OpenAI Jalapeño Chip: Custom AI Silicon Explained

Samantha: Welcome to The State of Tech, The European Edition, Wednesday July one, 2026. I'm Samantha Lawrence.

Bob: And I'm Bob Russell. Today: Google rebuilds search around AI-generated summaries, Ay-Double-You-Es pours four billion dollars into public sector AI, OpenAI unveils its first custom chip called Jalapeño, Taiwanese executives arrested over a twenty-two million dollar Nvidia chip smuggling ring, NASA sends a robot to rescue an ageing space telescope, and to close out, two things you can actually use today: the writing app Ghostwriter and the mapping tool OsmAnd. Let's start with Google.

Google rebuilds its search engine around AI-generated summaries with Gemini 3.5 Flash.

Samantha: Google is calling this its biggest search overhaul in more than twenty-five years. Starting today, the results page you have known since the late nineties is being replaced by AI-generated summary pages, powered by a new model called Gemini 3.5 Flash. Instead of ten blue links, you get a custom-written page that pulls the answer together for you.

Bob: And that is a genuine shift in how the web works. For decades, Google sent you outward, to newspapers, to Wikipedia, to small blogs. Now it increasingly answers on its own page. The publishers who depend on those clicks are watching this very nervously, because if the summary is good enough, most people never scroll to the source. The traffic that funds a lot of journalism and independent websites goes through Google, and Google is quietly changing the deal.

Samantha: There is also a quality question. AI summaries have famously told people to put glue on pizza or eat rocks. Google says Gemini 3.5 Flash is a big step up, but the model is now being asked to write the definitive answer to billions of queries a day. Any error scales instantly. And unlike a bad link, a bad summary looks authoritative because it sits under the Google logo.

Bob: For European users, this lands in a specific regulatory context. The Digital Markets Act already forces Google to treat rival services fairly in search, and the AI Act has transparency rules about AI-generated content. Brussels will want to know how these summaries are labelled, how sources are credited, and whether European publishers can opt out without losing visibility altogether. Expect complaints from news organisations within weeks. The bottom line is that the front door to the internet just changed shape, and most people will not have been asked.

Amazon Web Services commits four billion dollars to bring AI into the US public sector.

Samantha: Ay-Double-You-Es, the cloud arm of Amazon, is putting four billion dollars into deploying artificial intelligence across the US public sector. Half of that is aimed squarely at the intelligence community. There is a one billion dollar incentive programme to move US spy agencies onto Ay-Double-You-Es classified cloud infrastructure, and another billion for what the company calls Forward Deployed Engineering, essentially embedding thousands of Amazon's AI engineers directly inside customer offices.

Bob: That embed model is worth pausing on. It borrows a page from Palantir, whose engineers famously sit inside client organisations and build custom systems in real time. Amazon is now applying the same playbook, but at scale, and with generative AI as the product. The pitch to government agencies is clear: you do not need to figure out how to use AI, we will bring the engineers to your desk and build it with you.

Samantha: There is also a specific use case being highlighted, tackling antimicrobial resistance with AI, alongside the more predictable defence and intelligence work. So this is not only about spy agencies, it is about positioning Ay-Double-You-Es as the default AI partner for anything a government wants to do.

Bob: For Europe, this sets a benchmark that is hard to match. European cloud providers have been arguing for years that critical government workloads should stay on European infrastructure, both for sovereignty and for compliance with data-protection law. When Ay-Double-You-Es puts four billion on the table and offers to lend you its engineers, that argument gets harder to win on price alone. Expect fresh debate in Paris, Berlin and Brussels about how much European governments should rely on American hyperscalers for their most sensitive AI work.

Samantha: Quick interruption. If you listen to The State of Tech regularly, hit that like button and subscribe, that way you'll never miss an episode. Okay, moving on.

OpenAI reveals its first custom AI chip, called Jalapeño, and details GPT-5.

Bob: OpenAI has just unveiled its first in-house AI chip. It is called Jalapeño, and it is designed specifically for inference, that is, the part where a trained model actually answers your questions, as opposed to the training phase where it learns. Alongside that, the company shared new details on its upcoming GPT-5 model, including a score of just under fifty-eight percent on a benchmark called APB Pro.

Samantha: The chip is the bigger story here. Right now, OpenAI runs on Nvidia hardware, like almost everyone else, and pays a steep premium for it. Building your own inference silicon means you can tune the chip to exactly your model, cut power consumption, and reduce your dependence on a single supplier. Google has been doing this for years with its own chips, Amazon has its own line, and now OpenAI joins that club. It is a sign the biggest AI companies no longer trust the open chip market to serve them cheaply enough.

Bob: The name Jalapeño is a wink, but the strategy is serious. Every query to ChatGPT costs money in electricity and chip time. If OpenAI can shave even a fraction off that cost per query, at the scale of hundreds of millions of users, the savings run into billions per year. It also gives the company more control over supply, at a moment when Nvidia chips are the most contested commodity in tech.

Samantha: For European developers and businesses, the practical effect is that OpenAI's services should get faster and cheaper over time, which is good news for anyone building on the API. The concentration question gets sharper though. A single American company now designs its own model, its own chip, and increasingly runs its own infrastructure. European regulators looking at AI market power will note that the stack is getting more vertical, not less.

Taiwanese executives arrested in twenty-two million dollar Nvidia AI chip smuggling case to China.

Bob: Prosecutors in Keelung, in northern Taiwan, have detained three technology executives over the alleged smuggling of advanced Nvidia AI chips into China, Hong Kong and Macau. Among those detained is a vice president at Albatron Technology and managers from the Taiwan branch of Super Micro Computer, better known as Supermicro. The value of the smuggled shipments is estimated at around twenty-two million US dollars, and the arrests followed raids on twelve separate sites.

