a16z State of Markets II Finds AI Adoption Shallow While Capex Approaches $780 Billion

Andreessen Horowitz published its most comprehensive analysis of the AI investment cycle on September 30, 2026, and the central finding cuts against the dominant narrative: while $780 billion in capital is flowing into AI infrastructure this year, real-world adoption by consumers and enterprises remains strikingly thin. The State of Markets II report — featuring more than 100 charts and authored by David George, a general partner at a16z's Growth Fund — documents both the scale of the buildout and the gap between that buildout and measurable outcomes.

The report covers the first half of 2026 in public equity and private technology markets, framing the AI buildout not as a niche technology trend but as the primary driver of the entire US economic cycle.

Tech Has Become the Macro Cycle — Whether or Not AI Is Working

George's report argues that technology has replaced durable goods as the engine of the broader economic cycle — a framing he calls "Tech Is The Everything Cycle." According to a16z's analysis of S&P 500 earnings projections, the technology sector is responsible for approximately 76% of the index's full-year 2026 earnings growth, growing at roughly 3.8 times the rate of non-tech companies.

This concentration is historically unusual. Prior economic cycles were typically driven by housing, manufacturing, or energy. The 2026 version is driven by AI-adjacent infrastructure spending, even as AI's actual productivity effect on businesses remains largely unmeasured.

The distinction matters because it creates a structural dependency: the S&P 500's earnings trajectory is now materially tied to whether hyperscaler AI spending continues to expand. If AI buildout slows, the macro story changes quickly.

From Bits to Atoms: $780 Billion Flowing Into Physical Infrastructure

The original thesis for AI investment — from roughly 2021 through 2024 — assumed that value would accrue primarily in software: foundation model APIs, AI-enhanced SaaS products, and developer tooling. The 2026 data tells a different story.

The four major hyperscalers — Microsoft, Amazon, Google, and Meta — spent approximately $416 billion on AI infrastructure in 2025. The a16z report projects that figure will nearly double to roughly $780 billion in 2026, based on company guidance and analyst projections from Morgan Stanley. Those same analysts project the figure could exceed $1 trillion in 2027.

George describes this as a "bits to atoms" transformation: the free cash flow that hyperscalers generate from their software and cloud businesses is being converted into purchases of physical infrastructure — GPUs, high-bandwidth memory, networking interconnects, data center construction, and power generation. Earnings streams that once flowed primarily through software economics (near-zero marginal cost, high gross margins) are now being redirected into physical-asset categories with very different return profiles.

The immediate financial beneficiaries are not software companies. They are GPU manufacturers, memory suppliers, data center operators, and power infrastructure providers.

GPU Rental Rates Are Rising, Not Falling — and Jevons Explains Why

One of the most counterintuitive findings in the report concerns GPU pricing. As newer-generation chips like the NVIDIA B200 have entered the market, one might expect rental prices for older A100-based servers to fall as capacity is freed up. According to the a16z report, that has not happened: A100 GPU rental rates have held steady or increased, despite the availability of more powerful alternatives.

The explanation offered is a version of Jevons Paradox — an economic principle first articulated in 1865 to explain why improvements in coal-engine efficiency increased coal consumption rather than reducing it. Applied to AI inference: as the cost per AI query drops through more efficient chips, quantization, and model distillation, the number of queries rises dramatically. New applications that were previously cost-prohibitive become economically viable. Enterprise workflows that generated modest AI usage begin generating intensive AI usage. Consumer products that once made one or two AI calls per session begin making dozens.

The net result is that total GPU demand absorbs new capacity as fast as it arrives, keeping older-generation hardware in active use rather than retiring it. For companies with existing A100 infrastructure, this is significant: their assets have not been immediately stranded by the B200 generation.

This dynamic also reinforces the "bits to atoms" thesis. If demand reliably absorbs supply increases, the constraint on AI deployment is not algorithmic or software-based but physical: how fast can new GPU clusters, memory bandwidth, and power capacity be built.

Enterprise AI Deployment Is Wide but Almost Entirely Unmeasured

The adoption picture that emerges from the a16z analysis creates a significant analytical tension. On one hand, 69% of S&P 500 companies report having live AI deployments as of mid-2026, according to a16z's analysis of corporate disclosures — a figure suggesting that AI has crossed from early-adopter experimentation to mainstream enterprise rollout.

On the other hand, only approximately 2% of those companies have disclosed tracked metrics demonstrating AI's effect over time. A separate survey from April 2026 found that roughly 2% of US households were paying for AI products in any form.

These figures are not necessarily contradictory. Enterprises frequently deploy software without measuring ROI in the first year, and consumer AI adoption is still early. But the gap between a 69% enterprise deployment rate and a roughly 2% metric-tracking rate is large enough to raise a structural question: how much of the current infrastructure buildout is converting into demonstrable value, and how much is still in the exploratory phase?

