How is AI Different Than Other Technology Waves? (With Bret Taylor and Clay Bavor) [ACQ2]

2025-08-18 - source - Read full transcript

Ben Gilbert (host)David Rosenthal (host)Bret TaylorClay Bavor

Key insights

AI adoption is compounding on top of every prior infrastructure wave (PCs, internet, smartphones), letting new AI products reach mass scale far faster than any earlier technology.

Bavor notes the first website went live around 1991 but it took until 2002 to reach 10% weekly global usage, while ChatGPT reached roughly that same penetration in about 25 months and hit 100 million users in about two months - because it inherited a smartphone- and internet-connected population that PCs and the early web had to build from scratch.

ai-adoption-speed

If personal AI agents start acting on users' behalf, the entire internet demand funnel (awareness, interest, decision, action) could collapse, restructuring how demand generation, advertising, and brand relationships work.

Taylor argues that today's internet splits into demand generation (social/ads), demand fulfillment (search), and transactions (commerce), and that once agents - not people - are discovering and deciding on services, it becomes unclear who advertisers are even generating demand for, or which brands retain enough equity to avoid being commoditized by platform intermediaries.

ai-adoption-speed

Sierra charges customers only when an AI agent fully resolves an issue with no human involvement, rejecting seat-based and per-message/token pricing because none of the traditional SaaS pricing units map to the value an agent actually delivers.

Bavor explains they started from first principles - agents don't just make people marginally more productive like traditional software, they complete the job outright - so Sierra prices against completed outcomes (a resolved customer issue, an attached premium-delivery upsell, a completed sale) rather than seats or usage, deliberately taking on the performance risk itself.

ai-business-models

Taylor calls foundation models 'the fastest deteriorating asset of all time' and argues applied AI companies building their own frontier models is a bad bet, comparable to a SaaS startup building its own data center instead of renting cloud.

Clay Bavor points to the small set of 2023-era AI startups (Adept, Character.AI, Inflection) that tried to be foundation-model-first and 'ended differently,' and Taylor argues that spending tens to hundreds of millions of dollars pre-training a model whose competitive value may last only a week makes far less sense than licensing frontier models from the few labs that can amortize that capex at scale.

ai-business-models

The asset-light, high-margin software business model investors have rewarded for two decades is being replaced by AI companies carrying massive ongoing infrastructure capex, raising real barriers to entry alongside real financial risk.

Ben notes tech multiples were built on low-capex software economics that no longer hold; Bavor cites Google raising its infrastructure spend guidance from $75 billion to $85 billion the same day of the recording, and Taylor argues the capex is a genuine barrier to entry even as it makes DCF-style valuation harder to reason about.

ai-business-models

Sierra's engineering culture treats a bad AI output as a signal to fix the model's input context, not the output itself, because only fixing root context compounds into durable leverage.

Taylor frames the company as 'a machine to produce happy customers': when Cursor generates incorrect code, the team's rule is to diagnose what context the AI lacked that would have produced correct code, rather than hand-patching the bad result, since patching individual outputs doesn't scale.

ai-agents-and-labor

Both founders see AI turning intelligence from a scarce resource into a plentiful one, comparable to historical transitions in energy and food - a shift they expect to be net positive long-term but personally destabilizing in the near term.

Taylor draws the parallel to electricity and modern food distribution making once-scarce resources unremarkable, and argues that because many people (including the software engineers building this technology) tie their identity to their intelligence or expertise, the transition will feel uncomfortable even as it democratizes access to things like legal, medical, and educational expertise.

ai-agents-and-labor

New technology waves swing enterprise buying from 'best of platform' toward 'best of breed' because incumbents structurally struggle to adopt both the new technical architecture and the new business model at once.

Taylor argues procurement defaults to safe, bundled incumbent platforms once a category is commoditized, but genuinely new capabilities (like AI agents today) temporarily favor specialist vendors like Sierra because incumbents' software-as-a-service business model and technical stack aren't built for outcome-based pricing or agent architectures - a window he says lasts only until the new capability itself becomes commonplace.

technology-wave-parallels

The early-2010s Google-versus-Facebook social-network panic (Google's 'earthquake memo,' Google+) is offered as a direct historical parallel to today's AI race: incumbents treating an emerging wave as existential even when the specific bet ultimately fails.

