The $1 Trillion AI Buildout | State of Markets

David George, Sarah Wang, Alex Immerman, Santiago Rodriguez

AI-assisted summary of The $1 Trillion AI Buildout | State of Markets — original source. Check the original for full context.

Source published: 2026-09-30

Key ideas

Hyperscaler capital expenditures are projected to reach $780 billion in 2026 and exceed $1 trillion annually starting in 2027, driven by unyielding compute demand and multi-step agent workloads. (6:55)
AI startups captured 86% of total U.S. venture capital deal value in 2026, while the top six private tech giants reached an aggregate valuation of $2.4 trillion, surpassing the combined market cap of all prior-decade tech IPOs. (48:36)
Despite widespread experimentation—with 69% of S&P 500 firms deploying AI initiatives—only 2% currently track longitudinal bottom-line metrics, highlighting a significant application-layer monetization gap. (0:00)
Inference costs have fallen by 10x to 100x through context caching, fine-tuning, and smart routing, allowing software vendors to decouple customer pricing from raw token costs and improve margins. (26:26)

Detailed notes

## Background Published on September 30, 2026, as an audio podcast episode by Andreessen Horowitz (a16z) partners David George, Sarah Wang, Alex Immerman, and Santiago Rodriguez, this State of Markets briefing breaks down 25 key charts detailing the modern technology landscape. The discussion explores the macroeconomic drivers behind the approaching $1 trillion annual hyperscaler CapEx cycle, enterprise software adaptation, and the next wave of investments across robotics, defense, and physical infrastructure.

## TL;DR The a16z team argues that the current tech cycle is fundamentally driven by earnings rather than multiple expansion, with hyperscalers reinvesting virtually all operating cash flow into compute, power, and physical supply chains. At the application layer, enterprise adoption is shifting from basic prompt queries to complex multi-step agents, where context caching and model routing decouple software margins from token costs. While consumer AI subscription adoption remains in its early innings, specialized vertical software and private mega-caps continue to compound at historic rates. As a result, venture capital is increasingly concentrating in capital-intensive physical frontiers—including defense tech, autonomous vehicles, and robotics.

## Atmosphere & style The conversation carries a crisp, data-dense, and analytical tone typical of a high-level venture partnership debrief. The hosts interact with collegial ease, seamlessly tossing commentary back and forth to connect macro capital expenditure numbers with granular operating metrics from portfolio companies. Rather than relying on breathless hype, the dialogue feels grounded, pragmatic, and measured, balancing sober evaluations of enterprise ROI against a palpable, long-term bullish conviction in physical computing buildouts.

## Bulletin - `06:55` Hyperscaler capital expenditures are projected to reach $780 billion in 2026 and exceed $1 trillion annually starting in 2027, driven by unyielding compute demand and multi-step agent workloads. - `48:36` AI startups captured 86% of total U.S. venture capital deal value in 2026, while the top six private tech giants reached an aggregate valuation of $2.4 trillion, surpassing the combined market cap of all prior-decade tech IPOs. - `00:00` Despite widespread experimentation—with 69% of S&P 500 firms deploying AI initiatives—only 2% currently track longitudinal bottom-line metrics, highlighting a significant application-layer monetization gap. - `26:26` Inference costs have fallen by 10x to 100x through context caching, fine-tuning, and smart routing, allowing software vendors to decouple customer pricing from raw token costs and improve margins.

## Key takeaways - `05:27` Recent market expansion has been driven primarily by 15% corporate earnings growth rather than multiple expansion, with S&P 500 forward price-to-earnings multiples sitting below 20x. - `03:07` Tech infrastructure and enterprise R&D now account for roughly 55% of all non-residential capital spending in the United States, outpacing historical benchmarks like the 19th-century railroad expansion as a share of GDP. - `10:00` Hyperscalers are reinvesting nearly all operating cash flow into compute infrastructure, building up a combined cloud backlog of roughly $1.7 trillion across Microsoft, Alphabet, and Amazon. - `15:00` Counter to common concerns, data center expansion has lowered residential utility rates by 40 basis points for every 10% capacity growth by spreading fixed grid infrastructure costs across larger volumes. - `20:00` AI tooling adoption shows extreme power-user divergence, with median spend among the top 1% of users reaching 8x that of the top 10%, and AI-native startups spending up to 10% of headcount budget on tools. - `25:00` Enterprises generate far greater long-term ROI by directing AI spend toward net-new customer revenue rather than using engineering talent to rebuild internal systems of record for minor licensing savings. - `26:26` Agentic token consumption has increased 14x as applications move to multi-step reasoning, self-correction, and speculative execution across chained model calls. - `27:50` Smart routing and domain-specific fine-tuning enable applications to cut serving spend by 35% to 60% while maintaining latency and task performance. - `29:16` Consumer AI applications exhibit 'smiling' retention curves, where user retention dips initially but curves upward over time as underlying foundation models continuously improve. - `31:30` Autonomous purchasing agents threaten legacy high-margin search advertising franchises, prompting platforms like Instacart to integrate while Amazon restricts direct catalog access. - `36:30` Public SaaS growth has slowed, with fewer than five public software companies sustaining 30% revenue growth, leading to valuation bifurcation favoring cybersecurity and vertical AI specialists over horizontal tools. - `48:36` Venture capital is increasingly moving up the stack and into the physical realm, with robotics projected to potentially become larger than LLMs over a 3- to 5-year lag. - `50:00` Autonomous vehicles operating at safety profiles roughly 14 times safer than human drivers are poised to disrupt the broader passenger vehicle market beyond existing ride-hailing volumes.

## Detailed walkthrough ### `00:00` Overview: The 25 Key Charts and the State of Markets

The discussion opens by setting the scale of the current technological and economic wave. Across the globe, high-tech capital expenditure has reached unprecedented levels: eight of the top ten most valuable companies in the world are now American technology giants. Since the release of ChatGPT nearly four years prior, the broader market has soared 90%, generating an annualized return of 17%. While an instinctual reaction to such staggering returns is to assume the market is overheating and overdue for a painful pullback, the fundamentals tell a much more grounded story. The current environment marks a transition into what can be described as a "new age of atoms," where the demand for physical infrastructure goes far beyond data center walls into power generation, clean water, roads, grid modernization, and transit. Global infrastructure needs through 2040 are projected to hit an astounding $90 trillion.

At the enterprise level, a massive divergence has emerged. While 69% of S&P 500 companies actively discuss or experiment with AI initiatives, only about 2% have translated those efforts into measurable, tracked enterprise metrics. Despite that gap, real revenue and measurable savings are already flowing to the frontrunners, even though broader adoption remains in its infancy. David George introduces his Andreessen Horowitz growth team partners—Sarah Wang, Alex Immerman, and Santiago Rodriguez—to unpack 25 core slides from their annual State of Markets deck. Their presentation evaluates the macro drivers, hyperscaler capital expenditure, model progress, application layers, falling inference costs, and the question of whether this massive spending cycle is genuinely justified by underlying economics.

- **Global Infrastructure Spending:** Estimated at $90 trillion needed globally through 2040 across power, transit, water, and compute. - **Enterprise Adoption Gap:** 69% of S&P 500 companies cite AI initiatives, yet only 2% track longitudinal impact metrics on their bottom line. - **The Core Debate:** “Today's opportunity is so much around taking these capabilities and harnessing them to build reliable services. That's all great, but is it a bubble?”

---

`03:07` Tech as the "Everything Cycle"

David George kicks off the slide deck by declaring that technology has morphed into an "everything cycle." High-tech equipment, enterprise software, and research and development now make up approximately 55% of all non-residential capital spending in the United States. Rather than remaining isolated in Silicon Valley, technology has become the primary economic engine pulling the rest of the domestic industrial complex along with it, dictating investments in electrical utilities, specialized cooling, heavy concrete construction, and skilled physical labor. Tech firms now command roughly 40% of the aggregate equity value of the entire U.S. stock market, and eight of the world's ten most valuable public companies are American technology firms.

```mermaid flowchart TD accTitle: The Everything Cycle Capital Flow accDescr: Capital flows from high-tech R&D and hyperscaler CapEx into physical infrastructure, chips, and industrial demand.

subgraph TechCore["Tech Capital Engine"] A["55% of US CapEx: Tech, Software, R&D"] --> B["Hyperscalers & AI Labs"] end

subgraph PhysicalBuildout["Physical Infrastructure Needs"] B --> C["Power & Energy Grids"] B --> D["Cooling & Water Systems"] B --> E["Heavy Construction & Labor"] B --> F["Semiconductors & Foundries"] end

subgraph Downstream["Economic Transformation"] C & D & E & F --> G["Enterprise Automation & Services"] end ```

The sheer magnitude of this investment wave eclipses the historical benchmarks commonly cited by financial historians. David George points out that cumulative funding for foundation model developers alone has crossed $350 billion, and the current AI buildout as a percentage of gross domestic product has now surpassed the historic expansion of the American railroad network in the nineteenth century. Looking ahead across the next five to ten years, George and the team anticipate cumulative investment figures could easily expand by another 20x. Alex Immerman observes that this is the physical realization of Marc Andreessen’s original thesis: “Yeah, this does feel like the next chapter of Mark's software's eating the world. Over the last 15 years, software has transformed all these industries, but only a small fraction of the population could build it. None of us could. But today, look at us. We all have a handful of automations running every night.”

- **CapEx Share:** 55% of U.S. capital expenditures are now driven by tech equipment, software, and R&D. - **Market Dominance:** Technology equities comprise roughly 40% of total U.S. stock market capitalization. - **Historical Comparison:** Current AI infrastructure buildout has officially surpassed the 19th-century railroad boom as a percentage of GDP. - **Foundation Model Capital:** Model companies have collectively secured more than $350 billion in funding to date.

---

`05:27` Earnings-Driven Growth vs. Multiple Expansion

Addressing the prevailing market anxiety that AI is a runaway financial bubble, the team examines current equity valuations. While headline indices have continued to print all-time highs—rising approximately 20% in the recent cycle—underlying valuation multiples have actually declined by around 20%. The broad market rally is not the result of irrational speculative euphoria or multiple expansion, but is instead underpinned by resilient, fundamental corporate earnings growth. The price-to-earnings multiple for the S&P 500 currently sits below 20x forward earnings, while corporate earnings themselves have climbed roughly 15%.

Alex Immerman underscores the sharp contrast between today’s pricing and previous speculative peaks, such as the dot-com bubble of 2000 or the pandemic run-up in 2021. Back then, unprofitable companies with minimal revenue commanded forward multiples well over 100x PE. Today, high-margin market leaders underpin the indices, and deeply cyclical hardware and memory manufacturers supporting the AI backbone frequently trade at conservative valuations of only six to seven times forward earnings. Reflecting on the steady compounding since late 2022, Immerman notes: “And I think anytime there's been a 17% annualized gross for four years, the natural instinct is, well, that's got to come down, right? We're definitely in a hot period. But when you compare that to, as you said, the market trading for below 20 times earnings, growing 15%. This definitely feels a little different than maybe the 2021 period that was recent or the 2000 period when multiples in growth were not really going together.”

---

`06:55` Hyperscaler CapEx Approaching $1 Trillion and the "Model Buster" Demand

The scale of capital deployed by the five dominant hyperscalers—Alphabet, Amazon, Meta, Microsoft, and Oracle—is accelerating at a pace that repeatedly demolishes Wall Street models. In 2025, their combined capital expenditure stood at $416 billion; for 2026, projections have ballooned to approximately $780 billion. Internal forecasts and consensus estimates indicate their collective annual spend will cross $1 trillion annually starting in 2027. The hyperscalers are essentially reinvesting virtually all of their near-term operating cash flow back into the ground to ensure they do not run out of capacity.

```mermaid flowchart LR accTitle: Hyperscaler CapEx Trajectory accDescr: Capital expenditure from Alphabet, Amazon, Meta, Microsoft, and Oracle climbing toward one trillion dollars.

A["2025 CapEx: $416 Billion"] --> B["2026 Projected CapEx: $780 Billion"] B --> C["2027+ Run-Rate: >$1.0 Trillion Annually"] ```

Unlike historical infrastructure deployments that had to construct end-user distribution networks from scratch, this cycle builds directly on top of ubiquitous global smartphone ownership, broadband internet, and mature cloud ecosystems. Consequently, software applications can capture hundreds of millions of users overnight. Moreover, the transition from basic text prompts to complex autonomous agents has radically increased compute consumption. Sarah Wang explains that workloads are no longer confined to human typing speeds or isolated questions: “And I think it's an important point to double click on just because we've moved well beyond this model of occasional queries. Right. If you think about the rise of agents, you have parallel tasks, you have long running tasks. And so, I mean, Alex, you mentioned this previously, right? You have tasks going into the evening, during the day when you're doing other things. And so if you think about agents autonomously writing code, searching, carrying out these tasks, it just gives these platforms a much larger compute requirement than ever thought before.”

Santiago Rodriguez adds that market observers repeatedly fall into the trap of assuming investment rates must plateau, yet capacity continues to get absorbed immediately upon delivery. As an example of compute demand outrunning supply, the group highlights OpenAI’s aggressive capacity bets. A year prior, Sam Altman and CFO Sarah Friar were heavily criticized by market commentators for taking on reckless, multi-billion-dollar compute commitments; today, those commitments look remarkably prescient. Demand remains so insatiable that OpenAI recently had to pause sign-ups for its premium subscription tiers when compute bottlenecks emerged, proving that consumer and enterprise appetite for compute is effectively a "model buster" that shows no signs of saturation.

`10:00` The Hyperscaler J-Curve, Supply Constraints, and the New Age of Atoms

The discussion opens on the sheer scale of compute demand across the artificial intelligence supply chain. Every vendor, contractor, and operator echoes the same reality: demand continues to outpace available supply by a wide margin. In several critical segments of the data center supply chain, wait times have stretched out so far that customers cannot secure access to materials or physical equipment until 2028. At the same time, this spending is supported by verifiable downstream enterprise demand rather than pure speculation. Microsoft, Alphabet (Google), and Amazon have accumulated an estimated $1.7 trillion in combined cloud backlog. While heavy capital expenditures are currently depressing free cash flow across these platforms, market consensus anticipates a significant cash flow recovery beginning in 2028, with accelerating expansion in subsequent years.

```mermaid flowchart TD accTitle: Hyperscaler Capital Expenditure Cycle accDescr: Diagram showing how hyperscaler CapEx flows through supply chain physical buildouts into cash flow recovery. subgraph CapitalAllocation["Capital Outlay Phase"] A["$1.7T Combined Cloud Backlog (MSFT, GOOG, AMZN)"] --> B["Hyperscaler CapEx Surge"] B --> C["Free Cash Flow Compression (Near-term J-Curve dip)"] end subgraph PhysicalBuildout["Data Center & Physical Supply Chain"] B --> D["Physical Shells & Long-Lead Equipment (Backlogged to 2028)"] D --> E["GPU / TPU Procurement (Extended economic life)"] end subgraph CashFlowRealization["Economic Value & Recovery"] E --> F["High Margin Consumer/Enterprise Surplus"] F --> G["Projected FCF Inflection (2028 and beyond)"] end ```

Amazon's recent earnings call highlighted the economic mechanics governing this buildout, specifically the multi-year J-curve dynamic. Constructing data center shells and securing power takes significant upfront time and capital, followed by substantial chip purchases. However, the operational and economic lifespan of modern GPUs and TPUs has proven considerably longer and more durable than industry forecasters initially modeled. While public market investors have occasionally penalized hyperscalers for taking on debt to fund massive CapEx, the hosts note that hesitation has carried a steep penalty: “The ones who have blinked and been less aggressive have regretted it.” Every layer of the stack—from chip manufacturers commanding premium pricing to the cloud providers and software vendors serving downstream end users—continues to harvest high margins while delivering immense consumer surplus.

This surge in hyperscaler CapEx directly represents an order book for the broader physical economy. Software investing has traditionally centered on monetizing pure intellectual property through a "build once, sell infinitely" paradigm. Today, technology venture capital has pivoted into an industrial boom requiring mastery over physical supply chains, equipment financing, vendor contracts, capacity planning, and demand forecasting. This shift marks what the hosts identify as the "New Age of Atoms." Through 2040, global infrastructure investment requirements are projected to reach $90 trillion across data centers, power generation, water systems, roads, and transit networks. Portfolio companies illustrate this industrial transformation: - Anduril has constructed a massive manufacturing facility spanning 87 football fields. - Waymo is aggressively expanding physical depots across regional markets. - SpaceX is deploying an estimated $100 billion investment in Louisiana. - As the team observes, borrowing Elon Musk’s framing, founders emerging from Tesla and SpaceX succeed in this environment because they understand that “executing on the factory ends up being a massive competitive advantage as they scale.”

---

`15:00` Dispelling Grid Myths, Model Economics, and the Enterprise Adoption Gap

The hosts tackle prevailing public narratives regarding the environmental and utility impact of data center development, specifically the fear that large facilities consume localized water reserves, drive out residential communities, and cause consumer electricity bills to skyrocket. Contrary to standard assumptions, adding large, stable baseload customers can reduce residential energy expenses. Because an electric power grid operates as a shared fixed-cost asset composed of substations, transmission lines, and utility poles, data centers distribute those massive fixed capital costs over a significantly higher volume of consumed electricity units.

- A recent U.S. study established that for every 10% expansion in data center capacity within a utility territory, residential electricity rates declined by 40 basis points. - Real-world community case studies demonstrate shared benefits: former Goldman Sachs partner and Meta executive Dina Powell recently outlined on a podcast how Meta partnered directly with local Louisiana communities around their infrastructure site to actively lower regional electricity costs.

Turning to the model and application layers, inference costs are dropping rapidly, unlocking broad commercial viability for AI agents. Applying the economic logic of Jevons paradox—where efficiency gains sharply drive up total consumption—cost drops of two orders of magnitude are causing agent-driven workflows to become economically feasible across novel operational categories. Startups like Typeface (leveraging architectures like JEV) reflect this rapid compression of unit costs, providing an attractive environment for software margins and growth.

```mermaid flowchart LR accTitle: Enterprise AI Adoption Funnel accDescr: Funnel illustrating the drop-off from live AI pilots to longitudinal metric tracking in the S&P 500. A["Live AI Deployments (69% of S&P 500)"] --> B["Quantifiable Economic Impact (30% of S&P 500)"] B --> C["Longitudinal Metric Tracking Over Time (2% of S&P 500)"] ```

The scale achieved by the leading foundational model providers remains historically unprecedented. Annualized revenue metrics for OpenAI and Anthropic continue to outstrip the growth trajectories of the most successful enterprise software companies in software history. At the general partner offsites for Andreessen Horowitz, the revenue tracking slide comparing these leading labs against legacy software giants must be updated on a monthly cadence to keep up with actual revenue additions.

Yet despite these staggering top-line figures, overall enterprise market penetration remains in its infancy. Unlike the 2008 mobile shift—which simply required estimating the penetration of iPhone hardware among consumers—AI adoption requires measuring both user penetration and the expanding depth of wallet share per user over time. A look at enterprise integration within the S&P 500 demonstrates this massive execution gap: - **69%** of S&P 500 organizations have active, live AI deployments in production. - Only **30%** of those enterprises can demonstrate a clear, quantifiable operational or financial impact from those deployments. - A mere **2%** of enterprises actually track verified AI performance metrics longitudinally over time.

Most large enterprises remain limited to basic rollouts like Microsoft 365 Copilot. The major upside for the application ecosystem lies in converting superficial pilots into durable, core recurring workflows: “This gap between what the models can do and how they're being used is what makes me so excited about the application layer.”

