QUANTUM BEAT | JUNE 25, 2026 | YOUR DAILY AI BRIEFING
OPENAI BUILT ITS OWN CHIP IN NINE MONTHS USING ITS OWN AI, SLASHES NVIDIA COSTS IN HALF
Source: TechCrunch
Okay so here is something that should make Jensen Huang a little nervous over his morning coffee. OpenAI just unveiled Jalapeño, its first custom-built AI chip, designed and manufactured in partnership with Broadcom. It took nine months to build. They used their own AI models to help design it. And it reportedly cuts inference costs by roughly 50 percent compared to the Nvidia GPUs the entire industry has been running on.
Let that land for a second. OpenAI, a company that not long ago was begging Nvidia for more chips and paying whatever price was asked, has now gone and built its own silicon. In under a year. With AI doing some of the engineering work. That is either one of the most impressive things that has happened in tech in years, or the most audacious, depending on how the thing actually performs in the wild.
Jalapeño is an inference chip, meaning it is designed for the part of the process where you run a finished model and answer user queries, not for the brutal compute required to train a model from scratch. That distinction matters a lot. Training still requires Nvidia’s beastly hardware, and OpenAI is not walking away from that relationship anytime soon. But inference is where the money goes when you are serving hundreds of millions of ChatGPT users every day, and that is where custom silicon can really punch above its weight.
The chip was built around OpenAI’s specific workload needs. Greg Brockman, OpenAI’s president, explained the thinking on the company’s podcast shortly after the Broadcom partnership was announced last October. The idea was to find workloads that general-purpose chips serve poorly and build something tailor-made. Google has been doing this for years with its TPUs. Amazon has Trainium. Now OpenAI has Jalapeño, and if the 50 percent cost figure holds up at scale, this is a very big deal.
Think about what happens if OpenAI cuts its inference bill in half. It can lower prices for users. It can run more ambitious agentic products that require more compute per task. It becomes less dependent on a single supplier who has been, to put it politely, in a position of extraordinary leverage. The AI chip market has been Nvidia’s market for years, and everyone has accepted that as a fact of life. This announcement suggests that era is slowly ending, one custom chip at a time.
The initial deployment is targeting the end of 2026, which is not far off. The companies say they will scale from there. What nobody is saying out loud, but everyone in the industry is thinking, is that Broadcom just became one of the most important companies in AI without anyone really noticing. Nvidia got all the headlines. Broadcom got the contract. There is a lesson in there somewhere.
The most quietly remarkable detail in all of this is the meta-story. OpenAI used its own AI to help design a chip. The chip will then be used to run more AI. The loop is closing. At some point you start to wonder whether the chip designers of tomorrow are the models running today.
GOOGLE LOSES TWO MORE GEMINI ARCHITECTS TO ANTHROPIC AS THE TALENT EXODUS TURNS INTO A FLOOD
Source: TechCrunch / Bloomberg
At this point Google is less of an AI company and more of a very expensive training program for Anthropic and OpenAI. Jonas Adler and Alexander Pritzel, two researchers who played significant roles in building Google’s Gemini model, are leaving for Anthropic. This comes roughly a week after Nobel Prize winner John Jumper announced he was also heading to Anthropic, and days after Noam Shazeer, the co-inventor of the Transformer architecture and one of the most respected minds in the entire field, announced he was going to OpenAI.
We are not talking about one or two defections here. We are talking about a sustained, accelerating drain of some of Google’s most important AI talent, happening in public, all at once, in a span of about ten days. Adler was working on Google’s AI coding efforts. Pritzel was involved in the training of AI systems. These are not peripheral figures. These are people who helped build the models that Google is betting the company on.
So what is going on? The most obvious explanation is money, specifically the promise of pre-IPO equity. Both OpenAI and Anthropic are widely expected to go public in the not-too-distant future, and the opportunity to get in before that happens is an extremely compelling offer for someone sitting on Google’s relatively stable but hardly explosive compensation structure. You can make a very good living at Google for a very long time. You can also potentially become extremely wealthy by joining a company two years before it goes public at a valuation that the market has decided should be measured in the hundreds of billions.
But that is not the whole story. There is also the question of culture and autonomy. At a company the size of Google, even the most talented researchers are working within an enormous organization with a lot of inertia, a lot of committees, and a lot of competing priorities. At Anthropic or OpenAI, the work is still close enough to the frontier that individual researchers can feel the direct impact of what they are doing in a way that is harder to feel inside a conglomerate worth two trillion dollars.
Google is not sitting still. The company still has extraordinary resources and a lot of talent. Gemini is a real and competitive product. But the pattern here is undeniable. When a company starts losing not just one star researcher but a sequential series of them, all to the same competitors, all in the same short window, it suggests something systemic. The question Google needs to answer is not just how to retain the next Adler or Pritzel, but why the last four decided that somewhere else was a better place to spend their career.
Alphabet’s stock dropped 7 percent in recent days on news of the departures. The market is paying attention even if the company has not yet said much publicly about what is happening. At some point the silence becomes its own statement.
