AI, Infrastructure, and the Next Investment Cycle

Opening Frame: Technology as the Everything Cycle

The discussion opens with a striking observation from the a16z growth team: high-tech equipment, software, and R&D now account for roughly 55% of US capital spending, and tech makes up nearly 40% of the aggregate US stock market value, with eight of the top ten most valuable companies in the world being US tech firms. The speakers frame this as a broad economy-wide investment cycle that touches everything from software and models down to power construction and industrial capacity. They note the buildout has already surpassed railroads as a percentage of GDP, and they expect that numbers five to ten years out will look roughly 20x higher cumulatively. The core thesis is that “software is eating the world” has become a demand engine for compute, chips, power, and construction, because everyday users now run a handful of automations every night.

Is It a Bubble?

The team tackles the most common investor question head-on. Since the market has hit new highs, trading multiples have actually compressed: stocks are up roughly 20% while multiples are down roughly 20%, and the S&P 500 forward earnings multiple sits below 20x. The argument is that returns are being driven by earnings growth rather than multiple expansion, which is fundamentally different from the dot-com era, when top companies traded at around 100x P/E. Even cyclical memory companies trade at only six to seven times forward earnings. David George notes that the market has returned about 90% (roughly 17% annualized) since the release of ChatGPT nearly four years ago, which naturally triggers the instinct that it must come down. Yet with the market trading below 20x while growing earnings around 15%, the speakers argue the current period feels meaningfully different from 2000 or 2021, when multiples and growth were not going together.

The Hyperscaler CapEx Race

The hyperscalers (Alphabet, Amazon, Meta, Microsoft, and Oracle) are projected to spend about $780 billion on capex in 2026, up from $416 billion in 2025, with expectations of over $1 trillion annually from 2027. Sarah Wang points out that each prior computing platform supported a far larger population of users, and this cycle can move quickly because it sits on existing distribution from the internet, cloud, and mobile, allowing it to reach billions of users almost immediately. A key shift the team emphasizes is that usage has moved beyond occasional queries toward parallel and long-running agents that work through the night, autonomously writing code and searching. This creates a much larger compute requirement than anyone anticipated.

Forecasts for hyperscaler capex have been revised sharply upward in short succession, meaning spending that once looked like a ceiling several years out has become near-term. The team argues that the natural instinct to expect a flatline has repeatedly underestimated the strength of the trend. Alex Immerman and others describe demand for compute as insatiable. As an anecdote, they cite OpenAI’s compute commitments, which drew criticism a year or so ago but now look prescient, and the fact that OpenAI had to pause new subscriptions to its $200-per-month Pro plan. Everyone they speak with across the supply chain says some version of the same thing: demand outstrips supply, with certain data center components unavailable until 2028.

On the demand side, Microsoft, Google, and Amazon together hold about $1.7 trillion in combined cloud backlog, which the speakers present as high-quality evidence that customer commitments are building rapidly. Free cash flow is depressed during the buildout, but consensus forecasts show a recovery starting in 2028 with substantial growth after. They invoke Amazon’s J-curve explanation: data center shells take time to build, chips must be purchased, but GPUs and TPUs have a useful economic life that has proven longer than most expected. The team also argues that hyperscalers who blinked and were less aggressive in raising debt for capex have regretted it. The spot market pricing for existing GPUs is cited as another signal that any capacity brought online is commanding attractive rates, with vendors earning strong returns.

The speakers frame the economics through consumer surplus. Users get enormous value from AI, which explains the heavy usage. The vendors serving those users make good money, and even as chip prices have risen well above where they were 12 to 24 months ago, the chain still supports high margins and substantial surplus for end users. Viewed this way, hyperscaler capex is effectively someone else’s order book, and it has transformed a technology boom into an industrial boom spanning chips, power, and construction. Santi Rodriguez observes that this changes the venture job itself: traditional software investing was about monetizing IP built once and sold infinitely, whereas these capital-intensive businesses must manage financing, vendor relationships, capacity planning, and demand forecasting, which is a different expertise.

