Every ChatGPT answer has an electricity bill.

So does every Codex action. A long-running agent can repeat that cost hundreds of times before finishing one job.

On August 25, OpenAI published the first performance results from Jalapeno, its custom AI chip.

Jalapeno tells us which problem OpenAI wants to control: the cost of serving AI.

It is an inference chip. It runs trained models and produces answers. OpenAI will still use Nvidia and other partners for training and inference.

I am Alex. Welcome to ShortCu8 by Innov8.

Lets Dive Deep🐰🐰.

Read an AI company through three layers:

Physical layer: power, data centres, chips and networks.

Intelligence layer: models and the software used to serve them.

Distribution layer: apps, devices and access to users.

A model can lose its benchmark lead in a few months.

Control over these three layers is harder to copy.

OpenAI already controls its models, API, ChatGPT and Codex.

Jalapeno takes it deeper into the physical layer.

OpenAI designed the chip architecture. Broadcom handled the silicon implementation and networking. Celestica worked on the boards, racks and system integration.

So this is not a fully independent OpenAI factory. It is selective ownership.

OpenAI still uses chips and infrastructure from Nvidia, Microsoft, AWS, AMD, CoreWeave, Oracle and other partners. It is building where tighter control can improve its costs, while continuing to buy from the wider market.

In OpenAI's own tests, Jalapeno completed 1.5 to 1.9 times more AI work per watt at peak throughput and delivered 1.7 to 3.6 times lower end-to-end latency than the comparison systems.

Those are company-reported results, and the chip still has to prove itself at production scale. OpenAI plans to begin deploying it inside its infrastructure by the end of 2026.

Still, the reason for building it is easy to understand.

When an agent performs 50 steps, a delay gets repeated 50 times. Faster and cheaper inference can turn an expensive demo into a product people can use every day.

xAI does not have a custom chip like Jalapeno.

Its infrastructure advantage is Colossus, a giant cluster built with Nvidia GPUs.

xAI says it built the original system in 122 days, then doubled it to 200,000 H100 GPUs in another 92 days.

Its stack looks roughly like this:

Colossus -> Grok -> X

Colossus supplies the compute. Grok supplies the intelligence. X gives it built-in distribution.

This helps explain how a younger lab caught up so quickly. xAI built a huge physical base and connected it directly to an existing social platform.

OpenAI's approach is different, but the direction is similar. Both companies want more control over the machinery underneath their models.

Google has spent years building TPUs, data centres and networking alongside its models and products.

Its new TPU 8i is designed for inference and agent workloads. TPU 8t handles large training jobs. Gemini then reaches users through Google Cloud, Search, Workspace, Android and other Google products.

Meta also says it has deployed hundreds of thousands of its own MTIA chips. Those chips run inference across its apps, while Meta owns the models, Facebook, Instagram, WhatsApp and its growing device business.

OpenAI is trying to become the same kind of company: one that can improve the model, the hardware running it and the product using it at the same time.

Vertical integration can make AI faster and cheaper.

If OpenAI spends less power and money on each successful agent task, it could lower prices, offer larger limits or run more capable models without making every request painfully expensive.

It also changes how we should judge AI companies.

Do not look only at this month's benchmark.

Ask:

A lab with a brilliant model still needs someone else's chips, power and customers.

A vertically integrated company can use profit from one layer to support another. It can subsidise an AI product, reserve compute for itself and tune the whole system around its own workloads.

That is a stronger advantage than winning one benchmark.

Now go and build something great ♥

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Most e-commerce sellers are running their store across 6 to 10 separate tools — and spending more time managing software than growing their business. StoreClaw replaces your entire stack with one autonomous AI engine that monitors competitors, optimizes listings, automates marketing, and tracks real profit across Shopify, Amazon, and beyond.

It doesn't wait for you to ask. It runs 24/7 in the background, so you wake up to a full dashboard instead of a list of things you forgot to check.

Connect your store, and StoreClaw gets to work — no prompts, no complex setup, no six-app stack.

Free to start. No credit card required.

Try it for FREE today

#AD2

AI spend is now a major P&L line item—but most teams can't show what it's producing.

Harmonic Security maps AI activity to use cases and teams, revealing real productivity, shelfware, data risk, and adoption trends across approved and unapproved tools.

Give your board the data behind the return.

See Your AI ROI