
Nov 17, 2026 · 2h 51m
AI’s next phase depends on capital, infrastructure and specialized models
🔴 LIVE From TomorrowX | David Solomon, Brian Armstrong, Ron Biscardi, Gavin Baker, Alex Davis, Guy Oseary, Andrew Feldman, Keller Cliffton, Chantel Jeffries, Cornelius Schmitt, Chase Lochmiller
The broadcast connects AI’s investment cycle to cybersecurity, industrial capacity, autonomous delivery, agentic finance and the physical systems needed to scale it.
- 1AI’s biggest returns may come from applications built on infrastructure rather than from infrastructure and model providers themselves.
- 2Specialized models, open weights and model ensembles could improve competition, autonomy and defenses against AI-enabled cyberattacks.
- 3Scaling AI requires power, chips, data centers, manufacturing, regulation and new financial systems—not software alone.
Don't miss
Gavin Baker argues that Jensen Huang is effectively irreplaceable because of his leadership, supply-chain relationships and strategic role at NVIDIA.
The brief
Live from Austin’s TomorrowX Summit, TBPN moves from Goldman Sachs’s durable revenue strategy to the broader question of how institutions finance technological change.
David Solomon frames AI as a long investment cycle, while Gavin Baker argues that Jensen Huang’s leadership remains unusually difficult to replace at NVIDIA.
Investors debate whether private AI valuations are due for a reset, even as Marc Andreessen and others search for the next foundational company beyond OpenAI and Anthropic.
The physical bottleneck comes into view through Cerebras, Zipline and Crusoe: specialized chips, autonomous delivery, manufacturing, power and modular data centers all shape AI’s reach.
Brian Armstrong extends the thesis into agentic finance, describing payment controls and specialized models for software agents while warning that biology and quantum security remain hard problems.
The standout tension is clear: AI may transform finance, health and industry, but its gains depend on resilient infrastructure, public trust and disciplined capital allocation.
What was said on this episode
59 statements · 49 positive · 5 negative · 5 neutral
TomorrowX is currently the largest AI conference in the United States.
“this is... the biggest ai conference in the country right now”
Listen at 0:21
Online communities now exert more influence than ever before.
“the whole online community is much more influential now than ever before”
Listen at 3:45
TomorrowX will probably fill its venue next year.
“next year we're probably going to fill this place”
Listen at 6:15
Human connection will remain central to Goldman Sachs's business despite technology.
“our business is still a business of human connection. And that's not going to change.”
Listen at 9:46
Goldman Sachs operates approximately the fifth- or sixth-largest active asset-management platform.
“we now run what is probably the fifth or sixth largest active asset management platform”
Listen at 11:38
Goldman Sachs's management-fee revenue is currently growing above 10%.
“we're growing it at better than 10% at the moment”
Listen at 11:56
Approximately 60% of Goldman Sachs revenue is considered durable.
“approximately 60% of the revenues would be considered more durable revenues”
Listen at 12:24
Goldman Sachs intends to remain focused on institutional clients for now.
“I think we're going to stick to being an institutional firm for the moment.”
Listen at 14:28
Goldman Sachs will expand its client footprint over the next five to ten years.
“an expanded footprint is what you'll see over the next five to 10 years”
Listen at 15:13
Goldman Sachs is more secure now than seven months ago due to AI-assisted scanning and patching.
“the firm is more secure today than it was seven months ago”
Listen at 17:42
More technology-related system disruptions should be expected during rapid AI deployment and patching.
“we're going to have bumps like this. And it's concerning, and we should be cautious, but I think you're going to expect more.”
Listen at 18:33
A significant cyber incident may occur.
“I wouldn't be surprised if we did have a significant cyber incident in some way.”
Listen at 18:44
Cybersecurity conditions could become substantially more stable within five to seven years.
“I can certainly imagine a world five to seven years from now that's a lot stable and more durable in this context.”
Listen at 21:42
Goldman Sachs aims to lead major IPOs for important companies.
