
Nov 18, 2026 · 1h 26m
AI’s infrastructure boom collides with human agency
🤠 LIVE from TomorrowX Day 2 | Jon Gray, Scott DeBoer, Jed Finn, Scott Wu, Travis Kalanick
The episode connects AI’s capital demands, autonomous software, semiconductor supply, financial services, and leadership choices to a broader question of who controls the transition.
- 1AI infrastructure increasingly resembles an industrial buildout, with energy, manufacturing, defense, robotics, and finance moving together.
- 2Persistent software agents could own recurring engineering outcomes, but security, judgment, and organizational choices remain human responsibilities.
- 3Corporate crises and AI governance raise the same question: whether institutions preserve founder-led agency or surrender decisions to defensive systems.
Don't miss
Scott Wu’s description of persistent agents monitoring feedback, regressions, upgrades, and bugs reframes software engineering as an outcome they can continuously own.
The brief
Day two of TomorrowX treats AI as an industrial cycle rather than a software fad, linking infrastructure spending to energy, manufacturing, defense, pharmaceuticals, and robotics.
Jon Gray argues the buildout has healthier fundamentals than past speculative infrastructure booms, while still demanding responsible development and tangible community benefits.
Micron’s Scott DeBoer describes memory as a strategic bottleneck: AI is shifting semiconductor competition toward performance, customization, co-design, and new manufacturing hubs.
Morgan Stanley’s Jed Finn and Cognition’s Scott Wu show two paths for agents: regulated assistants with human approval and persistent systems that own software outcomes.
Travis Kalanick closes by connecting crisis management, succession, and AI governance, arguing that companies lose identity when leaders optimize for self-preservation over agency.
What was said on this episode
42 statements · 24 positive · 10 negative · 2 mixed · 6 neutral
AI infrastructure requires trillions of dollars in construction investment.
“In this case, trillions of dollars are needed to build.”
Listen at 2:19
- Jon Grayon U.S. compute and energy supplyNegative2:28
The United States currently faces severe compute and energy shortages.
“right now we seem to have a very profound compute shortage and an energy shortage in our country.”
Listen at 2:28
- Jon Grayon Infrastructure investment wavePositive3:48
The infrastructure investment wave will extend across defense, pharmaceuticals, manufacturing, and robotics.
“it's not just the AI infrastructure, it's defense, it's going to be pharmaceuticals, there's going to be more manufacturing build, robotics.”
Listen at 3:48
- Jon Grayon Model spending by large companiesPositive5:00
Large companies increased model spending twentyfold because of productivity gains.
“their spending on models is up 20-fold in the last year because of the productivity gains they're seeing”
Listen at 5:00
- Jon Grayon AI hyperscalersPositive6:14
Hyperscalers leading AI infrastructure have average $2.5 trillion market caps, low debt, and high cash flow.
“Today, this is being led by massive hyperscalers with market caps that average $2.5 trillion, who have very little net debt, who produce enormous amounts of cash flow.”
Listen at 6:14
- Jon Grayon AI data-center constructionPositive6:38
AI data-center construction is supported by contracted demand.
“In this case, you have contracted demand.”
Listen at 6:38
Constrained AI infrastructure supply may produce a healthier investment cycle.
“And that may turn out to be a better, healthier thing.”
Listen at 7:01
- Jon Grayon Global compute supplyNegative9:12
The compute shortage is global.
“It is a global phenomenon, the shortage of compute.”
Listen at 9:12
- Jon Grayon Personalized agents and physical AINegative9:17
Personalized agents and physical AI will intensify compute shortages.
“when we start to move to these personalized agents, to more of the physical AI, I think this problem gets tougher.”
Listen at 9:17
- Jon Grayon AI rental-application processingPositive11:02
AI can make rental-application processing 90% faster.
“you can process a rental application 90% faster.”
Listen at 11:02
Ancestry.com processes incoming family-history data thirty times faster with AI.
“Ancestry.com is 30 times faster processing your family history when it comes in.”
Listen at 11:07
- Jon Grayon AI usage costsNeutral13:19
Organizations will eventually become more sensitive to AI costs.
“I do think at some point, yes, more cost sensitivity will come into play.”
Listen at 13:19
- Jon Grayon AI-assisted drug discoveryPositive13:51
AI will generate many more potential therapies.
“this is going to lead to many more potential therapies”
Listen at 13:51
- Jon Grayon AI-enabled healthcarePositive14:02
AI will radically change future cancer, neurological, and pulmonary care.
“When we think about the future of cancer care, neurological diseases, pulmonary, I think it's going to be radically different.”
Listen at 14:02
- Jon Grayon Phase-three clinical trialsNegative14:37
Phase-three clinical trials cost one or multiple billions of dollars.
“Those trials cost a billion, multiple billion dollars.”
Listen at 14:37
- Jon Grayon AI-related job transitionsMixed16:10
Society should manage AI-related job transitions and provide skills training.
