Jan 26, 2026 · 1h 3m · a16z
The Biggest Bottlenecks For AI: Energy & Cooling
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this Andreessen Horowitz (a16z) presentation, General Partner David George analyzes the macroeconomic landscape of artificial intelligence, highlighting massive infrastructure capital expenditures, collapsing model inference costs, application unit economics, physical energy and cooling constraints, and venture capital growth strategies.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The host as informed peer 3.8 Guest teaching 4.8 Guest disagreement 1.3 The host pushing back 2.1
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 32:00 Rejection of enterprise stickiness rule
David directly challenges the conventional VC doctrine that enterprise software is stickier than consumer software, asserting consumer ChatGPT users are far more loyal than developers using switching-prone B2B model APIs.
Hardest push from the host ▶ 51:42 Challenging SPV and tender offer allocation qualityThe host directly confronts the guest on whether relying on broadly accessible tender offers and SPVs like OpenAI and Databricks compromises portfolio quality compared to securing allocations in competitive, hard-to-access names like Anduril or Flock.
Biggest teaching moment ▶ 8:50 Search traffic growth timeline comparisonDavid educates the host and audience on AI adoption velocity by highlighting that ChatGPT reached 365 billion searches in just two years, whereas Google took 11 years to hit the same milestone.
The host holds their own ▶ 20:57 Anticipating the thermal cooling bottleneckThe host demonstrates deep sector expertise by anticipating thermal cooling as the critical physical bottleneck following energy power generation before the guest introduces the topic.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| Macro Market Trends and AI Expansion | 0 | 4 | 1 | 0 | Monologue segment by the guest laying out macro tech market cap trends and $400B annual AI capex buildout. The host does not speak during this segment. | |
| Model Economics, Declining Costs, and Capabilities Escalation | 3 | 5 | 1 | 1 | Guest outlines model input cost declines and capabilities escalation before the host interjects with a meme anecdote and asks the guest to explain why this cycle won't crash like the early 2000s dot-com telecom bubble. | |
| Debunking Infrastructure Risks and Bubble Analogies | 0 | 6 | 1 | 0 | Guest responds in a extended monologue explaining how private debt and banking capital stabilize supply while demand growth for ChatGPT is 5.5x faster than Google search reached. | |
| Consumer Monetization Models and Search Traffic Shifts | 4 | 3 | 1 | 1 | Host demonstrates market knowledge by citing public company earnings calls showing referral traffic drops for companies like Groupon and IAC due to Google AI summaries. | |
| Energy Constraints, Nuclear Power, and Thermal Cooling Bottlenecks | 5 | 3 | 1 | 2 | Host fields an audience question about energy constraints and preemptively introduces liquid and thermal cooling as the next fundamental bottleneck after power generation. | |
| Evaluating AI Application Gross Margins and Retention Metrics | 5 | 6 | 2 | 3 | Host presses guest on AI application gross margin scrutiny and dependency on underlying model providers like Anthropic. Guest argues gross retention and customer acquisition ease matter more than temporary low gross margins. | |
| OpenAI Cash Burn, Pricing Power, and B2B vs. Consumer Stickiness | 4 | 6 | 2 | 2 | Guest reframing traditional SaaS metrics by arguing consumer AI subscriptions like ChatGPT are stickier than enterprise developer API integrations where model switching is easy. | |
| Application-Layer Revenue Durability and Deep Workflow Integration | 5 | 5 | 1 | 4 | Host directly asks guest to explain the specific underlying operational drivers that make certain AI application verticals stickier than others. | |
| Accelerated Growth Timelines and Evolution of Pricing Models | 4 | 5 | 1 | 2 | Host steers a rapid-fire Q&A segment regarding $100M ARR timelines and seat-based versus outcome-based pricing models. | |
| Private Market Dynamics and Late-Stage Venture Capital Strategy | 6 | 5 | 3 | 6 | Host challenges guest on portfolio construction, contrasting access to widely available tender offer rounds like OpenAI/Databricks versus hard-to-access companies like Anduril or Flock. | |
| Framework for Disrupting Incumbent Public Software Giants | 4 | 6 | 1 | 2 | Host introduces question regarding incumbent software vulnerability to AI, prompting guest to present a three-part framework for startup disruption. | |
| Early-to-Late Stage Collaboration and Sector Portfolio Sizing | 5 | 4 | 1 | 2 | Host shares internal firm metrics and guides discussion around late-stage growth collaboration across early-stage practice areas. |
Statements from this episode (41)
David George: Tech companies are staying private longer than ever
“Companies are staying private longer than ever.”
