The Great Unbundling: Why AI Is Destined to Disappear Into the Infrastructure

The most important thing to understand about artificial intelligence right now is that it is not a product. It never was. And the companies that fail to internalize this distinction will get eaten alive.

This argument sits at the heart of a provocative new essay from John Gruber's Daring Fireball, published Monday with the deceptively simple title "AI is a technology not a product" [1]. The piece arrives as the industry convulses with product-level reorganizations, enterprise adoption hits an inflection point, and the fundamental architecture of AI delivery to end users rewrites itself in real time. Gruber argues that the current wave of AI hype has systematically confused a profound technological capability with a discrete commercial offering—and that confusion now creates dangerous strategic blind spots across the entire tech landscape.

The timing is impeccable. Just days before Gruber's essay appeared, OpenAI announced yet another executive shakeup, with co-founder Greg Brockman formally taking control of product strategy in a reorganization designed to unify ChatGPT and the company's programming product Codex into a single core experience [3][4]. Meanwhile, a startup called Empromptu AI launched a platform that lets enterprises train custom models directly from their production workflows—no machine learning team required [2]. These three data points, taken together, tell a story far more interesting than any single announcement.

What we are witnessing is the beginning of the great unbundling of AI. The technology is prying itself apart from the products that currently contain it, and the implications for incumbents, startups, and enterprise buyers are staggering.

The Product Trap

Gruber's central insight is deceptively simple. "AI is a technology not a product," he writes, drawing a distinction that sounds obvious until you realize how thoroughly the industry has ignored it [1]. Technologies, he argues, are foundational capabilities that enable products. Electricity is a technology; the light bulb is a product. The internal combustion engine is a technology; the automobile is a product. TCP/IP is a technology; the web browser is a product.

The confusion arises because AI, unlike previous foundational technologies, arrived wearing a product disguise. ChatGPT looked and felt like a product. You opened a web page, typed a query, and got an answer. It had a user interface, a subscription tier, a brand. It was easy to mistake for a finished good. But ChatGPT actually represented a technology demo of unprecedented scale and sophistication—a glimpse of a capability that would eventually absorb into every other product on the planet.

This is not a semantic quibble. The distinction between technology and product carries profound strategic consequences. If you believe AI is a product, you build moats around the interface. You optimize for user retention within a single application. You measure success by monthly active users and subscription revenue. If you believe AI is a technology, you build for embeddability. You optimize for latency, cost-per-inference, and API reliability. You measure success by how many other products depend on your infrastructure.

The evidence increasingly favors the technology framing. OpenAI's reorganization, in which Brockman must unify ChatGPT and Codex, suggests that even the company that created the most recognizable AI "product" in history struggles with the product-technology tension [3][4]. Codex, which powers GitHub Copilot and other developer tools, is fundamentally an infrastructure play—an API that other products call. ChatGPT is a consumer application. Combining them into a single product experience acknowledges that the boundary between the two is artificial and unsustainable.

The Workflow as Training Data

If the product-technology distinction seems abstract, Empromptu AI's new Alchemy Models platform makes it brutally concrete. The company launched on May 14 with a proposition that fundamentally reimagines how enterprises should think about AI model development [2].

Here is the key insight: every query an enterprise AI application processes, every correction a subject matter expert makes to its output—that interaction is training data. Most organizations are not capturing it. The production workflows companies have already built generate a continuous signal that improves AI models, and that signal is disappearing [2].

Empromptu's platform captures that signal and uses it to train custom models without requiring a dedicated machine learning team. The company's pitch: enterprises already possess the most valuable asset in AI—their own operational data—but they lack the infrastructure to convert that data into model improvements. Alchemy Models aims to close that gap by treating production workflows as continuous training pipelines.

This directly assaults the product-centric view of AI. If you believe AI is a product, you buy a subscription to a model provider and treat the model as a black box. You optimize for prompt engineering and hope the model's capabilities align with your use case. If you believe AI is a technology, you build your own models from your own data, because you understand that the technology's value lies in its ability to customize and embed into specific operational contexts.

The numbers support the technology framing. Empromptu claims that 87% of organizations are not capturing the training signal from their production workflows [2]. That means the vast majority of enterprises use AI as a product—consuming it through APIs and chat interfaces—while leaving their most valuable data asset on the table. The companies that figure out how to capture that signal will have a structural advantage that no amount of prompt engineering can overcome.

The Brockman Reorganization and the Identity Crisis at OpenAI

The OpenAI reorganization provides a fascinating case study in the product-technology tension playing out at the highest levels of the industry. Greg Brockman, who co-founded the company with Sam Altman and Ilya Sutskever, has taken charge of product strategy in what both TechCrunch and Wired describe as the latest in a series of executive shakeups [3][4].

The specific mandate: unify ChatGPT and Codex into one core product experience [4]. On the surface, this sounds like a straightforward integration play. ChatGPT is the consumer face of OpenAI's language models; Codex is the developer-facing API for code generation. Combining them makes operational sense—why maintain two separate product teams for what is essentially the same underlying technology?

But the deeper story is more interesting. The unification effort implicitly admits that OpenAI's product strategy has pulled in two incompatible directions. ChatGPT is a product in the traditional sense: it has a user interface, a brand, a subscription model. Codex is a technology: an API that other companies embed into their own products. Trying to merge them into a single experience requires reconciling two fundamentally different business models, user expectations, and competitive dynamics.

