The increasing availability and commoditization of AI makes both opportunity and threat nearly limitless, while fluency and skill in operating systems becomes more important. Has the discourse moved here quickly enough?

You can now run the smallest Qwen model on a tablet. Laptop AI is a reality. Hugging Face has over three million models. Qwen model weights have been downloaded over three billion times. Generative AI is no longer a operating under a subscription-based monopolistic control model; it is now, most definitively, a choose your own adventure, just as its “abilities,” particularly in code, put the power of entire software companies in the hands of individual “developers” with persistent, effective prompting skills. The most consequential change in the constantly, rapidly shifting power balances world of artificial intelligence is now the loss of scarcity, even as the discourse and focus revolves around an overinflated kind of existential risk. Useful models are becoming available from more places, for more purposes, on more kinds of hardware. As access expands, the ability to apply them becomes a more important source of economic advantage. That presents a problem for companies whose business models depend on controlling access, and an opportunity for people who understand what to do with it.

Let’s be realistic: AI is dangerous, but not (yet) as an autonomous threat. It’s dangerous, no question, but power still resides firmly in the hands of human ambition, and it is now visibly changing entire spheres of influence. A recent thread on Twitter/X mused on the death of the hedge fund and the end of the quant career, one of the most lucrative in business, as many of the scarce skill set functions of those roles can now be performed more ably by AI. Yet much of the public conversation still treats AI as a contest among a few frontierslabs. Which company’s model has the best numbers? Who leads METR benchmarks? Which release will pull ahead next? These conversations, which dominated the discourse since inception of the AI chatbot, are no longer the significant questions.

AI is spreading into a much more complicated landscape: Chinese open-weight models, “reasoning” (the definition varying widely) systems, genomic models, filmmaking tools, coding assistants and small models that run on personal devices. The frontier lab remains part of that landscape. But the conversation is now about a generic assumption of AI dominance: which industries it will change irrevocably is now being charted on 4×4 matrices. The discourse has fallen behind reality and no longer provides a useful map of where we are at. It exited via the existential superintelligence off-ramp and shows no signs of reentering or even finding the freeway again. In fact, the freeway may no longer exist. It may be more like networks of Shinkansens, moving silently through new landscapes, implacable, not conscious, but faster than anything else, reality-altering nonetheless, and unstoppable.

Graphic of professions where AI will supplant next by Andrew Bailey

We have arrived, so soon, at the great AI fragmentation, and we are unable to see the forest for the trees. Never has a technology been so widely adopted, changed so much, yet been so massively misunderstood, not only by the people who govern its use, but by the intellectuals who drive the discourse about it, too. Its implications extend from software economics to geopolitical influence to the distribution of practical power. Where this leaves us is anyone’s guess, but one thing is clear: people who actually understand how to use it are already deploying that expertise, with most of the world engaged in a completely different conversation.

The Great Proliferation and Age of Abundance

In a matter of months, we have gone from a virtual monopoly to a universe of abundance. It was less than a year ago when DeepSeek shocked the Valley to its core. Developed at a fraction of the cost and with far more limited resources than anything from American labs, the model announced America was not only no longer alone, its formidable competitor was a a nimble, adept and very real threat. And now, here we are. Hugging Face’s public model catalogue displayed approximately 3.14 million repositories when checked for this article. Those are not three million independently developed foundation models: the catalogue includes fine-tunes, compressed versions and other adaptations. But adaptation is the point. A foundation can become the starting material for countless different applications, with different owners and different purposes.

Consider the distance between a general assistant and a genomic model. Meta’s Muse Spark is a multimodal “reasoning” model that asks for access to the entirety of our lives. Arc Institute’s Evo 2 models DNA, bringing language-model methods to biological sequences. These systems operate on different material and serve different needs. The category “AI” now contains so much variation that discussing its commercial future as if it were one product obscures more than it explains.

The same fragmentation is occurring in deployment. Chinese model families such as Qwen provide downloadable foundations for other developers. Google documents ways to run Gemma on Android and iOS. A suitable model can run on a personal computer or compatible mobile device, with hardware and memory determining what is practical. Useful capability increasingly exists beyond the remote service accessed through a chatbot window.

Chàbuduō, or what’s good enough?

The commercial question changes with that availability. A business needs a system that performs its task well enough, at a defensible cost, with acceptable control over its data. Winning an aggregate leaderboard is only one, increasingly irrelevant possible contribution to that decision. When a now generic relatively tiny open-weight Chinese models is chàbuduō or 差不多, a Chinese concept that has become a national (and beyond) ethos; not perfect, but it’s sufficient; not METR leading, but it’ll do and it does, how much does the frontier even matter? To how many? Enough to fund a two trillion dollar Anthropic IPO? Bloomberg among others has its doubts and is not optimistic about the prospective fate of American models, the American AI industry, or the consequent economic impact.

