The hottest argument in AI right now is not whether the models work. It’s whether anyone understands what happens when intelligence gets cheaper, usage explodes, and every business has to rebuild around abundance, verification, and risk.
For about five minutes, Silicon Valley convinced itself that cheaper models would mean cheaper businesses.
That story is dying in public.
The live argument on X today is not really about model evals, benchmarks, or who won the week. It is about what happens when intelligence becomes cheap enough to use everywhere, but not safe enough, reliable enough, or structurally integrated enough to make the rest of the business simple.
The signal is unusually clean tonight.
a16z is posting charts that amount to a blunt market verdict: token prices are down, token demand is up, and total token spend keeps hitting new highs. One of the posts calls it “textbook Jevons”. Another notes Google is now processing more than 3.2 quadrillion tokens per month, up 7x year on year. That is not a story about efficiency. That is a story about appetite.
At roughly the same time, DHH is posting that GPT-5.5 wrote the majority of a 30,000-line branch for a complicated QML conversion, and that agent work feels materially better than the previous generation. Sam Altman is out asking what problems people most want AI to solve, while OpenAI is shipping another Codex update designed to pull more of the actual desktop workflow into the model loop. Anthropic, meanwhile, says its Project Glasswing collaboration has already found more than ten thousand high- or critical-severity vulnerabilities in essential software.
That is the real shape of the moment.
Capability is improving fast enough to drive broad adoption. Cost is falling fast enough to remove hesitation. But operational complexity, infrastructure demand, and risk are rising right alongside them.
That means the next phase of AI is not “software gets cheaper”.
It is “intelligence gets cheaper, so everything else gets more intense”.
This is a very normal error. It is also the sort of error that gets board decks written by people who do not understand second-order effects.
When the unit cost of something useful drops sharply, you do not automatically spend less on it. Often you spend much, much more because you start using it in places that were previously uneconomic. That is the whole Jevons point. Efficiency does not just save money. It expands the set of worthwhile use cases.
We already know this from computing. Faster chips did not reduce total compute demand. Better broadband did not reduce traffic. Cheaper cloud did not end infrastructure bills. Every time a core input becomes more abundant, people stop rationing it and start redesigning around it.
That is what is now happening to intelligence itself.
For years, businesses bought software as a fixed workflow wrapped in a seat licence. The software did one narrow job. The human adapted to the software. AI flips that. Now the intelligence layer can adapt to the task, the file, the customer conversation, the codebase, or the back-office mess in front of it. That makes it more flexible than classic SaaS, but it also makes it much easier to overuse.
If intelligence becomes cheap enough, you do not ask whether a workflow deserves AI. You ask why it does not already have AI.
That is the line most operators still have not crossed mentally.
They are treating AI like an expensive plugin, when the economics are pushing it towards becoming a default utility.
You can see the pattern everywhere if you stop looking for tidy product narratives.
In coding, better agent performance does not reduce demand for model calls. It increases it. Once a model can do more than autocomplete and can actually push through bigger chunks of implementation, developers do not sit back and enjoy the savings. They widen the scope. They parallelise more tasks. They run more verification. They attempt migrations they would have postponed. They become more ambitious because the marginal cost of trying has collapsed.
That is exactly why DHH’s post matters. The interesting part is not that a model wrote a lot of code. The interesting part is that a credible operator is describing a threshold shift: work that previously felt too cumbersome or too unreliable for agents now feels increasingly viable. When that threshold moves, demand explodes because the set of “worth trying” jobs expands overnight.
In product workflows, the same thing is happening. OpenAI’s Codex updates are not merely feature releases. They are attempts to reduce the friction between real work and model context. The easier it becomes to pull app state, screenshots, text, and project context into the loop, the more often people will reach for the model. Again: lower friction does not calm usage. It accelerates it.
In security, the pattern turns darker. Anthropic’s Glasswing update is a reminder that capability lifts both productivity and threat volume. If models can help defenders discover high-severity vulnerabilities faster, good. They can also increase the rate at which software systems are probed, analysed, and stressed. The software industry has spent years operating on the assumption that vulnerability discovery is bottlenecked by scarce human attention. That assumption is now under attack.
This is why “AI gets cheaper” is such a misleading sentence.
Cheaper for whom? Under what architecture? After which knock-on effects? At what usage level? With which governance burden?
The model call may get cheaper. The system rarely does.
If the intelligence layer becomes abundant, the scarce things become more obvious.
Not raw capability. Not access. Not even, soon enough, model choice.
The bottlenecks shift to context, trust, orchestration, infrastructure, and permission.
Context matters because cheap intelligence without clean context is just fast confusion. Most companies do not have workflows. They have sediment. Legacy rules. Weird naming. Half-buried approvals. Institutional scar tissue disguised as process. Giving a model access to that mess does not magically produce leverage. It often produces very confident entropy.
Trust matters because the more often AI is used, the less realistic manual checking becomes. A lot of the current enterprise theatre is built on the fiction that humans will stay neatly “in the loop”. They won’t. Not at scale. Humans will be in the exception path, not the main path. The companies that admit this early will build better review systems. The rest will keep pretending their governance model is a person glancing at outputs and feeling responsible.