The New Age of Atoms

Global infrastructure investment needs are estimated at roughly $90 trillion through 2040, and the speakers stress that this extends well beyond AI and data centers to power, water, roads, and transit. They point to portfolio examples: Anduril’s massive manufacturing facility (described as 887 football fields in scale), Waymo aggressively expanding its depots, and SpaceX’s $100 billion investment in Louisiana. The thesis is that betting on the future now means building factories and infrastructure and creating skilled jobs across the country. David George recalls Elon Musk’s framing that “the factory is the product,” noting that many great founders in the a16z portfolio came out of Tesla or SpaceX because they learned that execution on the factory is a massive competitive advantage at scale.

The team addresses common myths about data centers. The claim that they drain America’s water and that wealthy people avoid living near them is set aside, along with the fear of soaring electricity bills. A recent US study found that for every 10% increase in data center capacity, residential electricity rates fell by about 40 basis points. The logic is that the grid is a shared fixed-cost system of poles, wires, and substations, so a large stable customer like a data center spreads those costs across more units of electricity. Sarah and Alex agree that investment in shared infrastructure can improve economics for connected households and businesses. They cite Meta’s Louisiana site, where the company worked with the community to lower electricity costs, as a concrete example of this not being mere lip service.

Model and Application Layer Trends

Adoption is described as “still extremely early” even as AI generates major revenue and savings. Costs are plummeting, which makes agents economical for a much wider range of work. The team invokes Jevons paradox, noting that two orders of magnitude differences in cost unlock new use cases. The combined annualized revenue of OpenAI and Anthropic is presented as the team’s favorite chart, one they update monthly because it goes stale so fast. The speakers compare that growth against the best software companies in history and find that the leading labs’ net new revenue has well surpassed theirs, showing both extraordinary scale and remarkable speed.

Sarah stresses that the AI shift is early in two dimensions: the percentage of the population using AI, and the share of wallet each user gives to AI products. This differs from the iPhone era, where the question was simply how many people owned one. The team cites several measures of the gap between capability and deployment. Live deployments at S&P 500 companies are at 69%. Quantifiable impact is at 30%. The ultimate barometer, a metric tracked over time, is only 2%. David George argues the next stage is moving from individual deployments to deeply influential, recurring workflows inside organizations. Anecdotally, most enterprises still touch AI mainly through Microsoft Copilot.

David frames the application layer as the bridge across that gap. Models know a lot about the world but very little about a specific company. He cites Revolut, which has a sophisticated engineering team yet still partnered with ElevenLabs, taking its best-in-class voice model and harnessing it to connect securely with customer accounts and banking workflows. The principle is that today’s opportunity lies in taking these capabilities and building reliable services.

Power Users and Enterprise Spend

The speakers highlight that power users are pulling away from the pack. Using Ramp-style spending data, the median AI vendor spend in the top 1% of companies is roughly eight times the level of the top 10%, and the top 1% alone spends almost as much as the 2% to 10% group combined. Anecdotally, within a16z’s AI-native portfolio companies, the top users spend between $7,500 and $9,000 per month, while median users spend closer to $200 to $400 (roughly one or two subscriptions), making the gap more than 20x. The team measures adoption by the share of spend on AI tools relative to headcount: relatively forward-leaning Fortune 500 companies are around 1%, while the most aggressive AI-build companies in the portfolio reach as high as 10%. Those numbers frame the core question of how early diffusion is and how deep it will go.

Sarah describes a common pattern. Enterprises get excited when they first buy Cursor or Cognition, but procurement and trial are just the beginning, and the runway to full adoption is massive.

Quantifiable Results and Where the Savings Go

The team presents public company case studies. Chime has reportedly reduced its cost to serve by over 10% a year for four years, compounding to nearly a 50% reduction, supported by a16z portfolio company Decagon. On the revenue side, Shopify’s AI Sidekick helped merchants get up to speed faster, raising the share of customers reaching five orders within 15 days of onboarding by 8%. Five orders is a metric Shopify tracks as a predictor of retention. ServiceNow reported over $1 billion in AI annual contract value and a ninefold increase in agentic deployments, showing that incumbents are monetizing AI as well as newer companies.

Sarah points out that lower cost to serve gives companies room to compete through better pricing, broader service, and reinvestment in growth, so the benefits eventually reach both customers and margins. Santi and David then raise a litmus test: where are companies putting incremental AI dollars? Building new products that drive revenue has unbounded upside, while cost optimization is a latent opportunity that remains available later. If a CEO prioritizes cost structure above revenue, that choice reveals what they see as the best opportunity. The team also observes that some engineering teams spent six months rebuilding internal systems of record to save a few hundred thousand dollars, and they expect better uses of those resources.