For suppliers, this is the uncomfortable arithmetic of commoditization. As more models become adequate substitutes for a particular task, buyers acquire alternatives. Alternatives constrain prices. They also make it harder to justify paying a premium for capabilities a workflow never uses.

A company classifying routine documents gains little from the most expensive reasoning model. A production studio will get far more value from a specialized system adapted to its own footage. A business handling sensitive material may want local operation and remain completely ignorant of which service scores higher on a general benchmark.

Models still differ. Training, running and maintaining them still costs money. Commoditization does not mean every model becomes identical or every inference to becomes free. But it means sufficient competition over the tasks for which customers actually pay may be reduced to unsustainable levels for companies like OpenAI and Anthropic, who still lead, but in a way that is increasingly unimportant.

That threatens a particular American economic proposition: that a small number of laboratories can sustain exceptional returns by remaining the indispensable suppliers of useful AI.

The technology can succeed spectacularly while that proposition weakens. More AI use does not automatically translate into durable pricing power for the companies that first popularized it. The benefits can accrue to customers through lower costs, to specialists through better applications, or to infrastructure suppliers through greater demand. The advantage in the corporate world, where the real money lies, becomes, as it always does during commoditization, relationships and deal making, amid the emergence of some specialization advantages.

This is not uniformly bad news for American industry. A US manufacturer using an inexpensive model can benefit from the same competition that pressures a US model provider. The interests of a laboratory, its investors and the wider economy should not be treated as interchangeable. But it is potentially disastrous news for the heavily leveraged American AI ecosystem, with visibly weakening players (Oracle, CoreWeave) and those whose ascendancy seems still unstoppable (NVIDIA, AMD).

China’s open-weight ecosystem adds a geopolitical dimension. A model that developers can download, adapt and incorporate into their own products can establish influence through adoption. It can shape the tools people learn, the systems they build and the technical foundations they continue to use. That influence need no longer depend on subscriptions from the original developer. Being widely built upon is itself a strategically advantageous position, one Chinese models are increasing fully competitive on, if not dominant.

The Shifting Sovereignty Discussion is Utility-Based, Less Nationalist

For governments and industries in other countries, these models create choices. Developing domestic AI capability no longer needs to begin with financing a national competitor advanced to the furthest part of frontier, though many may continue to go this route. It can begin with the ability to operate available models, adapt them to local needs, evaluate their performance and replace them in whole or in part as necessary.

The distinction between where a model originated and who controls its deployment is essential. Downloading weights to domestic infrastructure creates a different dependency from sending requests and data to a foreign-hosted service. Neither arrangement erases questions about provenance, licensing or maintenance. They distribute control differently.

The Emerging Societal Role of AI May Not be How it Appears

This is also where AI’s objectivity and possibility of democratization and its relationship with industries like social media and the legal/judicial professions becomes significant. In a recent Vox essay, Eric Levitz explored whether AI could counter some of social media’s fragmentation of shared reality. His argument concerned the different incentives behind useful answers and engagement-driven feeds. AI holds no inherent positions, has no ego, does not attach identity to responses (while not unbiased). This kind of broader objectivity is significant in an increasingly polarized world. The possibility of independent model operation adds another dimension: users can gain alternatives to a single company’s or “side’s” positions, tools, algorithms and editorial choices.

A social platform derives much of its power from controlling a shared environment: the audience, the feed, the rules of visibility. A downloaded model can operate outside its creator’s service. It can be adapted for another language, another profession or another institution’s needs.

No single company can control that entire ecosystem. Individual services remain controllable, and concentration can persist in chips, cloud infrastructure and distribution. But the utility of and ability to exercise useful, egoless, pattern-based computational capability is spreading beyond the organizations that originally developed it. A self-represented defendant with limited means has much greater resources and knowledge at their disposal than they did in the pre-AI era, nearly regardless of model sophistication.

Commoditization vs Specialization

Distributed access also does not guarantee distributed profits. Ben Affleck’s InterPositive illustrates both the opportunity and the tension.

In Netflix’s announcement of its acquisition, Affleck described developing a proprietary production dataset and models focused on visual logic and editorial consistency. The problems included missing shots, background replacements and incorrect lighting. His account centres on the practical requirements of filmmaking and the judgment needed to preserve a creative intention.

