with better timing.
5:14
SPEAKER_01: The timing piece does matter.
5:16
SPEAKER_01: I've seen guys miss a $2 swing on 100,000 barrels because
5:18
SPEAKER_01: the report they were looking at was from Tuesday and the market moved Wednesday.
5:22
SPEAKER_01: That's real money left on the table because the data was 48 hours stale.
5:26
SPEAKER_00: $2 on 100,000 barrels is 200 grand.
5:28
SPEAKER_00: That's not a rounding error.
5:30
SPEAKER_01: No, it is not.
5:31
SPEAKER_00: The other story I wanted to get into,
5:33
SPEAKER_00: the insurer piece,
5:34
SPEAKER_00: The Energy Gang had this whole segment
5:36
SPEAKER_00: on how insurance companies are now basically co-designing renewable energy
5:39
SPEAKER_00: and storage facilities for resilience.
5:42
SPEAKER_00: Hailstorms, wildfires, the physical threat layer.
5:44
SPEAKER_01: Which, look,
5:45
SPEAKER_01: the renewables side can have that conversation,
5:47
SPEAKER_01: but what caught my attention is the upstream implication.
5:50
SPEAKER_01: If insurers are getting that hands-on with facility design for solar and storage,
5:54
SPEAKER_01: how long before they start doing the same thing with well pads and tank batteries?
5:57
SPEAKER_00: You think they're not already?
5:59
SPEAKER_00: The insurance conversation in upstream has always been there, but it's been reactive.
6:02
SPEAKER_00: You build it.
6:03
SPEAKER_00: You insure it.
6:04
SPEAKER_00: You file a claim when something goes wrong.
6:06
SPEAKER_01: And that's how the whole industry approaches it.
6:09
SPEAKER_01: Now you've got the renewable side where the insurer is sitting at the table
6:11
SPEAKER_01: during engineering saying,
6:13
SPEAKER_01: if you orient the panels this way,
6:14
SPEAKER_01: your hail exposure drops and your premium drops with it.
6:17
SPEAKER_01: That's a fundamentally different relationship.
6:20
SPEAKER_00: So why isn't that happening in ENP?
6:22
SPEAKER_00: Is it a complexity problem?
6:24
SPEAKER_00: A solar farm is relatively standardized compared to a well pad in the Permian,
6:27
SPEAKER_00: where every location has different terrain, different proximity to population.
6:30
SPEAKER_01: Part of it is complexity.
6:32
SPEAKER_01: But upstream operators have historically viewed insurance as a cost center, not a design input.
6:37
SPEAKER_01: You don't call your insurer before you set a tank battery, you call your foreman.
6:41
SPEAKER_00: But if the insurer could model site-specific risk, fire, flood,
6:44
SPEAKER_00: seismicity,
6:45
SPEAKER_00: whatever,
6:46
SPEAKER_00: and tie that to premium reduction at the design stage,
6:48
SPEAKER_00: wouldn't that change the math for an operator?
6:50
SPEAKER_00: Especially someone running 50, 60 locations?
6:53
SPEAKER_01: It would change the math if the insurer actually understood upstream operations.
6:57
SPEAKER_01: In my experience, at that point it falls apart.
7:00
SPEAKER_01: I've had adjusters show up to a location and not know the difference between
7:02
SPEAKER_01: a separator and a heater-treater.
7:04
SPEAKER_01: You're going to tell me that person is going to co-design my facility?
7:07
SPEAKER_00: The gap is real.
7:08
SPEAKER_00: But that's exactly what the renewables side solved,
7:11
SPEAKER_00: because the assets are more standardized and the data is cleaner.
7:14
SPEAKER_00: A solar panel has a spec sheet.
7:16
SPEAKER_00: A well pad is a collection of decisions made by six different people
7:18
SPEAKER_00: over three months, half of which aren't documented anywhere.
7:21
SPEAKER_01: I've seen well pads where the AS built doesn't match the permit,
7:24
SPEAKER_01: the permit doesn't match the survey,
7:26
SPEAKER_01: and the survey was done by a guy who retired in 2019.
7:29
SPEAKER_01: You want to build an insurance model on top of that.
7:32
SPEAKER_00: I want to build a data layer underneath it so that five years from now,
7:34
SPEAKER_00: someone can build an insurance model on top of it.
7:36
SPEAKER_00: You've got to start somewhere.
7:38
SPEAKER_01: That's the most patient thing you've ever said on the show.
7:40
SPEAKER_00: Don't get used to it.
7:41
SPEAKER_01: The other piece of the insurer story that I think matters,
7:44
SPEAKER_01: and this connects to something we touched on a few weeks ago
7:46
SPEAKER_01: about surface rights getting complicated with these Amazon builds in West Texas,
7:49
SPEAKER_01: is that when an insurer starts dictating facility design,
7:52
SPEAKER_01: they're effectively adding another layer of constraint on where and how you develop.
