The post is an Ed Zitron essay arguing that AI is “slowing down” in the only way that matters financially. It says model improvements are getting less dramatic, demand will not grow fast enough to support the datacenter buildout underway, and labs such as OpenAI and Anthropic would need implausibly large revenue to service their compute commitments. The piece leans on recent signs like enterprise spend caps, local and cheaper models getting better, and Apple folding AI into the operating system to argue that frontier model vendors are losing pricing power before they have a durable business.
Most readers did not buy the essay as written. The dominant reaction was that Zitron’s tone is overheated and his timing record is weak. People pointed to repeated “bubble is bursting now” calls over the last two years that did not land, and said the article mixes real financial risk with maximalist claims about utility and progress that no longer fit the market. Even many skeptics of AI hype still thought he overreached by treating financial fragility as proof that the products are not improving or that demand is fake.
Where the conversation settled was narrower and more useful. There is real disagreement about how fast frontier capabilities are still improving, but far less disagreement that the current revenue expectations are extreme and that buyers are starting to manage token spend much more aggressively. Several commenters argued the likely end state is not “AI goes away” but “models become cheaper, more interchangeable, and embedded inside operating systems, clouds, and SaaS products.” In that world, the standalone labs may have less pricing power than their valuations imply, while incumbents like Google, Microsoft, Apple, and Amazon are better positioned because they control distribution, enterprise relationships, and other profit pools. The more bullish camp pushed back that spend caps are evidence of real demand, not weakness, and that coding agents alone already support budget levels that seemed absurd a year ago. Even that bullish view usually conceded the same structural point: useful technology does not guarantee that every frontier lab, datacenter project, or valuation survives intact.
Separate two questions in your own planning: whether LLMs are valuable to users, and whether today’s lab valuations and capex assumptions will hold. Expect continued adoption alongside pricing pressure, spend controls, and a likely shift of value toward integrated products, incumbents, and cheaper model tiers rather than pure model vendors.
Mostly skeptical of the article rather than of AI itself. Readers were annoyed by the author’s bombastic style and shaky timing, but a lot of them still agreed that AI valuations, capex plans, and revenue assumptions look frothy and vulnerable to a correction.
Key insights
01
Distribution does not guarantee user lock-in
Preinstalling an assistant inside the operating system is not enough to kill a product people already prefer. Gemini is bundled on Android and Microsoft put Copilot into Windows, yet many users still choose ChatGPT or Claude because habits and perceived quality beat default placement unless the built-in tool is clearly better or uniquely useful.
Do not assume OS integration will erase independent AI products. If you ship an AI feature through distribution alone, measure whether it changes default user behavior instead of treating install base as demand.
Several commenters argued that enterprise buyers will not want to buy raw model access forever. They will buy workflow products that hide the model behind a business outcome, much like Copilot inside Microsoft’s stack. That leaves labs competing as interchangeable suppliers while customers focus on integration, governance, and ROI. Claims that enterprise demand will absorb any amount of spend looked weak unless that spend turns into concrete software outcomes quickly.
If you sell into enterprises, build around a workflow and a budget owner, not around direct attachment to a specific frontier model. Expect model choice to become a procurement detail unless your product delivers a clearly differentiated result.
A detailed technical rebuttal said the article's financial conclusions depend on getting inference economics right, and that parts of that stack were described incorrectly. Routing, batching, prefix or key-value cache reuse, and tiering requests to cheaper models all materially change unit economics. If providers can segment workloads and keep expensive models for only the hard cases, the claim that they must keep chasing frontier intelligence just to sustain margins gets much weaker.
Treat any grand claim about AI gross margins with caution unless it explains the serving architecture. In your own cost models, separate prefill, caching, routing, and model tiering before extrapolating from headline token prices.
A strong line of argument was that frontier labs are being boxed in from below. Qwen, Gemma, DeepSeek, and other open or cheaper hosted models are now good enough for many tasks, especially when wrapped in the right harness. That does not make top models irrelevant, but it does make it harder to defend premium pricing across the whole market. As capabilities spread downward, the likely outcome looks more like commoditized compute and narrower high-end margins than winner-take-all software economics.
Continuously re-test whether your use case still needs the best model. Many production tasks can move to cheaper hosted or local options faster than strategy decks assume, and that can materially change vendor risk and gross margin.
The Uber-style budget caps became the clearest point where bulls and bears were talking past each other. One side sees caps as proof that companies finally hit the ROI wall. The other sees them as proof that employees want to spend far more than anyone expected a year ago. The useful read is that both are true. Demand is real, but finance teams are now stepping in, which means AI spend is moving from experimentation into normal software budgeting.
Watch for AI purchases to behave more like every other enterprise software line item. Growth can continue while usage gets fenced by approvals, dashboards, and per-seat or per-team limits.
A recurring theme was that even if labs keep leading on model quality, Google, Apple, Microsoft, and Amazon sit closer to where money is actually captured. They own devices, operating systems, browsers, search, enterprise suites, cloud contracts, and default user behavior. That positioning gives them more ways to bundle AI, trade margin for adoption, and make the model layer look like a feature instead of a standalone business.
Plan for value capture to drift toward the companies that already own customer access and workflow surfaces. If your product depends on selling raw intelligence, assume the channel owner can undercut you or absorb the feature.
Several practitioners said the obvious speedup in code generation does not cleanly translate into more finished product. Teams still bottleneck on planning, review, design, coordination, and defect ownership. Some described a pattern where coding agents create bursts of output followed by long cleanup cycles. The practical point was not that coding agents are useless. It was that local productivity anecdotes are a bad proxy for business value or for how much token spend a company can justify.
When evaluating coding tools, track cycle time to production, defect rates, and review load, not just code volume or prompt count. Budget based on delivered outcomes, because raw acceleration can hide downstream drag.
The strongest bullish pushback said spend caps like Uber’s are being misread. A year ago, spending even hundreds of dollars per employee each month on AI looked absurd. Now companies are setting soft caps in the low thousands because demand from engineers is overshooting budgets. That looks less like a dying market than a suddenly much larger total addressable market for coding and general-purpose agents.
Do not mistake budget controls for a collapse in usage. If you manage a product or investment thesis here, track willingness to pay at the team level before you infer macro slowdown from isolated caps.
Security research suggests frontier gains are still real
One rebuttal to the “progress is slowing” thesis pointed to Anthropic’s Mythos and the recent wave of vulnerabilities reportedly found in Firefox, OpenBSD, Linux, and OpenSSL. The claim was that this is not just nicer product packaging. It reflects a real jump in what top-tier systems can do for specialized work like vulnerability discovery, even if most users do not see those models directly.
Be careful about generalizing from consumer chatbot fatigue to frontier capability ceilings. In some domains, especially security, the gains may still be large enough to create new budgets and new competitors.
A more optimistic read of the consumer threat argued that the key revenue is not coming from casual chat users anyway. Heavy users doing hobby programming, design work, or agentic workflows are the ones paying meaningful amounts. Those buyers are less likely to churn just because Apple or Google bundles a baseline assistant. They are buying a power tool, not a default utility.
Segment AI demand by intensity, not by broad consumer versus enterprise labels. Premium revenue may hold up longer than expected if your target user depends on advanced workflows rather than generic assistance.