Agents, Inference Costs, and Economics

Agents are real and performing multi-step tasks that require many model calls, which helps explain the 14x growth in agent token usage on OpenRouter. The cost of those steps is falling through caching, which lets systems reuse background information rather than reprocessing it. Hebbia’s financial chat workloads became 10x cheaper to run, meaning the same budget now supports more work. Lower costs make it practical for agents to check their work and retry, opening up tasks where reasoning or tool use would have been prohibitively expensive, particularly where reliability is the deciding factor. Latency improvements accompany these cost gains.

Routing is another lever. Databricks uses a smart router that picks the appropriate model for each task, solving more problems at 35% lower cost than the strongest individual model. Fine-tuning is equally important: Harvey and Eleven Labs’ Eleven/Elise AI case showed a smaller fine-tuned model that was 60% cheaper and had much lower latency, making live audio use cases viable. David summarizes the application-layer economics: the relevant unit is the cost of getting the customer’s job done. If a company can decouple that from the underlying model, it can charge for outcomes and use routing to make the product economical, avoiding the most expensive model at each step. He expects margin improvements to follow for the best app companies.

Consumer AI: Early, Sticky, and Redefining Engagement

Consumer AI is still early. A recent survey found that just over 2% of US households pay for an AI subscription. David notes that subscriptions may not be the best monetization route for most households, but they have shown retention that is rare in consumer: curves that not only flatline but smile as the product improves. Santi adds perspective by comparing to Amazon Prime’s roughly 200 million households and Netflix’s 70 million, noting that AI subscription penetration is small by comparison. Josh Elman, a partner, offers a counterpoint definition: consumer AI is what people use in daily life, not necessarily what they expense, so the fact that 97% of AI-using households don’t pay is less surprising.

The engagement picture changes too. Facebook, Instagram, TikTok, and Snap average 30 to 60 minutes per day per active user, but the best AI assistants aspire to be always-on, proactive, and persistent, which may mean they won’t be where people spend the most time. Tracking engagement from outside becomes harder when agents work in the background, as Alex notes, because screen time no longer captures the value. The team recalls launching a deep research or agent task and then driving into the city, a reminder that the work happens without a screen.

Marketplaces, Advertising, and the Amazon-Instacart Moment

The team discusses the recent news that Amazon declined to participate with Muse (a consumer agent app), while Instacart said yes. For incumbent discovery platforms and marketplaces, the questions are how much incremental demand an AI agent can bring and how much of their profit pool depends on owning the customer relationship and discovery. Amazon’s advertising business, over $70 billion and very high margin, depends on users clicking ads on the website. If agents bypass that, the revenue is at risk. Instacart’s online grocery penetration remains relatively early, so the optimistic case is that agents remove friction, generating more orders and GMV that makes up for lost ad revenue. David notes that clicks used to be necessary to process an order, and that friction removal could let trips and dinners get booked that people wanted but never completed.

David also points out that Google’s search business has proven resilient, partly because high-value ad terms (insurance, hotels) are hard to automate today. That could change if AI can take action on users’ behalf. Santi offers a positive-sum view: advertising dollars may move to other channels, agents might create new form factors, targeting could improve, and lower friction could raise consumption, creating a flywheel for economic growth. The team agrees that both views can be true: profit pools can shift while the overall economy grows. Last week’s immediate market reaction was net positive across the ecosystem, with Meta’s gains far exceeding losses in the marketplaces.

The SaaS Bifurcation

Public software has shifted toward slower-growing, more profitable companies. Roughly 75% of the public software sample is profitable, but only about 30% grow at 20% or more, and at a 30% threshold fewer than five companies remain. By contrast, nearly every private company the team sees grows well above 30%. The speakers explain that the easiest place for IT budgets to come from is not new SaaS projects but reallocation from existing software spend toward AI. Investors and founders are focused on taking strong existing distribution and locked-in customers and applying AI products to accelerate revenue. The bar for dispelling the “SaaS apocalypse” fear is continued gross dollar retention around 98%, efficiency gains, and above all revenue growth acceleration showing offensive investments are working.