David George: AI companies are growing faster than any sector in history
“The AI companies are getting bigger faster than anything we've ever seen. The investment amounts are bigger than anything we've ever seen.”
David George: AI infrastructure spending will far exceed current projections
“These numbers are going to end up way bigger.”
George: Big tech annual capex run-rate has reached $400 billion
“Because I think just the big tech companies in their latest quarter, if you run rate their capex from the latest quarter I think it's like four hundred billion dollars of annual capex. And most of that is going into AI.”
David George: Big tech can easily absorb potential AI capex overbuild
“It turns out they're the best companies, you know, probably ever created you know, companies like Google, Facebook you know, Amazon and Microsoft, and they can bear you know, potential capacity overbuild and things like that.”
George: AI model input costs dropped over 99% in two years
“The cost of the inputs you know, of accessing these models has declined 99% or a little more than 99% over the last two years.”
George: Frontier AI model capabilities double every seven months
“The models have been improving in sort of frontier capabilities by a double factor every seven months.”
a16z View: AI will become a ubiquitous utility like electricity or Wi-Fi
“Our house view now is that AI is going to end up like, You know, electricity or wifi. Like if you're getting, if you're accessing, you know, electricity at somebody's house, you're not like, Hey, let me chip in, you know, a few pennies for, you know, sitting i…”
George: AI market value creation will exceed mobile and cloud's $10T
“And I think AI is going to be much larger because I think the impact on the economy is going to be much larger.”
George: US software spend is 1% of GDP vs 20% for payroll
“US software spend is like one percent of GDP. US white collar payroll is like 20% of GDP.”
George: Tech customers capture 90% of economic value while platforms capture 10%
“My rule of thumb is like, 90% of the value goes to the end customers and, you know, 10% of the value goes to the companies serving them, and it turns out that that's just a massive amount of market cap, you know, if you're the 10% that you're capturing.”
George: Banks and Private Debt Are Primary Funders of Private Capital
“The biggest funder of private capital is actually banks or private debt.”
George: Insurance Funding Signals Financial Stability for AI Infrastructure
“So maybe there's insurance companies that are kind of backdoor, you know, funding this build out, you know, that's a really good sign for the stability of the build out.”
George: ChatGPT reached 365 billion searches five times faster than Google
“The time to get to three hundred and sixty five billion searches on chat GPT was two years. The time for Google to get to three hundred and sixty five billion searches was 11 years.”
Over half of global internet users have already used AI tools
“It's probably well over half of the global internet population has used AI tools already.”
AI infrastructure will be utilized more predictably than early internet broadband
“That is heartening to me that the supply build out will be utilized. Maybe in a more predictable way than, you know, broadband in the early internet build out days just cause it, you know, it's built on the back of the previous infrastructure stuff.”
David George: AI Freemium Models Will Monetize Via Affiliate Advertising
“Probably end up with freemium products where they monetize through some form of advertising. It's hard to speculate now on what that would look like. I think it's probably some form of like an affiliate type thing that has like a dirty connotation because that…”
George: OpenAI has 30 to 40 million paying users
“There's, you know, whatever, probably 30 to forty million paying users today. The other platforms are kind of a rounding error relative to that. So maybe add another 10. So there's like, forty million people paying for this stuff today at some level. And, you …”
George: ChatGPT daily users average 30 minutes per day
“Active daily active users of chat GPT are ready today. Spend like 30 minutes at 20, 28, 29 minutes a day on the product. And, you know, to put that into context, I think Instagram's like 50 minutes a day and, you know, tick tock like Sadly, it's like 70 minute…”
George: Three Mile Island nuclear plant will be powered back up
“I think Three Mile Island's gonna get powered back up.”
