a16z Charts of the Week: Attention Struggles to Grow, Older GPU Rents Are Rising, and Where Is AI Infrastructure Still Stuck?

In this edition of a16z's weekly charts, a striking contrast emerges: Netflix's revenue continues to climb, yet the time people actively spend watching, listening, and reading has not meaningfully increased. Content platforms must fight over the same fixed pool of free time. Meanwhile, charts in the AI domain reveal a different set of dynamics—both new and older GPU rental rates sit higher than at the start of the year, select manufacturing sectors have failed to expand capacity accordingly, and cybersecurity share prices have run ahead of revenue growth.

Active Media Time Isn't Growing—Whose Hours Are Platforms Fighting For?

In its latest quarter, Netflix generated $12.87 billion in revenue, up 11.8% year-over-year. While top-line scale continues to reach record highs, the pace of expansion has slowed once again.

For Netflix, competitors include both rival streaming services and short-form video platforms. If you have an hour in the evening and spend forty minutes scrolling short videos, less time remains for an episode of television. a16z raises a central question: as more content vies for consumers' idle hours, who will keep signing up to pay for long-form video? Netflix is also experimenting with short-form formats and live sports. However, this revenue chart does not isolate variables such as pricing adjustments or subscriber scale, and it cannot alone prove that short video caused the decelerating growth.

Media consumption data from BCG offers a helpful benchmark for this battle over time.

BCG distinguishes between two modes of engagement: actively watching, listening, and reading versus letting content play in the background while doing something else. The "active media consumption" in the chart reflects the former. Under this definition, US per capita consumption recently stood at 25.3 hours per week, down from its prior peak of 26.5 hours; the compound annual rate of change across the entire period was approximately -0.5%. This represents an observational finding over a specific window and cannot be used to establish a biological ceiling on human attention.

While the aggregate total barely shifted, time allocated across media categories diverged significantly. From the start to the end of the period, category shares and compound annual growth rates in consumption hours were:

• Print media: Share fell from 19% to 9%, with hours declining at a CAGR of 10.5%;
• Social media: Share rose from 8% to 12%, with hours growing at a CAGR of 6.6%;
• YouTube: Share rose from 4% to 6%, with hours growing at a CAGR of 7.8%; gaming share also climbed from 7% to 9%, with hours growing at a CAGR of 2.1%;
• Video excluding YouTube: Share edged down from 41% to 40%, with hours declining at a CAGR of 0.7%; adding YouTube back in, total video consumption still grew at a CAGR of approximately 0.2%.

Thus, video as a whole has not meaningfully contracted; shifts have occurred primarily within video itself and across other media formats. For creators and app developers, flatlining active media time implies that winning incremental minutes increasingly looks like taking share from incumbents. Background playback can extend content into commutes, workspaces, and everyday routines, creating more opportunities for consumption, but background playback does not mean audiences are paying attention throughout.

Even Older GPUs Cost More Than at Year-Start, but Rising Rents Don't Equal Asset Appreciation

GPUs can be rented by the hour, with pricing reflecting what buyers are currently willing to pay for compute resources. Across five GPU models tracked by Ornn Data, rental rates all sit higher than at the start of the year—including the older A100.

This chart normalizes each GPU's year-start rental price to 100, comparing relative changes rather than absolute dollar figures. In end-of-period readings, the A100 stands at approximately 106 (up ~6% year-to-date); the H100 is at ~136; and the RTX PRO 6000 WS is at ~165. The H200 and B200 stand at ~264 and ~222, respectively—equivalent to 2.64 times and 2.22 times their starting baselines. This does not imply that an H200 costs more than twice as much in dollar terms as an A100, as each curve originates from its own baseline price.

a16z added a hypothetical reference line: if a model's 7-day moving average rental rate steadily depreciated to zero over six years from its first complete observation, would today's actual rental price sit above or below that line?

Dividing actual rental pricing by the hypothetical benchmark at the same point in time and subtracting 1 yields the chart's premium spreads: A100 sits 9% higher, H100 is 54% higher, H200 is 143% higher, B200 is 111% higher, and RTX PRO 6000 WS is 73% higher. These percentages use a different baseline from the year-to-date gains in the preceding chart.

This indicates that the current rental market is holding up stronger than a mechanical linear decline would suggest. However, it does not reveal how much a GPU could be sold for on the secondary market, nor does it factor in upfront acquisition, power, and operational overhead; it cannot determine net book value or whether hardware has broken even. Furthermore, both the A100 and H100 spent extended historical stretches well below the reference line, and recent rebounds do not erase earlier troughs.