David references his blog post “There are only two paths left for software,” noting that founders almost universally want to drive revenue growth higher. The team set a target of 10% plus acceleration, which is ambitious but seemingly doable given the budgets available for AI, and they expect a few portfolio companies to hit it within 12 to 18 months. The speakers stress that software is highly differentiated. Cybersecurity and observability have stood out, and vertical software has held up better than horizontal applications. The framework is to ask how AI changes the customer’s need for the product. Security demand is rising as agents access more systems, with the OpenAI-Hugging Face incident cited as an eye-opener, and CrowdStrike is called out as a beneficiary. Applications face varying degrees of workflow change, and Alex Immerman’s “chopping block of AI” idea captures which products are first at risk. Vertical AI companies such as Harvey, Abridge, and Ambience are among the fastest-growing private companies, because their value comes from orchestration and vertical workflows, not just the model.

The public market view is that the software index has returned to where it started the year, but it has bifurcated into perceived AI winners and losers. Simply improving cost structure is not enough. Winners are accelerating and capturing net new dollars, and those who fail to capture them are not riding the wave.

Stripe’s Renaissance and the Private Market

Stripe’s SaaS customer data offers a narrative violation: growth accelerated into 2026 across both young and mature businesses. Stripe calls this the “Renaissance,” which the speakers find fitting. Taken together, the top six private companies (Anthropic, OpenAI, Databricks, Stripe, Wealthfront, and Revolut, by last-round valuation) total about $2.4 trillion. That exceeds the combined market capitalization of all IPOs in the last decade excluding SpaceX, which totals about $1.7 trillion, and it is nearly the size of the Russell 2000. The speakers argue that companies are staying private longer so they can invest in growth and take larger swings with long-duration payback, rather than maximizing short-term profit.

On the IPO question, David says the founder is the asset class, and that for founder-led companies the IPO is just another financing event. Public markets bring greater scrutiny, as Meta’s stock falling below $100 during its heavy AR/VR investment period illustrates. Some companies benefit from public disclosure, though. Navan’s IPO produced a reacceleration as large enterprises came to trust it more. At some point, capital needs may exceed private-market pools.

Secondaries are another dynamic. Companies increasingly hold tender offers for employees, allowing liquidity without going public. Carta data shows only 58% participation, meaning employees are choosing not to cash out because they believe in their company. Alex and David frame tenders as a competitive tool against the liquidity appeal of public-company RSUs, and as a way to reset valuations frequently, which keeps private stock prices fresh and gives companies currency for M&A. The secondary discount to last-round price, which was meaningful in 2022 through 2024, is now essentially zero, signaling that recent valuations are fresh and that new investors are willing to pay the same price.

Venture Concentration and What’s Next

AI-related companies account for 86% of US VC deal activity in the 2026 snapshot, up from 65% in 2025. Yet the opportunity has broadened beyond model labs into enterprise apps, consumer apps, services, semiconductors, power, and defense. Sarah summarizes the thesis that the opportunity set has never been deeper or broader.

The team closes with areas they are excited about. Consumer agents that handle long-running, autonomous tasks could change what people spend time and money on. Robotics may ultimately be larger than LLMs, perhaps three to five years behind, with significant investment expected over the next five years. Autonomy is working: Uber and Lyft represent about 1% of US miles traveled today, and autonomous vehicles are described as 14 times safer than human drivers, with an expected order-of-magnitude expansion in coming years and roughly 17 million new cars sold annually in the US. AI and biology promise new drug discovery against debilitating diseases over the next decade. Personal health is an area where hyperpersonalized advice could finally synthesize one’s own data. Enterprise diffusion beyond coding remains very early. Finally, American dynamism is a retooling of defense and industrial capacity, where newer vendors such as Anduril, Shield AI, and Castelion currently capture less than 5% of military spending, which the team expects to grow dramatically. The overall message is that the buildout is massive and will be productivity-enhancing for the US economy.

• Tags:
• AI Agents
• Venture Capital
• Robotics
• AI Infrastructure
• Consumer AI
• Autonomous Vehicles
• Data Centers
• Defense Tech
• Private Markets
• Inference Costs
• Hyperscaler CapEx
• Enterprise AI Adoption
• SaaS Bifurcation
• AI Biology
• State of Markets