`20:00` The Application Layer Gap and the Power-User Pull-Away

Reflecting on a conversation with Ali Ghodsi of Databricks and Martine, the team highlights a core dynamic in software today: foundation models possess vast general knowledge about the world, but virtually no specific context about an individual enterprise. Bridging that chasm is the defining opportunity for application-layer companies. Even technologically sophisticated organizations face this hurdle. For example, Revolut—despite possessing an elite internal engineering team—partnered with ElevenLabs to harness their voice models, securely binding those capabilities into core customer accounts and banking workflows so customer support calls can resolve real account issues safely and efficiently. Today’s commercial upside lies in taking raw capabilities and engineering them into reliable, enterprise-grade services.

```mermaid flowchart LR accTitle: Enterprise AI Application Bridge accDescr: Diagram showing how application companies bridge general foundation models with private enterprise data to build reliable customer workflows.

subgraph Inputs["Inputs"] LLM["Foundation Models (World Knowledge)"] Data["Enterprise Data (Private Context)"] end

subgraph AppLayer["Application Layer Engine"] Bridge["Security, Routing & Workflow Integration"] end

subgraph Outputs["Enterprise Outcomes"] Prod["Reliable Services (e.g., Revolut Banking Voice Support)"] end

LLM --> Bridge Data --> Bridge Bridge --> Prod ```

At the same time, enterprise adoption remains in its infancy, yet a dramatic divergence is emerging between casual experimenters and intensive power users. Drawing on data from YipitData, the team notes that median AI vendor spend among the top 1% of users is roughly eight times higher than that of the top 10%—meaning the top 1% accounts for nearly as much spend as the 2nd through 10th percentiles combined. Inside fully "AI-pilled" or AI-native companies, the top tier of users consumes between $7,500 and $9,000 per month on AI tooling, compared to median users whose spend hovers between $200 and $400 monthly (equivalent to stacking one or two subscriptions), representing a greater than 20x gap between median and power users.

- Forward-leaning Fortune 500 corporations currently allocate roughly 1% of their headcount budget toward AI tooling. - The most aggressive, AI-native portfolio companies are spending as much as 10% of their headcount budget on AI tooling. - Procuring developer tools like Cursor or Cognition is merely day one; the operational runway between initial tool procurement and pervasive, org-wide adoption remains vast.

---

`22:45` Quantifiable ROI: Cost to Serve, Onboarding Velocity, and Incumbent Monetization

The market is finally moving beyond speculative proofs of concept into hard, quantifiable case studies reported by public companies. These real-world metrics fall neatly into distinct efficiency and top-line expansion vectors. On the cost optimization side, digital bank Chime publicly reported reducing its cost to serve by more than 10% year-over-year for four consecutive years—a compounding efficiency gain of nearly 50%, supported by portfolio company Decagon. This cost structure reduction provides massive competitive leverage: companies can offer sharper pricing, expand service boundaries, or recycle those operating margins directly back into growth engines.

```mermaid flowchart TD accTitle: AI ROI Dual Engine accDescr: Visualizing how AI drives both cost-to-serve reductions and top-line merchant acceleration.

subgraph CostEff["Cost Reduction Vector"] Chime["Chime Cost to Serve"] --> Decagon["Decagon Support Automation"] Decagon --> Margin["~50% 4-Year Compounded Drop (Reinvest into Price & Growth)"] end

subgraph RevVector["Revenue Acceleration Vector"] Shopify["Shopify AI Sidekick"] --> Merchant["Faster Merchant Onboarding"] Merchant --> Retention["+8% Reaching 5 Orders in 15 Days (Strong Retention Milestone)"] end ```

On the revenue generation side, Shopify rolled out its AI Sidekick to streamline onboarding for new merchants. The platform observed an 8% increase in the percentage of merchants reaching five orders within their first 15 days on the platform—a critical internal milestone, as merchants crossing that threshold demonstrate substantially higher lifetime retention. Incumbents are capturing upside just as aggressively as startups; ServiceNow, for instance, has logged over $1 billion in AI annual contract value (ACV) alongside a 9x increase in agentic workflow deployments across its customer base.

---

`25:00` The Capital Allocation Litmus Test: Growth vs. Internal Re-engineering

Where executive teams channel their next marginal dollar of AI capital expenditure serves as an essential litmus test for how large they believe their market really is. Drawing on strategic conversations with forward-leaning portfolio leaders like those at Stripe, the team contrasts companies investing in net-new, revenue-generating customer products with those pursuing internal cost reductions. Alex Immerman crystallizes this strategic divide:

> “Like the opportunity to drive revenue growth is unbounded to the upside. Whereas the opportunity for cost improvement, you know, yes, you could take that and reinvest it. But that's sort of a latent opportunity that will continue to exist.”

He goes on to question executive priorities:

> “And if you think that the best and highest use of your resources is to optimize your cost structure, what does that say about the revenue opportunity for you?”

Cost reduction itself contains important nuances. Lowering cost to serve can structurally transform unit economics and fuel competitive reinvestment. However, the team critiques misallocated engineering efforts from earlier cycles, where development teams spent months attempting to rebuild internal systems of record from scratch merely to save a few hundred thousand dollars in licensing fees—a poor application of scarce technical talent when unbounded product opportunities remain untapped.

---

`26:26` Agentic Workflows, Caching, and Inference Cost Deflation

Agentic usage is exploding as software shifts from single-turn prompts to autonomous, multi-step tasks. Because complex tasks demand iterative planning, multiple internal reasoning steps, and sequential model calls, agent token consumption on platforms like OpenRouter has surged 14x. However, the cost profile of executing these chained workflows is collapsing thanks to prompt caching and context re-use. Rather than processing static enterprise context and lengthy prompt chains from scratch on every step, cached background data dramatically lowers the marginal cost per call.

Hebbia provides a vivid example of this shift: by leaning into context caching, their financial analysis and chat workloads became 10x cheaper to run. A 10x to 100x cost decline fundamentally transforms software engineering behavior. When inference costs collapse, systems can afford to execute speculative execution, check their own work, run iterative self-correction loops, and leverage external tools without blowing up unit economics. Tasks previously deemed far too expensive or unreliable to automate suddenly become commercially viable, completely re-anchoring developer expectations around agentic reliability and latency.

---

`27:50` Smart Routing, Fine-Tuning, and Decoupled Software Margins

To optimize the frontier between latency, output quality, and cost, builders are adopting specialized model-orchestration architectures, notably smart routing and targeted fine-tuning:

```mermaid flowchart TD accTitle: Inference Optimization Strategies accDescr: Showing how Databricks smart routing and Harvey or EliseAI fine-tuning reduce inference cost and latency.

subgraph Routing["Smart Routing Strategy (Databricks)"] Req["User Request"] --> Router["Smart Router"] Router -->|"Light Task"| SmallM["Smaller / Faster Model"] Router -->|"Complex Task"| LargeM["Strong Frontier Model"] Router --> Save["35% Lower Cost + Higher Success"] end

subgraph FineTuning["Domain Fine-Tuning Strategy (Harvey / EliseAI)"] Base["Smaller Base Model"] --> FT["Task Fine-Tuning"] FT --> LowLat["60% Cheaper & Low Latency (Enables Real-Time Audio)"] end ```

- **Smart Model Routing:** Instead of defaulting to frontier flagships for every interaction, Databricks implements intelligent routing to dynamically match individual tasks with the optimal model. Their smart router successfully solves more tasks while slashing aggregate inference spend by 35% compared to directing all queries through the most capable individual model. - **Fine-Tuning Smaller Models:** Companies like Harvey (legal tech) and EliseAI (vertical real estate workflows) fine-tune smaller, open or domain-adapted models. This tactic lowered EliseAI's serving costs by 60% while drastically cutting latency, making interactive, real-time voice and audio applications feasible.

For application-layer software businesses, this architectural evolution decouples pricing from underlying compute. The true economic unit is no longer raw tokens, but successfully completing the customer's job. By charging the customer for a finished outcome while dynamically routing underlying tasks to cheaper, fine-tuned, or cached models, elite application companies are engineering massive software gross-margin expansion.

---

`29:16` Consumer AI: Subscriptions and the "Smiling" Retention Curve

Turning from the enterprise landscape to consumer software, the conversation highlights that consumer AI adoption is still in its earliest innings. A recent consumer survey revealed that only just over 2% of U.S. households currently pay for an AI subscription. While standard recurring monthly fees may not ultimately remain the primary monetization model for mainstream households over the long term, subscription mechanics have driven the vast majority of consumer AI revenue to date.

Crucially, the retention cohort dynamics of leading consumer AI products look radically different from standard consumer software. While conventional consumer applications suffer from cohorts that decay over time and eventually settle into a flatline, top consumer AI applications exhibit a "smiling" retention curve: retention dips initially, flattens, and then begins curving back upward. As foundation models rapidly improve in the background, users who previously dropped off return to re-engage with a noticeably more capable product, generating a compounding retention profile rarely observed in software history.

`30:00` Consumer AI Adoption, Retention Curves, and the Death of Screen Time

The conversation opens with an examination of consumer AI subscriptions relative to traditional consumer tech benchmarks. Looking at the data on household penetration, the current numbers for consumer AI subscriptions are strikingly small when contrasted with mature consumer giants: Amazon Prime counts over 200 million households, and Netflix reaches roughly 70 million. Alex Immerman references partner Josh Elman’s counter-definition of consumer AI: “consumer AI is what I use for my daily life and not necessarily what I expense.” From that perspective, it is hardly shocking that 97% of households have yet to adopt dedicated, paid consumer AI subscriptions.

```mermaid flowchart LR accTitle: Evolution of Consumer Engagement Metrics accDescr: Shift from legacy social media engagement tracked by active screen time to AI agent interactions running background tasks.

subgraph Legacy["Legacy Consumer Apps"] L1["Facebook / Instagram / TikTok"] --> L2["30 to 60 Mins/Day Screen Time"] L2 --> L3["Ad Impressions & Direct Clicks"] end

subgraph Agentic["Agentic AI Consumer Model"] A1["ChatGPT / Muse / Instinct"] --> A2["Background Deep Research Tasks"] A2 --> A3["Zero-Click Task Completion & GMV"] end ```

The team turns to engagement and retention metrics. While Santiago Rodriguez points out the classic preference for “smiling retention curves”—where cohorts re-engage at higher rates over time—the nature of user engagement is fundamentally decoupling from legacy metrics. Historically, mega-platforms like Facebook, Instagram, TikTok, and Snap defined success by capturing 30 to 60 minutes per day of active user screen time. Highly capable consumer AI applications, however, aim to be proactive and persistent, executing tasks quietly in the background rather than forcing users to stare at a screen. As one partner notes about measuring software performance: “Yeah, it's going to make our job harder when we can't, you know, look at engagement via outside in data because the agents are working on the background instead of us just looking at like screen time on a social media application.” An investor might kick off a complex research agent, set off on a drive, and return to find the job finished—a huge value unlock that registers zero traditional app screen time.

- Amazon Prime reaches more than 200 million households, whereas Netflix sits around 70 million households. - Legacy social apps capture between 30 and 60 minutes of daily active user attention, whereas autonomous agent workflows bypass active interface time entirely.

---

`31:30` Marketplaces, Agents, and the Vulnerability of Ad Profit Pools

Consumer discovery is facing a generational shift as emerging agentic platforms like Muse and Instinct rapidly attract query volume alongside ChatGPT, steadily siphoning search volume away from legacy engines. This transition forces consumer marketplaces and discovery engines to answer two critical strategic questions: 1. How much incremental demand and order volume can an AI agent direct to the platform? 2. How much of the existing profit pool relies on owning consumer discovery and monetizing customer touchpoints via advertising?

The contrasting strategies of Amazon and Instacart illustrate this tension. Amazon barred Muse from connecting directly to its catalog, whereas Instacart welcomed it. For both businesses (excluding AWS in Amazon's case), advertising revenue exceeds total operating profit. For Amazon, which already commands enormous consumer mindshare, connecting to an agent offers negligible incremental customer acquisition or gross merchandise value (GMV) to offset the risk of disintermediation. Conversely, online grocery penetration remains relatively low, meaning Instacart has significant runway to acquire net-new volume. The bull case for agents centers on eliminating transaction friction: if complex checkout funnels and manual searches are replaced by a frictionless agent executing in the background, overall transaction volume expands. Dinners get ordered, trips get booked, and uncompleted checkout carts convert to completed GMV.

Yet, this shift creates major uncertainty for monetization. Amazon operates an advertising business generating over $70 billion annually at extraordinarily high flow-through margins, underpinned entirely by users visiting the site and clicking sponsored search results. Google faces a related structural question, though its search moat has proven surprisingly durable over the past two years. Google’s most lucrative ad keywords—commercial intent queries like searching for insurance or hotel bookings—were initially insulated because generative AI tools could not reliably complete transactions. As David George observes regarding the vulnerability of search real estate: “Because they were things like, I need to buy insurance or find me a hotel in this city or things like that that you can very directly monetize. But that the AI was not yet capable of going to take action for on your behalf. If that changes, that could be a very different dynamic.”

- Amazon runs an advertising franchise generating over $70 billion per year, which represents high-margin flow-through that exceeds its non-AWS retail operating profits. - High-intent advertising categories (e.g., insurance, hospitality, travel booking) have historically insulated Google Search revenues because AI engines lacked reliable transactional execution capabilities.

---

`35:15` Positive-Sum Expansion vs. Relocating Profit Pools

Evaluating whether agentic commerce poses a net threat or benefit to the broader economy depends on whether one takes a zero-sum or positive-sum view. While market commentary often sounds alarms about marketplaces losing high-margin advertising real estate, there is a strong positive-sum counterpart: ad budgets may simply migrate toward agent-native formats or smarter targeting tools, while frictionless purchasing drives greater aggregate consumption.

The panel emphasizes that both dynamics can occur concurrently: specific corporate profit pools will undoubtedly be disrupted, but macroeconomic productivity and overall spending will expand. The broader financial markets have already signaled this optimism, as aggregate tech market capitalization has expanded net-positive: market capitalization gains in players like Meta have significantly outstripped the paper losses experienced by vulnerable discovery platforms. Over a three-year horizon, the true measure of success will be whether agentic tools generate measurable macroeconomic productivity enhancements that trickle down into higher discretionary consumer spending.

---

`36:30` The SaaS Growth Reckoning: Incumbent Playbooks and Growth Acceleration

The conversation pivots to the public enterprise software landscape, where financial profiles have drastically transformed. The public SaaS universe has migrated toward slower-growing, highly cash-generative operations. Today, roughly 75% of public software companies are profitable, but only 30% continue to grow revenue at 20% or higher. More dramatically, if the benchmark is raised to 30% year-over-year revenue growth, fewer than five public companies qualify. This stands in stark contrast to late-stage venture and private software markets, where nearly every prospective portfolio investment is expected to compound at rates well above 30%.

```mermaid flowchart TD accTitle: SaaS Incumbent Playbook for Growth Acceleration accDescr: The defensive and offensive levers enterprise software incumbents use to re-accelerate revenue.

subgraph Moat["Defensive Core"] D1["98%+ Gross Dollar Retention"] D2["Entrenched Customer Workflow"] D3["Established Enterprise Distribution"] end

subgraph Offensive["Offensive Strategy"] O1["Embed Native AI Agent Workflows"] O2["Tap Dedicated Enterprise AI Budgets"] O3["Target 10%+ Revenue Acceleration"] end

Moat --> Offensive ```

This deceleration reshapes conversations with public and late-stage SaaS founders. Incumbents maintain defensive foundations: customer gross dollar retention (GDR) routinely holds near 98%, proving that existing enterprise software accounts are not simply churning off to point-solution startups. However, enterprise software budgets are heavily captured by strategic AI line items, requiring incumbents to mount an aggressive offensive strategy. Founders are leveraging their pre-existing enterprise distribution and entrenched user data to ship proprietary AI features that re-accelerate top-line revenue rather than settling for pure margin optimization.

Referencing an earlier thesis regarding the divergent paths available to software companies in the AI era, the team notes that founders almost universally reject managing for terminal cash flows in favor of reigniting top-line compounding. Alex Immerman articulates the bar required for software companies to justify their multiples in an AI-driven budget landscape: “And so we had said, you know, let's target 10% plus acceleration, which is a high number. But with a magical product, given the budgets that are available to AI, is seemingly doable.” Over the next 12 to 18 months, leading enterprise incumbents are poised to reveal whether AI integrations can deliver that elusive double-digit growth inflection.

- Approximately 75% of public SaaS software companies are currently profitable. - Only 30% of public software businesses are expanding revenue at 20% or more year-over-year. - The count of public SaaS businesses growing at 30% or more has dropped below 5 companies. - Enterprise incumbents continue to post durable gross dollar retention figures hovering around 98%. - The strategic hurdle rate for SaaS reinvention centers on driving a 10%+ acceleration in annual top-line revenue growth by capturing net-new AI budget allocations.

`40:00` Public SaaS Bifurcation, Incumbent Resilience, and the Private Market Vertical AI Boom

The public software landscape has fragmented along stark lines rather than moving as a monolithic block. Cybersecurity and observability have emerged as major standouts, while vertical software has generally proven far more resilient than horizontal application software. As Sarah Wang frames it, assessing whether a software category holds up, climbs, or slips into secular decline comes down to examining how artificial intelligence reshapes the end customer's underlying necessity for that software. While AI-related cyber risks receive constant public attention, the proliferation of new autonomous agents and expanded software footprints creates massive net-new security surfaces and monitoring requirements. Rather than displacing incumbents, this environment is fueling increased enterprise demand for proven market leaders like CrowdStrike. In contrast, standard application software sits in the crosshairs of severe workflow redesign. Databricks CEO Ali Ghodsi frequently characterizes this tension by asking what sits first on the “chopping block of AI,” contrasting endangered workflows with the essential foundational infrastructure AI requires.

```mermaid flowchart TD accTitle: AI Impact on Software Categories accDescr: Diagram contrasting how AI dynamics create tailwinds for security and vertical software while threatening horizontal workflows.

subgraph Tailwinds ["Expanding Demand"] A["Proliferation of AI Agents"] --> B["New Attack Surfaces & Monitoring Needs"] B --> C["Incumbent Boost (e.g., Cybersecurity & Observability)"] D["Deep Industry Workflows"] --> E["Vertical AI Applications (e.g., Harvey, Abridge, EliseAI)"] end

subgraph Headwinds ["Workflow Disruption"] F["Generic Business Logic"] --> G["Horizontal SaaS Applications"] G --> H["'Chopping Block of AI' (Seat Compression & Automation)"] end ```

Alex Immerman underscores that these public market realities mirror what the team tracks in private ventures. As agents access deeper corporate architectures and execute actions across disparate databases, the OpenAI and Hugging Face security incidents served as an eye-opening wake-up call for CISOs, placing cybersecurity front and center. Simultaneously, specialized vertical AI businesses—including legal platform Harvey, medical scribe and documentation leader Abridge, and property management AI EliseAI—are putting up growth velocities that surpass any historical software precedent in their target markets. Enterprise buyers recognize that enterprise value stems from far more than raw foundational intelligence; it relies on orchestration layers and deeply integrated, industry-specific workflows that fit neatly around the models.

Reflecting on the volatility of public enterprise multiples, Santiago Rodriguez points out that the market began the year panicking over an imminent "SaaS apocalypse" before recovering back to flat baseline levels. Yet beneath that flat index lies total bifurcation between perceived AI winners and losers. As Santiago Rodriguez observes: “The public market sometimes simplifies things a bit too much, but to your point, the AI winners are not just the ones that can improve their cost structures.” Instead, market outperformance accrues strictly to businesses that accelerate topline revenue by seizing net-new customer spend rather than relying purely on internal automation efficiencies.

- Cybersecurity and observability software have substantially outperformed horizontal software due to an influx of agent-driven monitoring and security demands. - Public market bellwether CrowdStrike exemplifies legacy platform resilience amid expanding threat surfaces. - Fast-growing private vertical AI champions like Harvey, Abridge, and EliseAI are outpacing historical software growth benchmarks by tailoring specialized orchestration to complex industry workflows. - The broader public software index has recovered its early-year "SaaS apocalypse" losses, but trading multiples remain sharply bifurcated based on whether a platform can capture net-new software spend.