“we want to lead all these big IPOs for important companies”
Listen at 23:36
Private AI infrastructure markets will experience substantial turbulence in 2027.
“I see a lot of turbulence for the private markets in 27”
Listen at 24:45
The strongest AI companies will continue receiving funding.
“the best companies will continue to get funded”
Listen at 26:38
Slower AI capital formation may be healthy for the market.
“that might be a healthy thing”
Listen at 26:52
Sustained investment outperformance requires strong capabilities in chosen areas.
“to really be a successful investor and have better relative performance, you have to have great capabilities”
Listen at 28:27
U.S. reindustrialization should center on technology innovation.
“re-industrialization in the United States should be centered around technology innovation”
Listen at 30:18
The United States should maintain firm borders while attracting productive immigrants.
“we absolutely should have firm borders, but we also have to have great policies that get smart, productive people that want to be here”
Listen at 31:05
AI should be distributed broadly so people can participate.
“you want to distribute it as widely as possible so people can participate”
Listen at 32:37
Most AI value creation will come from applications rather than infrastructure, models, or hyperscalers.
“one of the big value creation around all this I don't think is going to be on the infrastructure build or the large language models or the hyperscalers. It's going to be on the application layer that's built over the next couple of decades.”
Listen at 33:31
NVIDIA's board should proceed slowly and gather facts before deciding on Jensen Huang's return.
“I think it's important that the board go slowly and thoughtfully and make sure it understands all the facts”
Listen at 36:14
Jensen Huang is irreplaceable as NVIDIA's leader.
“Jensen is irreplaceable.”
Listen at 41:05
A vibrant open-source AI ecosystem is needed for companies to control their own intelligence.
“We also need a vibrant open source ecosystem so that companies can own their own intelligence, own your weights, own your destiny.”
Listen at 43:42
Open-source model ensembles are important for defending against AI-enabled cyberattacks.
“the way you protect us. And I think that's another reason that open source is so important. If we don't have that, I think you're going to see way more of these hacks”
Listen at 46:31
Fast inference combined with model ensembles will prevail in cyber defense.
“the faster tokens with an ensemble of models is going to win”
Listen at 47:48
Cybersecurity will soon become defense-dominant as open-source AI proliferates.
“cyber is quickly going to become defense dominant”
Listen at 50:20
AI will extend human lifespans.
“I think we are going to see an extension of human lifespans. And that's from AI.”
Listen at 51:50
Disruptive has launched a $10 billion fund.
“A $10 billion fund.”
Listen at 53:14
Disruptive plans approximately ten $1 billion investments.
“10 $1 billion checks is the idea.”
Listen at 53:23
Down rounds for late-stage AI companies are probably forthcoming.
“the down round thing's probably real”
Listen at 57:55
Traditional venture capital firms will regain influence in AI financing.
“my prediction would be you're going to see a revenge of the vcs”
Listen at 58:30
Guy Oseary invested in Anthropic at $5 billion and OpenAI at $27 billion valuations.
“We did Anthropic at a $5 billion valuation. We did OpenAI at a $27 billion.”
Listen at 1:07:14
OpenAI and Anthropic were expected to become foundational AI companies.
“these were going to be the foundational AI companies that everyone's going to plug into”
Listen at 1:09:57
Approximately 75% of Cerebras data centers will be outside the United States in 2027.
“In 27, about 75%. of our data centers will be out of the U.S.”
Listen at 1:26:42
Inference workloads can operate effectively in distributed 50- to 100-megawatt facilities.
“For inference, 50, 100, just fine. We can put them in different locations.”
Listen at 1:29:50
Zipline has launched commercial drone delivery operations near Austin.
“We just launched Austin. We have a site that's now commercial and live in Round Rock”
Listen at 1:32:52
The United States can dominate autonomous delivery drones.
“the U.S. is and can dominate in this in this different, totally different class”
Listen at 1:36:07
Zipline became the first company approved to fly beyond visual line of sight in all 50 states.
