See pricing
The $1 Trillion AI Buildout | State of Markets
The a16z ShowAI

The $1 Trillion AI Buildout | State of Markets

a16z's growth team walks through 25 charts to argue the AI boom is backed by earnings, not hype, while enterprise impact is still early. Only 2% of S&P 500 companies track AI results over time.

September 30, 2026 · 54 min listen · 13 min read · David George, Sarah Wang, Alex Immerman, Santiago Rodriguez
0:00
–:––

Context

This episode is a walk through 25 charts from a16z's yearly "State of Markets" presentation, put together by its growth investing team. The four speakers are David George, Sarah Wang, Alex Immerman and Santiago Rodriguez. They cover the macro picture (markets, data center spending, compute demand), then what is happening inside companies and in the software and consumer markets. It matters to PMs because the same forces, falling model costs, agents, and slow enterprise adoption, shape what products are worth building and how they get priced.

The Big Idea

AI spending is backed by real earnings and real demand, but actual adoption inside companies is still early. The opening is in turning model capability into reliable services that customers use every day and measure.

The speakers see compute demand outrunning supply, while only a small share of large companies can show measured AI results over time. That gap between what models can do and how they are used is where they see the most room for product builders.

Key Insights

This boom runs on earnings

  • What: since the prior State of Markets, the market is up 90% (about 17% a year), yet stock multiples are down about 20% while stocks are up about 20%.
  • Why it matters: gains come from earnings, not from people paying more per dollar of profit. The S&P 500 trades below 20 times earnings.
  • Contrast: in the dot-com era, top companies traded at around 100 times earnings. Some cyclical memory companies today trade at six to seven times forward earnings.

Compute demand keeps beating forecasts

  • Hyperscaler capex (Alphabet, Amazon, Meta, Microsoft, Oracle) is about $780 billion in 2026, up from $416 billion in 2025. Expectations point to over $1 trillion a year from 2027.
  • Forecasts for this spending have been revised up again and again. Numbers that once looked like a ceiling became near-term plans.
  • Some data center supply chain items can't be bought until 2028.
  • Microsoft, Google and Amazon have about $1.7 trillion of combined cloud backlog, which is committed customer demand.
  • One sign of tight supply: OpenAI paused new subscriptions on its pro plans about two weeks before the recording.

Agents multiply compute use

The speakers say the world has moved past occasional queries. Agents run parallel and long-running tasks, even overnight, while the user does something else. Each task takes many steps, and each step is a model call. They cite a 14x rise in agent token usage on OpenRouter. This is why compute demand keeps rising even as each call gets cheaper.

Adoption is wide but shallow

  • Live AI deployments at S&P 500 companies: 69%.
  • Companies reporting quantifiable impact: 30%.
  • Companies tracking a metric over time: 2%.

Their read is that AI works, but most firms haven't moved from one-off deployments to recurring workflows they measure. Many large enterprises still mostly know AI through Microsoft Copilot.

Power users are pulling away

  • Using spend data from a vendor whose name is unclear in the transcript, median AI vendor spend among the top 1% of companies is roughly 8 times that of the top 10%.
  • Inside a16z's own AI-native portfolio companies, top users spend about $7,500 to $9,000 a month. Median users spend about $200 to $400. That is over 20 times more.
  • AI tool spend as a share of headcount cost: about 1% at forward-leaning Fortune 500 firms, as high as 10% at the most AI-heavy portfolio companies.
  • Buying a tool like Cursor or Cognition is described as the start of a long path to full adoption.

Falling costs open up more work

  • Hebbia's financial chat workloads became 10x cheaper to run, partly through caching, where the system reuses background information instead of reprocessing it.
  • Databricks routes each task to a fitting model. Its smart router solved more problems at 35% lower cost than the strongest single model.
  • Fine-tuned smaller models cut cost too. One company cited (the transcript name is unclear) made a model 60% cheaper with much lower latency, which made live audio use cases workable.
  • Lower cost lets an agent try, check its work, and try again, which opens use cases that were too expensive before. The speakers call this classic Jevons Paradox.

AI is moving profit and growth

  • Cost side: Chime reports cutting cost to serve by over 10% a year for four years, almost 50% in total. Decagon, an a16z portfolio company, supports this.
  • Revenue side: at Shopify, the share of merchants reaching five orders within 15 days of onboarding rose 8% after its AI Sidekick launched. Five orders is a point after which merchants tend to stick around.
  • Incumbents: ServiceNow reported more than $1 billion in AI contract value and a 9x increase in agentic deployments.