`43:10` The Stripe Renaissance, $2.4T in Private Mega-Caps, and the Founder-Led Strategy

Beneath public market skepticism, private transactional data paints a dramatically different operating picture. Stripe’s SaaS customer analytics demonstrate a surprising narrative violation: growth is reaccelerating across both early-stage startups and established mature enterprises, a phenomenon Stripe has labeled the software “renaissance.” This private growth engine is powering private tech valuations to historic scales. The combined valuation of the top six private tech giants—Anthropic, OpenAI, Databricks, Stripe, Waymo, and Revolut—has climbed to roughly $2.4 trillion. That total eclipses the $1.7 trillion aggregate market capitalization of all tech IPOs completed over the preceding decade (excluding SpaceX) and nearly rivals the entire Russell 2000 index. Santiago Rodriguez notes that these late-stage champions possess enough balance-sheet power to aggressively reinvest free cash flow into structural long-term growth rather than curtailing spending to cater to quarterly public earnings expectations.

David George elaborates on how a16z guides management teams on weighing private autonomy against the public markets, noting that the traditional IPO equation has completely transformed. As David George puts it: “One of the things that we've talked about and that I've written about is that the founder is the asset class at this point.” By remaining private longer, visionary founders are shielded from short-term public earnings scrutiny and can execute aggressive, capital-intensive bets that carry multi-year payback horizons. While public market outliers like Mark Zuckerberg and Elon Musk can muscle major capital reallocations through the public eye, Meta’s stock price dropping below $100 per share during peak Wall Street skepticism over AR/VR spending shows the intense pressure public scrutiny brings. In contrast, Databricks and Stripe have utilized private capital insulation to develop massive parallel product lines, translating bold capital expenditure directly into sustained topline revenue acceleration.

- Stripe internal platform data reveals accelerating software customer revenue across both emerging and legacy enterprise tiers, officially termed the SaaS “renaissance.” - The aggregate private valuation of Anthropic, OpenAI, Databricks, Stripe, Waymo, and Revolut stands at approximately $2.4 trillion. - Tech IPOs from the prior decade (excluding SpaceX) account for just $1.7 trillion in combined public market value, demonstrating how much enterprise value accumulation has shifted to private markets. - Founder-led governance allows mega-cap private companies to pursue expansive, high-conviction product bets without having their operational plans derailed by quarterly public market sentiment.

`45:45` The Mechanics of Staying Private: Public Trust, Employee Tenders, and Zero-Discount Secondaries

While remaining private confers distinct R&D advantages, transitioning to the public markets remains an important corporate financing milestone rather than an ultimate endgame. Companies increasingly operate with public-level financial rigor while staying private, implementing meticulous metric tracking across every division. In certain enterprise contexts, going public provides an undeniable boost in customer trust and institutional credibility. Sarah Wang points to business travel platform Navan, which unlocked corporate sales reacceleration after its public listing as large enterprise procurement departments gained comfort from audited public disclosures. Furthermore, when prolonged capital requirements outstrip private syndicates, accessing deep public market equity pools becomes inevitable.

```mermaid flowchart LR accTitle: Private Secondary Tender Offer Dynamics accDescr: Flowchart illustrating how private liquidity mechanisms function for employee retention and capitalization.

subgraph Capital ["Secondary Tender Event"] A["Company-Sponsored Tender"] --> B["Employee Retention & Liquidity"] A --> C["Updated Valuation Benchmark"] end

subgraph MarketOutcomes ["Operational Impact"] B --> D["58% Participation Rate (High Employee Conviction)"] C --> E["Zero Discount to Primary Round"] C --> F["Refreshed Equity Currency for M&A & Hiring"] end ```

To address the historic friction of staying private—namely, competitive employee compensation against the quarterly liquidity of public-company RSUs—private mega-caps increasingly orchestrate regular secondary tender offers. Analyzing secondary liquidity data from Carta, Santiago Rodriguez reveals an unexpected trend: employee tender participation rates average just 58%. Rather than rushing to cash out accumulated stock, more than 40% of eligible employee equity is intentionally held back by workers who retain deep personal conviction in their employers' trajectory.

David George and Santiago Rodriguez emphasize that frequent secondary rounds serve two strategic functions: keeping equity packages compelling against liquid public competitors, and continuously refreshing private share valuations to provide usable acquisition currency for M&A. Crucially, the steep 20% to 40% secondary discounts that plagued private transactions between 2022 and 2024 have completely evaporated. Secondary trades in top-tier companies are transacting at a flat zero percent discount to their primary rounds, driven by a deep bench of newly entered institutional capital eager to match original issue terms.

- Corporate transparency from public market listings can drive sales acceleration in trust-sensitive sectors, as demonstrated by enterprise adoption trends at Navan. - Carta benchmark data reveals average employee participation in structured private tender offers is only 58%, countering assumptions of widespread employee cashing out. - Structured secondary liquidity allows private firms to rival the net-tax quarterly liquidity appeal of public company RSUs. - Private secondary pricing discounts have narrowed from the heavy markdowns seen between 2022 and 2024 to nearly a 0% discount against the most recent primary preferred financing rounds.

`48:36` Venture Capital Concentration and the Emerging Frontiers of Robotics and Agents

The allocation of venture capital has reached an unprecedented state of concentration around artificial intelligence. Santiago Rodriguez highlights that AI-focused enterprises accounted for a staggering 86% of total US venture capital deal value in 2026, marking a swift climb from 65% in 2025. While market reporting focuses disproportionately on the top of the foundation model food chain—principally OpenAI, Anthropic, and Elon Musk’s xAI—the venture investable surface area has expanded significantly. Substantial venture dollars are dispersing into AI-native consumer applications, automated professional services, custom semiconductor development, grid power generation, and national defense tech.

Looking ahead, the investment team points to several high-conviction frontiers where they are deploying capital. Autonomous, long-running consumer agents represent a massive market unlock: by delegating complex, friction-laden tasks that consumers routinely avoid, agents create entirely new consumption opportunities and economic value. Beyond pure software agents, physical robotics stands out as an even larger technological paradigm shift. As David George states: “Robotics is an area where we're spending a lot of time and attention. We happen to think that it could be even larger than LLMs, but probably three to five years earlier.” Over the next five years, the convergence of advanced vision models, embodied physical systems, and autonomous mobility platforms is set to trigger a massive surge in private investment across physical automation.

- AI-related companies captured 86% of total US venture deal value in 2026, expanding upward from 65% recorded across 2025. - Beyond leading model providers, capital allocation is broadening across enterprise apps, consumer tooling, silicon design, power infrastructure, and defense systems. - Autonomous, long-running consumer agents are unlocking new economic value by automating friction-heavy tasks that users historically abandoned. - Robotics is projected to eclipse the total market footprint of software LLMs, with deployment timelines tracking roughly three to five years behind large language model adoption.

`50:00` Autonomy, AI in Biology, and Personal Health

Taking a broader step back from the pure software layer, the discussion shifts toward massive physical industries undergoing transformation, starting with transportation. While the tech world spends the vast majority of its time talking strictly about generative AI models, the global automotive and transportation sector represents one of the largest economic markets on earth. The progress in autonomous mobility has reached a tangible turning point: self-driving technology works today, and its expansion potential is immense. Rides taken in ride-hailing networks like Uber and Lyft currently account for only about 1% of all vehicle miles traveled in the United States. With autonomous vehicle fleets proving to be roughly 14 times safer than human drivers, that share is poised to expand by at least an order of magnitude in the coming years. Furthermore, out of the roughly 17 million new passenger vehicles sold in the United States each year, autonomous driving capability is expected to become standard across the board over the next decade.

At the intersection of artificial intelligence and the life sciences, AI and biology represent another monumental frontier. While researchers and machine learning teams inside frontier labs frequently highlight the potential of biological foundation models, the real promise centers on accelerating novel drug discovery and uncovering cures for some of the world's most debilitating and complex diseases. Alongside institutional biology, personal health is entering an unprecedented phase of hyper-personalization. For health optimizers and individuals tracking their own biomarkers, there has historically never been an integrated platform capable of ingesting exhaustive personal biometric and medical data to return tailored, highly actionable health guidance.

- Ride-hailing services such as Uber and Lyft currently account for approximately 1% of total vehicle miles traveled in the United States. - Autonomous driving networks are currently demonstrating safety profiles approximately 14 times safer than human drivers. - The U.S. auto market sells roughly 17 million new vehicles per year, with the expectation that new vehicles will transition to full autonomy over the next 10 years. - Key emerging application areas highlighted include AI-driven drug discovery for debilitating illnesses and hyper-personalized health intelligence based on aggregated individual data.

> "If you look at miles traveled by Uber and Lyft or in Ubers and Lyfts, it represents about 1% of miles traveled in the U S I think with, um, full networks of autonomous driving cars that are 14 times safer than human drivers."

`51:30` Enterprise Diffusion, American Dynamism, and the Productivity Boom

Beyond software engineering and developer tooling—which served as the initial proving ground for large language models—the diffusion of AI across the broader enterprise is entering its next phase of operational adoption. At the same time, substantial investment opportunities exist outside the pure AI domain, most notably across the "American Dynamism" ecosystem. This category focuses on retooling national critical infrastructure, defense, aerospace, and advanced manufacturing. Although venture firms have backed this sector aggressively over recent years, next-generation defense and aerospace contractors still receive less than 5% of overall defense and procurement expenditure. However, modern procurement needs and geopolitical pressures are expected to dramatically expand the market share captured by modern defense primes.

Wrapping up the state of the markets presentation, the team underscores an overarching optimism regarding the broader economic ramifications of the ongoing capital spending cycle. While the capital expenditure and physical infrastructure buildout across compute, power, and facilities is historically unprecedented, the downstream gains in productivity will provide the fundamental economic justification for the massive capital investment underway.

```mermaid flowchart TD accTitle: Expansion Horizons for Tech and Infrastructure Capital accDescr: Diagram illustrating three key growth frontiers including Enterprise AI, Autonomous Transportation, and American Dynamism defense retooling.

subgraph Physical ["Transportation & Autonomy"] A["17M U.S. Auto Sales / Year"] --> B["Full Autonomy (Next 10 Years)"] C["Ride-Hail (1% Miles Traveled)"] --> D["10x+ Miles via Autonomous Fleets"] end

subgraph EnterpriseBio ["AI Diffusion & Life Sciences"] E["Enterprise Software"] --> F["Diffusion Beyond Coding Agents"] G["AI x Biology"] --> H["Drug Discovery & Personalized Health"] end

subgraph DefenseStack ["American Dynamism"] I["Current Spend (<5% to New Vendors)"] --> J["Retooling via Anduril, Saronic, Castelion"] J --> K["Modern Defense & Industrial Base"] end ```

- Modern defense and hardware primes such as Anduril, Saronic, and Castelion currently capture less than 5% of total government procurement and defense spending, leaving substantial room for budget capture as requirements evolve. - The diffusion of enterprise AI is actively expanding beyond initial coding and developer copilots into general knowledge work, domain-specific workflows, and operations. - The massive scale of the AI buildout is projected to act as a primary catalyst for sustained macroeconomic productivity growth across the United States.

> "The build-out is massive. But we think it's going to be massively productivity-enhancing in the U.S."

## Notable quotes > `43:10` One of the things that we've talked about and that I've written about is that the founder is the asset class at this point. — David George

> `25:00` And if you think that the best and highest use of your resources is to optimize your cost structure, what does that say about the revenue opportunity for you? — Alex Immerman

## Mentioned Abridge, Alex Immerman, Ali Ghodsi, Alphabet, Amazon, Amazon Prime, Anduril, Anthropic, Carta, Castelion, ChatGPT, Chime, Cognition, CrowdStrike, Cursor, Databricks, David George, Decagon, Dina Powell, EliseAI, ElevenLabs, Elon Musk, Facebook, Goldman Sachs, Google, Harvey, Hebbia, Hugging Face, Instacart, Instinct, Instagram, Josh Elman, Lyft, Marc Andreessen, Mark Zuckerberg, Martine, Meta, Microsoft, Microsoft 365 Copilot, Muse, Navan, Netflix, OpenAI, OpenRouter, Oracle, Revolut, Russell 2000, Sam Altman, Santiago Rodriguez, Sarah Friar, Sarah Wang, Saronic, ServiceNow, Shopify, Shopify AI Sidekick, Snap, SpaceX, Stripe, Tesla, TikTok, Typeface, Uber, Waymo, xAI, YipitData

## Video description a16z’s David George, Sarah Wang, Alex Immerman, and Santiago Rodriguez unpack 25 key charts from the latest State of Markets presentation, from the scale of the AI infrastructure buildout to what adoption looks like inside companies today. They examine why rising markets have so far been supported by earnings rather than multiple expansion, why hyperscaler CapEx is approaching $1 trillion annually, and why demand for compute continues to outrun supply. They also look at the downstream effects of that spending across chips, power, construction, and physical infrastructure. State of Markets Then they move up the stack: OpenAI and Anthropic’s revenue growth, the gap between AI deployment and measurable enterprise impact, the rise of agents, falling inference costs, and what all of this means for SaaS. They close with where the team is spending time next, including consumer agents, robotics, autonomy, AI and biology, personal health, defense, and the continued diffusion of AI across the enterprise. State of Markets

[Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/7104380b-ad2f-4075-b459-cb07cd66a70b/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=7104380b-ad2f-4075-b459-cb07cd66a70b&feed=JGE3yC0V)

中文简报

1万亿美元AI大基建 | 市场现状 a16z团队认为，当前的科技周期从根本上是由盈利而非估值倍数扩张所驱动，超大规模云服务商几乎将全部经营现金流重新投入到算力、电力和物理供应链中。在应用层，企业端应用正从基础的提示词查询转向复杂的多步骤智能体，上下文缓存和模型路由技术使软件利润率脱离了对Token成本的依赖。尽管消费者端AI订阅的普及仍处于早期阶段，但专业垂直软件和非上市科技巨头仍在以历史性的速度实现复合增长。因此，风险投资正日益向资本密集型的实体前沿领域集中，包括国防科技、自动驾驶和机器人技术。在持续不减的算力需求和多步骤智能体工作负载的推动下，超大规模云服务商的资本支出预计在2026年达到7800亿美元，并从2027年开始每年超过1万亿美元。 (6:55) 2026年AI初创企业占据了美国风险投资总交易额的86%，而排名前六的非上市科技巨头总估值达到2.4万亿美元，超过了过去十年所有科技IPO上市公司的总市值。 (48:36) 尽管企业界广泛开展试验——69%的标准普尔500指数成份股公司部署了AI项目——但目前只有2%的公司追踪长期净利润指标，这凸显出应用层存在显著的变现缺口。 (0:00) 通过上下文缓存、微调和智能路由技术，推理成本已下降了10倍至100倍，使软件厂商能够将客户定价与原始Token成本解耦，并提升利润率。 (26:26)

Reading summary

Andreessen Horowitz growth partners argue that the multi-hundred-billion-dollar artificial intelligence buildout represents a fundamentally sound industrial cycle anchored in resilient corporate earnings and physical infrastructure, rather than a speculative valuation bubble. Earnings-Backed Valuation and Hyperscaler CapEx (5:27) The broader equity market rally has been driven by earnings expansion rather than multiple inflation. While major indexes have appreciated around 20% over the recent cycle, underlying price-to-earnings multiples contracted by roughly 20%, keeping the S&P 500 below 20 times forward earnings while corporate earnings grew approximately 15%. This dynamic contrasts sharply with the 2000 dot-com peak, where cash-burning firms frequently commanded forward multiples exceeding 100 times. Simultaneously, infrastructure investment by Alphabet, Amazon, Meta, Microsoft, and Oracle reached $416 billion in 2025 and is projected to hit $780 billion in 2026 before crossing $1 trillion annually starting in 2027. Unlike speculative capital, these investments are reinforced by a combined $1.7 trillion cloud backlog across Microsoft, Alphabet, and Amazon, alongside structural supply constraints that extend equipment waitlists into 2028. The Transition to Physical Infrastructure (10:00) Capital allocation in technology has transformed into an 'everything cycle' that dominates the physical economy, with high-tech equipment, software, and R&D accounting for 55% of all non-residential U.S. capital expenditures. The presenters project cumulative global infrastructure demands through 2040 to reach $90 trillion across compute, electrical power grids, water facilities, and transportation networks. The current infrastructure buildout as a percentage of GDP has already surpassed the expansion of the nineteenth-century American railroad network. This shift into physical manufacturing is visible among large private venture-backed firms: Anduril has constructed a manufacturing hub spanning 87 football fields, Waymo is building out regional vehicle depots, and SpaceX is executing a reported $100 billion investment in Louisiana. Furthermore, utility data indicates that data center baseloads can lower consumer costs; one cited U.S. study found that every 10% increase in utility territory data center capacity lowered residential electricity rates by 40 basis points by distributing fixed grid infrastructure costs. Enterprise and Consumer Software Economics (20:00) A significant adoption gap persists between superficial pilot programs and enterprise monetization. While 69% of S&P 500 companies run live AI initiatives, only 30% show quantifiable operational impact, and merely 2% longitudinally track verified performance metrics. Spending is heavily skewed: median tooling expenditures among the top 1% of enterprise users are eight times higher than the top 10%, reaching $7,500 to $9,000 per user monthly compared to $200 to $400 for median users. At the same time, inference cost deflation of 10x to 100x—unlocked by prompt caching and model routing—is expanding gross margins for application developers. Databricks' dynamic task routing reduced inference costs by 35%, while fine-tuning smaller specialized models allowed EliseAI to cut serving costs by 60%. In the consumer space, paid subscription penetration sits at just 2% of U.S. households, yet leading applications display unusual 'smiling' retention curves where cohorts re-engage over time as underlying foundation model capabilities improve in the background. Public-Private Disconnect and Emerging Frontiers A sharp bifurcation has emerged between public software vendors and late-stage private mega-caps. While only about five public SaaS companies continue to grow annual revenue at 30% or more, private transactional data from Stripe points to an underlying SaaS 'renaissance' of accelerating revenue across both startups and legacy enterprise customers. The combined valuation of the top six private technology firms—Anthropic, OpenAI, Databricks, Stripe, Waymo, and Revolut—has reached $2.4 trillion, surpassing the $1.7 trillion aggregate market capitalization of all tech IPOs over the past decade excluding SpaceX. Venture investment reflects this concentration, with AI enterprises capturing 86% of total U.S. venture deal value in 2026. The partners identify physical robotics as a frontier that could eventually exceed large language models in market size, while autonomous vehicle fleets—which current testing shows are roughly 14 times safer than human drivers—are positioned to expand far beyond the 1% of total U.S. vehicle miles currently captured by ride-hail services. The partners conclude that while the scale of the physical buildout and capital expenditure is unprecedented, the investment will be economically justified by downstream productivity gains. An open challenge remains whether enterprise software and legacy discovery marketplaces can successfully convert pilot workflows and ad-based profit pools into durable agentic business models without sacrificing incumbent margins.