Ben and Taylor recall Google reorienting company priorities around social for roughly three years out of fear of Facebook, a battle that in hindsight was 'a nothing burger' for the underlying threat it addressed - but Taylor argues the same all-in, existential posture is rational and recurring whenever companies face a wave of technology this size, AI included.

technology-wave-parallels

At startup scale, both founders stay deep in day-to-day operational detail - a leadership mode they contrast sharply with the cultural change management required to roll out AI adoption across an 80,000-person company like Salesforce.

Bavor says he is personally in the weeds of pricing proposals and contract language while Taylor checks in production code on weekends; Taylor adds that mandating tool adoption (e.g., using Cursor or ChatGPT deep research) is trivial at Sierra's size but nearly impossible to force consistently across a large, established workforce with entrenched habits.

startup-leadership

Taylor and Bavor describe their partnership as 20 years in the making, formed through Google's APM program, and argue that splitting a founder's emotional load between two trusted people makes the stress of building a company more bearable.

Taylor says he tried unsuccessfully to recruit Bavor to join him at every subsequent company since leaving Google in 2007, finally succeeding over a lunch in December 2022 right as ChatGPT launched; both describe being able to 'rant at the sky' to each other as a core reason they don't think they could have built Sierra solo.

startup-leadership

Media referenced

Studio Ghibli AI image trend - other - David cites the viral wave of Studio-Ghibli-style AI-generated images as an example of ChatGPT becoming a 'front door' distribution channel that surfaces new capabilities to mass audiences on its own

Companies

Sierra - Taylor and Bavor's company, co-founded in March 2023 and publicly launched in 2024; builds customer-facing AI agents and is the episode's central case study for AI business models
OpenAI - Maker of ChatGPT; Taylor is its board chairman. Cited repeatedly as the fastest-growing consumer product in history and the touchstone for foundation-model economics
Google - Both guests started their careers there as APMs under Marissa Mayer; extensive discussion of the Google-Facebook social rivalry, Google+, and Google's current AI capex increases
Facebook - Taylor was CTO there after Google; discussed as the other side of the early-2010s social-network rivalry with Google (Google+ vs. Facebook)
Salesforce - Taylor was co-CEO after Salesforce acquired his company Quip; used as the reference point for classic SaaS pricing that Sierra's model deliberately rejects
Quip - Taylor's earlier startup, acquired by Salesforce
Friend Feed - Taylor's startup before Quip, acquired by Facebook
Twitter - Mentioned in passing; Taylor was chairman of its board during its final period as a public company
Cursor - AI coding assistant Sierra's engineers use heavily; central to the 'fix the machine, not the output' example about correcting model context rather than individual bad outputs
ADT - Sierra customer; example of an agent troubleshooting alarm panels and mailing replacement batteries autonomously
SiriusXM - Sierra customer (Harmony agent); example of an agent sending a satellite signal to refresh a car's encryption keys with no human involved
OluKai - Sierra customer; example of an agent inspecting warranty photos and autonomously shipping replacement flip-flops, which caused office-wide celebration
Deepseek - Cited for publishing research on reducing the cost of training frontier-class models
Hugging Face - David references a prior Acquired conversation with CEO Clem Delangue's view that applied AI companies need their own foundation models, which Taylor strongly disagrees with
Harvey - Named as a peer vertical AI application company, building AI agents for the legal profession
Writer - Named as a peer vertical AI application company, building AI agents for marketing
Amazon Web Services - Used as the cloud-rental analogy for why applied AI companies should lease foundation models rather than build their own, and separately as an example of a durable infrastructure business with real barriers to entry
Microsoft - Cited as an example of a platform incumbent that fumbled mobile despite Windows Phone/Mobile being technically ahead, and later did well in cloud
Silicon Graphics (SGI) - Taylor recalls Google moving into SGI's former campus while SGI, then dying, still paid for food in the same cafeteria - used to illustrate how short tech company lifespans can be
Sun Microsystems - Facebook later moved into Sun's former campus after Sun's decline, paralleling the SGI story

