Listen on YouTube // Apple // Spotify
You can also watch/listen to the entire episode on X.
Bigger Than Apollo, Bigger Than the Highway System: Inside the $575B AI Infrastructure Bet
Hyperscalers are spending every dollar of free cash flow they generate. Then they’re borrowing more.
Meta, Google, and Oracle are now levered roughly 7-to-1 on a cash flow basis to fund data center build-outs nobody is certain they’ll fill. This year’s spend is the 5th largest infrastructure project in human history — bigger than Apollo, bigger than the Interstate Highway System, bigger than everything except the railroads and the two World Wars.
For every $1 of AI revenue the industry generates, it’s spending $12 on infrastructure. That’s a $575B bet.
We sat down with Tomasz Tunguz, General Partner at Theory Ventures, and one of the most prominent voices on data infrastructure for the past decade. He walks through what nobody’s priced in yet, and what changes if you’re a founder, an operator, or an investor.
Episode Highlights
0:00 - Intro & The Scale Nobody Anticipated
2:13 - Data Center CapEx Could Hit 5-7% of US GDP by 2030
5:21 - For Every $1 AI Companies Make, They Spend $12 on Infrastructure ($575B Bet)
6:20 - Market Share Capture vs. Margin Games: The Chicken Game Big Tech is Playing
9:48 - How the Data Stack & AI/ML Worlds Have Completely Fused
12:33 - Product-Market Fit is No Longer Binary: It’s Continuous Now
15:17 - How AI is Changing Venture Capital & Portfolio Management
17:09 - The Future: Image & Video Data is Going to Require MASSIVE Infrastructure
18:25 - Pattern Recognition Across Winning Companies (Domain Expertise is Key)
20:50 - Hot Take: Corporate Org Structure Will Transform in 5 Years
21:00 - Final Advice to Founders: Nobody Knows the Answer
Brought to You By: AngelList
GTMfund’s LP base spans from individual operators to institutional allocators, and AngelList has been instrumental in supporting all of them. They handle everything from investor onboarding and accreditation to distribution and tax documentation, creating a seamless experience across geographies and fund types. Plus, all of this is available on a single, modern platform.
For an LP-base like ours, with over 350 C-suite and VP-level operators, this kind of white glove service and seamless workflows is so important. It’s also instrumental that we support our institutional LPs that we’re fortunate to work with, and AngelList is able to do so every step of the way.
If you’re looking for a platform that can support any type of LP investing in your fund, learn more at www.angellist.com/gtmfund.
Inside the $575B AI Infrastructure Bet
AI Infrastructure Is Now the 5th Largest Project in Human History
This year’s data center build-out is the 5th largest infrastructure project in human history. Bigger than the Apollo program, bigger than the Interstate Highway System, bigger than everything except the railroads and the two World Wars.
Right now, data center spend is running at roughly 3.5% of US GDP. Tomasz thinks it hits 6–7% by 2030. The railroads peaked at about 5%.
“I don’t think anybody really appreciated the scale.”
The math underneath is more aggressive than most have priced in. For every $1 of AI revenue the industry generates, it’s spending $12 on infrastructure. That’s a $575B bet. Google is converting $75–90B in annual free cash flow into data center CapEx, every dollar of it, and borrowing on top. Meta is doing the same. Oracle is now levered roughly 7-to-1 on a cash flow basis.
“It’s crazy. People are really betting that they can win significant share over time.”
Tomasz is clear about how to read the spend: it’s a market-share game first, a margin game second. Whoever owns the most inference usage over the next 3–5 years gets to set the price for everyone else later. The metric that actually matters, in his words:
“The dominant metric that really matters is how much intelligence you can drive per watt of electricity.”
Anthropic is rumored to have very high gross margins. Some of the others, less so. The first wave is still wide open.
Foundation Models Have 35 Days to Beat the Next Release
From 2010 to late 2021, product-market fit was a binary thing. You either had it or you didn’t. Once you found it, the ratios — CAC, payback, NDR — were known. The job was just to raise the capital and execute.
That era is over.
“A foundation model company will develop a state-of-the-art model. They have 35 days to commercialize it before somebody else beats them.”
35 days. For a $5–10B investment.
And it’s not just foundation model companies. The same dynamic is moving up the stack into application software.