精炼阅读摘要

Andreessen Horowitz（a16z）增长投资团队指出，当前规模空前的人工智能资本开支并非脱离基本面的投机泡沫，而是由企业盈利增长和下游真实需求驱动的“一切周期”，正强力拉动以电力与电网为代表的物理基础设施全面扩张。盈利支撑的超级资本开支与物理瓶颈 (3:07) 科技投资已演变为驱动美国非住宅资本支出约55%的“一切周期”，科技股占美股总市值约40%。虽然市场四年内上涨90%（年化回报率17%），但标普500指数市盈率低于20倍，且受硬件与内存芯片制造商6至7倍低估值的平抑，整体估值倍数反而压缩约20%，显示涨势由企业约15%的盈利增长支撑，与2000年互联网泡沫存在本质区别。基础模型开发商累计融资已超3500亿美元，AI基础设施占GDP比重已超过19世纪美国铁路扩建潮。五大超大规模云服务商（Alphabet、Amazon、Meta、Microsoft、Oracle）资本支出从2025年的4160亿美元攀升至2026年预估的7800亿美元，预计2027年突破1万亿美元。微软、谷歌和亚马逊积压的云订单达1.7万亿美元。由于芯片与数据中心设备排单已延至2028年，这笔开支将近期的自由现金流压入J曲线低谷，但共识预期2028年起现金流将大幅反弹。至2040年全球基础设施建设资金需求预计达90万亿美元。关于数据中心推高民用水电成本的担忧，研究显示公用事业区内数据中心容量每增加10%，因固定资产折旧分摊扩大，居民电价反而下降40个基点。企业落地断层与推理降本驱动的软件架构重塑 (20:00) 标普500企业中已有69%部署AI试点，但仅30%能证明可量化的财务成果，实际进行长期指标追踪的仅占2%。不同于传统软件，企业重度用户（前1%）每月在AI工具上的支出达7500至9000美元，是中位数用户（200至400美元）的20倍以上。领先的AI原生公司将高达10%的人员预算用于AI工具，而传统财富500强仅投入约1%。量化成果开始在核心业务显现：Chime数字银行四年内将单客服务成本复合降低近50%，Shopify的AI助手促使新商家在前15天达成5笔订单的比例提升8%，ServiceNow也斩获超10亿美元的AI年化合同价值。推理端正经历两个数量级的成本通缩，催生从单次问答向自主智能体（Agent）的演进，OpenRouter平台的智能体Token消耗量暴增14倍。企业借助上下文缓存技术（如Hebbia将分析成本削减10倍）、智能路由（如Databricks动态调度大小模型降低35%成本）以及小模型微调（如EliseAI将服务成本降低60%并压缩延迟），成功将应用层软件的商业定价与底层算力成本解耦，不再按Token售卖，转而按业务成果交付收费，实现了软件毛利率的大幅扩张。消费者智能体变局与SaaS商业模式分化 (30:00) 消费级AI订阅仍处于导入期，美国仅略高于2%的家庭开通付费订阅，远低于Amazon Prime的2亿户和Netflix的7000万户。由于智能体转入后台自主处理复杂任务，衡量指标正从社交媒体时代30至60分钟的每日屏幕停留时间，转向“零点击”任务完成率与总交易额（GMV）。这一变革动摇了依靠搜索展示与广告变现的利润池：例如拥有超700亿美元高毛利广告业务的亚马逊拒绝接入智能体，而寻求增量订单的Instacart则主动开放接入。在公开软件市场，约75%的SaaS公司实现盈利，但年营收增速超过20%的仅剩30%，超30%增速的企业不足5家。尽管纯应用型软件面临席位削减与重构风险，但企业毛留存率（GDR）仍稳固在98%左右。网络安全与可观测性软件（如CrowdStrike）因智能体引入新的攻击面与监控需求而表现突出，垂直领域AI（如法律科技Harvey、医疗记录Abridge）则凭借行业工作流壁垒实现前所未有的扩张速度。一级市场超级集聚与物理世界自主化前沿 (40:00) 资本向头部极度集中，2026年AI相关企业占据美国风险投资总额的86%，较2025年的65%显著跃升。排名前六的未上市巨头（Anthropic、OpenAI、Databricks、Stripe、Waymo、Revolut）总估值达2.4万亿美元，超过过去十年所有科技IPO市值总和（1.7万亿美元，不含SpaceX）。非公开状态使管理层免于公开季度财报压力，能够实施周期长达数年的高强度资本再投资。Carta数据显示，私有巨头定期员工要约收购参与率平均仅为58%，表明员工惜售意愿强烈；且二级市场交易相对最近一轮优先股已抹平折扣，呈现0%折价。投资团队认为下一阶段高确定性机遇正向物理世界扩散。实体机器人技术在视觉模型与具身智能突破下，潜在规模可能超越大语言模型，其规模落地预计滞后大模型3至5年。自动驾驶已具备商业可行性且安全性达到人类驾驶员的14倍，预计未来十年全美年销1700万辆乘用车将全面标配自动驾驶，大幅拓展仅占全美车行里程1%的网约车市场。此外，AI驱动的复杂疾病新药研发、高度个性化健康管理，以及仅占传统采购支出不足5%的新一代国防科技（American Dynamism，如Anduril），均构成资本扩张的核心边界。 a16z投资团队总结认为，尽管数万亿美元级的基础设施与算力投资规模史无前例，但随着AI从单纯的代码生成扩散至更广泛的企业知识工作、物理机器人、生物医药及现代国防制造，下游释放的宏观生产力提升将最终验证这一轮重资本投入的商业合理性。