7:56
SPEAKER_01: If your insurance company says we won't cover a battery within 500 feet of
7:59
SPEAKER_01: a residence at a reasonable premium,
8:01
SPEAKER_01: that's a development constraint that looks a lot like a regulatory constraint,
8:04
SPEAKER_01: except it's coming from the private sector.
8:06
SPEAKER_00: And it's harder to fight because there's no comment period.
8:08
SPEAKER_00: There's no public hearing.
8:10
SPEAKER_00: Your insurer just sends you a rate sheet, and the economics either work or they don't.
8:13
SPEAKER_01: Which is why I think this trend is worth watching, even if it starts on the renewable side.
8:17
SPEAKER_01: Because the moment it crosses into upstream,
8:19
SPEAKER_01: it changes the land game,
8:20
SPEAKER_01: it changes where you put infrastructure,
8:22
SPEAKER_01: and it does it without anyone passing a law.
8:24
SPEAKER_00: And if you layer that on top of the state-level data center pushback we're going
8:27
SPEAKER_00: to talk about later.
8:28
SPEAKER_00: You've got private insurers constraining physical infrastructure on one side
8:31
SPEAKER_00: and state governments constraining compute infrastructure on the other.
8:34
SPEAKER_00: The map of where you can actually build things is getting smaller from both directions.
8:38
SPEAKER_01: What's worth noting is the framing this time is different from their earlier pauses.
8:40
SPEAKER_01: This isn't, we're worried the model is too smart.
8:42
SPEAKER_01: It's, we're worried about what the model can do to systems.
8:44
SPEAKER_01: Much more concrete problem.
9:04
SPEAKER_01: And I bring it up because I think people in our world hear OpenAI Pause Training
9:07
SPEAKER_01: and they either think, see,
9:08
SPEAKER_01: AI is dangerous,
9:09
SPEAKER_01: or they think, who cares, I'm trying to get my production reports to match.
9:13
SPEAKER_01: But there's a middle ground that actually matters for operators.
9:16
SPEAKER_00: The security surface.
9:18
SPEAKER_00: There's this episode I was listening to with Noam Schwartz.
9:20
SPEAKER_00: He's the CEO of a company called Alice,
9:22
SPEAKER_00: and he was breaking down where the actual vulnerabilities live
9:24
SPEAKER_00: in enterprise AI deployments.
9:26
SPEAKER_00: The risk isn't the model itself.
9:28
SPEAKER_00: It's the prompts,
9:28
SPEAKER_00: the tools the model has access to,
9:30
SPEAKER_00: the memory, the permissions, and especially agent-to-agent interactions.
9:34
SPEAKER_00: Every one of those is a door.
9:36
SPEAKER_01: Agent-to-agent meaning what fare?
9:37
SPEAKER_01: Two bots talking to each other?
9:39
SPEAKER_00: In functional terms, yes.
9:41
SPEAKER_00: Imagine you've got one agent that monitors well production and another
9:43
SPEAKER_00: that handles procurement alerts.
9:45
SPEAKER_00: If they can pass data to each other,
9:47
SPEAKER_00: and in a lot of these frameworks they can,
9:49
SPEAKER_00: then the security question isn't just, can an attacker trick the production agent?
9:52
SPEAKER_00: It's, can they trick the production agent into telling the procurement agent
9:55
SPEAKER_00: to do something it shouldn't?
9:57
SPEAKER_01: That's a chain of custody problem.
9:59
SPEAKER_00: A chain of custody problem that most people building these systems haven't even scoped yet.
10:03
SPEAKER_00: And I say that as someone who... Look,
10:05
SPEAKER_00: I built a prototype last quarter where I had two agents coordinating on lease data,
10:08
SPEAKER_00: and it didn't even occur to me to think about what happens if one of them gets
10:11
SPEAKER_00: a malformed input and passes bad instructions downstream.
10:14
SPEAKER_00: I just assumed the second agent would reject garbage.
10:16
SPEAKER_00: And?
10:17
SPEAKER_00: It did not reject garbage.
10:18
SPEAKER_00: It acted on the garbage with great enthusiasm.
10:20
SPEAKER_01: How much damage did it do?
10:22
SPEAKER_00: It tried to flag 11 leases as expired that weren't.
10:25
SPEAKER_00: Caught it in testing, thank God.
10:27
SPEAKER_00: But the point is, the security model for these systems isn't put a password on it.
10:31
SPEAKER_00: It's fundamentally different from anything we've dealt with in SCADA or traditional IT.