Increasingly, the specialization advantage lies not within the models but with companies’ own proprietary data: what can be extracted from it, learned from it, predicted from it. There is no contradiction in AI becoming more economically useful while parts of the model business become less attractive.

Value also comes from understanding the work closely enough to identify where a model can help. A filmmaker knows when lack of continuity breaks a scene, which adjustment preserves a performance and which technical obstacle consumes resources without improving the film.

That is expertise directing a tool toward a valuable outcome. Netflix’s acquisition also shows how value created through specialist application can subsequently be captured by an established company.

This combination merits different attention than another abstract argument about whether AI will replace creativity. People who understand a field can use models to change how work gets done within it. Their accumulated knowledge provides the requirements, the standards and the means of judging the result.

The same logic, applied to software

AI-assisted development lowers the effort required to turn some ideas into working applications. That expands the range of problems worth solving. A tool for one organization, one recurring task or one awkward gap between existing systems may become economical even when it would never justify a conventional software project.

It also puts pressure on products whose advantage consists largely of functionality others can reproduce. As implementation becomes cheaper, businesses have to ask what remains difficult to copy: trusted relationships, proprietary data, dependable operation, distribution or a deep understanding of the customer’s work.

There is still a difference between reproducing a visible interface and maintaining a reliable product. Authentication, integrations, unusual cases, support and ongoing change remain work. The economic shift is that more of the initial building can be accelerated, making the decisions around it more consequential.

This helps explain the widening gap between having AI and knowing how to use it. Many organizations still approach AI as a conventional software purchase. Select a vendor. Assign licences. Expect productivity to follow. Yet the same underlying model can produce very different value depending on how the task is defined, what information it receives, which tools it can use and how its output is checked.

Someone has to understand the problem well enough to describe it, recognize when the answer is wrong, and decide how the result should enter a real process. As the landscape becomes more complicated, choosing and combining tools becomes a capability of its own.

This opens a substantial opportunity for experienced practitioners willing to learn. A researcher can develop a better way to examine evidence. A public servant can redesign a cumbersome service process. A small-business owner can build a system around the way the business actually operates. Domain knowledge gives each of them a way to identify useful work and assess whether the technology performs it.

Experience alone is insufficient but increasingly important; it has to integrate with both experimentation and the ability to evaluate unfamiliar tools. But abundant models give expertise a much wider range of applications.The world is becoming AI-enabled while the ability to benefit from that change remains highly uneven. Access is one barrier. Time, training, organizational permission and practical skill are others. Making models available not only does not automatically overcome them, it makes skill and fluency ever more important and distinctive.

The Opportunities Are The Threats

For governments and businesses, that changes the investment question. Buying access is the beginning. Building the human and institutional capacity to apply it is where much of the potential value will be determined, a question that increasingly needs to be answered with a highly and rapidly evolving threat assessment. Yes, a former employee and Claude may now be able to build your entire proprietary software stack in a couple of weeks, but can they support, sell and innovate with it? These become the points of distinction and demarcation.

The great fragmentation makes the original promise of these systems more concrete. They are tools that humans can use to extend what they are able to do. As the tools become more available, the quality of their use becomes a larger part of the outcome.

Framing the Meaningful Discourse

The model is becoming easier to obtain. Knowing what to do with it is becoming the advantage. The discussion is no longer: who’s the best? but, what is possible, and when is good enough, good enough? And the questions around danger, the existential questions, become not: is AI going to become conscious and destroy us, but do we sufficiently understand its capabilities and limitations robot into place limits around its use so we do not destroy ourselves. We’ve lived with existentially threatening technology for a long time. But never one that was so widely distributed. When cooling systems in grocery stores can destroy millions in inventory, how do we protect ourselves from ourselves? Can we? The opportunity and threat of the most lies most profoundly and most urgently, here.

Sources and editorial notes

● Hugging Face model catalogue: approximately 3.14 million repositories at the time checked.

● Meta: Introducing Muse Spark: multimodal reasoning model; distinct from genomic modelling.

● Arc Institute: Evo 2: genomic foundation model.

● Qwen local deployment example: downloadable derivative and local-running instructions.

● Google: Deploy Gemma on mobile devices: Android and iOS deployment documentation.

● Eric Levitz, Vox: The internet fractured reality. AI might put it back together: different information incentives; the distributed-control argument is this article’s extension.

● Netflix: Why InterPositive Is Joining Netflix: Affleck’s first-person account of specialized production models and proprietary data. The draft does not assert an unverified open-weight lineage.