Detailed notes — zh

## Background 本期内容由 a16z Growth 团队的 David George、Sarah Wang、Alex Immerman 和 Santiago Rodriguez 于 2026 年 9 月 30 日在播客及网络平台发布。节目依托最新的《State of Markets》报告，通过 25 张核心图表深入剖析了规模逼近万亿美元的 AI 基础设施建设浪潮、企业内部的真实采纳现状，以及算力爆发对物理世界和应用层生态的深远影响。 ## TL;DR a16z 投资合伙人们系统论证了当前 AI 资本支出（CapEx）并非泡沫，而是由强劲的盈利增长与长达数年的积压订单所支撑的“全要素生产力周期”。讨论指出，超大规模云厂商的万亿美元基建正在将巨额资金注入电网、水利、芯片及重工业制造等物理实体，同时大模型推理成本的断崖式下跌与 Agent 架构正在催生新的需求飞轮。尽管大企业中能持续量化追踪 AI 价值的比例仅有 2%，但软件板块已出现剧烈的赢家通吃与二次加速分化，具身智能、自动驾驶与国防硬科技也将成为承接这一波生产力爆发的关键战场。 ## Atmosphere & style 整场对话洋溢着硅谷顶尖成长期投资人特有的敏锐、理性与高密度认知输出，风格紧凑干练，充斥着一线调研的实战数据与宏大视野。四位合伙人配合默契，在多轮观点抛接中既有对华尔街保守预测与“AI 泡沫论”的犀利反驳，也有对企业落地迟缓等现实痛点的冷静叩问。聆听过程宛如置身一场顶级风投机构的内部合伙人决策闭门会，充满了对“原子时代”工业复兴的激昂信念，令人真切感受到科技巨浪重塑物理世界时那种震撼人心的厚重感。 ## Bulletin - `05:00` 超大规模云厂商的资本开支正迈向万亿美元量级，五大巨头 2026 年 CapEx 预计达 7,800 亿美元，2027 年起年均将破 1 万亿美元，且微软、谷歌与亚马逊合计已积压 1.7 万亿美元云订单。 - `05:00` 本轮 AI 牛市受实际盈利而非估值泡沫驱动，标普 500 整体市盈率仍低于 20 倍，强周期存储芯片远期市盈率仅 6 至 7 倍，估值倍数在市场上行期间反而收缩了约 20%。 - `16:45` 企业端 AI 渗透呈现巨大反差：标普 500 企业中虽有 69% 部署了 AI 功能，但仅有 2% 建立了能跨周期持续追踪量化成效的业务指标体系。 - `40:00` 头部私有科技公司展现出空前的资本与体量集聚效应，Anthropic、OpenAI、Databricks、Stripe、Waymo 与 Revolut 六家公司总市值已达 2.4 万亿美元，超过过去十年所有美股 IPO 公司总市值总和。 - `45:00` 全美风投交易加速向 AI 赛道集中，2026 年统计切片中 AI 相关交易额已占全美 VC 总投资的 86%，较 2025 年的 65% 进一步飙升。 ## Key takeaways - `00:00` 科技正在驱动波及整个经济体的实体投资狂潮，高科技设备、软件与研发投入已占美国整体资本开支的 55%，科技板块占美股总市值近 40%。 - `05:00` 当前算力与数据中心的基建投资占 GDP 比例已正式超越美国历史上的铁路大扩张时期，但市场繁荣由坚实的企业盈利支撑而非估值泡沫。 - `07:30` 算力需求呈现出打破模型预测的贪婪状态，由于长运行 Agent 负载的兴起，算力供需缺口极大，供应链关键硬件排期已被预订至 2028 年。 - `10:00` 超大规模云厂商的大举投资遵循基础设施 J 曲线规律，虽然短期自由现金流承压，但预计在 2028 年随着订单交付和芯片较长经济寿命的释放将迎来爆发式复苏。 - `12:30` 科技巨头的资本开支本质上是实体工业供应链的客户订单簿，推动全球重返“原子时代”，基础设施到 2040 年的潜在投资需求高达 90 万亿美元。 - `14:50` 数据中心的大规模接入通过分摊公共电网的固定沉没成本，在实际案例与统计中反而促成了周边居民用电费率的稳步下降。 - `16:45` 杰文斯悖论在模型层再次上演：推理成本下降两个数量级不仅没有缩减支出，反而催生了指数级扩张的高复杂度 Agent 需求，推动 OpenAI 与 Anthropic 的营收曲线大幅超越传统 SaaS 巨头。 - `20:00` 应用层的核心使命是弥合通用基础模型与企业私有上下文之间的鸿沟，当前企业支出极化严重，前 1% 极客员工的 AI 月度开销是中位数员工的 20 倍以上。 - `25:00` 企业在 AI 上的增量投入正成为战略试金石：将预算投向能开拓新客群的增收场景具有无限上行空间，单纯用于降本增效则极易触及天花板。 - `25:00` 通过提示词缓存、模型路由与精调垂直小模型，Agent 的底层推理成本和延迟被大幅压低，推动应用层软件商业模式转向按任务交付结果收费。 - `29:15` 消费级 AI 订阅展现出罕见的“微笑留存曲线”，底层模型迭代不断唤醒流失用户；同时交互范式从“争夺屏幕使用时长”彻底转向“后台长时间自主运行”。 - `32:30` 具备跨平台执行能力的 AI 代理正在冲击传统双边市场与零售平台的商业模式，引发依赖广告流量闭环与拥抱增量 GMV 之间的战略博弈。 - `36:30` 公开市场 SaaS 呈现严重分化，头部软件企业依托 98% 的净金额留存率作为防守底座，必须通过引入高单价 AI 产品力争实现 10% 以上的收入增速二次加速。 - `40:00` AI 带来的系统安全与监控需求激增，反而大幅强化了网络安全与可观测性软件巨头的护城河，垂直行业应用增长亦显著快于通用横向 SaaS。 - `45:00` 创始人领导的独角兽倾向于长期停留在私募市场，通过要约收购满足员工流动性，从而规避公开市场短期业绩审视，专注于高回报周期的重注研发。 - `45:00` 具身智能机器人的潜在市场天花板被认为可能超越大语言模型，尽管其商业化阶段比 LLM 晚三到五年，但未来五年内将迎来基础设施投入的爆发。 - `50:00` 全自动驾驶车辆安全性达到人类驾驶的 14 倍，随着技术拐点到来，网约车出行里程渗透率有望提升一个数量级，并驱动每年 1700 万辆新车全面自主化。 - `51:24` AI 在生物制药、个人健康量化以及国防“美国活力”底层技术栈上的扩散尚处于早期，万亿美元规模的基础设施建设将从根本上全面提升宏观经济的实际生产力。 ## Detailed walkthrough ### `00:00` 技术驱动的“全要素周期”与宏观图景播客伊始，David George 携同事 Sarah Wang、Alex Immerman 和 Santiago Rodriguez 共同发布并探讨了 a16z Growth 团队年度《State of Markets》报告中的 25 张核心图表。David George 概述指出，当前科技正在驱动一场波及整个经济体的投资狂潮，不仅超大规模云厂商（Hyperscalers）几乎将其所有的近期经营现金流倾注于产能建设，更推动了覆盖芯片、电力、水资源、冷却技术、建筑及技术工人等物理基础设施的现代化重塑。 ```mermaid flowchart TD accTitle: AI投资驱动链条 accDescr: 展示从AI算力需求到下游物理基础设施投资的传导路径 subgraph Compute["算力需求层"] A["AI模型研发与Agent普及"] --> B["超大规模云厂商现金流投入"] end subgraph Infra["物理基建层"] B --> C["数据中心与芯片部署"] C --> D["电力、冷却与水网改造"] C --> E["建筑工程与熟练技术工人"] end ``` 针对整体经济层面的数据，团队指出科技在资本支出和市值集中度上均达到了惊人的历史量级： - 高科技设备、软件及研发投入目前约占美国整体资本开支（Capital Spending）的 55%。 - 科技板块已占美国股市总市值的近 40%。 - 全球市值前十大的公司中，有 8 家是美国科技企业。 - 全球至 2040 年的基础设施投资需求估计高达 90 万亿美元，涵盖范围远超数据中心本身，还包含电力、水利、道路与公共交通系统。 - 基础模型研发公司累计融资金额目前已超过 3,500 亿美元。关于企业内部落地的真实进展，David George 直言当前的局面呈现出强烈的反差：虽然标准普尔 500 指数成分股中已有 69% 提及或涉足相关探索，但如果观察最核心的长期追踪指标，“那其实仅仅只有 2%（that's actually only at 2%）”。AI 虽已展现出创造巨额营收与降本的潜力，但从整个企业界的采纳曲线来看，当下依然处于极其早期的阶段。 --- ### `05:00` 投资规模超越铁路时代：泡沫争论与盈利支撑随着资本开支持续飙升，市场对“AI 是否正处于泡沫”的质疑声不绝于耳。David George 将本轮 AI 基建与历史上著名的铁路建设潮进行对比，指出当前数据中心和算力的建设规模占 GDP 的比重，已经正式超越了美国历史上的铁路大扩张时期。面对市场见顶或泡沫破裂的担忧，团队用估值与基本面盈利的背离进行了有力反驳。David George 解释道，虽然市场触及历史新高，但股票估值倍数（Trading Multiples）实际上是下降的——美股涨幅约 20% 的同时，估值倍数反而收缩了约 20%，标普 500 的整体市盈率依然低于 20 倍。这与 2000 年互联网泡沫时期毫无业绩支撑、动辄上百倍 PE 的景象截然不同；即便是强周期性的存储芯片公司，目前的远期市盈率也仅在 6 至 7 倍左右徘徊。 Santiago Rodriguez 对此补充分析道：“当市场以年化 17% 的速度连续增长四年时，人们本能的直觉是‘这肯定该跌了，我们绝对处于过热期’（the natural instinct is, well, that's got to come down, right? We're definitely in a hot period. But when you compare that to, as you said, the market trading for below 20 times earnings... This definitely feels a little different than maybe the 2021 period that was recent or the 2000 period when multiples in growth were not really going together.）” 过去近四年里市场累计上涨 90%，其核心底座是实打实的盈利增长，而非虚高的估值泡沫。 Alex Immerman 则引用马克·安德森的经典论断指出，这正是“软件吞噬世界”的新章节：过去 15 年软件席卷各行各业，但当时全社会只有极少数专业人员能够编写软件；如今每个人每天夜间都在后台运行着数个自动化流程。软件生产与使用门槛的泛化，促使对底层算力、芯片、电力及施工的需求出现数量级式的跃升。 - 自 ChatGPT 问世以来的近四年内，美股大盘累计上涨 90%，年化收益率达 17%。 - 五大超大规模云服务商（Alphabet、Amazon、Meta、Microsoft 和 Oracle）的资本开支在 2025 年为 4,160 亿美元，预计 2026 年将跃升至 7,800 亿美元。 - 市场一致预期这些云巨头从 2027 年开始，每年的资本开支总和将突破 1 万亿美元大关。 --- ### `07:30` 破表的算力需求：Agent 演进与被低估的基建天花板谈及为何算力需求能持续超越供给，且资本开支能够如此迅速地转化为有效承载，团队指出根本原因在于分发渠道的先验性。与以往需要从零构建终端网络的计算平台不同，AI 浪潮直接嫁接在互联网、云计算以及智能手机数十亿级存量用户的现成网络之上。更为关键的转变在于计算负载模式的质变。Sarah Wang 深入剖析称，产业已经彻底告别了“偶发性对话问答（occasional queries）”的初级阶段： > “随着 Agent（智能体）的兴起，你面临的是并行任务和长时间运行的任务。正如 Alex 此前提到的，当你去做其他事情时，你的任务甚至会一直持续运行到深夜。如果你设想一下 Agent 自主编写代码、搜索并执行这些任务的场景，这直接赋予了这些计算平台前所未见的庞大算力需求。” 针对连续上调的资本支出指引，Alex Immerman 评论道，华尔街和经济学家在过去四年中一再将当期的投资数字视作“见顶平盘”的上限，但不断演进的模型和吞吐需求总是轻易打破预测。Santiago Rodriguez 更是形象地借用 David George 的口吻将算力需求定义为一种“打破模型预测的存在（model buster）”。 Santiago 顺势抛出了一个颇具戏剧性的行业轶事：大约一年前，Sam Altman 和 Sarah Friar 因大举激进承诺庞大算力采购而遭受外界猛烈批评，被斥为鲁莽冒进；然而时至今日，所有人不得不惊叹于他们当初决策的前瞻性。即便如此，由于底层算力瓶颈依旧严峻，OpenAI 甚至在几周前不得不一度暂停其 Pro 付费订阅通道的注册，这也从侧面印证了算力供不应求的极端现状。 ### `10:00` 算力供应链长周期锁定与巨头 1.7 万亿美元的云端积压订单在讨论算力基础设施的供需格局时，合伙人们一致强调，当前市场正在经历一种几乎“无法满足的贪婪需求”。他们从产业链的各个环节反复听到相同的结论：算力需求正在大幅超越供给能力。供应链甚至出现极端瓶颈，在数据中心建设的某些关键环节，客户现在下单甚至要排队到 2028 年才能拿到关键原材料或硬件设备。尽管这种结构性短缺让部分观察者对资本开支的可持续性产生疑虑，但在需求侧，超大规模云厂商（Hyperscalers）的基本面表现却提供了极为坚实的支撑。目前，微软、谷歌与亚马逊三家巨头合计拥有的云端积压订单（Cloud Backlog）已高达约 1.7 万亿美元，且客户的长期合约承诺仍在加速堆积。虽然这些科技巨头为了扩充算力服务能力进行了史无前例的重资产投入，导致当下的自由现金流承压，但华尔街与业界的共识预测均指向自由现金流将在 2028 年迎来强劲反弹，并在其后实现爆炸式增长。 ```mermaid flowchart TD accTitle: 基础设施投资 J 曲线与现金流兑现路径 accDescr: 展示超大规模厂商从前期重资本开支到长期云服务与现金流回收的演进过程。 subgraph Phase1["前期投入阶段（当下至2027）"] A["资本开支激增 （厂房外壳/电力接入/订购芯片）"] --> B["自由现金流（FCF）承压 供应链排期锁死至 2028 年"] end subgraph Phase2["产能交付与资产长效运营（2028及以后）"] C["1.7 万亿美元云积压订单转化"] --> D["GPU/TPU 展现超预期长经济寿命"] D --> E["自由现金流强劲复苏 利润率与现金流呈 J 曲线爆发"] end B --> C ``` 亚马逊在近期的财报电话会上对这一底层商业逻辑进行了极为透彻的阐释，生动勾勒出了这一基础设施投资的“J 曲线”动态。构建数据中心首先需要完成土地、物理外壳与电力配网等物理工程，这需要耗费固定的沉没时间与巨额前期资本；但一旦芯片和服务器进场点亮，GPU 和 TPU 的实际经济使用寿命往往比外界预期要持久得多。针对部分超大规模云厂商因举债进行资本支出而遭受公开市场惩罚的现象，对话者们指出，这与基础大模型实验室的演进逻辑如出一辙：在这场算力竞备中，凡是由于短期财务压力而选择犹豫、放缓甚至眨眼的玩家，最终都付出了惨痛的代价并深感后悔。更关键的是，这一波基础设施浪潮正在为终端创造极其庞大的“消费者剩余”（Consumer Surplus）：无论是终端消费者还是软件服务商，都在借助这些模型大幅提效并创造实际利润，而云厂商哪怕在芯片采购成本较一两年前大幅飙升的背景下，依然能凭借高定价和强劲利用率维持极高的利润率。 - 微软、谷歌与亚马逊三家巨头合计的云端积压订单达到约 1.7 万亿美元，展现出高质量的底层需求。 - 数据中心部分关键供应链环节的交付周期已被预订一空，最晚排期已延伸至 2028 年。 - 市场共识预测超大规模厂商在经历当前的资本开支重压后，自由现金流将于 2028 年触底回升并步入持续的高增长通道。 - 芯片（GPU/TPU）的经济使用寿命被证实显著长于市场早期的悲观折旧预期。 > “在当下这一轮大扩张期间，自由现金流确实受到了压制，但共识预测显示从 2028 年起现金流将开始复苏，并在那之后实现可观的增长。” —— 亚历克斯·伊默曼（Alex Immerman） --- ### `12:30` 科技繁荣演进为实体工业大浪潮与“原子时代”的回归对话随后深入探讨了超大规模厂商资本开支的本质：一家科技巨头的 CapEx，在本质上就是整个实体制造业供应链的“客户订单簿”。过去几十年来创造了人类商业史最高净利润的超级平台，正在将数千亿美元的真金白银反哺注入现实物理世界。这场原本发端于硅谷软件层的科技浪潮，如今已彻底蜕变为一场跨越芯片制造、电网能源、重型建筑与水务管网的宏大工业繁荣。a16z 投资团队明确表示，他们目前正将大量时间投入到服务这一工业大潮的各类硬件制造与数据中心供应链初创企业中。这也意味着风险投资的底层范式正在经历深刻重塑：传统软件投资的核心是“开发一次，无限复制分发”的知识产权变现，边际成本趋近于零；而工业与基础设施类企业则必须直面重型工程与供应链的残酷现实，包括复杂的债权融资、供应商博弈、产能预测以及对终端需求的精确调配。这种转变标志着全球正步入一个崭新的“原子时代（Age of Atoms）”。据测算，到 2040 年，全球基础设施建设的潜在投资需求将高达 90 万亿美元，而这一天文数字不仅涵盖 AI 数据中心本身，更涵盖电网、水资源处理系统、高速公路和公共交通网络等大动脉。投资物理世界与押注传统软件有着天壤之别，正如埃隆·马斯克（Elon Musk）多年前提出的名言——“工厂本身才是终极产品（The factory is the product）”。从特斯拉和 SpaceX 走出的优秀初创团队，往往在创业第一天就将极致的重工业工程制造与实体落地能力视作其核心护城河。这种大兴土木、重构工业基础的范式转移，不仅正在重塑投资版图，更为全美各地带来了遍及各州的高技术制造业工人就业机会。 - 到 2040 年，全球基础设施建设投资需求预计将高达 90 万亿美元，覆盖电力、水利、交通网络与 AI 基础设施。 - 国防科技独角兽 Anduril 正在全速推进其名为“Arsenal-1”的超大型制造工厂，占地面积相当于 87 个正规美式橄榄球场。 - 无人驾驶领军企业 Waymo 正在全美各大城市积极扩张其自研车辆的线下运维车棚与基地（Depots）。 - SpaceX 正在路易斯安那州等地斥资推进上千亿美元级别的重工业发射与制造基础设施建设。 > “当我们谈论超大规模厂商的资本开支时，应该把它看作是其他企业的订单簿。” —— 圣地亚哥·罗德里格斯（Santiago Rodriguez） > “从我们的视角来看，今天押注未来与过去已经截然不同；这意味着建造真正的工厂、扩展实体基础设施，并在全美范围内创造更多的高技能技术岗位。” —— 萨拉·王（Sarah Wang） --- ### `14:50` 破除数据中心迷思：公共电网固定成本摊薄与居民电费反降现象围绕数据中心的大规模铺开，社会舆论和媒体报道中充斥着诸多负面叙事与认知误区，例如“数据中心正在抽干全美淡水储备”、“高净值富人排斥居住在数据中心周边”，以及最广为人知的担忧：“数据中心的疯狂耗电将直接拉高周边居民的家庭电费账单”。对话者们直接对这些迷思进行了针对性反驳，指出数据中心的接入在很多情况下不仅不会抬高电价，反而能有效降低本地居民的用电费率。其底层的能源经济学逻辑在于：公共电网本质上是一个极其庞大的共享固定成本资产池，其中的电线杆、高压输电线和变电站等基础设施的建设与维护成本是相对固定的。当像数据中心这样拥有极高负荷因子、用电量巨大且需求极其稳定的超级大客户接入电网时，电网运营商得以将巨额的既有固定资产折旧与输变电运维成本，分摊到更多的用电单位（kWh）之上。简而言之，在一个共享固定成本系统内，总用电负荷的健康增加能够显著改善整体经济效益。这一论断并非纸上谈兵的理论推演，Meta 全球公共政策高管迪娜·鲍威尔（Dina Powell）在近期的访谈中就详述了其在路易斯安那州落地的超大型数据中心案例：通过与当地电力公用事业公司及社区通力协作优化电网设计，该数据中心项目的进驻最终促成了周边社区用电价格的整体下调，为当地企业和普通居民家庭带来了切实可感的基础设施红利。 | 常见数据中心公众舆论误区 | 底层工程与电网经济学现实 | 实际经验与研究数据佐证 | |:--- |:--- |:--- | | **电费飙升论**：认为算力中心虹吸电力，必然导致周边居民电价飞涨。 | **固定成本摊薄**：电网系统（线路、变电站）属于沉没固定成本，新增规模化稳定负荷能摊薄单位电力的输配电成本。 | 权威全美调研数据：数据中心容量每增加 10%，居民电费费率平均下降 40 个基点（0.4%）。 | | **社区冲突论**：认为数据中心是工业噪音与生态灾难，遭到本地住户全面抵制。 | **社区共建与能源基建反哺**：通过共担电网升级改造成本，巨头直接补贴并优化了本地的基础公用事业网络。 | Meta 在路易斯安那州的数据中心基地建设方案中，通过与本地社区合作直接降低了当地综合电费。 | > “全美近期的一项研究表明，数据中心容量每增加 10%，周边居民的用电费率就会下降 40 个基点。” —— 圣地亚哥·罗德里格斯（Santiago Rodriguez） --- ### `16:45` 杰文斯悖论与千亿美元模型层腾飞，企业落地的“2%”严酷现实话题转向大模型层与上层软件应用的最新演变。当前行业正处于一个戏剧性的十字路口：一方面，大模型调用与推理成本正以指数级速度断崖式下跌，催生了经典的“杰文斯悖论（Jevons Paradox）”——单位算力与推理成本下降两个数量级（99%），非但不会缩减整体算力支出，反而使得海量原本在财务上不可行的高复杂度 Agent（智能体）与自动化工作流变得极具经济可行性，从而撬动了指数级扩张的全新使用场景与总需求。在这一宏观红利推动下，头部模型厂商的商业化吸金能力正彻底刷新历史纪录。作为 a16z 历次合伙人会议（GP Offsite）中每个月都因数据狂飙而必须重绘的核心图表，OpenAI 与 Anthropic 的年化总营收正以惊人的垂直斜率攀升。若将它们新增年度经常性收入（Net New ARR）的扩张速度，与软件商业史上最璀璨的传奇 SaaS 巨头们相比，这两家头部大模型实验室的增长曲线已远远将 Salesforce、Workday 或 ServiceNow 等前辈甩在身后。 ```mermaid flowchart TD accTitle: 标普 500 企业 AI 部署现实漏斗 accDescr: 展现标普 500 企业在实际应用 AI 时，从试验性部署到最终能持续量化追踪商业价值之间的急剧衰减。 A["69% 标普 500 企业 完成了线上环境部署 (Live Deployments) （绝大部分仅停留于微软 Copilot 桌面试用）"] --> B["30% 标普 500 企业 产生了可被财务量化的业务成效 (Quantifiable Impact)"] --> C["仅 2% 标普 500 企业 建立了跨时间周期持续追踪的量化指标体系 (Metrics Tracked Over Time)"] ``` 然而，与模型层千亿美金估值与爆发式营收形成鲜明对照的，是企业软件落地端的极端初级与迟缓。2008 年移动互联网爆发时，分析师只需测算全社会拥有 iPhone 的人口渗透率；而 AI 时代的落地模型呈现出完全不同的复杂双重维度：既要考察尝试使用 AI 产品的用户比例，更要考量企业工作流中预算份额（Share of Wallet）与核心业务被渗透的深度。标普 500 指数（S&P 500）企业的真实数据充分揭示了这一鸿沟： - **第一层（部署率 69%）**：近七成的大型企业宣称已在生产或测试环境中上线了 AI 功能。但这在很大程度上只是流于表面，绝大多数企业目前唯一的切入口仅仅是采购了微软的 Microsoft Copilot 办公插件。 - **第二层（可量化成效 30%）**：在上线 AI 的企业中，仅有三成能够拿出明确量化的降本增效业务结果证明。 - **第三层（持续指标追踪 2%）**：作为衡量深度嵌入并重塑组织架构的核心风向标——建立跨时间周期、持续考核量化业务产出指标的企业，全美 500 强中竟然只有区区 2%。这种“模型能力的日新月异”与“企业实际落地停滞于浅尝辄止”之间的巨幅落差，恰恰是应用层创业者面临的最大结构性红利。从零散尝鲜的单点试验，走向真正深入企业深水区的核心高壁垒、周期性长效工作流，将是下一阶段 AI 软件公司诞生万亿美元市值的关键主战场。 - 当前智能体推理成本的大幅跳水促成了“杰文斯悖论”，两级数量级的降本正打开几何级数增长的自动化工作负载。 - OpenAI 与 Anthropic 展现出的净新增收入扩张斜率，已经全面超越了软件行业历史上任何一家标志性公有云 SaaS 公司。 - 标普 500 企业中仅有 2% 能够建立起长期跟踪 AI 业务价值的指标体系，表明 AI 在企业核心架构中的渗透才刚刚拉开序幕。 > “现在，如果你去看最严苛的终极晴雨表——也就是拥有跨时间周期持续追踪量化指标的企业占比——实际上只有区区 2%。” —— 萨拉·王（Sarah Wang） > “模型能力的高度与它们当下的实际应用程度之间所存在的巨大鸿沟，正是让我对应用层前景感到无比兴奋的原因。” —— 戴维·乔治（David George） ### `20:00` 应用层弥合认知鸿沟与头部超级用户的支出分化对话首先回顾了 Sarah Wang 与 Databricks 联合创始人 Ali Ghodsi 以及 Martin Casado 的一次交流。Ali Ghodsi 曾犀利指出：“AI 或许对整个世界了如指掌，但对你的公司却一无所知。”这一洞察精准点出了当今应用层创业公司的核心使命：弥合模型通用认知与企业私有上下文之间的鸿沟。例如，拥有极强内部工程团队的数字银行 Revolut，依然选择与语音 AI 公司 ElevenLabs 合作，将其顶尖的语音模型无缝集成到用户账户体系与银行风控工作流中。这样一来，当客户致电客服时，AI 能够兼顾安全性、顺畅度与效率，妥善解决具体业务问题。在企业级 AI 普及仍处早期的背景下，市场上正出现极端的“超级用户分化”现象。合伙人们引用了来自 YipitData 的统计数据：在所有采用 AI 的企业中，支出排名前 1% 的企业，其在 AI 供应商上的中位数支出大约是前 10% 企业的 8 倍；换句话说，排名前 1% 企业的支出总量几乎与排在第 2% 至 10% 的所有企业总和相当。在高度 AI 原生化（AI-pilled）的公司内部，这种分化同样悬殊。顶级员工每个月在 AI 工具上的花费高达 7,500 到 9,000 美元，而普通员工的中位数支出仅为每月 200 至 400 美元（大约相当于叠加上一两个订阅产品），两者差距超过 20 倍。 ```mermaid flowchart TD accTitle: AI支出在不同层级企业中的渗透差异 accDescr: 展示从传统500强到AI原生企业的AI支出占人力成本比例与绝对值断层 subgraph Scale["支出占人力成本比例"] F500["大型财富500强企业 约 1%"] AIPilled["最激进的AI原生企业 高达 10%"] F500 -.->|"扩散空间巨大"| AIPilled end subgraph InternalUser["企业内部单人月度支出"] MedianUser["中位数普通员工 $200 - $400 / 月 (基础订阅)"] TopUser["前 1% 极客员工 $7,500 - $9,000 / 月 (高频API与重度Agent)"] MedianUser -->|"超过 20 倍支出差"| TopUser end ``` - 大型传统企业（如前沿的财富 500 强）目前在 AI 工具上的开销约占其员工人力成本（Headcount）的 1%，而 a16z 投资组合中最激进的 AI 原生公司，这一比例已经达到 10%。 - 很多企业在初次采购 Cursor 或 Cognition 等 AI 编程辅助工具时兴奋不已，但这仅仅是漫长跑道的起点，从单点采购测试到企业全员深度融入仍有巨大距离。 --- ### `22:15` 可量化的 ROI：降本服务、收入飞轮与巨头的防御变现企业界正从概念验证阶段迈入具备可量化 ROI（投资回报率）的实际落地阶段。团队梳理了多家上市公司财报中的实打实数据：在降本端，数字银行 Chime 连续四年将服务成本（Cost to Serve）每年削减 10% 以上，复合降幅接近 50%，这一成效很大程度上得益于客服 Agent 公司 Decagon 的赋能。降低服务成本赋予企业巨大的竞争筹码——企业既可以给出更优惠的定价，也可以提供更全面的服务支持，甚至将节省的资本直接反哺至增长业务。在增收端，Shopify 推出的 AI 助手 Sidekick 展现了惊人的业务飞轮效应。数据显示，Shopify 上完成初始入驻并在 15 天内达成 5 笔订单的商家比例上升了 8%。在 Shopify 的商业模型中，“达成 5 笔订单”是商户长期留存的关键分水岭，一旦跨过该门槛，商户流失率就会锐减。与此同时，软件巨头并未坐以待毙。ServiceNow 在最新财报中披露了超过 10 亿美元的 AI 年合同价值（ACV），其 Agentic（智能体）部署量更实现了 9 倍的爆发式增长。这表明无论是敏捷的初创黑马还是老牌 SaaS 现存巨头，都正在这一波落地浪潮中找到各自的商业化路径。 --- ### `25:00` 战略试金石：增收与降本的权衡，以及 Agentic 架构下的推理成本断崖企业高管在 AI 上的增量投入方向正在成为检验其格局的试金石。