George: xAI built the largest data center four times faster than peers
“And one of the most remarkable things about what XAI did is they stood up the biggest data center at the time in, you know, like a quarter of the time that anyone else had done the same thing.”
George: Energy will be the primary AI bottleneck for five years
“I think energy ultimately in the next call it five years will probably be the bottleneck, and that's why we're so excited about nuclear and making investments in that area.”
a16z looks for 90%+ gross retention when evaluating software startup business models
“If you made me pick two top line stats to look at, to assess the business model, it would be gross retention rate... And so we look for things where like 90% plus customers are sticking around, and hopefully they're expanding their usage.”
George: a16z is lenient on AI startup gross margins expecting cost drops
“Relative to like mature SaaS apps, we probably are a little bit more lenient on assessing a company's gross margin today, because we strongly believe that their input costs are going to go down over time.”
George: OpenAI Has More Monetization Upside Than Downside Risk
“I think there's way more upside to monetize the base than there is risk of price pressure on, you know, today's thirty million people paying for it.”
George: Most of OpenAI's Cash Burn Is Driven by Research and R&D
“The actual, like most of the burn comes from research, like research, like R and D, you know, and so future investments.”
George: Portfolio Coding Startups Will Instantly Switch Models for Better Performance
“If there's a new coding model that comes along, that's better than the latest version from Anthropic, like, Our coding companies will just switch and it's pretty easy to do because it's an API call.”
George: Enterprise Rules and Integration Drive AI Application Retention
“The more stuff that gets integrated and the more company specific like kind of rules built around the model stuff you have, the stickier it's gonna be.”
George: Enterprises will not vibe code their own Salesforce replacements
“I don't think that like companies are going to vibe code up like their salesforce.com. That's just not worth it.”
George: Top AI startups hit $100M ARR four times faster than SaaS
“The top companies that we've seen you know, have gotten to ten million, then a hundred million, like, Four times faster or something like that.”
Software pricing models will not be radically overhauled within five years
“I'm low conviction that we end up, you know, five years from now with all the software companies monetizing in a completely different way.”
George: Aggregate unicorn market cap reached $3.5 trillion, growing 7x in a decade
“If you take the market cap of, you know, the private markets valued above a billion dollars, you know, we could argue are some of them overvalued, undervalued but that whole value in aggregate Is like three and a half trillion dollars, and that's like 11%, thr…”
George: Only 5% of public software companies forecast over 25% growth
“Something like five percent of software and internet public companies are forecasting 25% plus next 12 months growth. So, 95% of the public market universe in software and internet is growing less than 25%.”
George: a16z was the first outside investor in xAI after Elon Musk
“XAI. First outside money in beyond Elon.”
George: a16z expects a 3x to 4x return on Databricks investment
“I happen to think that the return profile of that recent one in Databricks is, is very attractive. But, you know, we don't want an entire portfolio of things that we think are like pretty safe two X's, you know, where I think the question is, can we make five …”
a16z Growth Fund Will Maintain Minimal Public Market Exposure
“Yeah, probably very few public companies in the fund. I mean, we'd have to have a really, really strong thesis and relationship with the management team to want to do that.”
George: AI Startups Need UI, Data, and Pricing Shifts to Beat SaaS Giants
“I think for startups to win, I think you need all three and that's just in the head-to-head stuff.”
George: a16z has not found a startup capable of dethroning Salesforce
“We haven't found the startup that's like got the killer idea for like dethroning salesforce.com yet.”
David George: 80% of a16z growth deals tie to early-stage relationships
“80% of the time when we make a new investment, That's not an early stage investment. One of our early stage folks has some preexisting relationship.”
George: Top a16z growth investments target 5x upside and 2x downside
“Where, you know, we have some, you know, absolute champion companies that we think actually have, like, a lot of room still to run where on the downside we think we would, if things really go poorly, like, we'd still probably make two times our money, and on t…”