The a16z Show

David George × Gavin Baker

AI demand accelerated through July and August, with usage concentrated among possibly sub-10 million heavy users and no leader identifying a worsening quantitative metric.Nebius suggests nine-to-ten-month paybacks on $50B-per-gigawatt builds, supported by 50-60% customer prepayments; undersupply through ’28 could lift token prices, while shifting capacity to training could cut lab revenue from $480B to $120B.

Ben Horowitz × Martin Casado × Raghu Raghuram × Erik Torenberg

a16z’s new AI infrastructure fund captures a founder migration into hardware, with top-founder hardware pitches rising from roughly 3–5% to “north of 20% or 30%.”Hyperscaler capex, booked-out GPUs and resale premiums support opportunities across chips, power and cooling, while grid shortages, regulation and uncontrolled agent spending remain constraints.

Elena Burger × Matt Bornstein × Simon Mo

vLLM has become the execution layer linking more than 1,000 open-weight model architectures to GPUs from NVIDIA, AMD, Google, Amazon, Intel, and others, making inference a strategic systems layer.Open weights increasingly offer controllable latency, data, security, and fine-tuning rather than merely cheaper tokens, while licensing and moderation pressures test whether the ecosystem can fund frontier development and trusted specialized use cases.

Marc Andreessen × Erik Torenberg

Andreessen sees AI three years into an 80-year revolution, with customer revenue validating unusually rapid adoption across an already-built global distribution network.Falling token costs, competing chips, smaller models, and US-China rivalry could expand demand dramatically, while state regulation remains a major unresolved risk.

Naveen Rao × Matt Bornstein

Unconventional is pursuing physical learning circuits, arguing that analog dynamics could complement digital computing for stochastic, time-dependent workloads rather than simply extend matrix multiplication.The thesis targets an energy wall: U.S. data centers use roughly 4% of the national grid, while some estimates call for 400 additional gigawatts over the next decade.Commercial significance depends on finding a manufacturable paradigm within five years, with TSMC partnership, a potentially very large analog prototype and unresolved AGI claims creating major execution risk.

Raghu Raghuram × Amin Vahdat × Jeetu Patel

AI infrastructure demand is outpacing deployment: Google’s seven- and eight-year-old TPUs remain at 100% utilization, while power, permitting, land, and supply chains may constrain trillions in committed spending for 3–5 years.Specialized silicon, distributed networking, and inference-native systems could deliver 10–100x efficiency gains, but bursty workloads and the need for full-stack co-design make utilization, architecture, and durable product differentiation central risks.

Erik Torenberg × Dylan Patel × Sarah Wang × Guido Appenzeller

Nvidia’s $5 billion Intel investment, following SoftBank’s $2 billion and the U.S. government’s $10 billion, could lower Intel’s cost of capital and redraw PC and data-center competition, though Patel says Intel still needs roughly $50 billion.Huawei has credible 7 nm designs and ambitious custom-HBM products, but HBM3 yields, etch capacity, and domestic volume remain unresolved as Nvidia’s upside depends on $450–500 billion of hyperscaler capex rather than further share gains.

Dylan Patel × Erin Price-Wright × Guido Appenzeller × Erik Torenberg

GPT-5’s router is an economic release, directing simple queries to mini models while reserving “ungodly amounts of compute” for transactions OpenAI could monetize through agentic commerce.Flat-rate subscriptions face heavy-user losses, while custom silicon threatens Nvidia mainly if demand stays concentrated among hyperscalers; powered sites, grid equipment, and Intel’s capital needs remain near-term constraints on broader deployment.

Martin Casado × Raghu Raghuram × Jeetu Patel

VMware’s cycle shows how a software abstraction can disrupt an incumbent before AWS retains the virtual machine and captures developers, a constituency VMware “had no idea how to work with.”Cisco’s reset targets market in nine months, $1 billion in three to four years, and 8–10 repeatable winners by combining protected two-pizza teams with scaled distribution.AI could expand infrastructure demand 100–1,000x as agents create sustained inference workloads, but vertical integration must remain open enough to include competitors such as Microsoft Teams.

Steph Smith × Martin Casado × Steven Sinofsky

DeepSeek R1’s permissive MIT-like license and released reasoning traces enable broad adoption and distillation into smaller models, shifting strategic advantage toward distribution and “AGI in your pocket.”Model value may migrate from benchmark leadership to stateful workflows as competitors catch up, while scale-out expands endpoints without eliminating hyperscale compute.The clearest policy signal is that chip and open-source restrictions did not prevent capable Chinese research, making faster domestic innovation and permissionless diffusion the unresolved US advantage.