以 Stripe 为代表的一流科技企业经常与 a16z 探讨这一课题：增量预算究竟应该投向能开拓新客群、带来无限收入上行空间的新产品，还是仅仅投向内部流程优化与成本削减？合伙人们直言： > “增收的机会在上行方向是没有天花板的，而降本的机会……那是一项无论何时都在那里的潜在机会。如果你认为对 AI 最好、最极致的利用仅仅是优化成本结构，这岂不是反向说明了你眼中的收入机会有多么匮乏？” 他们特别指出，半年前硅谷不少企业耗费极其昂贵的顶级工程师团队，内部重写底层记录系统（Systems of Record）去省几十万美元，这显然是对工程资源的巨大浪费。与此同时，以智能体为核心的 Agent 架构正迎来质的飞跃。OpenRouter 平台上的智能体 Token 消耗量实现了 14 倍的惊人增长。为了执行复杂任务，Agent 需要多步规划与多次模型调用，这原本会导致调用成本居高不下，但工程底层的技术革新正在解决这一难题： ```mermaid flowchart LR accTitle: Agent工程优化路径 accDescr: 展示提示词缓存、模型路由与精调小模型如何降低Agent推理成本与延迟 subgraph Optimization["工程优化维度"] Cache["提示词缓存 Prompt Caching (复用背景上下文)"] Router["智能模型路由 Smart Routing (按需分派任务)"] FineTune["小模型微调 Fine-Tuning (专精任务下放)"] end subgraph Impact["商业与技术结果"] Hebbia["Hebbia 金融检索 成本降低 10 倍"] Databricks["Databricks 路由引擎 表现超越单一大模型且成本省 35%"] Voice["Harvey / EliseAI 语音场景 成本降 60% 且实现极低延迟"] end Cache --> Hebbia Router --> Databricks FineTune --> Voice ``` - **提示词缓存（Prompt Caching）**：系统无需在每一步从头解析海量背景数据，金融 AI 平台 Hebbia 借助缓存技术将其金融对话与分析工作流的运行成本降低了整整 10 倍，让既有预算能支撑起十倍的工作量。 - **智能模型路由（Smart Routing）**：Databricks 开发的路由器能依据任务难度将请求分发给最合适的模型，综合表现优于最强单体模型，且综合成本下降了 35%。 - **小模型微调（Fine-Tuning）**：如法律 AI 平台 Harvey 以及物业 AI 平台 EliseAI，通过精调垂直小模型，不仅使成本暴跌 60%，更压低了端到端延迟，使得原本因时延问题而难以落地的实时语音交互场景完全变得切实可行。 - **商业模式重塑**：软件应用层公司的度量衡正在从单次 Token 计费转变为“完成客户交付任务的成本”。只要产品能够搞定任务并按结果收费，底层通过路由组合更便宜的模型，就能沉淀出惊人的毛利率提升空间。 --- ### `29:15` C端 AI 订阅的破局点与罕见的“微笑留存曲线” 话题随后转向消费级（C端）市场。尽管当前全美仅有略高于 2% 的家庭为 AI 产品支付每月订阅费用，整体渗透还处于极早期，但其表现出的黏性已超出传统软件的认知范畴。在典型的消费级软件产品生命周期中，用户留存曲线通常是在随时间推移逐渐下滑后走平（Flatline）；然而，合伙人们观察到，顶级生成式 AI 产品的订阅留存呈现出极其罕见的“微笑曲线”（Smiling Retention Profile）——曲线在触底后随着时间推移反而逐步反弹上扬。这一异象的本质原因在于，底座模型与底层产品的进化速度极快，当产品体验、能力与生态不断自我革新时，大量流失或沉睡的老用户会被全新解锁的功能所唤醒，重新回到产品矩阵中持续付费。 ### `30:00` 消费者 AI 的渗透率、使用时长悖论与代理范式转移在探讨订阅模式能为付费用户带来多大实际价值时，团队首先关注到图表右下方所呈现的惊人反差：当前消费者 AI 的订阅渗透率其实极低。相比于传统消费级互联网平台——如拥有超过 2 亿美国家庭用户的 Amazon Prime，或是覆盖 7000 万家庭的 Netflix——AI 订阅的覆盖面才刚刚起步。虽然团队合伙人 Josh Elman 曾提出过一个耐人寻味的定义：“消费级 AI 是指我用于日常生活、而不一定拿去公司报销的产品”，因此 97% 的美国家庭尚未订阅付费 AI 并不令人意外，但这恰恰说明该市场蕴含着巨大的扩张空间。 ```mermaid flowchart LR accTitle: 从屏幕时长到后台代理的交互范式演进 accDescr: 展示传统社交消费应用以屏幕时长衡量，而现代AI应用转向后台自主代理运作的对比过程。 subgraph 传统范式["传统消费互联网平台"] A["高频人工交互"] --> B["每日停留 30-60 分钟"] B --> C["展示型/点击型广告变现"] end subgraph AI范式["自主代理与后台计算"] D["一次性指令/复杂任务委派"] --> E["后台长时间持续运行"] E --> F["直接达成端到端结果"] end ``` 针对用户参与度的衡量标准，团队展开了深入探讨。在传统的 Web2 时代，Facebook、Instagram、TikTok 和 Snapchat 这类大型消费级平台的共同特征就是极高的屏幕停留时间，活跃用户每天平均会花费 30 到 60 分钟。然而在自主代理（Agentic AI）时代，这一评估指标正在彻底失效： > “对于下一代应用，我希望它们具备极高的持久性、全天候在线并且超级主动，以至于未来它们可能根本不是我花最多时间停留的地方。” 这种范式转移给投资人通过外部数据衡量参与度带来了巨大挑战。例如，用户触发一个深度研究任务或复杂代理任务后，可能直接合上电脑开车去旧金山市区，在这期间计算任务一直在后台疯狂运转并产生巨大的实际价值，但传统的屏幕使用时长（Screen Time）却记录为零。 - **Amazon Prime 家庭覆盖数**：超过 2 亿户家庭。 - **Netflix 家庭覆盖数**：约 7000 万户家庭。 - **Web2 平台日均使用时长**：活跃用户每天停留约 30 至 60 分钟。 - **全新产品形态分流搜索**：Muse、Instinct 以及 ChatGPT 正在持续从传统搜索引擎手中抢夺真实搜索与购买意图。 --- ### `32:30` 代理电商对传统存量平台的商业模式冲击：GMV 扩张与广告利润池的博弈随着 AI 代理具备跨平台执行能力，现有的消费级发现平台与双边市场（Marketplace）正面临严峻的战略抉择。团队指出，现存平台眼下必须自问两个生死攸关的棋局问题：第一，AI 代理能为我带来多少增量需求与订单？第二，我当前的利润池中有多少比例是依靠牢牢把控“客户关系”与“商品发现流程”来实现的？这种战略分歧在巨头之间已经显现：Amazon 对接入 Muse 这类外部 AI 代理断然回绝，而 Instacart 则持完全开放的欢迎态度。背后的核心逻辑在于商业模式的结构差异：如果剥离 AWS，Amazon 零售与 Instacart 的广告业务收入规模都超过了各自的整体营业利润。对 Amazon 而言，其拥有超过 700 亿美元且利润率极高的广告业务，这完全建立在用户登录网站、主动浏览并点击广告的前提之上；若由外部代理代劳，不仅几乎无法带来任何增量客户，还会直接削弱广告变现壁垒。相反，Instacart 面对的生鲜杂货线上化渗透率仍处于早期阶段，通过代理引入增量订单的价值远大于潜在的流量损失。团队同时提出了一种正和博弈的乐观视角：过去繁琐的多步骤结账往往导致大量交易流失，若能全部委托给代理，原本因嫌麻烦而放弃的旅行预订、今天晚上的聚餐外卖都将顺利下单，整个生态系统的交易总额（GMV）将大幅提升。 > “在 Amazon 上投放的那 700 亿美元广告预算可能只是流向了不同的渠道，或许会以另一种产品形态直接投放到各个代理上，或者找到了更好的受众定向方式；随着交易摩擦降低，我们可能会看到更多的实际消费，这形成了一个推动经济增长的正向飞轮。” - **平台战略分歧**：Amazon 选择筑墙保护自有广告闭环，Instacart 选择接入第三方 AI 代理以获取增量订单。 - **广告利润核心地位**：在发达市场，主流平台从每位用户身上可获取约 200 美元以上的收益，且零售广告业务支撑了核心平台的营业利润。 - **Google 搜索的韧性与隐忧**：两年前外界预测 Google 搜索将被颠覆，但其高变现商业搜索词（如“购买保险”、“预订特定城市酒店”）极具韧性，因为早期的 AI 尚无法代客下单履约；然而一旦具备自主行动力的代理成熟，这一稳固格局将被彻底打破。 - **宏观溢出效应**：资本市场近期的净市值变动仍呈现正向反馈，Meta 等巨头的市值增长幅度远超垂直交易市场可能遭受的潜在估值折损。 --- ### `36:30` 公共 SaaS 市场的盈利分化与现有软件巨头的二次加速路径讨论转向二级市场软件板块时，团队剖析了一组令人触目惊心的结构性数据：当前的公开市场 SaaS 样本中，约 75% 的公司已经实现盈利，但仅有 30% 的公司收入年增速能维持在 20% 以上。如果将增速门槛提高到 30%，二级市场中达标的软件公司绝对数量已经不足 5 家；相比之下，一级市场中团队接触和评估的优质软件企业增速普遍远超 30%。 ```mermaid flowchart TD accTitle: SaaS 企业的防御性壁垒与进攻性加速路径 accDescr: 展示现有软件企业如何利用存量分发渠道和高客户留存率，结合AI产品实现收入再加速。 A["存量高粘性客户 (~98% 净金额留存率 GDR)"] --> B{"战略应对路径"} B -->|"防守端"| C["降本增效与持续盈利"] B -->|"进攻端"| D["注入高单价/高价值 AI 新产品"] D --> E["目标：实现 10%+ 的收入增速二次加速"] ``` 这一现象反映出增量 IT 预算正在被全面抽调去支持 AI 项目，导致传统 SaaS 项目的新签扩张明显放缓。面对这种局面，在二级市场持有 SaaS 资产并深度参与公司治理的 a16z 团队，正与被投创始人密集讨论如何把握破局之道： > “我们与他们以及创始人重点探讨的核心在于：如何利用你现有的分发网络——在很多情况下，这种网络极度强大，牢牢锁定了那些绝不会轻易离开的客户——并向他们提供极其吸引人的全新 AI 产品，从而带动你的收入增速再次走高。” 传统顶尖 SaaS 企业高达 98% 的净金额留存率（Gross Dollar Retention, GDR）构成了坚不可摧的防守底色，粉碎了一年前市场上“AI 将瞬间消灭所有传统软件”的恐慌情绪。但在防御稳固的前提下，企业必须展现出强劲的进攻能力。团队在此前提出的核心行动准则是：利用当前充沛的 AI 采购预算，依托真正具有突破性的产品体验，力争在未来 12 到 18 个月内实现 10% 以上的收入增速二次加速（Re-acceleration）。团队强调，绝不能将整个软件板块一刀切地视为估值下滑和增长停滞的弃儿，优胜劣汰的分化才刚刚拉开序幕。 - **二级市场 SaaS 现状**：样本中约 75% 盈利，仅 30% 增速在 20% 以上，增速超过 30% 的企业屈指可数（不足 5 家）。 - **防御性资产特征**：优质头部 SaaS 维持着接近 98% 的净金额留存率，基本盘未受结构性侵蚀。 - **核心考核指标**：进攻端以拉动 10% 以上的收入增速再加速为战略基准线，验证 AI 增量变现的兑现能力。 ### `40:00` 软件市场的 AI 大分化与“SaaS 复兴” 在软件市场内部，AI 带来的冲击正呈现出极度分化的格局。网络安全（Cybersecurity）和可观测性（Observability）板块表现格外抢眼，垂直行业软件（Vertical Software）的抗跌与增长表现也显著优于通用横向应用（Horizontal Applications）。Sarah Wang 指出，评估软件企业是屹立不倒甚至加速爆发，还是陷入长期衰退，核心在于 AI 如何从底层重塑客户对该产品的真实需求。虽然 AI 带来的网络安全风险已广为人知，但随着更多软件和智能体（Agents）的部署，市场对新型安全与监控机制的需求激增，这反而在客观上强化了 CrowdStrike 等在位领军企业的护城河。相比之下，传统横向应用层正面临不同程度的工作流重构；正如 Databricks CEO Ali Ghodsi 常讨论的“AI 断头台（chopping block of AI）”框架，企业必须看清哪些工作流正首当其冲被砍掉，哪些环节又是 AI 时代的刚需基础设施。在私募市场端，这种逻辑得到了充分印证。Alex Immerman 和 Santiago Rodriguez 强调，随着 AI Agent 开始跨系统调用并执行更复杂的实际操作，系统安全成了不可逾越的红线（OpenAI 与 Hugging Face 的安全事件便是典型警钟）。而在垂直 AI 应用领域，诸如 Harvey（法律）、Abridge（医疗）和 EliseAI（不动产管理）等初创公司的增长速度，已经远超其所在行业历史上的任何前辈。这表明客户深刻认识到：单纯的基础大模型并非全部，如何将模型编排进特定行业严苛的工作流应用中才是价值所在。回看年初所谓的“SaaS 末日（SaaS apocalypse）”恐慌，公开市场软件指数如今虽已反弹至年初水平，但内部早已彻底分化为“AI 输家”与“AI 赢家”。与公开市场的悲观论调不同，底层真实的企业运营数据打破了旧有叙事。Stripe 的 SaaS 客户数据显示，无论是初创企业还是成熟公司，其业务增长在 2026 年均呈现再加速态势，Stripe 将这种强劲反弹形象地称为“SaaS 的复兴（The Renaissance）”。与此同时，企业保持私有化（Stay Private）的周期也在大幅拉长。以 Anthropic、OpenAI、Databricks、Stripe、Waymo 和 Revolut 这六家私有巨头为例，按最新一轮估值计算，其总市值已达到惊人的 2.4 万亿美元。这一数字不仅超过了过去十年所有 IPO 企业的市值总和（剔除 SpaceX 后的 1.7 万亿美元），甚至逼近了容纳数百位基金经理的罗素 2000（Russell 2000）指数的总市值体量。 ```mermaid flowchart TD accTitle: AI 驱动的软件市场分化格局 accDescr: 展示 AI 如何驱动安全可观测性、垂直应用与横向通用软件的分化，以及后期巨头的私有化演进 subgraph Winners["受 AI 顺风驱动的赢家"] A["网络安全与可观测性 (CrowdStrike 等在位巨头)"] B["垂直行业 AI 智能体应用 (Harvey / Abridge / EliseAI)"] end subgraph Pressured["面临工作流重构压力的板块"] C["通用横向 SaaS (处于 AI '断头台' 的被替代风险中)"] end subgraph PrivateGiants["私有化长跑巨头 (总市值 2.4 万亿美元)"] D["OpenAI / Anthropic"] E["Databricks / Stripe"] F["Waymo / Revolut"] end Winners -->|"捕获净新增预算，估值坚挺"| G["SaaS 复兴 (The Renaissance)"] Pressured -->|"工作流遭重塑，面临剧烈重估"| G PrivateGiants -->|"推迟 IPO，支持长期研发下重注"| G ``` - “以最新一轮估值计算，Anthropic、OpenAI、Databricks、Stripe、Waymo 和 Revolut 这六家头部企业加起来市值约为 2.4 万亿美元；这已经超过了过去十年全部 IPO 公司合并市值的总和（排除 SpaceX 后的 1.7 万亿美元）。” - 软件指数重回年初点位，但市场逻辑出现结构性两极分化：“AI 赢家不仅是能改善成本结构的公司，更是那些能实质性加速并捕获市场上净新增预算的公司。” ### `45:00` 上市抉择、员工要约收购与 86% 的 AI 投资版图面对是否进行 IPO 的抉择，David George 深入拆解了 a16z 给创始人制定的战略逻辑。在当前的市场环境下，创始人本身就是一种独立的核心资产类别。优秀公司之所以选择长期停留在私募市场，关键原因在于创始人可以在不受公开市场每季度审视的压力下，进行回报周期更长、体量更庞大的激进产品豪赌。尽管扎克伯格（Mark Zuckerberg）和埃隆·马斯克（Elon Musk）能在公开市场力排众议顶住压力——例如 Meta 曾因全力押注元宇宙而导致股价跌破每股 100 美元低谷——但普通上市公司往往难以承受这种级别的公众质疑。相反，像 Databricks 和 Stripe 这样的非上市公司，正是依靠在私募阶段进行的大手笔新产品研发，最终实现了收入的再加速。当然，IPO 依然是重要的融资与信用工具。有些企业业务运行极其透明严密，其管理指标甚至比上市公司还要规范；而像企业差旅平台 Navan 这样的公司，走向公开市场后反而赢得了大型企业客户的深度信任，进而带动了业务再加速。另一方面，随着这些企业的资金需求日益庞大，即便是深度雄厚的私募资本市场，终究也会迎来承载极限。为了解决员工流动性并对抗大厂每季度发放限制性股票（RSU）的薪酬优势，二级市场要约收购（Tender Offers）已成为顶级独角兽的标配。有趣的是，根据 Carta 统计的要约收购数据，员工实际参与要约套现的比例仅有 58%。这直接打破了“内部员工急于兑现走人”的刻板印象，证明绝大多数员工对公司的长期前景抱有高度信念。此外，相比 2021 至 2024 年间二级市场普遍存在的深幅折价，如今优质资产在二级市场的交易折价率已基本归零，新入场的资金完全愿意按最新一轮融资的原价买单。视线转向整个风险投资大盘，AI 资本集中的趋势正达到历史顶峰：在 2026 年的数据切片中，AI 相关公司占据了全美 VC 交易总额的 86%，较 2025 年的 65% 进一步飙升。虽然市场焦点大多集中在 OpenAI、SpaceX、xAI 和 Anthropic 等明星公司身上，但实际的投资机遇早已从单纯的横向模型扩展至底层电力、半导体制造、物理防御基础设施以及垂直行业应用。最后，团队展望了未来最令人振奋的前沿赛道：除了能全天候自主执行复杂繁琐任务的长期运行消费者智能体（Long-running consumer agents），团队目前倾注最多精力的领域是具身智能机器人（Robotics）。a16z 团队认为，机器人领域的潜在市场天花板甚至可能超越大语言模型（LLM），尽管它目前的商业化发展阶段要比 LLM 晚三到五年，但在未来五年内，它必将迎来爆发式的基础设施投入与增长。 - David George 谈及创始人治理与资本市场时总结道：“对于由创始人领导的公司而言，我们讨论并写过的一点是，在当下这个节点，创始人本身就是一个资产类别。” - 谈及公开与私有市场的本质差异时，David George 强调：“在我们看来，与创始人的沟通方式是：IPO 只不过是另一次融资事件；核心在于成为上市公司相比留在私有公司，你究竟能获得什么？” - AI 赛道资本虹吸效应显著：AI 相关交易在全美风险投资中的资金占比从 2025 年的 65% 跃升至 2026 年统计切片的 86%，投资外延已全面渗透进企业级软件、消费级终端、半导体、能源电力以及国防工业等关键实体支柱。 ### `50:00` 自动驾驶、AI 生物医药与企业扩散的前景自动驾驶技术已经迎来了真正见效的关键拐点。退一步审视整个汽车与交通出行产业，这原本是全球规模最庞大的行业之一，但在当下生成式 AI 唱主角的舆论场中，它的受关注度反而被相对掩盖了。目前，通过 Uber 和 Lyft 等网约车平台完成的出行里程，仅占全美总行驶里程的 1% 左右。然而，随着全自动驾驶车队的逐步落地，自动驾驶汽车在安全性上展现出了惊人的优势——其安全性是人类驾驶员的整整 14 倍。演讲者预测，在全自动驾驶网络铺开后，这一出行渗透率在未来几年内将迎来至少一个数量级的跃升；加之美国每年销售的新车数量高达 1700 万辆，预计在未来 10 年内，这些新车都将全面迈向自主化。在自动驾驶之外，AI 与生物医疗（AI × Bio）的交叉融合同样展现出巨大的变革潜力。无论是在各大研发实验室的底层技术探讨中，还是在前沿产业落地层面，利用 AI 进行新型药物发现、攻克全球范围内最具致坚性和致残性的顽疾，都是未来 10 年极具确定性且令人振奋的突破方向。与生物科技相辅相成的是个人健康（Personal Health）领域的智能化重塑。对于关注自我量化与极限健康管理的群体来说，过去一直缺乏一个综合性的渠道，能把个人全维度的身体与生理数据汇集在一起，并输出高度超个性化（hyper-personalized）的健康干预与行动建议；AI 正在彻底打破这一壁垒。此外，AI 在企业端的扩散早已超越了单纯的辅助编程，正全方位渗透到各类企业工作流中，带来深远的生产力重塑。 - **出行里程渗透空间**：Uber 与 Lyft 目前仅占美国总出行里程的约 1%，未来在全无人驾驶车队推动下预计将实现 10 倍（数量级）以上的增长。 - **驾驶安全性飞跃**：完全自动驾驶车辆的安全记录表现为人类驾驶员的 14 倍。 - **美国新车年销量**：全美每年新车销售量约为 1700 万辆，未来 10 年预计将全面转向自动化与自主化。 - **个人健康数据汇聚**：AI 使得将零散的个人生理指标汇聚并生成超个性化建议成为可能。 > “Plus there's 17 million new cars sold per year in the US. And I think over the next 10 years, those will all be autonomous.” ```mermaid flowchart TD accTitle: 自动驾驶技术驱动下的交通出行变革 accDescr: 展示自动驾驶技术提升安全性并推动网约车与新车市场全面自主化的路径 subgraph Safety["核心技术突破"] A["全自动驾驶车队网络"] -->|"安全性达人类驾驶 14 倍"| B["极高安全与可靠性"] end subgraph Impact["市场规模扩展"] B --> C["网约车里程渗透率提升 (从当前 1% 增长一个数量级)"] B --> D["新车市场全面自主化 (美国每年 1700 万辆新车)"] end ``` --- ### `51:24` “美国活力”重构与万亿 AI 基建的宏观经济效应除了纯粹的 AI 赛道，团队重点布局的另一大核心方向是“美国活力”（American Dynamism）底层技术堆栈的全面重构。a16z 长期以来作为该领域的重度投资人，积极参与重塑国家关键能力与基础设施的过程。当前，诸如 Anduril、Saronic 或 Castelion 等新一代防务与硬科技供应商，在整体相关支出中所占的份额依然微乎其微，甚至不足总支出的 5%。随着现实地缘政治与安全需求的深刻演变，这一比例预计在未来将迎来爆发式攀升。总结全篇，无论是自动驾驶、生物制药、个人健康，还是国防与国家级硬科技重构，未来的创新机会呈井喷之势。尽管围绕 AI 展开的基础设施建设投资规模达到了惊人的近万亿美元级别，但这种前所未有的 CapEx 扩张绝非泡沫盲动，而是将从根本上大幅提高美国的整体生产力，为宏观经济注入极其强劲的增长动能。团队在总结中对未来技术驱动的繁荣表达了坚定的乐观态度，并强调了在多维度持续重注前沿创新的投资决心。 - **新一代供应商占比**：在国防与“美国活力”相关总支出中，Anduril、Saronic、Castelion 等新兴创新厂商占比尚不足 5%，成长空间极为广阔。 - **宏观生产力红利**：万亿美元级别的 AI 基础设施重资产投入，其终极落脚点在于全面提升美国宏观经济的实际生产力效率。 > “The build-out is massive. But we think it's going to be massively productivity-enhancing in the U.S.” ## Notable quotes > `12:30` 当我们谈论超大规模厂商的资本开支时，应该把它看作是其他企业的订单簿。 — 圣地亚哥·罗德里格斯（Santiago Rodriguez） > `45:00` 在当下这个节点，创始人本身就是一个资产类别。 — David George > `45:00` IPO 只不过是另一次融资事件；核心在于成为上市公司相比留在私有公司，你究竟能获得什么？ — David George > `16:45` 模型能力的高度与它们当下的实际应用程度之间所存在的巨大鸿沟，正是让我对应用层前景感到无比兴奋的原因。 — 戴维·乔治（David George） > `25:00` 如果你认为对 AI 最好、最极致的利用仅仅是优化成本结构，这岂不是反向说明了你眼中的收入机会有多么匮乏？ > `12:30` 从我们的视角来看，今天押注未来与过去已经截然不同；这意味着建造真正的工厂、扩展实体基础设施，并在全美范围内创造更多的高技能技术岗位。 — 萨拉·王（Sarah Wang） > `30:00` 对于下一代应用，我希望它们具备极高的持久性、全天候在线并且超级主动，以至于未来它们可能根本不是我花最多时间停留的地方。 ## Mentioned David George, Sarah Wang, Alex Immerman, Santiago Rodriguez, Marc Andreessen, Sam Altman, Sarah Friar, Elon Musk, Dina Powell, Josh Elman, Ali Ghodsi, Martin Casado, Mark Zuckerberg, a16z, Alphabet, Amazon, Meta, Microsoft, Oracle, OpenAI, Anthropic, SpaceX, xAI, Waymo, Databricks, Stripe, Revolut, ElevenLabs, Cursor, Cognition, Decagon, Chime, Shopify, Sidekick, ServiceNow, Salesforce, Workday, CrowdStrike, Anduril, Saronic, Castelion, Carta, Navan, Hebbia, Harvey, Abridge, EliseAI, OpenRouter, Netflix, Instacart, Muse, Instinct, ChatGPT, Uber, Lyft, YipitData, Hugging Face, S&P 500, Russell 2000 ## Video description a16z’s David George, Sarah Wang, Alex Immerman, and Santiago Rodriguez unpack 25 key charts from the latest State of Markets presentation, from the scale of the AI infrastructure buildout to what adoption looks like inside companies today. They examine why rising markets have so far been supported by earnings rather than multiple expansion, why hyperscaler CapEx is approaching $1 trillion annually, and why demand for compute continues to outrun supply. They also look at the downstream effects of that spending across chips, power, construction, and physical infrastructure. State of Markets Then they move up the stack: OpenAI and Anthropic’s revenue growth, the gap between AI deployment and measurable enterprise impact, the rise of agents, falling inference costs, and what all of this means for SaaS. They close with where the team is spending time next, including consumer agents, robotics, autonomy, AI and biology, personal health, defense, and the continued diffusion of AI across the enterprise. State of Markets [Source](https://mgln.ai/e/1344/afp-848985-injected.calisto.simplecastaudio.com/3f86df7b-51c6-4101-88a2-550dba782de8/episodes/7104380b-ad2f-4075-b459-cb07cd66a70b/audio/128/default.mp3?aid=rss_feed&awCollectionId=3f86df7b-51c6-4101-88a2-550dba782de8&awEpisodeId=7104380b-ad2f-4075-b459-cb07cd66a70b&feed=JGE3yC0V)