10:36
SPEAKER_01: And this is where I get skeptical about the pace of deployment.
10:38
SPEAKER_01: Because you've got these open standards coming out.
10:40
SPEAKER_01: What were they, MCP, A2A?
10:43
SPEAKER_00: Multi-agent communications protocol and agent-to-agent protocol.
10:46
SPEAKER_00: The Agentic AI Foundation is pushing these as the way to make agents interoperable at scale.
10:50
SPEAKER_00: And conceptually, I'm on board.
10:52
SPEAKER_00: You need standards.
10:53
SPEAKER_00: You can't have every vendor building their own handshake.
10:56
SPEAKER_01: But standards assume everyone's implementing them the same way, and they won't.
11:00
SPEAKER_01: So in the meantime,
11:00
SPEAKER_01: you've got operators who are being told,
11:02
SPEAKER_01: deploy agents,
11:03
SPEAKER_01: they'll handle your workflows,
11:04
SPEAKER_01: and the security framework underneath is still being written by committee.
11:07
SPEAKER_00: I'd push back on the implication that we should wait, though.
11:10
SPEAKER_00: Because the alternative is what we have now.
11:12
SPEAKER_00: People copy-pasting data between three different systems and emailing Excel files
11:15
SPEAKER_00: with the subject line, Final V3 Real Final.
11:18
SPEAKER_01: I'm not saying wait.
11:19
SPEAKER_01: I'm saying know what you're deploying.
11:21
SPEAKER_01: Because right now there's a gap between what the agent demos show and what
11:23
SPEAKER_01: the agent actually has permission to touch in a production environment.
11:26
SPEAKER_01: And if you skip that scoping step...
11:28
SPEAKER_00: You get my 11 phantom expired leases.
11:31
SPEAKER_01: Or worse,
11:32
SPEAKER_01: you get an agent that has right access to something it shouldn't because
11:34
SPEAKER_01: the IT team provisioned it the same way they'd provision a new employee
11:37
SPEAKER_01: and nobody thought about whether a software agent should have the same permissions
11:40
SPEAKER_01: as a person.
11:41
SPEAKER_00: That is the core issue.
11:42
SPEAKER_00: The trust gap.
11:44
SPEAKER_00: So the guardrail has to be in the permissions, not in the agent's common sense.
11:47
SPEAKER_00: On the lease prototype?
12:11
SPEAKER_00: Three times as long.
12:12
SPEAKER_00: I spent a week building the thing
12:13
SPEAKER_00: and three weeks figuring out what it shouldn't be allowed to do.
12:16
SPEAKER_00: For a prototype, that's fine.
12:18
SPEAKER_00: But if you're trying to roll that out across an organization
12:20
SPEAKER_00: with 40 different data systems,
12:21
SPEAKER_00: half of which have their own permission models that don't talk to each other.
12:24
SPEAKER_01: You're describing every mid-cap operator I've ever worked for.
12:27
SPEAKER_00: And this is where the data center ban conversation gets interesting, too.
12:31
SPEAKER_00: There was a piece this week about state-level opposition to new data center builds,
12:34
SPEAKER_00: and a Princeton professor,
12:36
SPEAKER_00: Arvind Narayanan,
12:37
SPEAKER_00: was making the point
12:37
SPEAKER_00: that banning data centers doesn't actually slow down AI development.
12:41
SPEAKER_00: The compute just moves somewhere else.
12:43
SPEAKER_01: You can't ban math.
12:44
SPEAKER_01: But what a ban does is change where the infrastructure sits.
12:47
SPEAKER_01: And if you're an operator in a state that's pushing back on data center builds,
12:50
SPEAKER_01: and you were counting on that data center load to justify your gas production...
12:54
SPEAKER_00: Then your demand forecast just got a political variable in it that you didn't model.
12:57
SPEAKER_01: Right.
12:58
SPEAKER_00: Which connects to what we were talking about last week with pipeline constraints.
13:01
SPEAKER_00: You've got molecules that can't get to the plant,
13:03
SPEAKER_00: and now you might have plants that can't get built in the state where the gas is.
13:06
SPEAKER_00: The infrastructure map is getting squeezed from three directions at once.
13:09
SPEAKER_01: And here's what I don't hear anyone in the AI space talk about.
13:12
SPEAKER_01: The data center opposition isn't just nimbyism.
13:14
SPEAKER_01: These things use a staggering amount of water.
13:16
SPEAKER_01: They generate heat.
13:17
SPEAKER_01: They need grid connections that sometimes don't exist.
13:20
SPEAKER_01: When a county says, we don't want this, they're not being irrational.
13:23
SPEAKER_01: They're doing the math on what their water table looks like in 10 years.
