The battle for the AI stack.

The next AI leader may be the company that makes the whole system feel effortless. Over the past two years, the contest has expanded from model quality to the layers that turn intelligence into useful work. Our hypothesis: OpenAI’s most durable advantage could be a premium integrated experience.

In November 2022, OpenAI joined two things that had rarely worked so well together: a capable language model and an interaction almost anyone could understand. Write a question. Read the answer. Ask again. ChatGPT was fine-tuned from GPT‑3.5 for dialogue; its training and its interface worked together. The innovation was technical, but also ergonomic.

According to OpenAI’s retrospective, one million people used ChatGPT within five days. That number records adoption. The cultural shock, in our reading, came from the gesture: interacting with a language model suddenly resembled something humans already knew how to do. The name itself carried the instruction. Chat.

The contest of 2024–2026 follows from that opening. Once several companies can offer a good model in a chat window, where does the next advantage come from? Our answer begins below the screen, and ends above it.

Electricity powers compute. Compute runs models. Models supply tokens to a harness: the software that manages context, tools and execution. An interface gives an operator controls. The operator directs work towards an artefact: a contract, a spreadsheet, a programme or an image. These are functions, not seven sealed compartments. Products overlap them, and feedback travels both ways.

The strategic question is therefore larger than “who has the best LLM?” It is: which dependency can a company control, which relationship can it own, and at which layer will someone keep paying it?

The digital experience rests on physical capacity: power, land, grid connections, cooling and machines. In April 2026, the IEA projected global data centre electricity demand rising from around 485 TWh in 2025 to 950 TWh in 2030. This includes other computing workloads; it is not a measurement of LLM consumption alone. It is also a projection, not an accomplished outcome.

Our strategic reading is simple: a provider promising fast, always-available intelligence eventually has to care about what happens far below its API. A better model cannot compensate for a site that cannot obtain power. Moving down the stack is partly a way to secure the conditions under which the product can exist.

In October 2025, OpenAI and Broadcom announced a programme for 10 GW of custom accelerators and systems, with deployment planned from the second half of 2026 through 2029. In June 2026, they unveiled Jalapeño, OpenAI’s first custom inference accelerator. At the announcement, engineering samples were running in the lab; initial deployment was planned for the end of 2026. These milestones establish a direction, not a completed industrial rollout.

Designing a chip around the models, memory traffic and serving patterns can connect hardware choices to product latency. OpenAI does not become a semiconductor manufacturer by doing this: Broadcom and other partners industrialise the platform. But the frontier lab is now helping design the floor on which it stands.

NVIDIA moves in the other direction. CUDA already gives its hardware a software ecosystem. Dynamo, introduced in 2025, organises inference serving; the Nemotron Coalition, announced in March 2026, brings partners together to develop open models. Our interpretation: making models easier to obtain and deploy can strengthen demand for the infrastructure beneath them. NVIDIA and OpenAI are therefore partners at one level and potential competitors at another.

DeepSeek R1, released in January 2025, and Qwen3, released that April, made strong models available for independent deployment. Qwen documents several serving engines and local applications. The model can travel beyond its original chat interface. Here “open weights” is the useful term: publishing parameters under a permissive licence does not necessarily publish the entire training process and its data.

This competition moves back and forth. Stanford’s AI Index 2026 records the Arena gap between leading closed and open-weight models narrowing sharply in 2024, then widening with new proprietary releases. In March 2026, it reports 1,503 points for Claude Opus 4.6 and 1,454 for GLM‑5. Arena measures preferences in a particular evaluation setting. Those 49 points do not measure the quality of an entire professional workflow.

Price pressure is also real. Stanford’s AI Index 2025 reports that the API price of reaching a fixed GPT‑3.5-level MMLU threshold fell from $20 to $0.07 per million tokens between November 2022 and October 2024. This is a benchmark-specific comparison of inference prices, not the price of a ChatGPT subscription or proof of equal agent reliability. A capability can become cheaper while the frontier moves.

The strategic effect matters more than a snapshot ranking. A credible substitute gives application builders another negotiating position, another deployment option and another cost structure. It does not guarantee that replacing the model will preserve tool use, context handling or reliability. The next layer has work to do.

A language model can suggest an action. Something else must execute it, return the result, retain the right context and decide how to continue. That surrounding system is the harness. Instructions, tool definitions, permissions, state and recovery become part of the product. Two products using the same underlying model can therefore produce very different work.

OpenAI documents one shared Codex harness behind its web, CLI, IDE and desktop experiences, connected through an App Server. Anthropic’s account of long-running agents shows why a simple loop needs environment setup, progress records and tests to stay effective over time. Reproducing a chat transcript or an API request reproduces only part of this machinery.

Openness also appears here. Anthropic launched MCP in November 2024 to standardise connections to tools and data, including in Claude Desktop. Parts of the Codex harness are open source. The opposition between closed frontier labs and an open ecosystem is therefore an incomplete map: the same actor can defend a proprietary model while opening the connections around it.