“There will be challenges. And as a society, we've got to make sure job transitions, skills training— we've got to be sensitive to this.”
Listen at 16:10
- Scott DeBoeron AI memory technologyNeutral18:11
AI has shifted memory development toward performance and customization over raw bit growth.
“the past few years, especially with AI, More focused on performance and customization, less bit growth, raw bit growth in terms of technology”
Listen at 18:11
- Jed Finnon Private and public marketsPositive34:46
Private and public markets should be regulated as a disclosure continuum.
“the difference between private and public ought not to be a binary distinction.”
Listen at 34:46
- Jed Finnon DevinBotPositive37:41
DevinBot can persistently own outcomes rather than merely execute tasks.
“these persistent agents that can actually own outcomes instead of tasks.”
Listen at 37:41
- Jed Finnon DevinBotPositive37:57
DevinBot has memory, manages projects, and delegates work at scale.
“it has memory, it can take on a full project, it can really, really delegate and run things en masse.”
Listen at 37:57
AI pretraining runs will grow from billions to hundreds of billions of parameters or scale units.
“these will get bigger and bigger and will go from billions to tens of billions to hundreds of billions.”
Listen at 45:23
AI models will eventually outperform humans at creating interesting math problems.
“the LLM and all the AI models over time will actually be much better than we are at coming up with interesting creative math problems.”
Listen at 50:00
- Jed Finnon AI applicationsPositive51:42
AI will solve drug discovery and develop novel technologies.
“I think we will.”
Listen at 51:42
- Jed Finnon AI agent adoptionNeutral54:14
AI agents diffuse first through clean, verifiable, well-scoped tasks.
“if you have the clean kind of verifiable, well-scoped tasks, that was going to work first”
Listen at 54:14
- Jed Finnon AI adoptionNegative54:54
AI adoption remains constrained by a major diffusion bottleneck.
“I think there's still for sure just like a massive diffusion bottleneck for all the other things that we talk about in AI.”
Listen at 54:54
- Jed Finnon Recurring software maintenancePositive56:24
Recurring software maintenance tasks should become autonomous.
“all of the latter tasks kind of should be self-driving”
Listen at 56:24
- Jed Finnon DevinBotPositive57:06
Persistent Devin agents can autonomously monitor and fix recurring software issues.
“you can actually just have a persistent Devin bot that just goes and owns that”
Listen at 57:06
- Jed Finnon Agent-built personalized softwarePositive58:58
Personalized software built by agents will become more common.
“this form of like, I'll call it like personalized software, I think we're going to see a lot more of.”
Listen at 58:58
- Jed Finnon Vendor data sharingPositive59:49
Companies should carefully limit data shared with software vendors.
“being very thoughtful about what kind of data you're sharing with vendors”
Listen at 59:49
AI will not make it optimal for companies to build all software internally.
“I don't think that is the case that we're at the full SaaSpocalypse build all software in-house thing.”
Listen at 1:00:34
Kalanick experienced coordinated corporate attacks weekly for six months.
“For me, it was 6 months of ops. Yeah, one a week.”
Listen at 1:06:16
Kalanick recommends against responding to corporate attacks by retreating into a shell.
“I wouldn't recommend it.”
Listen at 1:10:55
Organizations facing corporate attacks should establish a strategic war room.
“you need to get a war room up.”
Listen at 1:10:58
Nvidia should not be handed over to executives focused on bureaucracy.
“it's not letting Nvidia go to the suits.”
Listen at 1:13:58
Apple’s post-Jobs leadership transition was planned.
“First, the Apple transition was a planned one.”
Listen at 1:14:51
A replacement Nvidia CEO would naturally resist Jensen Huang’s return.
“that is a natural resistance to Jensen coming back.”
Listen at 1:16:16
Humanity must preserve agency over how AI is used rather than accept an AI overlord.
“the future of humanity is on the line. Do we want an overlord, or do you want to have— do you want to use, consume, utilize, benefit from AI on your terms?”
Listen at 1:19:40
Employee walkouts are more effective than petitions for influencing corporate leadership.
“petitions are nice. Yeah, walk out better.”
Listen at 1:21:30
Atoms’ mining business will be several times larger within twelve months.
“we're probably, you know, a huge number multiple bigger over the next 12 months than we are today on mining”
Listen at 1:22:32
Atoms will build a food-robot factory in Pittsburgh in Q1.
“The factory that makes the food robots is going up in Pittsburgh Q1.”
Listen at 1:23:04
Data-center opposition will not prevent infrastructure construction everywhere.
“you're not gonna have a situation where every single state and every town in every state is like, I don't want this.”
Listen at 1:23:55
Data centers will be built in jurisdictions willing to host them.
“Somebody's down. Yeah. So it's going to go where it needs to go”
Listen at 1:24:05
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.