13:27
SPEAKER_00: Honestly,
13:27
SPEAKER_00: the AI industry has done a terrible job of making the case for why
13:30
SPEAKER_00: a community should want a data center.
13:32
SPEAKER_00: The pitch is basically jobs and tax revenue.
13:34
SPEAKER_00: And the jobs number is usually like 40.
13:36
SPEAKER_00: Forty.
13:37
SPEAKER_00: For a facility that draws as much power as a small city.
13:40
SPEAKER_00: So the skepticism makes sense.
13:42
SPEAKER_00: But the solution can't be ban them everywhere because the demand for compute isn't going away.
13:46
SPEAKER_00: The other thing I wanted to hit was this NVIDIA earnings situation
13:49
SPEAKER_00: because it matters for the trajectory of all of this.
13:51
SPEAKER_01: The big quarter.
13:53
SPEAKER_00: Historic quarter.
13:54
SPEAKER_00: Record-breaking revenue driven entirely by AI demand.
13:57
SPEAKER_00: And there's been all this talk about whether we're in a compute bubble,
13:59
SPEAKER_00: whether the spending is sustainable.
14:01
SPEAKER_00: And NVIDIA just answered that question with a revenue number that made everyone stop arguing.
14:05
SPEAKER_01: Until the next quarter anyway.
14:07
SPEAKER_00: But I'd also say,
14:08
SPEAKER_00: the operators who bet everything on data center demand and then
14:10
SPEAKER_00: a state moratorium hits,
14:11
SPEAKER_00: or an insurer says they won't cover a facility within a certain distance
14:13
SPEAKER_00: of something,
14:35
SPEAKER_01: Those operators are going to wish they'd kept some optionality.
14:37
SPEAKER_00: Which is basically the story of energy for the last hundred years.
14:40
SPEAKER_00: Bet on growth.
14:41
SPEAKER_00: Hedge against politics.
14:42
SPEAKER_01: PAC, that might be the most accurate one-sentence summary
14:45
SPEAKER_01: of this industry I've ever heard on this show.
14:47
SPEAKER_01: So here's where I keep landing this week.
14:49
SPEAKER_01: You've got insurers now actively involved in how renewable
14:51
SPEAKER_01: and storage facilities get designed,
14:53
SPEAKER_01: helping spec out hail resistance, wildfire hardening, all of that.
14:57
SPEAKER_01: And on the other side, you've got states throwing up moratoriums on data centers.
15:01
SPEAKER_01: Those two things are about to collide in a way nobody's talking about.
15:04
SPEAKER_00: Walk me through how you connect those.
15:06
SPEAKER_01: Because the insurer problem and the siting problem are the same problem.
15:09
SPEAKER_01: Risk assessment with incomplete data.
15:12
SPEAKER_01: An insurer is trying to figure out whether a battery storage facility
15:14
SPEAKER_01: in West Texas can survive a derecho.
15:16
SPEAKER_01: A county commissioner is trying to figure out whether a data center is going
15:19
SPEAKER_01: to drain the aquifer.
15:21
SPEAKER_01: Both of them need the same kind of answer,
15:22
SPEAKER_01: a forward-looking model of physical risk tied to a specific location,
15:26
SPEAKER_01: and neither of them has good tools for it.
15:28
SPEAKER_00: I follow your reasoning, but let me challenge one piece.
15:31
SPEAKER_00: The insurer at least has actuarial tables.
15:33
SPEAKER_00: They've got decades of lost data.
15:35
SPEAKER_00: The county commissioner has... what?
15:37
SPEAKER_00: A water study from 2019 and a consultant's PDF?
15:40
SPEAKER_01: That's my point.
15:41
SPEAKER_01: The data asymmetry is enormous.
15:43
SPEAKER_01: And it matters because the decisions coming out of that asymmetry are massive.
15:47
SPEAKER_01: A moratorium isn't a zoning tweak.
15:48
SPEAKER_01: It freezes hundreds of millions in capital deployment.
15:51
SPEAKER_01: And the insurer pulling coverage or pricing it at punitive levels does
15:54
SPEAKER_01: the same thing from the private side.
15:56
SPEAKER_00: So you're saying there's a use case for something
15:58
SPEAKER_00: that synthesizes physical risk data,
16:00
SPEAKER_00: water tables,
16:00
SPEAKER_00: grid capacity,
16:01
SPEAKER_00: weather exposure,
16:02
SPEAKER_00: seismic,
16:03
SPEAKER_00: and gives both the insurer and the local authority
16:05
SPEAKER_00: a shared picture they can actually argue from?
16:07
SPEAKER_01: Not shared.
16:08
SPEAKER_01: I'd pump the brakes there.
16:10
SPEAKER_01: The insurer and the county don't need to see the same dashboard.