Source quotes

[12:30] 圣地亚哥·罗德里格斯（Santiago Rodriguez）: 当我们谈论超大规模厂商的资本开支时，应该把它看作是其他企业的订单簿。 [45:00] David George: 在当下这个节点，创始人本身就是一个资产类别。 [45:00] David George: IPO 只不过是另一次融资事件；核心在于成为上市公司相比留在私有公司，你究竟能获得什么？ [16:45] 戴维·乔治（David George）: 模型能力的高度与它们当下的实际应用程度之间所存在的巨大鸿沟，正是让我对应用层前景感到无比兴奋的原因。 [25:00] 如果你认为对 AI 最好、最极致的利用仅仅是优化成本结构，这岂不是反向说明了你眼中的收入机会有多么匮乏？ [12:30] 萨拉·王（Sarah Wang）: 从我们的视角来看，今天押注未来与过去已经截然不同；这意味着建造真正的工厂、扩展实体基础设施，并在全美范围内创造更多的高技能技术岗位。 [30:00] 对于下一代应用，我希望它们具备极高的持久性、全天候在线并且超级主动，以至于未来它们可能根本不是我花最多时间停留的地方。

Source quotes

[43:10] David George: One of the things that we've talked about and that I've written about is that the founder is the asset class at this point. [25:00] Alex Immerman: And if you think that the best and highest use of your resources is to optimize your cost structure, what does that say about the revenue opportunity for you?