Model inference volume offers another lens. OpenRouter operates as an aggregation platform routing calls across diverse models, while tokens represent the granular units processed and generated by language models. Platform metrics show volume gains across both open-weights and closed-weights models. (Open weights means model weights are accessible, but does not automatically imply that all code, training data, and licenses are fully open source.)

Weekly token consumption on OpenRouter accelerated visibly after spring, paused briefly in late summer, and turned upward again, recently reaching 160–170 trillion tokens per week. The platform's disclosed breakdown between volume share and spend share reveals a stark divide:

• Open-weights models account for approximately 72% of token volume, but only about 25% of total spend;
• Closed-weights models generate under 30% of token volume, but command roughly three-quarters of spend; their volume share also saw a modest uptick in the most recent week.

This indicates that within the blended usage observed on the platform, open-weights models carry a lower cost per token on average. Traffic share cannot be directly equated with revenue share, nor does token share equate to compute share, since the computational workload required to process a token varies widely across architectures. Together, rental rates and inference volumes support a16z's assessment of resilient demand. Yet both remain localized indicators—they neither capture aggregate industry spend nor specify which models are running on which underlying GPUs.

AI Infrastructure Demands More Hardware, but US Factories Aren't Expanding in Lockstep

When order books fill up, manufacturers can build new facilities and install machinery to expand potential capacity; alternatively, they can run existing equipment more intensively to boost capacity utilization. The volume actually manufactured is physical output. Both capacity expansion and higher utilization can drive output growth, but four critical industrial sectors have taken distinctly different paths.

Federal Reserve industrial data and BLS price statistics highlight four divergent patterns:

1. Computers and peripheral equipment: Capacity grew 16%, utilization rose relatively by 26%, output expanded 46%, and producer prices increased 33%. Both adding physical capacity and utilizing existing lines more intensively contributed to output.
2. Communications equipment: Capacity surged 110%, relative utilization fell 11%, yet output still rose 87%, alongside a 19% increase in producer prices. With capacity more than doubling while operational utilization softened, a16z suggests that pockets of capacity may have become redundant.
3. Semiconductors: Capacity expanded 77%, utilization held roughly flat, output rose 76%, and producer prices climbed 34%. Output gains in this category were mainly accompanied by an expansion in production capacity.
4. Electrical equipment: For this sector providing critical apparatus to utility grids and distribution networks, producer prices surged 69%, while capacity contracted 9%. Even with a relative 13% gain in utilization, output still fell 7%. The coexistence of steep price inflation and shrinking domestic capacity is the sharpest anomaly among the four groups. However, due to differing subsector classification scopes, one cannot directly multiply capacity and utilization changes to deduce output here.

Imports represent the other side of supply. Adjusting Census Bureau import values by the BLS import price index isolates volume expansion from price changes; comparing this with domestic output indices shows that real imports across all four equipment categories grew far faster than domestic production.

• Computer peripherals: Domestic output grew 46%, while price-adjusted real imports surged 366%;
• Communications equipment: Domestic output rose 87%, while real imports climbed 175%;
• Semiconductors and components: Domestic output expanded 76%, while real imports grew 131%;
• Electrical equipment: Against a 7% contraction in domestic output, real imports increased by 69%.

These growth figures show that overseas supply expanded far more rapidly, but they do not reveal what share of the total market imports now control, nor do they prove that foreign supply fully met outstanding demand. a16z outlines several potential explanations: capital costs, investment risks, or concerns about competing with foreign producers may have deterred domestic firms from expanding alongside demand. The available dataset cannot yet determine which factor dominates.

Labor data points to another bottleneck: in several manufacturing segments, weekly hours worked per employee grew faster than headcount.

In nonmetallic mineral products, hours per worker rose 4.7% while headcount increased by only 0.4%; in chemicals, the figures were +4.2% and +0.4%; and in plastics and rubber, they were +3.6% and -0.9%. Across these sectors, existing employees logged longer shifts while net payroll additions remained minimal or negative. Yet this pattern does not hold everywhere: in electrical equipment, headcount growth reached 4.0%, slightly outpacing its 3.8% increase in hours; fabricated metals (+1.7% headcount vs. +0.9% hours) and machinery (+1.4% headcount vs. +0.7% hours) also represented exceptions where headcount grew faster than hours per worker.

Gross manufacturing compensation rose 5.5% year-over-year; job changers saw gains around 10%, roughly 5 percentage points higher than job stayers. Rising overtime and pay gains for job switchers align with employers competing for labor, but cannot alone prove that labor shortages were the sole driver. Longer individual hours also do not equate to higher output per hour worked. These remain sector-wide aggregate figures that cannot isolate how much labor demand stemmed specifically from AI projects. a16z frames them as critical labor supply signals when building data centers and related infrastructure; whether individual projects face constraints will depend on actual hiring and delivery on the ground.