16:13
SPEAKER_01: They need to be working from the same underlying data set so that when
16:15
SPEAKER_01: an operator shows up and says,
16:17
SPEAKER_01: we want to build here,
16:18
SPEAKER_01: the conversation isn't two parties
16:19
SPEAKER_01: with completely different numbers yelling past each other.
16:22
SPEAKER_00: Which is what happens right now.
16:24
SPEAKER_00: The operator comes in with a site assessment that says everything's fine,
16:27
SPEAKER_00: the county pulls out a different hydrological study that says the opposite,
16:30
SPEAKER_00: and the insurer is running a third model that disagrees with both of them.
16:33
SPEAKER_00: Nobody's clearly wrong.
16:34
SPEAKER_00: They're just looking at different vintages of different data.
16:37
SPEAKER_01: Different vintages?
16:38
SPEAKER_01: That's the phrase.
16:39
SPEAKER_01: I had a situation last year, not data centers, but a saltwater disposal well permit.
16:44
SPEAKER_01: The operator had water data from the state survey that was 18 months old.
16:47
SPEAKER_01: The landowner's attorney had a private well test from six months ago showing different levels.
16:52
SPEAKER_01: And the regulatory filing referenced a third dataset entirely.
16:55
SPEAKER_01: Three parties,
16:56
SPEAKER_01: three versions of the water table,
16:57
SPEAKER_01: and every single one of them was technically sourced.
17:00
SPEAKER_00: And that's a relatively simple permit compared to siting a facility that's going
17:02
SPEAKER_00: to draw megawatts and millions of gallons.
17:04
SPEAKER_01: Orders of magnitude more complex.
17:06
SPEAKER_01: So what's the AI angle here?
17:08
SPEAKER_01: Because I'll be honest, this is one where I actually want to hear the pitch.
17:10
SPEAKER_00: Let me back up before I get into what we built.
17:13
SPEAKER_00: The core problem isn't analysis.
17:15
SPEAKER_00: People know how to analyze water data or grid capacity if you hand them a clean dataset.
17:19
SPEAKER_00: The problem is assembly.
17:20
SPEAKER_00: You've got USGS water data,
17:22
SPEAKER_00: state well logs,
17:23
SPEAKER_00: no AA weather data,
17:24
SPEAKER_00: grid interconnection cues from the ISO, county land records, FEMA flood maps.
17:29
SPEAKER_00: And none of them talk to each other.
17:31
SPEAKER_00: Different formats, different update cycles, different spatial resolutions.
17:35
SPEAKER_00: Some of it's API accessible.
17:37
SPEAKER_00: Some of it's genuinely PDFs on a state website that hasn't been redesigned since 2008.
17:41
SPEAKER_00: I've
17:41
SPEAKER_01: seen those websites.
17:43
SPEAKER_00: Everyone has.
17:44
SPEAKER_00: So the AI capability here isn't make a prediction about the water table.
17:48
SPEAKER_00: It's take 14 disconnected public and private data sources,
17:51
SPEAKER_00: normalize them spatially and temporally,
17:53
SPEAKER_00: flag where they conflict, and present a unified risk picture for a specific parcel.
17:57
SPEAKER_00: That's the actual hard problem,
17:59
SPEAKER_00: and it's the kind of messy data integration task
18:00
SPEAKER_00: where these tools deliver real value.
18:03
SPEAKER_01: Who's the user, though?
18:04
SPEAKER_01: That is when I get stuck.
18:06
SPEAKER_01: Is it the operator doing site selection,
18:08
SPEAKER_01: the insurer underwriting the facility, the county reviewing the application?
18:12
SPEAKER_00: But the insurer criteria aren't public.
18:13
SPEAKER_00: No insurer is going to hand over their risk model.
18:15
SPEAKER_00: They don't have to hand over the model.
18:17
SPEAKER_00: They just need to publish the thresholds.
18:38
SPEAKER_00: We won't cover a facility within X distance of a floodplain or
18:41
SPEAKER_00: where average groundwater depth is below Y feet or where
18:44
SPEAKER_00: the grid interconnection queue is longer than Z months.
18:46
SPEAKER_00: Those aren't proprietary insights.
18:48
SPEAKER_00: They're underwriting guidelines that the brokers already know informally.
18:51
SPEAKER_00: Making them explicit and machine-readable just saves everyone 18 months of back and forth.
18:55
SPEAKER_01: You're describing a world where insurers voluntarily standardize their risk criteria.
18:59
SPEAKER_01: I've been in this business long enough to know how that usually goes.
19:02
SPEAKER_00: But look at what's already happening.
19:04
SPEAKER_00: The Energy Gang episode this week was genuinely about insurers collaborating
19:07
SPEAKER_00: with developers on facility design.
