a16z State of Markets Takeaways: AI Dollars Flow into Hardware as Software Enters Its "Prove-It" Era

How Tech Evolved from a Single Sector into a Major Driver of Economic Growth

For a long time, in the eyes of traditional investors and macroeconomists, the tech sector looked more like a high-growth, asset-light "niche vertical." Following the financial crisis, alongside the rise of mobile internet and cloud computing, the software industry enjoyed more than a decade of rich valuation premiums thanks to near-zero marginal costs and virtually limitless room for market expansion. In earlier economic cycles, demand for purchasing and building durable goods such as housing and automobiles often served as the primary growth engines.

In a16z's view, technology's driving role in the economy is strengthening significantly. Statistics indicate that tech investments now account for approximately 55% of U.S. capital expenditures (spanning high-tech equipment, software, and R&D spending). The report argues that AI infrastructure buildouts can be seen as a major force driving investment in today's economy; simultaneously, nearly every industry is adopting digital tools, making it difficult to find traditional businesses running purely on pen and paper, with enterprises transitioning toward technology to varying degrees.

According to public market financial reporting, tech's existing profit base already approaches half of total S&P 500 profits. Using a base period from over a decade ago, the tech sector's cumulative realized profit growth is roughly 3.8 times that of non-tech sectors. An even more tangible change comes from the incremental growth: among the S&P 500's expected profit growth for the entire current year, tech companies contribute about 76%. The 76% figure refers to the contribution to new profit growth, while tech's share of total profits approaches half: the former answers where growth is coming from, while the latter answers who is already earning the most.

Regarding whether current market conditions represent a bubble, the report points out that broad market indices reaching fresh all-time highs are primarily supported by corporate earnings growth. Valuation multiples for the tech sector (such as price-to-earnings ratios) have actually compressed by around 20% compared to their five-year averages, indicating that stock price gains have been accompanied by actual profit expansion. However, while profit growth supports current market levels, this does not entirely rule out the possibility of a bubble, nor does it eliminate the risk of future earnings misses or valuation pullbacks; if subsequent earnings growth slows or capital investments fail to deliver the expected returns, the market will still face the test of repricing.

Profits Earned by Big Tech Are Flowing into Chips, Power, and Data Centers

Over the past decade, "software is eating the world" was Silicon Valley's loudest creed; asset-light operations, high gross margins, and low maintenance costs made software companies the darlings of the market. Yet the most defining feature of the current cycle is precisely a massive migration of capital attention "from bits to atoms."

Running AI models requires chip compute, network transmission, heat dissipation, and power supply, which translates software demand into massive orders for physical equipment. Tech giants continue to expand infrastructure, driven by unprecedented downstream demand signals for compute: frontier foundation model companies are seeing rapid growth in annual recurring revenue (ARR), while remaining performance obligations and cloud order backlogs (Backlog / RPO) at leading cloud vendors have also doubled year-over-year.

ARR is merely an annualized revenue run-rate projected at a specific point in time, not equivalent to actual recognized revenue for that year; order backlogs represent future performance contracts rather than cash already collected, and these cloud backlogs also include traditional enterprise needs rather than exclusively AI demand.

Based on these massive demand signals, the hyperscalers tracked in the report are continuously channeling their abundant operating cash flows into hard capital for physical infrastructure.

The scale of this capital expenditure has repeatedly exceeded Wall Street estimates. Data shows that cloud vendors' total capex reached $416 billion in the prior year, is estimated to surge to $791 billion this year, and is heading toward the trillion-dollar mark over the next two years. These figures track overall cloud vendor capex and do not consist entirely of dedicated AI spending. In terms of capital flows, accounting book profits and remaining post-investment cash are two different metrics: operating cash flow generated from daily business, minus capital expenditures such as facilities and equipment, leaves free cash flow, which can be freely used for dividends, share buybacks, or risk mitigation. While the giants' operating cash flows remain robust, their share of free cash flow has dropped sharply in the short term due to the rapid expansion of capex. Institutions project that this cash flow pressure will persist until around 2028, with its subsequent recovery contingent upon actual realization of cloud revenues.