FACEBOOK WANTS TO BE YOUR AI MANAGER NOW, LAUNCHES STANDALONE APP TELLING CREATORS WHEN TO POST AND HOW TO REPLY TO COMMENTS
Source: TechCrunch
So Facebook would like you to know that it is no longer just a place where your aunt shares recipes and your uncle argues about politics. It is now also your personal AI growth coach, and it has built a whole standalone app to prove it.
Meta announced on Wednesday that it is reinventing Creator Studio as a dedicated AI companion app for creators. The app gives you a feed of daily priorities when you open it, which apparently include reviewing your newest post’s performance, tracking your progress toward goals, and figuring out which comments need a reply. The built-in AI assistant will tell you when to post, analyze what your audience actually wants, and even draft replies to comments in what it describes as your own tone. You get to approve those drafts before they go out, which is a nice touch, though one imagines the approval step gets skipped more than it should once people get busy.
The strategic logic here is pretty clear. Meta is watching creators spend time in ChatGPT brainstorming content ideas, using third-party analytics tools to understand their audiences, and generally looking everywhere except Facebook for help running their Facebook presence. That is not a good look. If the AI tools live inside Meta’s own ecosystem, creators have less reason to go outside it, and Meta gets to keep them engaged and generating content that keeps users on the platform.
It is also part of a broader pattern. Mark Zuckerberg has reportedly told employees that AI-driven efficiency gains mean the company can now build more apps than it has historically, and they are certainly following through. In the past few months alone, Meta launched Forum, which is basically Reddit inside Facebook, a disappearing photo app called Instants on Instagram, and now this creator companion app. At some point it starts to look less like a product roadmap and more like a fire hose.
Whether creators actually want this is a different question. There is a very specific type of creator who would genuinely benefit from an AI that tells them when to post and how to respond to their audience. There is also a very specific type of creator who will find the whole thing slightly unnerving, because having an algorithm tell you how to talk to the people who follow you starts to feel like the algorithm is running your creative life, not just supporting it.
That said, the commenting feature is probably the most immediately useful thing in here. Going through hundreds of comments on a popular post is genuinely time-consuming and tedious. An AI that can surface the most important ones and draft thoughtful replies is actually solving a real problem. Whether people trust it enough to use it is the question nobody can answer until creators actually get their hands on it. The app is currently being tested with select creators, so the rest of us will have to wait and see whether the tone matching holds up or whether everyone ends up sounding like a chatbot that read one too many motivational quotes.
CORPORATE AI BILLS ARE EXPLODING AND PANICKED EXECUTIVES ARE NOW RATIONING THE SAME TOOLS THEY DEMANDED EMPLOYEES USE OR LOSE A PROMOTION
Source: TechCrunch
This one is almost too good. Not long ago, Accenture, one of the largest consulting firms in the world, was apparently so enthusiastic about getting its employees to use AI that it built internal leaderboards to track and celebrate who was using it the most. Use more AI, get promoted. That was the message. The era of what people in the industry have started calling tokenmaxxing was in full swing, and corporate America was all in.
Fast forward a few months and you have leaked audio from an internal Accenture meeting in which the company’s agentic AI strategy lead is telling employees to stop spending so many tokens on basic tasks like converting PDFs into presentation slides. The same company that was essentially threatening employees with career consequences for not using AI enough is now pleading with them not to use too much of it on stuff that does not matter.
The technical problem here is actually pretty interesting. AI tools billed on a per-token basis get expensive very fast when employees start using them for everything, including things that do not actually require AI and where the cost-benefit math is frankly terrible. Converting a PDF to a slide deck is a good example. A junior employee can do that in half an hour. An AI can do it in thirty seconds. But if you are paying per token and your employees are doing this all day every day across an organization of hundreds of thousands of people, the bill adds up to something that makes the CFO’s eye twitch.
And this is not just an Accenture problem. Reporting from VentureBeat and others has shown that Microsoft has been canceling Claude Code licenses in certain divisions because the per-engineer API costs were running between five hundred and two thousand dollars a month, and some teams were burning through their annual AI budget by April. Companies that were previously asking employees to use AI constantly are now scrambling to figure out which uses actually produce enough value to justify the cost and which ones are basically very expensive habit formation.
There is something almost philosophically interesting happening here. The AI industry spent two years convincing everyone that the only thing limiting business transformation was not having enough AI adoption. Then everyone adopted it as fast as possible, and it turned out the limiting factor is actually return on investment, which is an old-fashioned concept that has not gone away just because the technology is exciting.
The broader picture is what some analysts are calling the AI selloff, where memory chip makers and other AI-dependent businesses have taken a beating in recent days as the market starts to ask harder questions about whether all this spending is producing commensurate returns. The companies selling AI infrastructure are still doing extremely well. The companies buying it are starting to have difficult conversations about whether the investment is paying off and whether their employees are using it in ways that actually move the business forward or just in ways that feel productive.
Accenture’s Kwak put it bluntly in that leaked audio: AI spend is becoming unpredictable, and leadership at the CFO, COO, and CIO level is asking whether they are getting value from what they are spending. That question was always coming. The surprising part is how quickly the tokenmaxxing era ended once it arrived.
Quantum Beat publishes daily. All stories sourced from credible outlets. No AI slop, just the news.