19:09
SPEAKER_00: They're already in the room.
19:10
SPEAKER_00: The question is whether the information they're sharing
19:12
SPEAKER_00: in those rooms gets structured in a way that's reusable,
19:14
SPEAKER_00: or whether it stays trapped in project-specific consulting engagements that cost
19:17
SPEAKER_00: a fortune and help one site.
19:19
SPEAKER_01: Failure modes, because this sounds clean, and clean makes me nervous.
19:23
SPEAKER_00: Biggest one is data freshness.
19:25
SPEAKER_00: Aquifer data that's two years old might as well be fiction
19:27
SPEAKER_00: if there's been significant drilling activity in the basin since then.
19:30
SPEAKER_00: And most of these public datasets update annually at best.
19:33
SPEAKER_00: So you build a beautiful integrated risk map and it's stale before anyone acts on it.
19:37
SPEAKER_01: Which is the same problem we talked about last week with pipeline capacity monitoring.
19:41
SPEAKER_01: The alert that shows up after the operator's already
19:43
SPEAKER_00: moved on to the next fire.
19:45
SPEAKER_00: Same structural issue.
19:46
SPEAKER_00: The system is technically correct about yesterday's conditions.
19:49
SPEAKER_01: And the approval constraint is real.
19:51
SPEAKER_01: A county commissioner isn't going to accept a risk assessment from a tool
19:53
SPEAKER_01: the operator brought to the meeting.
19:55
SPEAKER_01: That's like a defendant bringing their own polygraph results.
19:58
SPEAKER_00: So the system has to be credible to multiple parties, which means transparent sourcing.
20:03
SPEAKER_00: Every number on that risk map needs to trace back to a specific dataset,
20:06
SPEAKER_00: a specific date, a specific methodology, no black boxes.
20:10
SPEAKER_00: The moment a county attorney can't audit where a water table number came from,
20:13
SPEAKER_00: the whole thing's dead.
20:14
SPEAKER_01: And ROI, who pays for this?
20:16
SPEAKER_00: The operator, initially.
20:18
SPEAKER_00: Because the operator's the one burning capital on failed site assessments.
20:22
SPEAKER_00: If you can screen out bad sites before you spend money on them,
20:24
SPEAKER_00: the savings are immediate and concrete.
20:26
SPEAKER_00: One avoided six-month permitting dead-end pays for the tool many times over.
20:30
SPEAKER_01: What about the community side?
20:32
SPEAKER_01: Because I keep coming back to what we were saying about counties doing the math
20:34
SPEAKER_01: on their water table in 10 years.
20:37
SPEAKER_01: If an operator shows up with a tool that says your water's fine,
20:39
SPEAKER_01: that's not going to land well regardless of how good the sourcing is.
20:43
SPEAKER_00: The useful version isn't your water's fine.
20:45
SPEAKER_00: It's here are the five data sets we can both look at.
20:47
SPEAKER_00: Here's where they agree.
20:48
SPEAKER_00: Here's where they conflict.
20:49
SPEAKER_00: And here's what additional data collection would resolve the conflict.
20:53
SPEAKER_00: A structured conversation starter, not a verdict.
20:55
SPEAKER_01: That's a version I could actually see working.
20:57
SPEAKER_01: I was expecting you to tell me we should automate the county commissioner.
21:00
SPEAKER_00: Give me a few more episodes.
21:02
SPEAKER_01: The part that makes this interesting to me genuinely is the insurer angle.
21:05
SPEAKER_01: Because right now, the insurer shows up after the facility's designed.
21:09
SPEAKER_01: They look at the specs,
21:10
SPEAKER_01: they look at the location, and they either quote a premium or they walk away.
21:14
SPEAKER_01: If you can get insurer risk criteria into the site selection process upstream,
21:18
SPEAKER_01: you potentially avoid the scenario where
21:20
SPEAKER_01: a fully designed facility can't get coverage.
21:22
SPEAKER_01: And I've seen that happen,
21:23
SPEAKER_01: not with data centers,
21:24
SPEAKER_01: but with midstream infrastructure in areas with seismic concerns.
21:27
SPEAKER_00: And the seismic thing is relevant because,
21:30
SPEAKER_00: you remember a few episodes back,
21:31
SPEAKER_00: the concern about
21:32
SPEAKER_00: whether heavy industrial activity near compute facilities creates physical stability
21:36
SPEAKER_00: problems?
21:37
SPEAKER_00: That's an insurer question too.
21:39
SPEAKER_00: If you're siting a data center adjacent to active drilling,
21:41
SPEAKER_00: the insurer needs to understand that risk profile,
21:43
SPEAKER_00: and right now there's no standardized way to present it.
