Summary

Amin Vahdat sees the AI buildout as “100x what the internet was,” while Jeetu Patel calls it the internet, space race, and Manhattan Project “all put into one.” Patel says the buildout is being grossly underestimated because geopolitical, economic, national-security, and speed imperatives are arriving together.
Google has seven TPU generations in production, and its seven- and eight-year-old TPUs still run at 100% utilization. Vahdat says power, land transformation, permitting, and supply-chain delivery are the constraints; even if trillions are committed, buyers may be unable to “cash all those checks” for 3–5 years.
Power scarcity is already dictating where data centers get built and how they connect. Patel says enterprises remain early in the required re-racking and power-density transition, while Cisco has launched “scale-across” systems intended to let facilities act as one logical data center even when separated by a distance rendered in the transcript as “8 900 km.”
The next computing stack will be co-designed from silicon through software and will be “unrecognizable” within five years. Vahdat calls this a “golden age of specialization”: for certain computations, TPUs deliver 10–100x the efficiency per watt of CPUs, but even the best teams need 2½ years to move a specialized architecture from concept into production.
Networking is becoming both a primary bottleneck and a force multiplier for scarce compute power. Workloads can swing by tens or hundreds of megawatts between computation and communication, yet the most expensive network capacity may be needed only 5% of the time—an unresolved architecture and utilization problem.
Inference efficiency is improving by 10x and 100x, but users consume the gain by demanding smarter models and longer autonomous runs. “Intelligence per dollar” rises while total cost can still increase; prefill, decode, reinforcement learning, and inference-native infrastructure therefore create different hardware, memory, latency, and networking tradeoffs.
AI-assisted engineering is moving from demos into migrations that previously looked economically impossible. Google once estimated a Bigtable-to-Spanner migration at “seven staff millennia” and abandoned it; Vahdat says AI has since assisted instruction-set migration across Google’s codebase. Patel hopes Cisco’s 25,000 engineers can reach 2–3x productivity within a year, while warning that tools dismissed today must be retested within four weeks.
Patel’s startup warning is that “thin wrappers” around another company’s model will have short-lived durability. He favors models tightly coupled to product feedback plus dynamic routing between a product’s models and foundation models; Vahdat expects agents and productive image-and-video systems—not merely better text chat—to become transformative over the next 12 months.

Deep dive

1. Demand is outrunning the industry’s ability to deploy capital

Vahdat’s historical comparison is categorical: “I’ve seen nothing like this. I’m fairly certain no one’s seen anything like this.” The internet buildout was large; AI is “100x what the internet was,” with potentially comparable upside.
Google has seven TPU generations in production, yet its seven- and eight-year-old units remain at 100% utilization. Customers accept whatever capacity exists, while use cases are turned away because there is no room for them.
Patel says enterprises are still nascent in rebuilding data centers for much higher power requirements per rack, while hyperscalers and neo-clouds are further along. Because power is scarce in any single location, data centers are being built where power is available rather than bringing power to existing sites.
Vahdat accepts the projected trillions of dollars of spending but doubts the industry can “cash all those checks” quickly. Power, land transformation, permits, and supply-chain delivery could constrain deployment for 3–5 years; hardware can be purchased just in time, while space and power assets have depreciation cycles of 25–40 years.