Orchestration matters because one good answer is not the end state. Real work increasingly looks like chains of delegated tasks, verification passes, external tool calls, retrieval, escalation, retries, and audit logs. That means the valuable layer is not only the frontier model. It is the operating system around the model.
Infrastructure matters because all of this hits the physical world. Stratechery’s data centre point is not some side quest. It is the bill coming due. Everyone wants AI in the abstract. Far fewer want the substations, land use, cooling, and grid politics required to support it. The digital fantasy always ends in concrete, power, and permits.
Permission matters because the old software world assumed bounded tools. The new one increasingly pushes towards agents that can touch systems, move money, patch code, talk to customers, or make operational decisions. The hard question is no longer “can the model do it?” It is “what are we willing to let it touch without blowing our feet off?”
That is where the next serious companies will differentiate.
Not by saying “we use AI”.
By deciding where autonomy is worth more than control, and where it absolutely is not.
There is a reason the discourse feels twitchy.
The old software business model was easy to understand. You sold seats. You expanded accounts. You maybe layered in usage. The customer paid for the software and did the work themselves.
AI breaks that symmetry.
If the software actually does more of the work, users do not necessarily want to pay per seat in the same way. They may pay for outcomes, usage, approvals, throughput, or business impact. Or they may expect the baseline software price to stay roughly where it is while the vendor absorbs much higher compute and support complexity. Neither outcome is especially cosy for incumbents.
That is why “AI features” are not enough. A thin assistant bolted onto the old product only postpones the reckoning. The deeper question is whether the product’s economic model still makes sense when intelligence is part of the operating cost of serving every user interaction.
This is also why the startup opportunity is real, but narrower than the breathless takes suggest.
Yes, AI lowers the cost of starting things. a16z is right to note the rise in solopreneurship. Yes, agents widen what one strong operator can ship. Yes, there will be real new companies built on top of cheaper intelligence.
But the easy conclusion, that this means software margins stay magically pristine while output goes vertical, is nonsense.
Some categories will become excellent businesses. Some will become brutal utilities. Some will briefly look like software and then get repriced as labour arbitrage with a shorter shelf life.
The winners will be the companies that understand where intelligence is commoditising and where system design still compounds.
A lot of boards and leadership teams are approaching AI like an unusually noisy buying decision.
Which model.
Which vendor.
Which pilot.
Which compliance wrapper.
Which line item.
That is too small.
The real question is whether the company is prepared to rebuild around abundant intelligence the same way earlier generations rebuilt around abundant bandwidth, abundant mobile access, and abundant cloud compute.
Here is the blunt version.
If intelligence is cheap, you should expect:
more tasks attempted,
more software generated,
more customer interactions handled automatically,
more internal documents queried,
more edge cases surfaced,
more vulnerabilities discovered,
more infrastructure consumed,
more governance required,
and more competition from people who can now do far more with smaller teams.
That does not mean every AI spend is wise. Plenty of it is wasteful. Plenty of “agent” products are still wrappers with a confidence problem.
But the direction of travel is obvious.
The organisations that win will not be the ones that ask, “How do we save 20% with AI?”
They will be the ones that ask, “What would our business look like if intelligence were abundant, supervision were expensive, and speed were now the default expectation?”
That is a much harder question. It is also the only one worth asking.
That debate is over, except in corners of the internet that enjoy being wrong slowly.
The live debate now is economic and organisational.
Can businesses survive the inversion that happens when intelligence becomes cheap but everything around it becomes more demanding? Can software pricing survive a world where usage explodes and value is harder to map to seats? Can security practices survive a world where vulnerability discovery is no longer constrained by small expert teams? Can infrastructure politics survive a world where every productivity gain turns into a bigger hunger for compute?
Those are real debates. Serious people are now having them in public. That is why tonight’s X chatter matters.
The important thing is not that a few well-followed accounts posted about token demand, coding agents, or vulnerabilities. It is that those signals are pointing in the same direction.
The industry is leaving the demo phase.
We are entering the abundance phase.
And abundance, contrary to the sales pitch, is rarely neat.
Because in the space of a few hours, the conversation tightened around the same underlying truth from multiple angles: a16z on Jevons and token demand, DHH on materially better agent output, OpenAI on pushing models deeper into the work surface, Anthropic on the security consequences of stronger systems, and Stratechery on the economic and physical bottlenecks that show up next. Different posts, same message: cheaper intelligence is not a cost-saving footnote. It is a structural change.
a16z: Charts of the Week, including posts on cheaper intelligence, rising token demand, and new highs in spend: https://www.a16z.news/p/charts-of-the-week-reflections-from
a16z on X: “Textbook Jevons: Token prices are down, token demand is up.” https://nitter.net/a16z/status/2057876678534447487#m
Anthropic on X: Project Glasswing update claiming more than 10,000 high- or critical-severity vulnerabilities found in essential software. https://nitter.net/AnthropicAI/status/2057909102542549503#m
DHH on X: GPT-5.5 materially improving complicated agent work. https://nitter.net/dhh/status/2057906669158309913#m
Sam Altman on X: “new codex ships today!” https://nitter.net/sama/status/2057559714788258003#m
OpenAI Developers on X: Codex Appshots and deeper desktop context. https://nitter.net/OpenAIDevs/status/2057530207976989179#m
Stratechery: The Data Center Veto / Agent Economics roundup. https://stratechery.com/2026/the-data-center-veto/