To support such large-scale infrastructure construction, tech giants that previously carried minimal debt have begun raising more capital through debt markets (including on- and off-balance-sheet financing). For lenders, the overall return on invested capital (ROIC) of tech giants remains higher than borrowing costs, and the investments back physical assets capable of generating steady cash flows. Overall ROIC measures the return across all invested capital of a firm, including mature businesses; whether new AI projects are cost-effective depends on how much cash the individual projects themselves can generate.

More importantly, this wave of physical investment extends well beyond semiconductors. High-bandwidth memory (HBM), optical transceivers, liquid cooling systems, high-voltage transformers, and commodities like copper all face supply chain tightness. In practice, while GPUs are delivered relatively on schedule, lead times for transformers and gas turbines can stretch for several years, leaving the industry facing the real-world bottleneck of chips arriving without timely power hookups. Furthermore, infrastructure construction has driven wage premiums for specialized electrical and mechanical installation roles. On the power side, a regional study covering historical data from 2019 to 2024 noted that data centers, as steady large-load consumers, helped absorb the grid's high fixed costs; in states with rapid data center capacity growth, residential retail electricity prices on average experienced some degree of decline. However, this is an empirical observation from a specific historical window and regional sample, and one cannot infer that prices will broadly drop across all regions or in the future.

The drivers behind this wave of physical investment do not stem from AI alone: grid upgrades driven by the adoption of electric vehicles and solar power, the reshoring and rebuilding of manufacturing driven by geopolitical considerations, automation demand spurred by an aging population, and steadily rising defense spending are all driving physical investment. Unlike AI infrastructure's focus on compute and power, these sectors serve distinct real-world needs across transportation, manufacturing, labor substitution, and defense. What they share is a departure from software's asset-light logic, relying heavily on grid upgrades, plant construction, heavy equipment procurement, and on-site engineering and installation. AI infrastructure is merely the most prominent representative; multiple physical investment forces are jointly driving capital toward the "world of atoms."

Why Older GPUs Can Still Be Rented Out When New Chips Arrive

Amid continuous expansion in compute investments, a common bearish concern in the market is accelerated chip depreciation: if high-performance chips are replaced by new architectures every 3 to 4 years, will previous-generation hardware rapidly lose market demand, triggering massive write-downs on cloud vendors' balance sheets?

The operational logic of the real-world market has broken this linear assumption. Empirical data from the compute rental market shows that unit rental prices for previous-generation chips, rather than falling off a cliff as new products launch, have remained stable or even rebounded under the surge of demand.

According to tracking data, performance across GPU generations diverges in the rental market but retains value overall: the B200 commands an hourly rental rate of around $5.69, the H200 around $3.29, the H100 around $2.63, and the earlier A100 around $1.59. Rental rates for the A100 have not seen sustained declines throughout the year, reflecting that earlier-generation hardware continues to offer rental value and viable use cases amid current compute demand.

Being rentable, however, does not mean an investment has broken even. A GPU's true return on investment depends on the initial acquisition price of the chip, actual utilization rates, data center power consumption, routine operations and maintenance costs, and remaining useful life. While stable rental rates ease concerns over accelerated hardware depreciation, actual commercial returns still require accounting for all of these cost factors.

The report invokes economics' Jevons Paradox to explain this mechanism: when technological efficiency improves and unit costs decline, the lower barrier to entry can stimulate more application demand, thereby expanding overall resource consumption. This is a demand transmission mechanism, not an iron law guaranteeing explosive growth.

At the model inference level, demand for Tokens (the basic metric for text processed by models) has grown with particular prominence. On the model routing platform OpenRouter, 7-day average Token usage driven by Agent calls demonstrated rapid growth of more than tenfold, reaching approximately 7.3 trillion Tokens at the end of the period. Compared to single query-and-response interactions from human users, Agents consume far more Tokens during multi-step execution and long-context calls.

Notably, an increase in gross Token volume does not directly equate to an identical increase in actual computational load borne by server hardware. In OpenRouter's data, cached prompts (Prompt Caching) accounted for 86% of Tokens consumed by Agents, compared to only 29% for human users. Caching reuses intermediate computational results obtained from earlier processing of the identical prompt prefix. Imagine an Agent handling tasks continuously: every step carries the same task instructions and reference documents, along with the latest operational outputs. Without caching, the model must repeatedly process that identical content; with prompt caching, existing results can be reused, reducing redundant compute. Because newly appended content and model-generated responses still require active computation, Token usage and actual compute load do not have a one-to-one correspondence.