2. AI is forcing a full-stack reinvention, not a return to mainframes

Raghu Raghuram’s provocation was that NVIDIA is bringing back the mainframe. Vahdat pushed back: a pool of 16,384 GPUs or a 9,000-chip TPU pod remains dynamically divisible—one workload might need 256 chips, another 100,000—so scale-out software survives.
Vahdat expects the hardware-to-software stack to be “unrecognizable” in five years. Google’s earlier transition paired commodity clusters with Bigtable, Spanner, GFS, Borg, and Colossus; this cycle will likewise require hardware and software to be co-designed.
Patel’s organizational extension is that suppliers spanning “the physics to the semantics” must “work like one company” while remaining an open ecosystem rather than a walled garden. He points to the deep design partnerships required between companies and hyperscalers before products are even contracted.
Vahdat calls this a “golden age of specialization.” For certain computations, a TPU is 10–100x more efficient per watt than a CPU, but the best teams still face a 2½-year speed limit from concept to production. He expects more specialization once the current pace slows, because the power, cost, and space savings are too large to ignore.
He also says geography will shape architectures: China’s 7-nanometer chips could be paired with abundant power and engineering resources, while other regions might favor more power-efficient 2-nanometer designs. Regulatory frameworks and regional expansion could therefore produce different architectures.

3. Networks must serve extreme bursts without becoming stranded assets

Vahdat says bandwidth inside a building is becoming a primary bottleneck, while networking itself consumes relatively little power: “Spend a little bit here, get way more there.” Known communication patterns may also permit designs that do not require the full generality of a packet switch.
The harder problem is burstiness. Utilities can see tens or hundreds of megawatts of power demand stop for network communication and then restart for computation; large-scale pretraining may require enormous networks only 5% of the time, before the newest chips migrate to different sites.
Patel calls networking the “force multiplier”: every kilowatt saved moving a packet can be reassigned to a GPU. He expects scale-up, scale-out, and scale-across fabrics, plus inference-native infrastructure rather than training systems merely repurposed for inference.
Cisco has launched silicon, a chip, and a system for scale-across networking so that two data centers can act as one logical data center when separated by a distance rendered in the transcript as “8 900 km.” The design reflects the dispersion caused by power scarcity.
His competitive warning was unusually direct: if systems are only wrappers around Broadcom, “you’ve got a monopoly that’s going to be a very predatory one.” Cisco’s strategic case is silicon choice and diversity at high consumption volumes.

4. AI coding wins are real, but adoption requires continual retesting

Google once calculated that moving from Bigtable to Spanner would consume “seven staff millennia”; the opportunity cost produced the verdict, “Long live Bigtable.” Vahdat now points to AI-assisted instruction-set migration from x86 to ARM across Google’s codebase, making it agnostic to future architectures such as RISC-V.
He also says Google has carried out migrations such as TensorFlow to JAX internally with AI assistance, many factors faster, while acknowledging that some tasks still exceed the tools’ capabilities.
Patel reports strong results in code migration using Codex Cloud, Cursor, and some Windsurf; debugging, especially through CLIs; and zero-to-one front ends. Older code, particularly further down the infrastructure stack, is harder.
Patel says the challenge is more a cultural reset than a technical one. He told 150 distinguished engineers to assume the tools will be dramatically better within six months and to revisit failures within four weeks rather than shelving them for 6–9 months. He hopes Cisco’s 25,000 engineers reach 2–3x productivity within a year.
Sales preparation, legal contract review, and product marketing are also producing good results. Patel says a first ChatGPT-based competitive-analysis draft is often better than starting from a blank page.

5. Durable products will couple models, feedback, and routing

Pressed not to predict merely that models improve, Vahdat said models and agent frameworks are becoming “scary good.” The ability to have work “go quite right for quite long” could be transformative over the coming 12 months.
Patel’s founder advice: do not build “thin wrappers around models that are other people’s models.” Durable products will tightly couple a model to the product so that product feedback improves it, while using foundation models for other tasks and dynamically routing between them.
Patel also says Cisco has momentum after being viewed as a legacy or has-been company. He promises innovation across silicon, networking, security, observability, the data platform, and applications, and invites collaboration with startups.
Vahdat expects images and video to repeat text models’ journey from novelty to utility: two-and-a-half to three years ago, text models were good for amusing haikus; the next phase should make image and video models productivity, educational, and learning tools rather than novelty generators.

Is AI Stalling Out? Cutting Through Capabilities Confusion, w/ Erik Torenberg, from the a16z Podcast Google DeepMind Developers: How Nano Banana Was Made