Apr 23, 2026 · 36m · catalyst
Inside Google’s massive AI capex (live)
0:00 / 0:00
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this live episode of Catalyst, host Shayle Kann interviews Google's Chief Technologist for AI Infrastructure, Amin Vahdat, exploring how Google manages its massive $175B+ capital expenditure through data center scaling, grid integration, demand flexibility, and full-stack hardware-software co-design.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Shayle holds 38.1% of the talking time here. How this is scored →
Shayle as informed peer 5.9 Guest teaching 5.1 Guest disagreement 2.1 Shayle pushing back 3.4
speaking balance: gold is Shayle, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 26:48 Refusal to pick a single bottleneck
Vahdat explicitly rejects Kann's forced ranking heuristic despite repeated prompting, maintaining that power, chips, and labor are simultaneously maxed out.
Hardest push from Shayle ▶ 26:29 Forcing the ranking questionKann refuses to accept Vahdat's diplomatic non-answer and reframes the question with a hypothetical budget increase from Sundar Pichai to demand a definitive answer.
Biggest teaching moment ▶ 13:00 The economics of two nines versus four ninesVahdat provides a detailed breakdown of software service costs versus compute costs, demonstrating why customers willingly accept 3.65 days of annual downtime for double compute capacity.
Shayle holds their own ▶ 30:07 Deconstructing the edge compute thesisKann uses technical deduction to challenge edge data center demand, arguing that if safety-critical compute stays on-device in vehicles, non-critical latency can easily route back to distant hyperscalers.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Shayle as informed peer | Guest teaching | Guest disagreement | Shayle pushing back | Why |
|---|---|---|---|---|---|---|
| Data Center Scale: Training Demands vs. Inference Locality | 6 | 5 | 1 | 2 | Kann opens with detailed macro CapEx context comparing Google's budget to national GDP and Vogtle nuclear plant costs, then questions inference scale requirements. Vahdat explains how older gigawatt training clusters naturally lifecycle into serving capacity, while Kann notes the minimum viable footprint. | |
| Reassessing Data Center Reliability and Availability Standards | 6 | 6 | 2 | 3 | Kann asks why data center reliability standards are kept so rigidly high despite immense CapEx and generator supply chain costs. Vahdat educates on the math of four nines versus two nines availability, showing why internal customers prefer double compute over high uptime. | |
| Behind-the-Meter Power, Grid Integration, and Demand Response | 6 | 6 | 2 | 4 | Kann probes the risk of stranded generation assets from behind-the-meter bridge power deployments. Vahdat explains Google's gigawatt demand response agreements with utilities and willingness to brown down during peak grid stress. | |
| Mid-Show Sponsor Messages: Bloom Energy, Engie, and Energy Hub | 6 | 5 | 1 | 2 | Following mid-show sponsor reads, Kann questions whether data centers function as dynamic microgrids and whether Google's extreme vertical integration gives it an unfair advantage. Vahdat confirms microgrids and software workload balancing are crucial co-design elements. | |
| Evaluating AI Expansion Bottlenecks: Chips, Power, and Construction | 5 | 3 | 5 | 6 | Kann pushes Vahdat to rank the biggest rate limiter among power, chips, and labor/construction. When Vahdat resists picking one, Kann pushes back and forces him to answer with a hypothetical budget scenario, though Vahdat still insists all three are simultaneous bottlenecks. | |
| Physical AI Infrastructure: Autonomous Vehicles, Edge, and Robotics | 6 | 5 | 2 | 4 | Kann tests the physical AI and robotics architecture, arguing that putting safety-critical compute on-device diminishes the need for localized edge data centers. Vahdat agrees for autonomous vehicles but notes factory robotics could retain edge setups. | |
| Driving Long-Term CapEx Efficiency and Data Center Density | 6 | 6 | 2 | 3 | Kann explores data center density and cost reductions, noting his assumption that linear density was already maxed out. Vahdat explains how the massive power disparity between disks and accelerators forces building-level specialization, and corrects Kann regarding Google's workload smoothing. |
Statements from this episode (16)
Google's 2026 CapEx is 5-7x the entire US transmission spend
“We spend about 25 or thirty-five billion dollars a year in capex on transmission. Electricity transmission infrastructure in the United States, so this is five to seven times that amount, just from Google, just in one year.”
Vahdat: AI inference does not require gigawatt-scale data center capacity
“You don't strictly need a gigawatt of capacity to be able to do useful work. You probably don't even need a hundred megawatts of capacity. It gets a little bit more interesting than that because of, let's say, co-located compute and storage and networking and …”
Vahdat: Interactive AI models will require geographically distributed inference infrastructure
“So as these services become more interactive, as they become more efficient, and that, that is still going to be a journey. We're not there today. You're going to want to have geographic locality. That's also going to impact reliability, because again, you can…”
Vahdat: AI footprint will favor many medium data centers over gigawatt hubs
“It'll really come down to geographic locality and probably a medium number of medium-sized data centers. Sorry for the whatever lack of precision there, but medium number of medium-sized data centers augmented with a small number of large data centers.”
Vahdat: The AI industry should accept lower-reliability power delivery
“No, it is not intrinsic and we should be thinking about lower reliability power delivery overall.”
Vahdat: Many Google data centers target four nines of availability
“Many of our data centers aim for four nines of, I mean, minutes of downtime a year maximum, which as you said has a large amount of costs associated with it.”
Vahdat: Google trades uptime for capacity in customer co-design
“Without saying too much, it's happening. I would say there's actually the co-design there with our customers at Google has been one of our sources of significant efficiency.”
Google reached 1 GW in utility demand response agreements by March 2026
“In March, we actually hit a significant milestone in agreements with utilities for a gigawatt of demand response across our fleet.”
Google uses behind-the-meter power while waiting years for grid transmission
“We like behind the meter if it means that we get the capacity up most quickly, but we're always going to look to invest with the utilities to bring the transmission. Maybe it's a year after. Maybe it's two years after, right?”
Vahdat: The AI data center community is underinvested in microgrid software control
“The microgrid and the software control here is going to be absolutely key. And this is a place where I think we as a community are under invested today.”
Vahdat: Google co-designs TPUs directly with power sources, buildings, and Gemini
“In other words, for us, for let's say our TPUs, we co-design them with a building. We co-design them with the power generation source. We co-design them with the DeepMind team that builds Gemini models. So it's the software above, the models above that, the ch…”
Vahdat: AI scaling bottlenecks constantly rotate between labor, power, and chips
“When delivering the end-to-end, we unfortunately don't have the luxury of focusing on a single limiter. I would say very sincerely and honestly at 10 a.m. It's labor, at noon it's power, and at two p.m. It's chips every single day.”
Vahdat: Safety-critical physical AI compute will have to run on-device
“Without talking about any specific use case, I believe that a lot of it's going to have to be on device and dedicated to that use case.”
Vahdat: Dominant compute costs force data centers to sacrifice fungibility
“If compute is not the dominant portion of your cost, you actually want to have flexibility and fungibility. When compute becomes a more dominant portion of your cost, you now actually are thinking, okay, what am I going to do for this year, next year, and the …”