Within the model ecosystem, open-weights models are capturing an increasing share of total Token consumption, and overall Token spending is expanding, while frontier closed-source models maintain pricing power on high-difficulty tasks. Older models have likewise avoided rapid obsolescence: different tasks require different capabilities, and as long as a less expensive model can accomplish the task, users still have a compelling reason to use it.

Model Routing delegates this selection to software systems: choosing models based on task specifications, making trade-offs between cost and completion quality. In Databricks' internal evaluation of coding tasks, intelligent routing achieved a task completion rate slightly higher than the benchmark models in the chart while lowering per-task costs. It demonstrates an optimization method whose specific impact depends on the task and model combination.

This links costs back to demand: when individual tasks become cheaper, workloads previously abandoned due to prohibitive expenses become economically viable. If enterprises broaden the scope of automation as a result, more tasks can drive up total Token usage and compute demand. Whether efficiency gains boost aggregate demand ultimately depends on whether enough new applications emerge.

AI Is Being Used, but Returns Have Yet to Be Broadly Accounted For

Despite enormous investments in underlying compute, real-world adoption indicates that AI implementation still exhibits a "broad reach but relatively shallow depth" dynamic. While a substantial share of enterprises has deployed AI features into active use, fully converting those deployments into publicly verifiable financial returns remains in its infancy.

Tracking the AI implementation progress disclosed by S&P 500 companies reveals three tiers that answer distinct questions:

• Deployment Tier (answering "is it in active use?"): 74% of companies claim to have an AI strategic roadmap, and 69% have deployed specific AI features and disclosed their deployment status;
• Benefits Tier (answering "what business results has it delivered?"): Companies able to present quantifiable commercial results in earnings reports or public filings fall to 29%;
• Continuous Tracking Tier (answering "can results be consistently tracked and measured?"): Companies capable of consistently reporting ongoing tracking metrics stand at just 2%.

These percentages track public disclosures. Companies that have not published tracking metrics may still be evaluating them internally; verifiable evidence in public disclosures remains scarce for the market.

Looking at internal depth of usage, the tool adoption gap across enterprises is equally striking. OpenAI's research shows substantial divergence in tool adoption across corporate cohorts. Plugins are add-ons that connect to external tools, while Skills are reusable workflows and procedures; the denominator in the table below represents weekly active enterprise users:

Note: With OpenAI itself serving as an independent benchmark group, its Plugins and Skills adoption rates reached 95% and 93%, respectively. Adoption rates here reflect the breadth and depth of advanced tool usage within enterprises and should not be equated directly with business returns or project success rates.

Actual usage intensity across enterprises is also highly uneven. In OpenAI's platform sample, indexed against the report's chosen baseline, output Token volume for the top 10% of enterprises across industries grew to 17.1x, compared to 2.1x for the median 10%. This compares output Token growth rather than providing a comprehensive census of market-wide compute consumption.

Adoption rates answer "how many have started using it," while usage intensity answers "how much each company uses." While large numbers of companies remain at a shallow stage of usage, a handful of heavy enterprise users already generate massive volumes of calls; thus, shallow aggregate adoption and rapid platform usage expansion can comfortably coexist.

Several companies have published specific business results. Below are self-reported enterprise case studies compiled in the report:

• Amazon and Meta (E-commerce shopping and advertising performance): Amazon self-reported in earnings disclosures that customers shopping with Alexa assistance spent over 40% more per order on average than non-users (a cross-sectional cohort comparison rather than a randomized controlled trial), while its Ads Agent drove a 6% decline in cost-per-acquisition (CPA); Meta Chief Financial Officer Susan Li disclosed that Advantage+ drove a 13% incremental increase in purchases and a 16% lift in add-to-cart conversions.
• ServiceNow (Enterprise workflows and contract value): Chairman and CEO Bill McDermott disclosed that ServiceNow AI annual contract value (ACV) surpassed $1 billion in the second quarter, with Agentic deployments expanding 9x in 9 months.
• Affirm (Credit underwriting and transaction facilitation): Affirm disclosed in company announcements that its model approved applications previously rejected by legacy systems (including borrowers with limited credit history and no FICO score), facilitating an additional 3.4% in completed transaction volume, with these incremental loans outperforming comparable expansion cohorts.
• Lemonade and Chime (Claims expenses and servicing costs): Lemonade stated in its shareholder letter that automation across claims operations brought its loss adjustment expense ratio (LAE ratio, the proportion of expenses spent investigating and settling claims, distinct from the loss ratio) down to a low of 5%; Chime CEO Chris Britt stated that continuous AI efficiency gains drove cost-to-serve per customer down by 10% annually for four consecutive years.