This is not just an organizational challenge. It is an existential one. OpenAI's valuation and market position depend on the belief that AI is a product—that the company can build a sustainable business around a proprietary interface that users pay to access. But the technology's trajectory suggests otherwise. As models become cheaper to run, easier to customize, and more widely available through open-source alternatives, the value shifts from the model itself to the data and workflows that surround it.

Brockman's reorganization can be read as an attempt to hedge this bet. By unifying ChatGPT and Codex, OpenAI is trying to create a platform that spans both the product and technology layers—a single infrastructure that supports both consumer applications and developer integrations. Whether this is strategically coherent or a recipe for internal conflict will depend on execution, but the very fact that it is necessary tells us something important about the industry's direction.

The Enterprise Reality Check

The enterprise market is where the product-technology distinction becomes most consequential. For the past two years, enterprises have been buying AI products—subscriptions to ChatGPT Enterprise, licenses for GitHub Copilot, access to various model APIs. These purchases have been driven by a combination of genuine productivity gains and FOMO, with budgets allocated to "AI initiatives" that often lack clear ROI metrics.

But the enterprise buying pattern is shifting. The Empromptu launch signals a growing recognition that the real value of AI lies not in consuming models but in building them. The production workflows companies have already built generate a continuous signal that improves AI models [2]. The companies that capture that signal will have models uniquely tailored to their operations, their data, their customers.

This is a fundamentally different value proposition than buying a subscription to a general-purpose model. A general-purpose model is a commodity. It can answer questions, generate text, and write code, but it cannot optimize your supply chain, predict your customer churn, or automate your specific regulatory compliance workflows—at least not without extensive customization. The companies that invest in capturing their own training signal will have models that can do those things, because the models will train on the data that matters most to their business.

The implications for the current AI vendor landscape are stark. Companies that sell AI as a product—whose primary value proposition is access to a model—face a future in which their offering becomes increasingly commoditized. Open-source models are closing the gap with proprietary ones. Inference costs are plummeting. The barriers to entry are falling. The moat is not the model; the moat is the data and the workflow integration.

This is why Empromptu's pitch is so strategically significant. The company is not selling a model. It is selling the infrastructure to build models from your own data. That is a technology play, not a product play. And it aligns with the direction the industry is heading.

The Infrastructure Layer and the Invisible AI

If Gruber is right that AI is a technology, not a product, then the most successful AI companies of the next decade will be the ones that make AI invisible. They will embed AI into existing products and workflows so seamlessly that users never think about the AI at all.

This is already happening in specific domains. Google Search uses AI to understand queries and rank results, but users do not think of Google Netflix uses AI to recommend content, but users do not think of Netflix that way either. The AI layer is invisible because it is working.

The same pattern will play out across the enterprise. Companies will not buy "AI products." They will buy supply chain optimization software that happens to use AI. They will buy customer service platforms that happen to use AI. They will buy developer tools that happen to use AI. The AI will be a feature, not the product.

This is the future that Empromptu is betting on. By enabling enterprises to train custom models from their production workflows, the company positions itself as infrastructure for the invisible AI era. The models will train on operational data, embed into operational systems, and optimize for operational outcomes. Users will never interact with the models directly. They will just see better results.

The OpenAI reorganization suggests that even the most prominent AI company recognizes this trajectory. By unifying ChatGPT and Codex, Brockman is trying to create a platform that can serve both the visible AI market (consumer chat interfaces) and the invisible AI market (developer APIs and embedded models). The question is whether one company can successfully play both roles, or whether the product-technology tension will eventually force a split.

The Hidden Risk the Mainstream Media Is Missing

The mainstream coverage of AI has focused on a narrow set of narratives: the race between OpenAI and Google, the potential for AGI, the regulatory debates in Brussels and Washington. These stories matter, but they miss the deeper structural shift underway.

The real story is that AI is following the same trajectory as every other foundational technology. It starts as a standalone product because that is the easiest way to demonstrate the capability. It gets absorbed into infrastructure because that is where the value ultimately resides. The companies that understand this trajectory will invest in data capture, workflow integration, and model customization. The companies that do not will find themselves selling a commodity in a market where the margins are disappearing.

The Empromptu launch signals that the infrastructure layer is being built. The OpenAI reorganization signals that the product layer is being rethought. The Gruber essay signals that the strategic framing is being clarified. These three signals, taken together, point to a future in which AI is everywhere and nowhere—a technology so thoroughly embedded into the fabric of computing that it ceases to be a distinct category.

The winners in this future will not be the companies with the best models. They will be the companies with the best data, the best workflows, and the best integration. The models will be table stakes. The moat will be everything else.

This is the argument that Gruber makes, and it is the argument that the industry is only beginning to understand. AI is a technology, not a product. The sooner everyone internalizes that distinction, the sooner we can stop pretending that chat interfaces are the end state and start building the infrastructure that will make AI truly transformative.

The unbundling has begun. The question is not whether AI will disappear into the infrastructure. It is which companies will build the infrastructure that makes the disappearance possible.

---

References

[1] Editorial_board — Original article — https://daringfireball.net/2026/05/ai_is_technology_not_a_product

[2] VentureBeat — Enterprises can now train custom AI models from production workflows — no ML team required — https://venturebeat.com/data/enterprises-can-now-train-custom-ai-models-from-production-workflows-no-ml-team-required

[3] TechCrunch — OpenAI co-founder Greg Brockman takes charge of product strategy — https://techcrunch.com/2026/05/16/openai-co-founder-greg-brockman-reportedly-takes-charge-of-product-strategy/

[4] Wired — Greg Brockman Officially Takes Control of OpenAI’s Products in Latest Shake-Up — https://www.wired.com/story/openai-reorg-greg-brockman-product/