Samantha: This is one of the biggest enforcement actions we have seen since the United States tightened its export rules on advanced AI chips. Those rules are designed to keep the most powerful Nvidia processors out of Chinese hands, on the argument that they can be used for military AI and mass surveillance. But the reality is that a black market has emerged, with chips being routed through third parties and shell companies to end up in Chinese data centres.

Bob: The involvement of executives from established firms is significant. This is not a couple of resellers with a suitcase, it suggests smuggling networks operating through legitimate corporate structures, which is exactly what Washington was worried about. Supermicro has said it is cooperating with authorities.

Samantha: Europe is watching this closely because European exporters face similar dilemmas with dual-use technology. The Netherlands, home to lithography giant ASML, has been at the sharp end of chip export controls for years. This Taiwan case will likely feed into ongoing discussions in Brussels about how tightly to police AI chip and equipment exports, and how to prevent European companies from becoming the next weak link in the enforcement chain. Expect fresh guidance from Dutch and German authorities in the coming months.

NASA launches a robotic rescue mission to save the ageing Swift space telescope.

Bob: NASA has begun a genuinely unusual mission. It is sending a robot into orbit to physically rescue the Neil Gehrels Swift Observatory, a space telescope that has been hunting gamma-ray bursts for over twenty years and is now slowly falling out of orbit. The robot is called Link, built by a US startup called Katalyst, and it launched aboard a Pegasus XL rocket that itself is dropped from a modified jet at altitude.

Samantha: The plan is genuinely science fiction made real. Once in orbit, Link has to find the Swift telescope, manoeuvre alongside it, and grab hold using three movable arms. Then it will tow the telescope up to a higher, stable orbit where it can keep working. The whole mission costs around thirty million dollars, aimed at saving a telescope worth about two hundred and fifty million. It was originally scheduled for June thirtieth but got pushed to today because of weather at the launch site.

Bob: What makes this important beyond the specific telescope is the precedent. Until now, satellites and telescopes have been essentially disposable, once they run out of fuel or drift out of position, they burn up. If robotic servicing works reliably, that changes the economics of space. You can extend missions by years, avoid the cost of replacement, and start thinking about repair rather than replacement.

Samantha: European space agencies have their own interest in this. Ee-Es-Ay, the European Space Agency, has been developing its own orbital servicing programmes, including missions to remove space debris. A successful Link operation gives everyone a working template. And with tens of thousands of satellites now in orbit, most of them commercial, the market for on-orbit servicing could become a real industry. European startups in Germany and Luxembourg are already positioning for it. Today's mission is the proof of concept the sector has been waiting for.

Ghostwriter turns any device into a calm, distraction-free writing space with markdown built in.

Bob: To close out, two things you can actually use today. First up, Ghostwriter. It is a free, open-source writing app for anyone who wants a clean space to just get words down. No formatting toolbars in your face, no autosaving to some corporate cloud, just a page and your text.

Samantha: Ghostwriter is built around markdown, which is a simple way of formatting text using little symbols instead of buttons. Two asterisks around a word makes it bold, a hash sign in front of a line makes it a heading, and so on. Once you get used to it, it is genuinely faster than clicking around a menu, and it works everywhere. The app has a focus mode that dims everything except the sentence you are writing, and a typewriter mode that keeps your current line in the middle of the screen.

Bob: It is popular with journalists, novelists and students who want to escape the distractions of a full word processor. Reviewers highlight the live preview, so you can see how your markdown will look when rendered, and the built-in word count and reading-time estimates. It runs on Windows and Linux, is completely free, and the files you create are plain text, meaning they will still open on any device in twenty years time.

Samantha: For European users, there is no account to create, no data leaves your machine, and it works fully offline. If you have been meaning to start writing something longer, a book, a thesis, a long essay, and every other tool keeps interrupting you, Ghostwriter is worth twenty minutes of your afternoon. Download it from ghostwriter dot dev.

OsmAnd gives you full offline maps and navigation for Europe, without tracking.

Bob: Second Try This is OsmAnd. It is a mapping and navigation app built on OpenStreetMap, the community-made map of the world. Where Google Maps needs a connection and quietly harvests your movements, OsmAnd lets you download entire countries to your phone and navigate fully offline, with no account required.

Samantha: The strength of OsmAnd is depth. It has driving, cycling, hiking and public transport modes, with turn-by-turn voice guidance in dozens of languages. Hikers particularly rate it, because the OpenStreetMap data includes tiny footpaths, mountain huts and viewpoints that commercial maps often miss. For cyclists, it shows elevation profiles and cycle-specific routing that avoids busy roads.

Bob: The free version lets you download several map regions, which is plenty for most people. There is a paid version for unlimited downloads and extras, but you do not need it to get started. It works on Android and iPhone, and there is a desktop version too.

Samantha: For European travellers, this is a genuinely useful tool. Roaming charges on the continent are usually fine within the European Union, but the moment you cross into Switzerland, the United Kingdom, or the Balkans, mobile data gets expensive fast. Download the countries you are visiting before you leave, and you have full navigation without spending a cent. Find it in your app store as OsmAnd, spelled Oh Es Em Ay En Dee.

Samantha: Today we covered Google rebuilding search around AI summaries, Ay-Double-You-Es investing four billion dollars in public sector AI, OpenAI's first custom chip Jalapeño, the Nvidia chip smuggling arrests in Taiwan, NASA's robotic rescue of the Swift telescope, and two things to try yourself: Ghostwriter and OsmAnd.

Bob: Want to know more or react? Visit stateoftech.eu or email us at info@doorzetters.net.

Bob: State of Tech, the tech world in 15 minutes.