On the consumer side, clear divergence exists between AI usage rates and paying penetration. Data indicates that as of April, the share of U.S. households paying for personal AI subscriptions stood at roughly 2.2%; in a separate survey as of May, average monthly spend among paying households was approximately $31. These two surveys involve distinct samples and timeframes and cannot be conflated. This data tracks paid subscriptions, excluding households using free AI tiers. Paying penetration remains low, yet a cohort of users is already willing to pay continuously; niche heavy-user demand and the absence of broad consumer monetization can exist simultaneously.

Software Is More Profitable, So Why Are Valuations Falling?

Alongside breakthroughs by large models in code generation, UI construction, and workflow automation, capital markets were briefly swept up in a panicked narrative: the "death of software" (SaaSpocalypse). Markets feared that if AI could instantly generate custom tools from natural language, traditional SaaS (Software-as-a-Service delivered via subscription models) with seat-based pricing would be completely dismantled.

a16z uses "SaaS-Prove-It" to describe the new mandate facing software: companies must prove not only that they can continue growing, but also that AI adds product value customers remain willing to pay for.

The valuation multiple contraction experienced by the software sector was driven fundamentally not by the threat of AI substitution alone, but by an internal repricing tied to shifts in macroeconomic interest rates. During the zero-interest-rate policy (ZIRP) era, cheap capital subsidized an expansion model that traded losses for hypergrowth; as interest rates normalized and capital costs rose, the software industry broadly undertook a self-correction, trading growth rates for profitability.

Data from public software companies reveals this shift: among public software firms surveyed, roughly 75% reported positive operating income during the period, but overall growth slowed, with only about 30% maintaining annual revenue growth rates above 20%. While revenue growth has decelerated, it is showing signs of bottoming out and stabilizing, with operating leverage at most companies remaining steady or improving. Operating leverage here means: when revenue expands, certain operating expenses do not need to grow proportionally, allowing more incremental revenue to flow through to operating profit. In terms of valuation multiples (the price level the market is willing to pay for each dollar of revenue or profit), software companies capable of sustaining high growth continue to trade close to historical averages, below their ZIRP-era peaks; the multiple compression seen across the broader sector is primarily because far fewer companies can sustain high growth than before.

More importantly, resilience within software displays stark structural divergence:

1. Cybersecurity and Observability (tools for monitoring system health, troubleshooting, and logs) lead significantly: As of the report's chart cutoff, the market-cap-weighted basket in this segment delivered a year-to-date share price return of approximately +80%, notably outperforming other software cohorts. Another valuation chart in the report also highlights multiple premiums for security companies. Price returns measure how much a stock has appreciated, whereas valuation multiples measure what the market will pay per dollar of revenue or profit; the two must be evaluated separately.
2. Vertical SaaS experienced less valuation compression than horizontal SaaS: Compared to generic collaboration or productivity tools, vertical software is typically more deeply embedded in industry-specific workflows, leveraging domain-specific proprietary data and trusted delivery requirements, thus facing relatively less multiple compression pressure. Whether customers continue to pay still hinges on whether it reliably solves real-world business problems.

Regarding the software industry's outlook, the report lays out two divergent scenarios.

One is a replay of traditional print newspapers: shares of publicly listed newspaper publishers began falling years before actual operating profits collapsed. If AI can directly generate bespoke tools, undermining the foundational premise of seat-based licensing, software companies could face valuation resets or disruption even if current revenues and operating leverage appear resilient.