“Zipline became the first company to get full regulatory approval to fly in all 50 states beyond visualization site in the U.S.”
Listen at 1:38:24
At least three strategic partners each buy over one million daily deliveries of Zipline capacity.
“We now have more than three, we have three strategic partners and more that we'll soon announce that are each buying more than a million deliveries a day of capacity from Zipline.”
Listen at 1:40:23
Zipline will soon operate its delivery network continuously.
“We'll soon be operating 24-7.”
Listen at 1:42:15
Each Zipline vehicle can complete approximately 20 to 50 deliveries daily.
“you can think of each vehicle as doing anywhere from 20 to 50 deliveries a day.”
Listen at 1:42:21
Zipline employs approximately 700 engineers.
“Zipline is, I think today it's probably... 700 engineers”
Listen at 1:48:53
Zipline's flight volume will exceed all other U.S. airlines combined next year.
“flight volume will be bigger than all other airlines combined by next year.”
Listen at 1:50:07
Building a new drone-delivery industry requires a full-stack company.
“the goal has always been, I mean, I guess coming back to the main point, you know, the goal has always been, you just got to be a full stack company”
Listen at 1:52:09
Nighttime drone operations are easier for Zipline because competing aircraft are less active.
“Night is much easier for us, interestingly, because guess who isn't flying at night?”
Listen at 1:53:00
Valor materially supported Zipline by embedding engineering staff to solve manufacturing problems.
“Valor like deployed quality engineers and process engineers and manufacturing engineers into Zipline. And they like sat on the line with us for months on end, helping us solve these hard operating problems.”
Listen at 1:55:47
Human technological ambition has expanded dramatically over the past thirteen years.
“the scope and scale of human ambition. has grown so dramatically.”
Listen at 2:00:16
Managed fine-tuning revenue is currently much smaller than frontier-lab revenue.
“You know, I think it's much smaller than frontier lab revenue today.”
Listen at 2:39:37
Fine-tuning can produce better results for customers’ domain-specific objectives.
“And what we're seeing is there are customers that are able to get better results on... things that they care about.”
Listen at 2:40:07
Use of private datasets with fine-tuned models will continue growing and expanding.
“We think that's going to be a trend that's going to continue to grow and expand as people think about using these rich private data sets to apply to general use cases.”
Listen at 2:40:24
Iceland has ultra-low-cost geothermal power.
“It sort of has this ultra-low-cost geothermal power.”
Listen at 2:40:55
Iceland’s cooling environment enables very low data-center PUEs.
“So you get very low PUEs when you're managing data infrastructure.”
Listen at 2:41:14
Iceland’s subsea connectivity supports energy-intensive AI compute for remote customers.
“ai is a much better manifestation of that trade right you're moving data over subsea cables we have three redundant subsea cables that interconnect into iceland and then you're sort of doing the energy-intensive compute workload on managed clusters of GPUs and then shipping the end result”
Listen at 2:43:19
Canada is a compelling market for AI infrastructure.
“Canada has always been a very compelling market.”
Listen at 2:44:30
Modular data centers can use excess grid power where transmission is limited.
“And it's also very flexible in the sense that... There's actually a lot of areas of the grid where there is excess power, and there may be limited transmission.”
Listen at 2:47:53
Distributed modular AI infrastructure is important for national security.
“We also think it's an important thing from a national security standpoint.”
Listen at 2:48:37
Distributed AI infrastructure is more secure and resilient than concentrated infrastructure.
“And we think it's just a better, secure, distributed infrastructure solution.”
Listen at 2:49:10
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.
Books & mentions
Some links are affiliate links — PodLume may earn a commission if you buy.

The Bible
It is discussed alongside other texts during the conversation about values and diverse perspectives in AI.

Tao Te Ching
It appears in the discussion of foundational texts and competing value systems.

Quran
It is mentioned as another major text relevant to the episode’s discussion of diverse value systems.
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.