21:46
SPEAKER_01: So the outcome,
21:47
SPEAKER_01: the useful human-approved outcome as you like to say,
21:49
SPEAKER_00: Is an operator who walks into a county hearing with a site assessment that
21:52
SPEAKER_00: the county's own consultants can verify against the same source data
21:55
SPEAKER_00: that addresses insurer criteria explicitly and that identifies
21:58
SPEAKER_00: the specific unknowns rather than papering over them?
22:00
SPEAKER_00: Better shared information de-risks these conversations, not better sales pitches.
22:05
SPEAKER_01: And the human approval layer is the county vote.
22:07
SPEAKER_01: That doesn't get automated.
22:08
SPEAKER_00: Never should be.
22:10
SPEAKER_00: The decision is political, and it should be political.
22:12
SPEAKER_00: The data assembly is what should be automated because doing it manually takes months
22:16
SPEAKER_00: and produces inconsistent results.
22:17
SPEAKER_00: And by the time it's done, the aquifer data is another year older.
22:20
SPEAKER_01: Which brings us full circle to freshness.
22:22
SPEAKER_00: Always does.
22:24
SPEAKER_01: So here's where I've landed this week.
22:26
SPEAKER_01: The freshness problem isn't just about aquifer data for citing decisions.
22:29
SPEAKER_01: It's the whole chain.
22:30
SPEAKER_01: You've got crack spreads at $102.
22:32
SPEAKER_01: You've got transit disruptions that have been running for months.
22:34
SPEAKER_01: And the people who should be synthesizing that into a coherent picture are,
22:38
SPEAKER_01: what, updating spreadsheets quarterly?
22:40
SPEAKER_00: If that.
22:41
SPEAKER_00: And the thing that kills me is the data exists.
22:43
SPEAKER_00: The tanker rate data, the refining margin data, the transit disruption timelines.
22:47
SPEAKER_00: It's all out there.
22:48
SPEAKER_00: It's just sitting in 14 different formats across six different vendor platforms
22:51
SPEAKER_00: and nobody's stitching it together in a way that an asset team can actually act on
22:55
SPEAKER_00: in real time.
22:56
SPEAKER_01: Same pattern we talked about with pipeline capacity monitoring a few weeks back.
22:59
SPEAKER_01: The alert shows up after someone in the field already noticed the pressure drop
23:02
SPEAKER_01: and called it in.
23:03
SPEAKER_00: And this is that pattern, but at the macro level.
23:06
SPEAKER_00: An operator making capital allocation decisions.
23:08
SPEAKER_00: Do I accelerate completions?
23:10
SPEAKER_00: Do I hold?
23:10
SPEAKER_00: Do I hedge differently?
23:12
SPEAKER_00: They're working off a picture that's already stale by the time it reaches their desk.
23:15
SPEAKER_00: The crack spread moved.
23:17
SPEAKER_00: The shipping cost moved.
23:18
SPEAKER_00: The basis differential moved.
23:20
SPEAKER_00: And the summary they're reading is from last Tuesday.
23:22
SPEAKER_01: I will say some of the bigger shops have trading desks that stay on top of this.
23:26
SPEAKER_01: The midsize operators,
23:27
SPEAKER_01: the guys running 1,500,
23:28
SPEAKER_01: 2,000 barrels a day,
23:30
SPEAKER_01: who don't have a commodities analyst on staff,
23:32
SPEAKER_01: they're making decisions based
23:33
SPEAKER_01: on whatever their marketer tells them plus whatever they read that morning.
23:36
SPEAKER_00: And that's the build.
23:37
SPEAKER_00: Not a trading desk replacement.
23:39
SPEAKER_00: Nobody's suggesting that.
23:40
SPEAKER_00: But a monitoring layer that pulls in the five or six data streams
23:42
SPEAKER_00: that actually matter to a midsize operator's hedging
23:45
SPEAKER_00: and capital timing decisions normalizes them,
23:47
SPEAKER_00: flags when something's moved outside a band they care about,
23:50
SPEAKER_00: and puts it in front of them before their weekly planning meeting.
23:52
SPEAKER_00: Not after.
23:53
SPEAKER_01: I've seen operators try this with dashboards before.
23:56
SPEAKER_01: They pay some consultant to build a beautiful thing with 12 charts,
23:59
SPEAKER_01: and six months later,
24:00
SPEAKER_01: nobody's looking at it because half the feed's broke and nobody maintained
24:03
SPEAKER_01: the connections.
24:04
SPEAKER_00: That integration brittleness is real.
24:06
SPEAKER_00: But it's a design failure, not a concept failure.
