Owning the layer below

Three moves this week, one logic: control of the layer beneath you.

On Wednesday, OpenAI took delivery of the first samples of Jalapeño, its first custom chip, designed end to end in nine months with Broadcom. The headline number, per Broadcom CEO Hock Tan, is roughly 50% lower cost per inference than the AI GPUs everyone else rents. Read that again. The most capital-intensive company in private technology, the one projecting $14 billion in losses this year, just decided the cheapest path forward runs through silicon it controls. That is the move of the week, and it is not isolated. The same seven days saw the US government keep two of Anthropic's most capable models switched off by export-control order, and a fresh survey confirm that 97% of enterprise developers now use AI coding tools while only 30% govern them. The pattern underneath all three: everyone is racing to control the layer beneath them, and the layer most enterprises have neglected is the one that decides whether their AI still runs in eighteen months.

OpenAI Builds Its Own Chip, and Redraws the Compute Map

Early samples are already running production workloads in the lab, including a GPT-5.3 coding model. Initial deployment is targeted for late 2026 at gigawatt scale, with Microsoft as the primary deployment partner. Why it matters: inference, not training, is the recurring cost line on the income statement. Training is a capital event. Inference is the meter that runs every time a customer sends a prompt.

For a company staring at $14 billion in losses, halving the cost of serving every token is the most direct lever available on unit economics, and Hock Tan put the savings at roughly 50% versus typical AI GPUs in an interview with Bloomberg.

The labs that can afford custom silicon are quietly pulling the floor of the cost curve down beneath everyone still paying market rates for Nvidia hardware.

With Jalapeño, OpenAI now spans the full stack: model, kernels, serving systems, and the chip underneath. That is the same vertical-integration play Google ran with TPUs and Amazon ran with Trainium and Inferentia. For enterprise leaders, the consequence is structural. If the largest model providers serve their own models on their own chips at half the cost, the price you negotiate for AI is increasingly set by infrastructure you cannot see and cannot buy.

And the economics are not as clean as the announcement suggests: reports indicate Broadcom required Microsoft to guarantee purchase of 40% of the first production run to anchor the line. That is concentration stacked on concentration, a custom chip whose viability already leans on a single buyer.

Sources: OpenAI · CNBC · Bloomberg · Tom's Hardware

Black Duck's June survey of 831 enterprise developers found 97% now use AI coding tools, but only 30% have full governance over them. Governed teams were 55% more likely to report a major improvement in efficiency, and 90% of all teams reported problems with AI-generated code. The capability arrived everywhere. The controls arrived almost nowhere, and the data says the controls are where the return actually is.

Source: Black Duck

The portability you are told to protect may be the thing the chip takes away

This edition argues for owning your cost curve and staying portable. Here is the complication. The 50% figure is a Broadcom executive's claim on early lab samples, not measured production performance, and Jalapeño does not deploy until late 2026. More important, custom inference ASICs earn their savings precisely by being tuned to a narrow set of models and serving patterns. The cheaper the silicon, the more specialized it tends to be, and the more specialized it is, the less it resembles the open, swappable substrate that portability depends on.

Google's TPUs and Amazon's accelerators followed the same logic: the cost advantage lives inside infrastructure you cannot buy and cannot port to. So the enterprise advice splits in two. Portability at the application layer is still worth protecting. But the real cost advantage is migrating into silicon that is, by design, the opposite of portable. You can keep the freedom to switch providers. What you are quietly losing is the freedom to match their economics.

Sources: Tom's Hardware · Bloomberg

Every story this week is about who controls the layer below.

OpenAI reached down into silicon, because the layer it could not control, the cost of serving its own models, was the one bleeding it out. The US government reached into a deployed model and flipped it off, demonstrating a power most contracts assumed no one had. And enterprises reached for AI coding tools at near-total adoption while seven in ten reached right past the governance that the same survey says doubles the return.

In the field, the projects that survive contact with production are not the ones with the best model. They are the ones where someone owns the unglamorous layer underneath: the cost-per-token model, the fallback path when a vendor disappears, the review process for code a machine wrote. The exciting layer is always the model. The layer that decides whether you are still running in eighteen months is the one beneath it.

So here is the practical takeaway from a week of vertical-integration headlines. Build for the day your preferred model gets switched off, whether by a regulator, a price change, or a provider's roadmap. June stopped that from being hypothetical.

The teams that own the freedom to swap any layer of their stack are the ones still standing when the layer below them moves. And it will move.

Let's get to production,
AK