Original transcript

[0:00] Eight of the top 10 valued companies in the world are U.S. tech companies. Since JotGPT came out almost four years ago, the market's up 90%, which is 17% annualized. The natural instinct is, well, that's got to come down. We're definitely in a hot period. This puts us in a new age of atoms. Global infrastructure investment needs are estimated at $90 trillion through 2040. This goes way beyond AI and data centers. It includes power, water, roads, transit. at S&P 500 companies, that's at 69%. Now, if you go to the ultimate barometer, which is a metric tracked over time, that's actually only at 2%. AI is generating major revenue and savings. On the other hand, adoption is still extremely early. Today's opportunity is so much around taking these capabilities and harnessing them to build reliable services. That's all great, but is it a investment cycles in modern history but is the spending getting ahead of the economics today david george sarah wang santiago rodriguez and alex immerman walk through 25 key charts from our latest state of the markets presentation they start with the macro picture why markets have risen even as multiples have come down how hyperscaler capex is approaching 1 trillion dollars a year and why demand for compute continues to exceed supply. Then they look at what's happening inside businesses. AI adoption is growing, but measurable enterprise impact is surprisingly early. At the same time, power users are pulling away. Agents are consuming more tokens. Inference costs are falling. And the fastest moving companies are finding ways to turn those improvements into both growth and better margins. And finally, they look at what comes next. Welcome back to the A16Z Podcast. I'm David George. I'm here with my colleagues Sarah Wang, Alex Emmerman, and Santiago Rodriguez. Welcome back to the A16Z Podcast. I'm David George. I'm here with my colleagues Sarah Wang, Alex Emmerman, and Santiago Rodriguez. Today, we are walking through 25 key slides from our latest state of markets presentation. The earnings behind the markets rise, the scale of the AI build out. the evidence of growing adoption, and what this cycle means for hardware, software, and the next generation of private companies. We'll explain what the charts show and discuss what we're seeing inside businesses along the way. So you can follow along whether you're watching or listening. So Sarah, Alex, Santee, thanks for joining me. Of course. Happy to be here. Okay, so... The day this podcast goes live, we will be releasing our state of markets presentation. So this is a now yearly tradition from our growth team today. where we synthesize the biggest trends in tech, AI, infra, and markets. Today, we're just picking out a subset of interesting slides and having a discussion about them. So I would direct you to look at the whole version, which has filled with a lot of nuggets. The areas that we're going to discuss today... are macro, so where we'll talk about capex, data centers, accelerating demand at a high level, And then on the ground takeaways where we will discuss models, apps, and some vertical deep dives. Technology is driving an economy-wide investment boom, with rising earnings supporting market gains and AI demand pushing infrastructure spending far beyond earlier forecasts. Hyperscalers are investing all of their near-term operating cash flow to build out this capacity for demand that continues to outstrip supply in almost every case that we see. They are channeling this investment into chips, power, cooling, construction, skilled labor, and much more. This expansion coincides with a broader need to modernize physical infrastructure and, creating opportunities across industries and potentially lowering shared costs for businesses and households. So with that, let's jump in. Okay. First slide, tech is the everything cycle. So some of this may be a little bit obvious, but the numbers are somewhat striking at this point, right? So high-tech equipment, software, and R&D now account for roughly – 55% of U.S. capital spending, which is just a staggering number. Tech is driving the investment cycle, obviously, across all areas of the economy, from software and models to power construction and industrial capacity. Tech is almost 40% of the aggregate value of the whole stock market in the U.S., and eight of the top 10 valued companies today. in the world are U.S. tech companies. So, This is a broad story. You're seeing it in CapEx. This is heavily covered in the data center build out. It's obviously heavily covered in the amount of capital that's been going in to fund the model development. model companies together raising I think over $350 billion at this point. And just to put it into very deep historical context, because this is probably the analogy that we've seen the most of, [5:00] Buildout just surpassed railroads as a percentage of GDP. So exciting times, massive buildout. We all happen to think that, If you fast forward five to seven, maybe 10 years from now, we'll be looking at these numbers and they'll probably be 20x higher. Cumulatively. So, Yeah, this does feel like the next chapter of Mark's software's eating the world. Over the last 15 years, software has transformed all these industries, but only a small fraction of the population could build it. None of us could. But today, look at us. We all have a handful of automations running every night. And so if you think about the software demand, that means a lot more compute, chips, power, construction. And so it's no surprise to see the majority of investment now covered in tech. Okay, so one of the questions that we get all the time from various audiences is, Okay, that's all great, but is it a bubble? So the market has reached new highs, and at the same time that it has reached new highs, the trading multiples of the market are actually down. So stocks are up about 20% while multiples are down about 20%. So what that means is the performance is driven by fundamental earnings, not increased multiples. The S&P 500 earnings multiple is below 20 times. So if you just start with that, these are in most cases very high-quality businesses. world had their massive stock run-ups based on increases in their trading multiples and would trade in many cases for 100 times PE. That's not what's happening here. In contrast, some of the memory companies, which again are very cyclical, are trading for, call it six times, seven times forward earnings. So very, very different. And I think there's an important double click here where since ChatGPT came out, Almost four years ago, the market's up 90%. which is 17% annualized. And I think anytime there's been a 17% annualized gross for four years, the natural instinct is, Well, that's got to come down, right? We're definitely in a hot period. But when you compare that to, as you said, the market trading for below 20 times earnings. growing 15%. This definitely feels a little different than maybe the 2021 period that was recent or the 2000 period when multiples in growth were not really going together. Yeah, totally agreed. So I mentioned the scale of the CapEx build-out on the context of the railroads. If you just look at the hyperscalers, Alphabet, Amazon, Meta, Microsoft, and Oracle, Their CapEx in 2026 is about $780 billion. That's up from $416 billion in 2025. And all expectations point to them spending over a trillion dollars annually from 2027. So the pattern recognition is each of these computing platforms has supported a much larger population of users and uses. In this case, this build-out can happen so quickly and demand can still outstrip supply. is driven by the fact that there's already existing distribution. This trend is built on top of the internet and cloud computing, obviously, and mobile phones. And so sort of immediately could reach... billions of users in contrast to previous technology cycles. Yeah, and I think it's an important point to double click on just because we've moved well beyond this model of occasional queries. Right. If you think about the rise of agents, you have parallel tasks, you have long running tasks. And so, I mean, Alex, you mentioned this previously, right? You have tasks going into the evening, during the day when you're doing other things. And so if you think about agents autonomously writing code, searching, carrying out these tasks, it just gives these platforms a much larger compute requirement than ever thought before. Yeah, absolutely. Totally agree. This one is another one that just points to what is happening in the CapEx side of things. Successive forecasts for the five largest hyperscalers CapEx have moved sharply higher. in pretty short succession. So spending that once looked like a ceiling, a number of quarters or years out, has become near term as the demand for the compute keeps expanding. Yeah, look, if you look at the chart here, I think at any point in time, I think natural instinct is just to say, like, the investment will kind of flatline from here. Yet for the last four years, I think, Like as an economy, we keep underestimate just the strength of the trend. And as Sarah said, with model developments, usage of our installed capacity keeps being taken up. So right now, the numbers that David pointed to are the current estimates. Yeah, I mean, maybe to use DG's language, I think we'd all call the demand for compute a model buster at this point. I think one anecdote that's telling that we've all experienced is Sam Altman, Sarah Fryer. They got a lot of flack a year or so ago for their massive compute commitments. They were being reckless and aggressive. And I think at this point, like everyone would say they're incredibly prescient with that decision. And even so, two weeks ago, we all saw that they had to pause new subscriptions on their pro plans. showing up with a $2,000 service. No thanks. Pretty amazing. [10:00] insatiable demand that we're experiencing here. Yeah, absolutely. Yeah, pretty much everyone we talk to at every stage of the supply chain is telling us the same thing, some version of the same thing that demand outstrips supply. There are certain elements in the data center supply chain where you can't get access to materials or products until 2028. And so this is not softened. So I would say at the same time, you can look to the hyperscalers and see some evidence of high quality, you know, business on the demand side that you can hang your hat on. Right. So Microsoft, Google, and Amazon have, you know, about 1.7 trillion of combined cloud backlog. together. And those customer commitments are building rapidly while the platforms invest heavily in the capacity to serve them. So while right now free cash flow is depressed during this build out, you know, consensus forecasts are showing recovery from 2028 and substantial growth thereafter. You know, you could look to Amazon's, actually Amazon's latest earnings call, where they did a really good job of explaining this sort of J curve dynamic where the useful life of GPUs is actually pretty, or TPUs, is actually pretty long. And so, you know, you have to build out the shell, the data center, you know, that's a certain amount of time you have to buy the chips. But, you know, those will have a very useful economic life for a long period of time. And it's been longer than I think any of us expected. Yeah, I know some of the hyperscalers have frankly gotten dinged for raising debt for CapEx. But, you know, similar to the model lab dynamic, I think the ones who have blinked and been less aggressive have regretted it. I know on the podcast recently that you did with Gavin, Microsoft came up. But I think this is an issue across the board. the hyperscalers also earning an attractive return. Yeah, it's a key point. You know, this is one of the things that we talk about all the time is like each one of these successive waves just creates a tremendous amount of user or consumer surplus. And so, What's actually happening right now, at least as far as we can tell, is, you know, consumers and users get a tremendous amount of value out of this. Like, that's why they're using it so much. You know, the... the vendors who are serving those users are making very good money. And then you go all the way down every level of the stack to the chips where you have to pay much higher than you did 12, 18, 24 months ago to get access to them. And yet you could still make very high margins and create a tremendous amount of surplus out of the users. And as we talk about this CapEx for the hyperscalers, we should think of it as someone else's order book. So these big platforms with historically the largest profits are pouring their cash back into AI infrastructure. That's putting pressure on their free cash flow short term, as we just talked about. But it's been a boon for chip orders, for power, for construction. And that's why this broader technology boom has become an industrial boom. And we I mean, we've been spending more of our time looking at businesses serving this industrial boom and everywhere along the data center supply chain. I think it's interesting because. Traditionally, our world was more about monetizing existing IP, build once and then sell infinitely. But with these businesses, there's many more complexities that the companies need to manage, like financing, managing vendor relationships. predicting capacity and predicting demand. And we've seen, you know, there are some teams that excel at that and some teams that have struggled or kind of taken a pause. And it's a little bit of a different expertise than we typically spend our time with. This puts us in a new age of atoms. Big numbers on this page. Global infrastructure investment needs are estimated at $90 trillion through 2040. Importantly, though, this goes way beyond AI and data centers. It includes power, water, roads, transit. We're seeing this across our portfolio. You can see this at Anderil with their massive manufacturing facility. 87 football fields. Waymo, they're, you know, aggressively expanding their depots. And of course, you know, SpaceX with their $100 billion investment in Louisiana. Betting on the future today from our perspective isn't, you know, what it exactly was in the past. It was in the past, as Santi said. It means building factories, expanding infrastructure, more skilled jobs and across the country at large. Thank you. Investing in physical infrastructure is very different than like, you know, investing in software and Elon coined the term, you know, the factory is the product many years ago. And it's no surprise that a lot of the great founders that we've backed have come out of Tesla or SpaceX because they just learned that, you know, executing on the factory ends up being a massive competitor. [15:00] advantage as they scale. So there are many myths that we hear all the time about data centers. They're draining all of America's water. Rich people don't want to live near them. And then, you know, one of our favorites, my electricity bill is going to go through the roof. It may seem counterintuitive, but a data center can help lower your electricity rate. A recent U.S. study showed that for every 10% increase in data center capacity, residential rates went down by 40 bps. You should think about it as a power grid is a shared fixed cost base, poles, wires, substations. And so a large stable customer, like a data center, can help spread those costs across more units of electricity. Simply put, more demand across a shared system is a positive. Yeah, I totally agree with what you're saying. Investment in shared infrastructure can also... improve the economics for the households and businesses connected to it. And I don't think that point gets enough media attention. Dina Powell went on a podcast recently from Meta and talked about the Louisiana site that they did where they worked with the community to lower electricity costs. And so there's real examples of that today. It's not just lip service. Totally agree. So now we're going to talk about the trends that we're seeing at the model and application layer. On the one hand, as we've sort of previewed, AI is generating... major revenue and savings, On the other hand, adoption is still extremely early. And then, you know, I think on top of this, the trend is very much so that costs are plummeting. And, you know, we previewed agents already, but Agents are actually economical now for a much wider range of work. And of course, this is changing even as we speak. New innovations like JEV, by newer labs like TypeSafe are taking this to a greater extreme. You're starting to see things like... two orders of magnitude cost differences really impact the number of use cases. It's sort of classic Jevons paradox, which is why it's very aptly named. And, you know, I'd say all in all, this is an awesome setup for both the model layer and the app layer. These improvements are creating real opportunities for both growth and profitability among software companies. And that's not to say a very much expanding range of industries. And we're also seeing more companies just reach enormous scale in the private markets. So this is not a new chart, but is one of my favorite charts, which is the combined scale of revenue for OpenAI and Anthropik. And this is one that we've had sort of in every GP offsite, as far as I can remember, the last few years. And we have to keep updating it every month because it just sort of gets out of date that quickly. I think the point here, you can see in the numbers, OpenAI, Anthropic, they're combined... annualized revenue has climbed to an extraordinary degree. We love to show this relative to the greatest software companies in history. And you can see that in the right that if you look at the estimates for revenue added for the best software companies ever built, the estimates for net new for the leading labs has well surpassed that. So it's really a striking combination of not only to other platform shifts, the interesting thing about the AI shift is we're not only early in terms of percentage of population or percentage of users that use AI, but still very early on kind of the share of wallet gains within those users. So it's a little bit different than like, you know, in 2008, we would model percentage of people with an iPhone. Here, we need to think about percentage of people that will use AI products and then to what extent they're using the AI products. Yeah, I think that point that we're still so early is such an important one. And there's a bunch of metrics in here that I think elucidate that. One is, you know, if you think about live deployments at S&P 500 companies, That's at 69%. If you're AI pilled, as we all are, maybe something closer to what we would all expect. But then if you move toward quantifiable impact, which is probably a good metric of how far along they are in their deployment. That's 30%. Now, if you go to the ultimate barometer, which is a metric tracked over time, that's actually only at 2%. So, We believe that AI is delivering results, but there is a huge amount of room to actually deepen its use inside the organization and track those results over time. And so moving from these individual deployments that you're seeing to really deeply influential recurring workflows is the next stage, frankly. And I would just say that most of the enterprises that we speak with anecdotally, they're exposure to AI is still mostly with Microsoft Copilot, which just shows you how far they have to go. Yeah, this gap between what the models can do and how they're being used is what makes me so excited about the application. [20:00] there. Sarah, you did this great conversation with Ali Goetzi at Databricks with Martine, and he talked about how the AI can know a lot about the world, but know very little about your company. And we're seeing this in our application layer companies where they're bridging this gap. And like one example that has stuck out to me over the last few months is if you were to take Revolut, who has a really sophisticated engineering team, they still had to partner with Eleven Labs and take Eleven's, you know, best in class voice model and harness it to connect securely with customer accounts and banking workflows so that when I as a customer call, I can have my, you know, problem resolved smoothly, efficiently and, you know, securely. Yeah. Today's opportunity is so much around taking these capabilities and harnessing them to build a reliable services. So in addition to the fact that, enterprise adoption is still quite early for AI. The other really remarkable trend that we wanted to put some stats to is just that the power users are totally pulling away. So you can think of AI spending rising generally across companies, but The power users, as we talked about, the most intensive users are spending much, much more. So we looked at some YIPIT data for this. And if you look at sort of median AI vendor spending in the top 1%, that's roughly eight times the level of the top 10%. What's impressive about that is it's almost as much as the 2% to 10% combined is basically the 1%. Yeah. But just you got these power users that are clearly early adopters. Yeah, absolutely. companies, which you could say are... pretty much fully AI native or AI pilled, you're seeing the top users spend anywhere from call it seven and a half to 9,000 a month. And the median users, again, for the most AI native companies, probably closer to the cost of like a monthly subscription of, you know, 200, maybe you're stacking two subscriptions on top. So 200 to 400. So over 20 times the spend, if you think about median versus sort of top users. Yeah, that, you know, we talked to our own portfolio companies. And, you know, one of the questions that we discuss is like, you know, how much adoption do percentage of money that they're spending on AI tools compared to headcount. As an example, you know, I think like relatively forward leaning large Fortune 500 type companies are probably today at like 1%. And, you know, the most forward leaning, you know, AI pilled companies in our portfolio can be as high as like 10%. And so all of these are just questions of like, how early are we into diffusion and how deep will that diffusion go? Yeah, another example we've seen within our portfolio, you know, many enterprises are really excited when they, you know, buy cursor or cognition for the first time. But the reality is that's just like the beginning of the journey. The runway from just procuring and trying these tools to like full adoption is massive. And I mean, Sarah, you touched on this early, but we're starting to see quantifiable case studies. And, you know, we've highlighted a couple here of public companies that actually report on the metrics that, where they've seen meaningful improvements with AI. You know, two that we call out, one on the cost side, for example, Chime has reported that they've reduced their cost to serve by over 10% a year for the last four years. You know, compounded, we're talking about almost a 50% reduction in cost to serve, you know, happily supported by one of our portfolio companies, Decagon. But just in general, many initiatives to lower the cost to serve. You know, in the revenue side, an interesting one that I found was Shopify here, who... They launched their AI Sidekick and this AI Sidekick helps merchants get up to speed much, much faster. So the percentage of customers that reach five orders within 15 days after onboarding has grown 8%. And that is kind of a metric that Shopify tracks is once you reach five orders, you're going to stick around and you're going to retain on Shopify. So it's been a meaningful tailwind to their business as well. Yeah, absolutely. And I think to your point on just the cost to serve coming down, right, if you think about the advantages that lower cost to serve gives you, you just have more room to compete. So you have better pricing, you know, maybe broader service, you can reinvest in growth. And so I think the benefits are reaching the customers, not only the customers, but also the margins, and then they rotate that back in. it, but the incumbents have done a pretty nice job on monetizing this as well. And ServiceNow is a good example, right? They've reported more than a billion in AI ACV and actually a 9x increase in agentic deployments. So you're kind of seeing it across the state. [25:00] back. incumbents to newer companies. Yeah, one of the interesting things that we've talked about a lot, and again, some of our most forward-leaning companies like Stripe have discussed with us, is where are they actually putting their incremental AI investment dollars? Are they putting it toward things like building new products for customers that could drive higher revenue, Or are they putting it toward optimization or efficiency gains on the cost side. And this is sort of a litmus test for us of like, where are the founders or CEOs of these companies seeing the most amount of opportunity? Like the opportunity to drive revenue growth is unbounded to the upside. Whereas the opportunity for cost improvement, you know, yes, you could take that and reinvest it. But that's sort of a latent opportunity that will continue to exist. And if you think that the best and highest use to optimize your cost structure. What does that say about the revenue opportunity for you? And even within cost optimizations, there's different flavors, right? Like one is, you know, lowering your cost to serve, which allows you to reinvest and like makes you a better business. I think six months ago, our industry was really focused on like rebuilding systems of record internally. And if your engineers are focused on, you know, rebuilding a system of record to save a couple hundred thousand dollars, I expect there to be better uses of like those resources. Yeah, seeing massive agentic usage. Agents are here. They are performing tasks. Tasks require multiple steps, which require multiple model calls. That helps explain the 14x growth in agent token usage on OpenRouter. Those steps, though, are becoming more affordable. We're seeing caching. So the system can reuse background information instead of processing it from scratch each time. I've seen Hebbia take advantage of this. They've seen their financial chat workloads become 10x cheaper to run. The same budget that they had before can now support more work. Yeah, I think your broader point is just that lower costs make it practical now for an agent to try, check its work, try again, and you can just open up a... a ton of tasks where maybe reasoning or tool use would have been just prohibitively expensive. And I think this is especially important in cases where reliability is maybe the de facto reason you would or would not use an agent. And again, you know, we sort of talked about type safe previously, but imagine what happens when you go an order of magnitude cheaper, two orders of magnitude cheaper. You know, I think the amount of use cases that open up are actually unimaginable. improvement in latency. Yeah, absolutely. Yeah, so companies are definitely thinking more and more today about how to optimize latency or performance at large with cost. So, Take Databricks as an example. They're leveraging routing, choosing the appropriate model for each specific task to improve performance and cost. And so their smart router performed better. It solves more problems at 35% lower cost than the strongest individual model. Another approach we're seeing a lot of right now is fine tuning. We've seen this with Harvey. We've seen this with Elise AI. tuned a smaller model. So it became much more affordable, 60% cheaper, but also had way lower latency. So live use cases from an audio perspective became tenable. And so these engineering gains are making AI more useful, but also more affordable. Yeah. And I think that if you sort of roll that up, um, And think about for application business, the relevant unit for them is really the cost of getting the customer's job done. And if you can decouple that, right, you get the customer's job done, you charge for that, and you can use better routing to make the product more economical. That's pretty magical because now you're not using the most expensive model for each step, but you are getting the job done. And that's going to I think we're going to see more margin improvements in the best app companies. It's a nice segue. It's a nice segue, yeah. We've been talking mostly about usage inside of the enterprise, but important to talk about consumer, given obviously that's... the place where we've had some of the largest outcomes and where we spend all of our time. No, we think it's still pretty early. There was a recent survey that showed that only just over 2% of U.S. households actually have a paying subscription for AI. Maybe subscription is not going to be the best place to monetize the majority of households, but it has been so far. And the interesting thing about the subscription revenue is it has some of the best retention we've to see a retention profile that not only flatlines, but actually smiles as, you know, the product improves and people come back to it. But that's what we've seen with AI, which just demonstrates [30:00] the value that the subscription is being, is able to deliver to the, to the person who pays. By the way, just on subscription, what struck me about these numbers, especially if you look at the bottom right chart, is just how small they are. Like think about Amazon prime. That's over 200 million household. Um, Netflix. Netflix is a great one. 70 million households, right? I had to look it up. I'm like, is that K correct? And so totally agree that it's just getting started. Yeah. I mean, our partner, Josh Elman, does have a definition I like that would be counter to that a little bit, which is consumer AI is what I use for my daily life and not necessarily what I expense. So in his view, he would say like, it's maybe not super surprising that 97% of households You are right, Santi. I do like smiling retention curves. We do like to see users coming back more and more. You know, another common characteristic of large consumer platforms is time spent. So if you were to look at Facebook, Instagram, TikTok, Snap, they all get 60 or 30 to 60 minutes per day from their active users. O2 application, like I, I hope that they are going to be incredibly persistent and always on and super proactive such that they may not be the place where I spend the most time in the future. Yeah, it's going to make our job harder when we can't, you know, look at engagement via outside in data because the agents are working on the background instead of us just looking at like screen time on a social media application. But it is important. I mean, I remember like when the You know, we would set off a deep research task or an agentic task and then, you know, drive up to the city or something. And that's not striving to ISA. That's not screen time being tracked as would have been in prior generations of consumer. Yeah, totally agree. It's early, but it is a really new and exciting time in consumer. Muse and Instinct are no doubt the new kids on the block, but they've grown really quickly. Those, along with ChatGPT, are taking more and more of my queries away from traditional search. But it's really a dynamic time. And so if you are an existing consumer discovery platform, an existing marketplace, you really need to be thinking about your chess moves right now. There's a couple questions I'd be asking. First, how much incremental demand can I get from an AI agent? How many more orders can they bring me? And two, how much of my business, how much of my profit pool comes from owning the customer relationship and discovery? Amazon said, you know, no, no, thank you to Muse. But Instacart said, yes, please. And if you think about Amazon X, AWS and Instacart and look at their advertising revenue, it outpaces all of their operating profit. And so advertising revenue, owning the customer relationship is incredibly important. But for Amazon, you know, how many incremental new orders or certainly customers am I going to get from connecting to Muse? Not that many. Look at Instacart, online penetration of groceries, still relatively early, a lot more orders to go get. And so the optimistic possibility here is that there's going to be a lot more orders coming from... uh, from these applications. Historically, I've had to click all these different ways through to process an order. If I offload that to an agent, all of a sudden, there's no clicks and hopefully more GMV. The trip I once wanted to book but didn't gets booked. The order that we all wanted to place for dinner tonight happens. So I'm optimistic there's going to be a lot more GMV and that'll make up for some of this lost advertising revenue. Yeah, it's sort of this question, though, that's open of how does it get compensated for, right? So to your point, you know, Amazon has an over 70 billion advertising business that's extremely high margin flow through that, you know, is totally predicated on the fact that consumers go to the website and click on the ads. And if they don't do that anymore, you know, what happens? of the internet platforms. Absolutely. Um, you know, they each make, call it 200 bucks plus per, per users in the U S and developed world. Um, um, In the case of Meta, it's an entertainment application. We'll see what happens with that. It's probably a little bit safer. In the case of Google, it's funny. The whole talk track around Google... two years ago when all of this just happened was, oh my gosh, what's going to happen to Google? [35:00] search business. And it turns out it's been really resilient. Part of the reason it's been really resilient is because their very high monetizing ad terms were somewhat safe. Because they were things like, I need to buy insurance or find me a hotel in this city or things like that that you can very directly monetize. But that the AI was not yet capable of going to take action for on your behalf. If that changes, that could be a very different dynamic. Yeah. The other thing, just to add, I think the narrative last week was a little bit like negative sum, where it was very negative, you know, like, look at these marketplaces that are going to be impacted. I do think like the positive sum view of this is, number one, maybe that $70 billion that's spent on advertising on Amazon just finds a different channel. and you know, goes elsewhere. Maybe spend directly on the agents through a different form factor or just finds better ways to target people. And then two, I think, you know, with the lower friction, we might just see more consumption, like people might buy more, you know, and that's just like a positive flywheel that drives economic growth. And I think a little bit we're locked into this, oh, this is bad for profitables, but actually might be just good. Yeah. Totally. Both could be true, by the way. Yeah. That it could be profit pools, but it could be additive overall to the economy. I think we all would be disappointed if we looked back three years from now and there's not a meaningful productivity improvement, actually, at the overall macroeconomic level, which would drive consumer spending higher in the case of advertising. But it might just be that profit pools disappear from certain places and get relocated. The immediate reaction has been net market cap positive across the ecosystem. The meta gains have far exceeded any of the losses from the marketplaces. point. Yeah, going back to our discussion on software and moving maybe more to the public side, I think the big trend that jumps out at you from these charts is that the mix has really shifted toward slower growing, more profitable companies. Roughly 75% of the public software sample here is profitable and only 30% is growing at 20% or more. Which is a staggering number that only 30% of the public software companies are growing at 20% plus. And not only 20%, I mean, that's 30%. If you draw the line at 30%, we've discussed like the absolute number of companies is less than five when, you know, in the private markets, basically every company we see and spend time with is growing well in excess of 30%. incremental new dollars on new SaaS software projects. So, you know, I think, We're investors in some of the SaaS software companies, including the public markets. And the conversations that we're having with them and the founders are very focused on is, okay, how do I take my existing distribution, which in many cases is very, very strong, you know, locked in customers who, you know, to your point earlier, Santi, wouldn't go anywhere and apply really interesting new AI products to them where I can drive my revenue growth higher. post things like 98% gross dollar retention, which was kind of always the sticking point for why they were so attractive. as investments continue to drive efficiency, like we've talked about, but most importantly drive, you know, revenue growth acceleration, which shows that they're sort of, safe from a defensive standpoint, but that the offensive investments that they're making are actually going to drive their business to be better. I thought your point in the blog post that you wrote a couple months ago on two paths was really interesting as a rule of thumb. Maybe the finer point of the percentages. Yeah. You know, the more I talked to founders after it, the more I felt like, you know, almost everyone that we talked to is like, yeah, I'm trying to drive revenue growth higher. And, you know, it's it's it's it's clearly, I think, the predominant path for the types of companies that we're investors in and that we've backed over the years. Again, I think, you know, maybe relative to a year ago when the fear was like, OK, everyone's like that is clearly not what's happening in the market. But, you know, the onus is high on delivering revenue growth acceleration. And so we had said, you know, let's target 10% plus acceleration, which is a high number. But with a magical product, given the budgets that are available to AI, we're going to be is seemingly doable. And so we'll see. I think that there are a few of them that we are close to where we'll see that over the next 12 to 18 months. You know, as with all things, you can't lump every company into the software bucket and sort of, you know, talk about multiples coming down and sort of growth decline as well. In fact, I think [40:00] Software has been very differentiated. And if you look at this chart here, you can see that cybersecurity observability have really stood out. And vertical software as well has generally held up much better than the horizontal applications. And I think the framework for... how we think about what's been, you know, what's held up or even gone up nicely versus, um, in more secular decline, if you will, is that really you think about how AI changes the customer's need for that product. And I think the cybersecurity risk associated with AI has been well publicized. But as you think about more software and agents creating new security and new monitoring needs, this is creating greater demand, honestly, for the incumbents in these markets. And you can see the call out on CrowdStrike on the right hand. And then, of course, on the flip side, applications are facing different degrees of workflow change. Right. And it's sort of, you know, if you think about, you know, one of our top CEOs, Ali Goetze, likes to talk about this chopping block of AI and like what's first in the chopping block. And, you know, you kind of think about on the flip side, you know, what is actually needed for AI and you pair those dynamics together. And I think, you know, it explains a lot of what we're seeing in the public markets. Yeah, I think a lot of this is pretty intuitive with what we're seeing on the private market side. So as you alluded to, Sarah, on the security front, agents are accessing more and more systems, taking more actions. The open AI hugging face incident was an eye-opening event for many. So security is paramount. And then as we look at the vertical AI companies, these are some of the fastest growing companies we're seeing in the private markets, right? Harvey, Abridge, Elise. than any of the precedents ever have in their industry. And that's a reflection of the customers recognizing it's not just the model, which we talked about earlier, but how you orchestrate that, how you build the application around it. And vertical-specific workflows command these needs. And I mean, we spend a lot of time looking at public markets, less so for investments, but we underwrite exit multiples, right? So we track that pretty closely. And I think if we were to recap the year, what's happened? You know, first couple months, there was the SaaS apocalypse. about what kind of businesses would do well. And if you fast forward to today, the software index is actually back to, you know, where it started the year but it's really bifurcated into like some companies that are deemed AI losers and some companies that are deemed the winners and The public market sometimes simplifies things a bit too much, but to your point, the AI winners are not just the ones that can improve their cost structures. They're the ones that are actually... accelerating and capturing some of this net new dollars up for grabs that if you're not really capturing, then you're probably not riding the wave. Yep. Yeah, you know, I think on the flip side, we wanted to showcase some private company operating data that gives us a view of what customers are actually doing. And there's a little bit of a narrative violation in Stripe's SaaS customer data, which actually shows growth accelerating into 2026 across both young and mature businesses. Stripe calls it the renaissance. Oh, the renaissance. I actually quite like the renaissance. I like renaissance. Yeah. It's very good. It's very good. We've been talking about public markets a bit. Sarah talked about private markets, SaaS acceleration, We get a lot of questions around, you know, companies staying private for longer. So if you actually look at the top six companies today, you know, Anthropic, OpenAI, Databricks, Stripe, Waymo, and Revolut by last round valuation, they add up to about $2.4 trillion. This is more than the combined market cap of IPOs we've seen in the last 10 years, excluding SpaceX, which adds up to $1.7 trillion. So just the amount of activity that happens within our market, again, six companies, 2.4 trillion, it's almost as big as the Russell 2000, which is, you know, a big index in the public markets where there's. hundreds of public managers that spend most of their time in. So, you know, it's increasingly exciting to just spend time in these late stage champions that can keep investing in growth more so than maybe maximize short-term profit. But David, you can talk a little bit about how we advise companies on the IPO and when to stay private versus public. Yeah, look, I mean, the IPO, especially for founder-led companies, one of the things that we've talked about and that I've written about is that the founder is the asset class at this point. And so the bet that we make and part of the reason that it's been a benefit for some of these companies to remain private is they can many times take bigger swings in the private markets that have longer duration paybacks. Zuck and Elon are kind of obvious exceptions to the rule in the public markets. But the way you see it manifest in the numbers in companies like Databricks or Stripe, is massive, massive new bets in new product areas. And you can see revenue acceleration that happens as a result. And so you can do that in the public market, [45:00] hits, but it's going to catch greater scrutiny. Obviously, Meta's stock price at the Nader got below $100 a share when people were very, very skeptical about their investments in AR, VR. For us, the way we talk about it with our founders is just the IPO is another financing event. And what do you get out of being a public company relative to what do you get out of being a private company? But clearly, you can reach big scale. in the private markets. And there's similar companies that, you know, already talk about running themselves like they were a public company, you know, with like, high focus on efficiency and just basically like every metro being tracked. There's also some companies that benefit from maybe like public disclosing of financials and earning customer trust, both at the enterprise level and on a consumer level. I mean, Navan went public and we've seen a reacceleration just on the basis of some of the larger enterprises actually trusting them more because they're a public business. Yeah. And then, of course, there are examples where, you know, capital needs over time could be very, very large. And so the capital pools the private markets are very large and can serve the needs of many of these companies. But at some point, you know, the capital needs may even get too big for the private markets. That's the public markets, you know, in the private markets, one of the things that we we spend a lot of time in, too, is secondaries. So there's two dynamics at play. You know, one is we see an increasing amount of companies holding tenders for the employees. So if they choose to stay private, you know, allowing employees to get liquidity outside of the public markets. And an interesting data point show here is, you know, participation in tenders struck by Carta has only been 58%. So this is employees choosing not to take liquidity because they have so much conviction in the performance of their business. So it's the opposite of things. Sometimes there's a narrative of like employees cashing out. What we're seeing is actually employees choosing not to cash out because they believe in their company that was 100%. Exactly. It'd be hard. Exactly. Yeah, the tender offers that happen, you know, we're obviously like participating in these and leading these pretty frequently. So we think they're a good thing. They do serve two purposes. One, you know, for employees and prospective employees. it is sometimes hard to compete with the liquidity appeal of public markets, RSUs that hit your account on a net tax basis every quarter. So, you know, it is a, it is a, you know, a weapon of competition for the private markets that want to compete for employees or retaining their employees in the, with, with, with the public market companies. The other side of it is, We're advocates of more frequent resetting of your valuation, and so that keeps your sort of stock price, you know, private stock price, fresh for a number of reasons. One, it's easier to talk about that with employees and prospective employees. But two, you know, in the event that you want to do M&A, like many of our companies have done, you know, you have fresh currency that you can use. Yeah. No, maybe just the other dynamic, I think people talk about the secondary discount quite a bit, because that's a dynamic that we've seen the last five years for maybe coming out of 2021 and 22, 23, 24, the median discount to the price too high and it was out of date. But actually what you're seeing today is when there are transactions in secondary markets, the discount to the last round price is basically zero, which just speaks to these valuations of which companies have raised that are fresh and there's always net new investors willing to pay that same price, which wasn't the case for the last couple of years. Yeah. Great point. So, This is a fun slide that says everything is AI computer. So AI-related companies account for 86% of US VC deal activity in this 2026 snapshot, up from 65% in 2025. So obviously AI has become the dominant destination for venture dollars. Yeah, but beneath that AI label, the opportunity has broadened a lot, right? So enterprise apps, consumer apps, but also services, semiconductors, power, defense. of the attention goes to OpenAI, SpaceX, AI, and Anthropic. But beyond that, our opportunity set has never been deeper and broader. Last, we thought it would be fun just to talk about some of the areas that we're very excited about right now. Obviously, we touched on some of the consumer work that will now be done by agents. But you can call it long-running agents. You can call it autonomous agents, heavy-use consumer. If you take the perspective of these, that you could have consumer agents do all the tasks that you wouldn't want to do to give you things that you otherwise wouldn't have spent time or money on. It's very appealing. The distribution of this could happen pretty quickly. Robotics is an area where we're spending a lot of time and attention. We happen to think that it could be even larger than LLMs, but probably three to five years earlier. And so we think over the next five years, this is going to be a massive area of investment and excitement. Autonomy. [50:00] here. Self-driving works It's a very exciting time. If you just take, you know, a step back and talk about the auto industry and transportation industry, this is one of the biggest industries in the world that we probably don't talk about as much because we spend so much time talking about AI right now. But, you know, if you look at miles traveled by Uber and Lyft or in Ubers and Lyfts, it represents about 1% of miles traveled in the U S I think with, um, full networks of autonomous driving cars that are 14 times safer than human drivers. We expect that to expand by at least an order of magnitude in the coming years. Plus there's 17 million new cars sold per year in the US. And I think over the next 10 years, those will all be autonomous. AI times bio, this is a super exciting area. You know, obviously the folks in the labs are talking about this. But, you know, new drug discovery, solving some of the, you know, most debilitating illnesses or diseases in the world, I think is a promise that we all are very hopeful for, that we'll see a lot of progress over the next 10 years. personal health. This is another one that I'm very excited about for myself, you know, as a health maxer. But, you know, there's not been a great place or avenue to take all of the information about yourself, put it into somewhere and get very hyper personalized advice. And then, you know, lastly, diffusion into the enterprise beyond coding. This is one of the, you know, the ones into the enterprise, um, that I think it's going to be, you know, super, super exciting. Um, and then lastly, um, you know, there's another area that is not inside the AI bucket, which is, you know, what we call American dynamism, but, um, the sort of retooling of the entire, um, sort of American dynamism stack we've, you know, we've invested, we've been large investors in this area for a while. Um, you know, it's still just a small fraction, less than 5% of overall dollars spent, is sort of newer vendors like Andrel or Sironic or Castellian. And we expect that to grow dramatically as needs change. So, so many areas of excitement. Obviously, we've been very active in this AI space. And we're optimistic about the effect that it's going to have on the overall economy. The build-out is massive. But we think it's going to be massively productivity-enhancing in the U.S. So, it's a blast to hang out with you guys. Thank you. episode of the A16Z podcast. If you liked this episode, be sure to like, comment, subscribe, leave us a rating or review, and share it with your friends and family. For more episodes, go to YouTube, Apple Podcasts, and Spotify. Follow us on X at A16Z and subscribe to our substack at a16z.substack.com. Thanks again for listening, and I'll see you in the next episode. As a reminder, the content here is for informational purposes only. It should not advice or be used to evaluate any investment or security and is not directed at any investors or potential investors in any A16Z fund. Please note that A16Z and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com forward slash disclosures. .