This is the classical early-cycle risk pattern in technology investment: infrastructure build-out precedes provable demand. The current situation — roughly $780 billion in annual AI infrastructure spending against an enterprise adoption base where outcomes are almost entirely self-reported and unquantified — is not yet a crisis, but it is a condition worth monitoring.

Software Isn't Dead — It's Bifurcated

The "SaaSpocalypse" thesis held that AI would commoditize software products, destroying SaaS revenue multiples across the board. The actual 2026 data presents a more nuanced picture.

Approximately 75% of publicly traded software companies were profitable as of mid-2026, up sharply from the roughly 30% that were profitable in 2021. Only about 30% are growing revenue at 20% or faster — a steep decline from the frothy 2021 era. The old growth-at-any-cost model is largely gone, but the industry has survived by prioritizing profitability.

The valuation bifurcation is the more instructive finding. Infrastructure software — the category encompassing cloud infrastructure tools, database platforms, and networking software that enables the AI buildout itself — trades at roughly 9.1 times revenue. Horizontal software tools (productivity, general-purpose workflow, collaboration) trade at approximately 2.7 times revenue. The spread between these categories represents the market's verdict on which software benefits from the AI buildout and which is threatened by it.

Consumer AI Agents and the Advertising Reckoning

One of the sharpest implications in the report concerns the ad-revenue model underpinning Amazon, Google, and Meta's businesses. A new category of consumer-facing AI agents — tools that can research, compare, and purchase products on behalf of users — is beginning to disrupt how commercial intent flows through digital platforms.

The threat mechanism is structural: advertising revenue depends on inserting sponsored content at the moment a consumer expresses purchase intent. An AI agent that routes a consumer directly to the best product match, bypassing search results and sponsored listings, removes the ad-insertion point entirely. According to a16z's companion analysis published the same day, Amazon generated approximately $69 billion in advertising revenue in 2025, while Alphabet and Meta combined for nearly $500 billion in advertising revenue from controlling digital distribution. Those figures represent the financial stack at risk if AI agents become a mainstream shopping interface.

The corporate responses to this threat have diverged sharply. When Meta launched its Muse personal AI agent in September 2026, Amazon moved to block Muse from shopping on its marketplace, citing concerns that it accessed customer data without authorization and did not identify itself as an automated agent. Instacart and Shopify, by contrast, welcomed Muse integrations — a difference that reflects the structural distinction between ad-dependent platforms and transaction-fee-dependent platforms. For Shopify, an AI agent that converts shopping intent into a purchase is a revenue event regardless of the interface. For Amazon, whose advertising business contributed more to operating income than its own reported operating profit excluding AWS in 2025, the calculus is very different.

Read more: Meta Muse tops App Store charts then faces Mac zero-day security flaw

VC Returns Are Polarizing at Historic Extremes

The venture capital picture in the report underscores a theme running through the entire analysis: AI-era outcomes are concentrating at the extremes.

For 2024 vintage funds, according to industry data from Burgiss and Cambridge Associates cited by a16z, the top 10% of funds by net internal rate of return are generating 40.5% IRR. The median fund is at -3.3%. The bottom quartile is at -14.6%.

For vintages from 2021 through 2023 — the period of peak AI hype and valuations — distributions to paid-in capital remain near zero across the board. Limited partners who committed capital during that period are still waiting for meaningful liquidity. Late-stage write-downs from 2021 peak valuations continue to weigh on unrealized portfolio values.

In August 2026, a16z closed its Machine Age Fund at $1.1 billion, specifically targeting AI physical infrastructure: chips, memory, networking, storage, data centers, and robotics. The fund's thesis is a direct expression of the "bits to atoms" argument — that the next major value capture in AI will be in physical assets rather than software abstractions.

Read more: Andreessen Horowitz launches academy as AI reshapes the entry-level talent market

What Comes Next: Agents, Robots, and the Measurement Problem

The a16z report identifies several areas where the firm expects significant AI development through the remainder of the decade: consumer-facing AI agents, robotics, autonomous vehicles, AI in biology and drug discovery, AI in health, and AI in defense.

These categories share a common characteristic: they represent AI moving from language tasks to physical-world tasks, from digital-only to embodied operation. The shift from inference that produces text and images to inference that controls robots or navigates vehicles represents a qualitatively different infrastructure requirement — one where latency, reliability, and physical cost tolerances differ significantly from the data center workloads of 2024.

The measurement problem identified in the current report will likely persist and intensify in these categories. When AI operates in a physical environment, establishing causation between AI deployment and business outcome is harder, not easier. A robot performing a warehouse task is measurable; a recommendation algorithm improving a drug discovery process is substantially more difficult to quantify cleanly.

The more immediate question hanging over the AI cycle is whether enterprise and consumer adoption will accelerate fast enough to justify the $780 billion infrastructure buildout underway in 2026. The a16z data suggests the infrastructure is arriving faster than the measurable demand. History offers both reassuring and cautionary precedents for that condition — the internet infrastructure buildout of the late 1990s being the most instructive comparison the current moment invites.