The alternative is that AI boosts returns for asset-light enterprises. If software companies leverage AI to lower delivery and operational costs or add capabilities customers are eager to pay for, they can simultaneously expand margins and enhance product value. The report illustrates this possibility using return on equity (ROE, the ratio of net income to shareholders' equity) trends across asset-light versus asset-heavy businesses: if firms can generate higher net profits via AI without requiring a proportional expansion in equity, asset-light ROE could expand further. However, the ROE improvements shown in the chart incorporate consensus analyst forecasts, representing a modeled projection rather than realized or measured causal proof.

Startup Funding Centers on AI; Returns Concentrate in a Few Massive Outcomes

In venture capital and startups, capital concentration toward a select few avenues is pronounced. Some of the most valuable tech companies are staying private longer ("Private-for-Longer"), growing into massive enterprises such that more valuation appreciation occurs prior to an IPO. Consequently, the window for the general public to participate in this growth by purchasing public equities arrives much later. In private secondary markets (where existing shareholders trade shares rather than companies raising primary capital through new issuances), premier private companies trade at minimal discounts or even command valuation premiums; this does not mean all private companies escape discounts, as transactions in the chart continue to exhibit price divergence and discounts.

Valuation markups primarily increase book value; hard cash can only be realized upon selling shares or achieving exits, and the actual cash distributed to fund LPs ultimately depends on closed transactions and distributions.

Capital is noticeably clustering around top-tier players and AI. Data indicates that among total U.S. venture deal value this year, AI-related companies accounted for 86%, while non-AI companies accounted for just 14%. On the exit side, the power-law distribution is exceptionally stark: U.S. M&A and IPO exit data for the first half of the year reveals that the top 1% of transactions captured 84% of total exit value, and the top 10% captured 94%.

At the operating level, much like mature enterprises, most ordinary startups have generally traded growth for profitability following the rise in capital costs; cohorts that successfully raised funding recently have maintained faster growth. Leading AI applications tracked in the report grew especially fast: within the top tier with base ARR between $5 million and $50 million, median year-over-year growth reached 423% (expanding to roughly 5.2x original scale); in the larger cohort with base ARR between $50 million and $200 million, median year-over-year growth reached 150% (expanding to roughly 2.5x original scale). These figures reflect a cohort of top-tier AI applications and do not represent the probability of success for average startups or the universal performance of the AI sector as a whole.

Alongside extreme capital concentration, investment themes are also branching into vertical niches that previously saw little venture backing:

• Professional Services: Traditional high-cost manual services like legal, tax, and accounting have become key frontiers for VC allocation into AI applications.
• Defense and Manufacturing Tech: Under fixed defense budget constraints, procurement and investment priorities are pivoting toward small, low-cost, mass-producible systems, allowing a given budget to acquire and deploy greater volumes of hardware, though these lower-cost assets cannot simply be equated to the standalone performance of expensive, heavy legacy systems.
• Commercial Case Studies in Physical Settings: In public safety, automated license plate readers (ALPR, such as Flock mentioned in the report) help understaffed police departments uncover investigative leads via vehicle data; related crime reductions reflect pre- and post-installation observational comparisons rather than established causal proof. In residential leasing, EliseAI collaborated with voice providers to deploy regional-accented leasing assistants; company self-reported data indicated a 5.28% increase in tour conversions (a relative percentage increase rather than percentage points), a 7.8% lift in call engagement, and a 7.8% increase in median words spoken (more words spoken does not equate to higher conversions). Given the lack of public randomized control trial details, these figures are self-reported and cannot be taken as definitive causal conclusions.

Operating metrics from private platforms are also helping observers track broader market shifts: OpenRouter reflects Token usage trends across its own routing platform, while Stripe reflects SaaS revenue flows across its billing platform. Each platform covers different customer profiles and business models, showing trends specific to its own sample. For example, Stripe's data conversely suggests that SaaS revenue growth across its customer base is beginning to reaccelerate.

Concluding the report, the authors look toward future frontiers including compute-heavy consumer experiences, embodied robotics, full autonomy, AI-driven biology and healthcare, and deep penetration into non-coding enterprise workflows. a16z assesses that the next wave of demand will continue expanding in these directions; what will materialize and how much users will pay will ultimately be answered by real-world products and business execution.