Cybersecurity AI Units Are Growing—Is Top-Line Revenue Keeping Up?

AI can empower adversaries while multiplying the models, applications, and Agents that enterprises must defend, leading to elevated growth expectations for cybersecurity. a16z constructed a market-cap-weighted basket of cybersecurity and observability constituents within the software ETF IGV to track their performance against the broader software sector. (Observability tools help organizations monitor how systems are operating and diagnose failures; this basket includes more than pure-play cybersecurity vendors.)

Year-to-date cumulative returns for this basket peaked around 110%–115% before pulling back to roughly 83%, compared with 0%–2% for the broader IGV software ETF. Despite the recent pullback, the basket maintains a substantial lead year-to-date.

Has this equity rally translated into accelerating top-line revenue growth? The chart compares consensus expected year-over-year revenue growth for the latest quarter against growth from an earlier baseline quarter. Because this incorporates consensus market expectations, it should not be conflated with reported earnings results.

Several vendors in the chart illustrate this divergence. The following figures are visual estimates, with growth rate differences expressed in percentage points:

• Okta shares gained approximately 120% year-to-date, yet its expected revenue growth rate decelerated by 1.5 percentage points compared to the baseline;
• Fortinet posted an approximate 115% return, with growth accelerating by roughly 5 percentage points;
• CrowdStrike gained nearly 95% in share price, but its growth rate change was a marginal +0.5 percentage points;
• Cloudflare (NET) gained 50%, while its expected growth rate slipped by 2 percentage points;
• Palo Alto Networks (PANW) saw expected growth accelerate by roughly 19 percentage points alongside a ~90% share price gain, standing out as a notable exception; however, a16z notes that this acceleration was largely driven by acquisitions rather than purely organic AI expansion.

Meanwhile, company disclosures around emerging AI offerings do show rapid expansion across specific metrics:

Palo Alto reported quarter-over-quarter customer growth exceeding 100% for Prisma AIRS, while tracked Agentic traffic grew over 800% across nine months. SentinelOne reported year-over-year annual recurring revenue (ARR) growth of approximately 200% for Purple AI and Prompt Security, while CrowdStrike's AIDR ARR grew roughly 200% quarter-over-quarter. SailPoint disclosed that AI-driven pipeline expanded over 100% since June. (Prisma AIRS secures AI models, applications, and Agents; Purple AI assists security analysts; Prompt Security protects enterprise AI adoption; AIDR detects and responds to AI-related risks, securing prompts, applications, and Agent interactions.)

Beyond pipeline metrics, SailPoint also disclosed data closer to actual monetization: annual spend among existing customers adopting AI solutions expanded by more than 60%. While this represents an encouraging indicator of customer spend expansion and was cited in the company's Q2 FY27 business updates, the disclosure did not specify the exact baseline used to calculate the uplift, nor is it sufficient to attribute the entire increase solely to AI.

In Q2 FY27, AI-related solutions accounted for >30% of net new ARR, representing a tangible contribution to incremental annualized recurring business. Under company definitions, net new ARR measures the net sequential change—ending ARR minus beginning ARR for the quarter—rather than gross contract value signed. This >30% figure represents an incremental contribution share (meaning for every 100 units of net new ARR added, AI-related offerings contributed more than 30 units); it is not a 30% growth rate, nor is it 30% of total ARR, let alone 30% of total company revenue.

Interpreting these metrics requires first clarifying where the measured growth actually occurs. Traffic measures network activity; pipeline represents potential unclosed opportunities; bookings are defined company by company and can diverge substantially from recognized revenue in both amount and timing; and ARR annualizes recurring business at a single point in time, which is not equivalent to revenue recognized during the quarter. Customer counts, network volume, and contracts can all signal traction, but they represent fundamentally different operating metrics. Incremental contribution share and percentage growth rates also follow distinct logics, just as a 200% quarter-over-quarter gain and a 200% year-over-year gain measure entirely different durations.

Thus, while these disclosures provide tangible evidence of traction closer to monetization and net additions, they do not prove that overall company revenue is accelerating; the absolute scale of these emerging product lines remains critical. AI may create net-new demand, or it may lead customers to swap existing security budgets into newer tooling—the ultimate net effect will depend on broader corporate performance. a16z's takeaway for this basket is specific: markets have priced in elevated expectations, and vendors will now need to validate that optimism through recognized revenue growth.

The original newsletter opened with an independent security briefing by Joel de la Garza and Zane Lackey titled "Patch or Die", previously sent to portfolio CEOs. It sits outside the four core recurring chart topics discussed above.