“If you develop something unique, it’s very easily copied. So you have to keep pushing. That’s what we mean when we say product-market fit is continuous.”
The framework most early-stage founders were taught is broken. PMF isn’t a milestone you hit and then scale from — it’s a state you have to defend, every week, against a competitor who just shipped a model release that ate your differentiation.
This changes what founders should hire for, what investors should underwrite, and how operators should think about category capture.
“And the buyer demands are also changing, because the buyers are starting to understand what they want.”
The product moves under you. So does the buyer.
You’re Selling to Two Buyers Now: The Human and Their AI Agent
“The head of Carta was saying they’re no longer investing in their website or their mobile app. No new products available. They will all be free agents.”
This goes against the entire conventional B2B distribution playbook.
The reason it makes sense is the buyer just changed. The person evaluating software now uses an agent to do most of the early-stage research. The agent reads the docs, the pricing page, the comparison content. By the time the human shows up, the agent has already shortlisted.
“You now have two different constituencies to market to. The first is the head of engineering. The second is the agent of the head of engineering.”
That agent is now a member of the buying committee, sitting alongside the head of engineering, the head of AI, the head of legal. Three humans and a software agent, all weighing in on the same decision. If you’re selling enterprise, you’re selling to all four.
And the way you sell to each is different. Humans respond to brand, design, and emotional positioning. Agents don’t.
“Agents don’t respond to emotion, at least not yet. Right now it’s just pure text, raw markdown, statements of facts and clarity.”
Most teams are still writing for the human and assuming the agent reads the same content. It doesn’t. The agent is parsing facts. The human is responding to taste. You need both layers, written for both audiences.
AI Agents Are Being Benchmarked on Persuasion
AI agents are being benchmarked on their ability to persuade humans.
There’s a public benchmark called Giving for Good. An AI agent engages a real person in conversation about a charity, learns what they care about, and tries to convince them to donate. The benchmark scores propensity to donate and amount donated.
“The more effective it is at convincing you, the higher it scores.”
This sits inside a larger framework Tomasz uses for what humans will and won’t outsource to AI:
“If you were to ask AI what’s the best pair of running shoes to buy, you’d probably trust the recommendation. Best laptop? Probably. Best car? Maybe, even a $20–30k buying decision. But enterprise sales is much more complex.”
The line where AI buying stops is the line where trust between humans still matters. And that line is moving up fast. Three years ago it was at $50, now it’s at $30K, enterprise is next.
The implication for founders is the new distribution channel isn’t ads or SEO. It’s being the recommendation the agent makes. And right now, almost nobody is optimizing for that.
PR is part of the answer too. The press is willing to write about AI in a way they never wrote about software, and agents are reading the press to build their recommendation sets. Channel partnerships now kick in at low-single-digit ARR — historically that threshold was $15–30M. Distribution is moving earlier and faster than founders realize.
Today’s Org Chart Won’t Survive the Next Five Years
The current ratio inside most companies: 5% executive leadership, 75% middle management, 20% individual contributors who actually ship the work.
“In five years, it won’t look anything like that.”
He’s predicting a re-imagination of every role. The forward-thinking leaders he’s watching are hiring generalists, not specialists. Vercel replaced nine senior engineers with one AI agent and a part-time engineer. SDR and BDR roles are being fully automated, and Tomasz is clear that’s a one-way change. Data teams now report to the head of engineering, which used to be unthinkable.
When AI takes execution off the table, what’s left is judgment. Middle management — the layer whose job was to coordinate execution — is the layer most exposed.
Tomasz’s ultimate advice:
“Nobody really knows the answer to anything. This is a tremendous period of experimentation. The best thing you can do is just jump in with two feet and figure it out yourself.”
The GTMnow Podcast - subscribe to support
It would mean the world to us if you subscribe to the show on your platform of choice. This helps support our growth, which in turn helps us continuously bring on bigger guests and deliver the content that you want to hear most.
The GTMnow Podcast shares how the best in tech build, scale and invest.
Visit gtmnow.com for more episodes, The GTMnow Newsletter editions, and other content.
GTMnow is run by GTMfund, an early-stage venture firm made up of 350+ go-to-market executives from the fastest-growing companies.