Techniques and frameworks

outcome-based (resolution-based) pricing - Sierra charges customers only when an AI agent fully resolves a customer issue with zero human involvement, rejecting seat-based or per-message/token pricing because neither maps to the value an agent actually delivers
fix the machine, not the output - Sierra's internal engineering philosophy: when a tool like Cursor produces bad code, the fix is correcting the context/inputs that produced the bad output, not patching that one output, because only the former compounds into leverage
distillation - Taking a very high-parameter-count model and producing a smaller model that retains most of its quality at much lower inference cost
best of platform vs. best of breed - Enterprise buying pattern where mature/commoditized technology categories favor safe, bundled incumbent platforms, while genuinely new technology waves temporarily favor specialist best-of-breed vendors until the new capability itself becomes commoditized

Summary

Acquired hosts Ben Gilbert and David Rosenthal sit down with Bret Taylor and Clay Bavor, longtime Google APM-program friends who are now co-founders of Sierra, an AI customer-agent company. Both guests helped extensively with Acquired's two-part Google history, and the conversation ranges from vocabulary (will "agent" become the era's lasting noun, the way "app" did?) to the deepest strategic question of the AI wave: is this genuinely transformational, or is it just better, more powerful software riding the same curve as every prior wave? Taylor and Bavor land on "both" - a shift as significant as the internet or industrial revolution, layered on top of infrastructure (smartphones, the internet, cloud) built by the previous waves, which is why ChatGPT and Sierra can grow faster than any earlier technology in history.

Much of the episode is a case study of Sierra's own choices. Rather than seat-based or usage-based SaaS pricing, Sierra bills only when an agent fully resolves a customer issue with no human involved - "resolution-based pricing" - which Taylor and Bavor say forces the company to bear performance risk but also perfectly aligns their incentives with customers' outcomes. They pair this with a firm view that applied AI companies should not build their own foundation models: Taylor calls frontier models "the fastest deteriorating asset of all time" and argues licensing from a small number of scaled foundation labs is the only viable strategy, pushing back directly on a claim made by Hugging Face's Clem Delangue in an earlier Acquired episode. The conversation also covers the return of heavy infrastructure capex to a software industry that spent two decades prizing asset-light margins, using Google's same-day $10 billion capex guidance increase as a live example.

The pair connects Sierra's engineering culture to this same discipline: when an AI coding tool produces a bad output, the team's rule is "fix the machine, not the output" - correct the model's context rather than patch individual mistakes, since only the former scales. They extend the idea to labor and identity more broadly, framing AI as making intelligence plentiful the way electricity and modern agriculture made energy and food plentiful, a shift they expect to be net positive but personally destabilizing, especially for technologists whose self-worth is tied to expertise.

Taylor and Bavor use their own careers - Bavor's 18 years at Google across ads, Workspace, Google Labs, and AR/VR; Taylor's run through Google Maps, Friend Feed, Facebook CTO, Quip, and Salesforce co-CEO - to draw a broader theory of enterprise technology cycles: markets swing from "best of platform" toward "best of breed" whenever a new wave arrives, because incumbents are structurally slow to adopt both new architectures and new business models, and swing back only once the new capability becomes commoditized. They close on leadership and partnership: both describe staying deep in day-to-day detail at Sierra's small scale, contrast that with the much harder cultural work of driving AI adoption across a large organization, and reflect on a 20-year effort to finally found a company together, sparked by a December 2022 lunch that happened to land the week ChatGPT launched.

Notable Quotes

"Fix the machine, don't just fix the output of the machine." - Bret Taylor

"It's a very, very expensive carton of milk." - Clay Bavor, on building your own foundation model

"I think probably all of those things are true... this could be great in 10 years but really hard over the next two years." - Bret Taylor

"We're building a machine to produce happy customers." - Clay Bavor

"I've been trying to work with Clay unsuccessfully every single day since I left Google in 2007." - Bret Taylor