Mental Models & Frameworks

Capex is someone else's order book

When big platforms pour cash into AI infrastructure, that spending lands as orders for chips, power and construction. This is why the speakers say the tech boom has become an industrial boom. Use it to ask who is paid by a spending wave, not only who spends.

Where incremental AI dollars go

  • The speakers use this as a litmus test: when a company adds AI budget, does it go to new products for customers or to cost efficiency?
  • Revenue growth is unbounded upside, while cost savings are a limited, latent opportunity.
  • If a company decides the best use of dollars is optimizing costs, what does that say about its revenue opportunity?
  • Even within cost work, lowering cost to serve is better than, for example, spending engineers on rebuilding an internal system of record to save a few hundred thousand dollars.

Price the job, not model

For an application company, the unit that matters is the cost of getting the customer's job done. If you charge for the job and route each step to the right-sized model, you do not need the most expensive model at every step. The speakers expect this to improve margins at the best app companies.

Two paths for software

  • David George's piece "There are only two paths left for software" is cited as the frame.
  • The path he says most founders pick: use existing distribution and locked-in customers to add AI products and drive revenue growth higher.
  • The rule of thumb in that piece is to target about 10% or more growth acceleration. They call it a high bar, but doable with a strong product given the AI budgets available.

Who wins and loses in software

The framework: ask how AI changes the customer's need for the product.

  • Cybersecurity and observability held up well, since agents create new security and monitoring needs. CrowdStrike is called out.
  • Vertical software has held up better than horizontal apps.
  • Applications facing big workflow change are first on what Ali Ghodsi calls the "chopping block of AI".

Trade-offs & Nuance

Incumbent profit pools versus agents

  • Amazon has an advertising business of over $70 billion that depends on shoppers clicking ads. If agents place orders with no clicks, that revenue is at risk.
  • The optimistic case: ad spend finds a new channel, and lower friction means more orders and more total spending.
  • The speakers say both can be true. Profit pools can shrink in some places and still grow the overall economy.
  • The test they suggest for a marketplace: how many new orders could an AI agent bring, and how much of your profit comes from owning discovery? They contrast Amazon, which declined to connect, with Instacart, which agreed, since grocery is less penetrated online.
  • Google search has held up partly because high-value queries, like buying insurance, were ones AI could not yet act on for the user. If that changes, the dynamic could shift.

Subscriptions may understate consumer AI

Only a little over 2% of US households pay for an AI subscription, compared with Amazon Prime at over 200 million households and Netflix at about 70 million. Josh Elman's view is that consumer AI is what you use in daily life, not what you expense, so 97% not paying is not a surprise. Retention is strong: they describe "smile curves", where retention rises as the product improves.

Staying private versus going public

  • Private: companies can take bigger, longer-payback bets with less scrutiny. The speakers point to Databricks and Stripe launching new product areas.
  • Public: it can build trust, for example Navan reaccelerated after going public as large enterprises trusted it more. Capital needs may also grow beyond what private markets can fund.
  • David George says the IPO is "another financing event" and that "the founder is the asset class".

Is it a bubble?

The speakers argue no, citing earnings-led gains and committed demand. They also note free cash flow is depressed during the buildout, with consensus expecting recovery from 2028. They add that long GPU useful lives, longer than they expected, support the economics (a point they tie to an Amazon earnings call).

Common Mistakes

Mistake: Treating all software as doomed

Early in the year there was a "SaaS apocalypse" narrative. The software index later returned to its start-of-year level, but split into perceived AI winners and losers. About 75% of public software companies in their sample are profitable, but only 30% grow 20% or more. The speakers say the winners are those that accelerate growth, not only those that cut costs.

Mistake: Assuming spending has peaked

The speakers say the economy has underestimated this trend for four years. Their stated view is that the buildout will likely look about 20x larger cumulatively in five to ten years.

Practical Application

Track AI impact over time

Following the adoption data (69% deployed, 30% with quantified impact, 2% tracked over time), pick one AI feature or workflow you own and define a metric for it that you log monthly. Moving from one-off deployments to measured recurring workflows is what the speakers say comes next.

Pair cost and growth metrics

Use Chime and Shopify as models. Chime reduced cost to serve more than 10% a year. Shopify tied its AI assistant to a retention-linked milestone (five orders in 15 days). Pair a cost metric and a growth metric when you justify an AI investment.

