About 91% of households pay for at least one streaming service. About 55% pay for cloud storage. Only about 2% pay for AI. Those numbers are not close. Paid household AI is still a rounding error next to streaming money, and the gap is large enough that it reads less like a victory lap and more like a bubble check. GPUs already run hot. The household checkbook still says almost nobody is writing a check for AI.
The Household Receipts
A comparison shared by co-founder Nick puts the early stage in plain terms. Only 2% of households pay for AI. By comparison, 25% pay for SiriusXM, 55% pay for cloud storage, and 91% pay for at least one streaming service. People who have looked at the real receipts land in the same place: barely even 2% of U.S. households were paying for some AI service. The report behind those figures was run in April. Adoption may have moved since then, but a rise all the way to 3% would still be surprising. Even at the slightly higher figure of 2.2%, more households paying for AI still means almost everyone is not. About 98% of households are not paying for an AI service yet.
Among the households that do pay, the monthly average spend is only $31. That is less than many households already spend on streaming. The feed sells peak AI. The checkbook still says this is early. Someone who does about half of their work inside Codex, often with multiple agents running, and who follows model-release benchmarks, still reads those numbers as a reminder. The people paying attention can feel behind and stressed. The same numbers say they are on the cutting edge, and that the wider world still has little idea what is coming. The same pattern shows up in ordinary life: kids scattered across the U.S. and around the world, and still very little AI penetration in day-to-day use.
What the Andreessen Horowitz Market Report Shows
The household comparison sits inside a broader state-of-the-market report from Andreessen Horowitz, one of the largest venture capital firms in the United States. The firm released a second edition of that report. The receipts in it are not a reason to treat every chart as a finish line. Consumer paid penetration is the story, and the story is that AI is early.
The report also frames tech as the everything cycle. Cumulative earnings growth for the S&P 500 shows tech earnings growing 3.8 times more than non-tech. From 2010 through 2026, the top tech companies have exploded since COVID, with a sharp rise in tech earnings. Tech contributed 76% of S&P 500 annual earnings growth. A lot of people look at that and call it an AI tech bubble. The more interesting split in the report is between hyperscalers and semiconductors. Hyperscaler free cash flow has effectively become semiconductor free cash flow. Semiconductor results are exploding. Hyperscaler free cash flow is dropping off a cliff. That spike is what has been driving the pain in RAM and GPU costs since 2022.
The AI build-out has caused a surge in demand for industries that used to look sleepy, cyclical, and capital-intensive, including semiconductors. That demand has been funded in large part by the historical massive profits of the world’s largest tech companies. Corporate spending is what is propping up AI growth right now. It has not pushed down into residential households in any meaningful way.
Companies Know They Need AI. Few Know Why.
On the enterprise side, consumers are not simply the laggards. The report says 69% of S&P 500 companies have a live AI deployment, while only 2% disclose a metric they track over time. That gap is the point. It is still early in figuring out how companies will use AI inside existing businesses and how that use will generate revenue. Everybody knows they need to be deploying AI. Only 2% know exactly why they are deploying it.
Explosive household growth has not started, partly because AI has not really been integrated into everyday products yet. Gemini is one counterexample on the product side. It has not become the best frontier model, but it has been integrated into more than 800 applications in the Google stack. In one household, most Gemini use happens inside Google Search, where the chat is opened directly from the search results. The standalone app is more useful if history is kept, but the search integration is what is actually getting used.
GPU Obsolescence Looks Overstated
Reports of GPU obsolescence have been greatly exaggerated. The common assumption was that GPUs would only be effective for three to four years. A 3090 and a 3090 Ti, now five- and six-year-old GPUs, are still fully effective and useful. The 3090s are still rock solid. These cards were not really designed to be run that long. Usefulness out to 10 years is a reasonable prediction from here, and it will be interesting to see whether a real falloff shows up. So far, it has not.
If demand is this early and the cloud invoice is already this loud, renting every brain cell forever is not a great plan. One practical way in is local. OpenMonoAgent.ai is an open-source repository that can be installed and run locally. A free giveaway of small inference boxes is meant to let people run that stack on their own hardware and own the entire stack. A new feature that makes it more useful is set to be announced this week.
AI as Infrastructure You Own
Startuphakk builds custom software for companies, including custom AI solutions, and works with teams that want to be part of the AI cycle rather than spectators to it. Spencer Thomason, fractional CTO and founder, has 25 years of software development and a decade of executive leadership, including time at organizations such as GoDaddy, SRP, and Wells Fargo. The pattern he describes is not mainly a technology problem. Most companies have a leadership problem, and they pay for it in missed deadlines, failed integrations, and AI investments that delivered nothing. Bad technology decisions do not only waste money. They can kill companies.
The businesses winning right now are not the ones chasing the latest AI trends. They are the ones that built on solid engineering and treated AI as infrastructure they own, control, and integrate into software that actually works. That means database architecture, API design, system integration, and scalable infrastructure built the way it has been done for more than 25 years. When AI belongs in a solution, it is designed into the architecture and run in the company’s environment, not bolted on, not wrapped around someone else’s API, and not left on a vendor’s cloud or someone else’s pricing schedule.
OpenMonoAgent.ai is the proof point for that model: a terminal-native AI coding agent running entirely on local LLMs, with zero API costs, zero telemetry, and full ownership. It is growing because serious engineers recognize real infrastructure when they see it. The same standard is what a fractional CTO engagement is meant to deliver without a full-time executive cost: strategic architecture, hands-on delivery, and clear accountability, rather than slide decks that gather dust.
Conclusion
The takeaway from the household numbers and the Andreessen Horowitz report is the same. Paid consumer AI is still around 2%, average spend is about $31 a month, and corporate deployment is far ahead of corporate measurement. Semiconductors are absorbing the build-out that hyperscaler profits are funding. Older GPUs are lasting longer than the three-to-four-year story suggested. Integration into real products is only beginning. We have not even begun to see the explosive growth yet. That is what early looks like.