24:09
SPEAKER_00: If you build it assuming the feeds will break,
24:11
SPEAKER_00: because they will,
24:12
SPEAKER_00: and you design the system to flag its own staleness,
24:14
SPEAKER_00: to say, hey, this tanker rate data hasn't updated in 72 hours.
24:18
SPEAKER_00: Here's what I'm showing, but don't trust it.
24:20
SPEAKER_00: That's a fundamentally different product than a pretty dashboard
24:22
SPEAKER_00: that silently shows you last month's numbers in this month's wrapper.
24:25
SPEAKER_01: A dashboard that tells you it's lying to you.
24:27
SPEAKER_00: More useful than one that doesn't know.
24:29
SPEAKER_01: Fair enough.
24:30
SPEAKER_01: So the human approval boundary here, where does the person stay in the loop?
24:33
SPEAKER_00: The system assembles the picture.
24:35
SPEAKER_00: It flags the movements.
24:37
SPEAKER_00: It can even draft scenario language.
24:39
SPEAKER_00: If crack spreads stay above 90 for another two weeks,
24:41
SPEAKER_00: here's what that implies for your net back in the Midland Basin.
24:44
SPEAKER_00: But the hedging decision,
24:45
SPEAKER_00: the capital timing decision,
24:47
SPEAKER_00: the call to accelerate or defer, that stays with the operator.
24:51
SPEAKER_00: Always.
24:51
SPEAKER_01: Because the system doesn't know they've got a rigged contract that penalizes them
24:54
SPEAKER_01: for releasing early or that their bank's borrowing base review is in six weeks
24:58
SPEAKER_01: and they need to show a certain production trajectory.
25:00
SPEAKER_00: The context that makes the decision a decision.
25:03
SPEAKER_00: That's human.
25:04
SPEAKER_00: The assembly of the inputs that inform the decision, that's what should be automated.
25:08
SPEAKER_00: Because doing it manually means you're either slow or you're missing something.
25:11
SPEAKER_01: And the insurer piece we were just talking about fits the same framework.
25:14
SPEAKER_01: Assemble the inputs, flag the conflicts, present it for human judgment, don't make the call.
25:19
SPEAKER_00: Same architecture, different domain.
25:21
SPEAKER_00: Site selection, hedging strategy, regulatory prep.
25:24
SPEAKER_00: The pattern's identical.
25:26
SPEAKER_00: Structured assembly, transparent sourcing, human decision.
25:29
SPEAKER_01: The operators I know who'd actually pay for this are the ones who've been burned by
25:32
SPEAKER_01: a stale number.
25:33
SPEAKER_01: You make one hedging call based on a basis differential that moved three days ago
25:37
SPEAKER_01: and nobody told you.
25:38
SPEAKER_01: That's an expensive lesson.
25:39
SPEAKER_00: And you only need to avoid that once to justify the cost of the tool for years.
25:43
SPEAKER_01: So if you're listening
25:43
SPEAKER_01: and thinking this sounds like something your operation actually needs,
25:46
SPEAKER_01: whether it's the market monitoring piece,
25:48
SPEAKER_01: the site assessment assembly,
25:49
SPEAKER_01: the pipeline capacity alerts we've talked about before, these aren't hypothetical.
25:53
SPEAKER_00: We build this.
25:54
SPEAKER_00: That's what we do.
25:55
SPEAKER_00: Reach out siliconandshale at gmail.com.
25:58
SPEAKER_00: Tell us what decision you're trying to make faster or with better data,
26:01
SPEAKER_00: and we'll figure out what the right automation looks like for your operation.
26:04
SPEAKER_01: And we'll tell you where the human stays in the loop, because that's the part most vendors skip.
26:08
SPEAKER_01: I've got to jump.
26:09
SPEAKER_01: I've got a call with a title company in 20 minutes that's probably going
26:11
SPEAKER_01: to age me another five years.
26:13
SPEAKER_00: Before you go,
26:14
SPEAKER_00: if your company sells to operators,
26:16
SPEAKER_00: landmen,
26:16
SPEAKER_00: engineers,
26:17
SPEAKER_00: anyone making decisions in the energy space and you want to reach the people
26:20
SPEAKER_00: who actually use the tools they buy, we're looking for sponsors.
26:23
SPEAKER_00: Siliconandshale at geemail.com.
26:26
SPEAKER_00: We'll put your product through its paces before we say a word about it.
26:29
SPEAKER_01: That's not a throwaway line either.
26:30
SPEAKER_01: We will.
26:31
SPEAKER_00: Go to your title call.
26:33
SPEAKER_00: Try not to find any more $102 chasms in there.
26:35
SPEAKER_01: At this point, I'd settle for a clean chain of title.
26:38
SPEAKER_01: That feels ambitious enough.