Match model size to each step

Map the steps in your AI feature and check whether every step needs your most expensive model. Databricks gained performance and a 35% cost cut from routing, and a fine-tuned smaller model made live audio workable for one company. Lower unit cost can raise margin, or fund tasks you could not afford before.

Test your product for agent traffic

If your product is a marketplace or discovery tool, ask two things: how many extra orders could an agent bring, and how much profit depends on owning the customer relationship. Run the same check on any ad-funded flow that relies on clicks.

Build the wrapper around the model

Revolut still partnered with ElevenLabs for voice, then connected it to customer accounts and banking workflows so a customer's call could be resolved securely. List what your product adds around a model, such as company data, security and workflow, since that is the layer the speakers say customers pay for.

Questions to Consider

  • Which of your AI features could you measure monthly for a year, so you are in the 2% of companies that track results over time rather than the 69% that only deploy?
  • If your product's revenue depends on users clicking ads or browsing, what happens to that revenue when an AI agent completes the purchase for the user?
  • If the cost of one AI step fell 10x, which tasks you rejected as too expensive would you build?
  • When your company adds AI budget, does it go to new customer-facing products or to internal cost cutting, and what does that say about your growth outlook?

Bottom Line

The a16z team argues AI spending is supported by earnings, committed demand and falling costs, not just hype. Adoption inside companies is still shallow, so the open opportunity is building reliable, measured services that turn model capability into daily use.

Resources Mentioned

ResourceTypeWhy it was mentioned
State of Markets presentationReportThe yearly a16z growth team deck that the 25 charts come from, released on the day of the episode.
"There are only two paths left for software" by David GeorgeArticleLays out the framework for how software companies respond to AI, including the 10% growth acceleration target.
OpenRouter token usage dataDataSource of the 14x growth in agent token usage.
AI vendor spend dataDataUsed for the gap between top 1% and top 10% AI vendor spending.

Concepts to Explore

Jevons Paradox

When something gets cheaper, total use can rise instead of fall. The speakers apply it to model costs: a 100x drop in cost per task makes many more tasks worth doing.

Smile retention curves

A retention curve that flattens and then rises as the product improves and users return. The speakers say AI subscriptions show this, which is rare in consumer products.

Model routing

Sending each task to the model that fits it best on cost, speed and quality. Databricks cites a router that beat the strongest single model at 35% lower cost.

Case Studies Mentioned

Revolut voice support with ElevenLabs

Revolut has a sophisticated engineering team, but still partnered with ElevenLabs and used its voice model. It connected the model to customer accounts and banking workflows so that a customer calling in gets a secure, smooth resolution. The lesson in the speakers' words: the opportunity is in harnessing model capabilities to build reliable services, because an AI model knows a lot about the world and little about your company.

Chime cuts cost to serve

Chime reports cutting cost to serve by over 10% a year for four years, close to 50% cumulatively. Lower cost to serve gives room for better pricing, broader service, and reinvestment in growth.

Shopify Sidekick and early merchants

Shopify launched an AI assistant called Sidekick to help merchants get started faster. The share of merchants hitting five orders within 15 days of onboarding rose 8%. Shopify treats five orders as a sign a merchant will stay.

Instacart versus Amazon on AI agents

The speakers describe Amazon saying no to connecting with one AI agent, while Instacart said yes. Their reasoning is that Amazon would gain few new orders and risks its ad business, while Instacart has more online grocery orders to win.

Tools & Products

Tool / ProductWhat it doesWhy it was mentioned
Cursor and CognitionAI coding toolsExamples of tools enterprises buy first, which is only the start of adoption.
Microsoft CopilotAI assistant for workSaid to be the main AI exposure at most enterprises the team talks with.
OpenRouterModel routing platform with public usage dataSource of the 14x growth in agent token usage.
ElevenLabsVoice AI modelsUsed by Revolut for customer voice support.

People to Follow

David George

Leads a16z's growth investing and wrote "There are only two paths left for software". Argues that the founder is the asset class.

Sarah Wang

a16z growth investor who discussed enterprise adoption, public software trends and private-market data. She held a conversation with Databricks CEO Ali Ghodsi on how AI knows the world but not your company.

Alex Immerman

a16z growth investor who covered consumer AI, retention curves and the AI-driven shift in marketplaces.

Santiago Rodriguez

a16z growth investor who covered private company scale, tender offers and secondaries.

Notable Quotes

"The founder is the asset class at this point." (David George)

"The IPO is another financing event." (David George)

AI PM course

Everyone hears the same episodes.
Few can do what they describe.

Start for free