In our view, looking so closely at models can obscure the precision of the ChatGPT and Codex desktop applications. ChatGPT made dialogue intuitive; Codex extends the question to sustained work. Its February 2026 app brought parallel agents, persistent projects, worktrees and diff review into one surface. The value is in the transitions: seeing what is happening, finding the right work, inspecting a change and continuing without rebuilding the context.

For us, the beauty of these applications lives in coordinated behaviour as much as in pixels: a model, an execution environment and controls that fit together. Someone may reverse-engineer an individual feature. Maintaining the coherence of hundreds of interactions while models and tools change is a continuing engineering task.

The difficulty remains visible. Which model, how much reasoning, which documents, which connector? GPT‑5 introduced automatic routing between immediate and deeper responses; OpenAI’s 2026 memory work moves some context management into the product. We read these changes as attempts to absorb configuration. They do not establish that the need for user judgement has disappeared.

Moving up the stack means approaching the work that someone actually buys. In June 2026, OpenAI introduced role-specific Codex plugins, including investment banking. In September, Astra for Law combined a model, a legal search index, specialist instructions and ecosystem plugins. Its initial research corpus focuses on US law. The vertical offering is already a configuration of several layers, not simply a model with a professional label.

It is too early to declare these initiatives a success or a failure. OpenAI’s enterprise AI report itself describes a wide gap between advanced and median users, alongside organisational barriers. Knowing how to generate a legal argument is different from knowing which authorities govern a Swiss mandate, which documents may be accessed and what the partner expects to sign. A role plugin can supply a starting point. A profession also has local standards, exceptions and accountability.

This is where we place ogram’s contribution: helping the human operator configure the environment, connect the relevant material, encode professional standards and assess the result. The user does not need to become a computer scientist. The system still needs someone to make the right expertise operational.

At the top, the artefact settles the question. A spreadsheet must calculate correctly. A contract must fit the mandate. An image must meet the brief. Our inference is that a cheap token can produce an expensive error, while a more expensive system can be economical if it reduces review and rework. The relevant comparison is the total cost of an acceptable result.

Anthropic shares much of this model–harness–application ambition, with its own emphasis on Claude, Claude Code and connections through MCP. xAI connects Grok to compute infrastructure and distribution through X. Its October 2026 subscription announcement bundles Grok, Cursor and X, while retaining third-party models in Cursor under separate billing. Integration and modularity can coexist within the same commercial offer.

Google starts from a different position. Its TPUs give it custom compute; Gemini provides models; Search and Workspace offer established distribution. The October 2026 Gemini at Work announcement puts agents inside existing work surfaces. Google can bring intelligence into the place where the document, the email and the calendar already live. It is competing for a different route to the user.

The pieces remain separable. A provider can change its accelerators. An application builder can change its model. A team can adopt another harness or another interface. Anthropic’s Managed Agents architecture explicitly separates session, harness and sandbox so their implementations can evolve independently. These boundaries are real engineering choices.

Separable does not mean interchangeable at no cost. Hardware needs compatible runtimes and kernels. A different model may need different prompts, tool schemas or context management. A new interface changes habits and permissions. Nor does a modular market give a ChatGPT user the right to insert any LLM into the proprietary app. Provider architecture, developer choice and end-user controls are three different things.

A component can become replaceable while the quality of the assembly becomes more valuable.

We think OpenAI has a credible path to becoming a reference for people who want advanced capability, responsiveness and carefully designed interaction, with price as a secondary concern. An AI of luxury, if the phrase means paying for the quality of the whole experience. The advantage would live across the model, the harness, the serving infrastructure and the application.

The Apple analogy is useful at this level: coordinated choices can spare the user a succession of technical decisions. It remains an analogy. It does not establish an equivalent market structure, permanent leadership or superiority on every dimension. Google, Anthropic and modular builders can also produce excellent integrated experiences.

“OpenAI will never compete on cheap” would go too far. It launched ChatGPT Pro at $200 a month in 2024, but also announced ChatGPT Go worldwide at $8 a month in the US in January 2026, alongside free access, and released open-weight gpt-oss models in 2025. Its custom hardware programme also aims to lower costs. The stronger thesis concerns where its differentiation could be most defensible, not which price tiers it is allowed to serve.

This hypothesis is testable. It weakens if modular alternatives consistently match the whole experience at a lower total cost, if professional workflows remain too difficult to configure, or if another integrated provider does the job better. It strengthens if users willingly keep paying for continuity, speed and less operational friction as individual components become easier to replace.

That is the dance visible across the stack: move down to secure capacity, move up to reach the work, open some boundaries to encourage adoption, integrate others to make the product coherent. The winner will be judged at the top. What artefact arrived, under which constraints, and how much effort did it take the human to get there?

Sources checked on 23 October 2026. Product announcements establish availability, architecture and stated intentions; they do not independently establish commercial success or universal superiority. Strategic interpretations and the premium-experience forecast are the author’s hypotheses.