Three weeks I've been saying the chips and the concrete are the story — and today OpenAI named a silicon chip after a pepper. Jalapeño. Okay. I'm listening. This is the AI Daily Briefing. Today: a model company becoming a chipmaker, Groq's $650 million bet, and Blackstone dropping thirty billion in Japan like it's nothing. We start with Jalapeño. First actual hardware all week. Finally something I can argue with instead of a capex slide. Tap follow so the next episode finds you. Here's Ars Technica:
OpenAI, the company behind ChatGPT and Codex and the models those tools utilize, and Broadcom, an established silicon supplier, have announced a new chip called Jalapeño, designed specifically for large language model inference in data centers. The chip is intended to be deployed at large data centers, both companies claim this is just the first generation in a long-term project that will see chips refined over time.
OpenAI just announced a chip. Jalapeño, built with Broadcom, an ASIC designed from scratch for LLM inference. For three weeks I've been saying the inference stack is the story, and now the model company is putting a piece of it into silicon. This adds a third axis to the corporate-versus-geopolitical story. The company that owns the model is reaching down into silicon, and neither the hyperscaler frame nor the coalition frame really covers it. And here's the line that matters: OpenAI says 'performance per watt substantially better than current state-of-the-art.' The technical report? Sometime in the next few months. So we've got a name, a claim, and zero published cost-per-token against an H100 at scale. Nine months from design to announcement is genuinely fast for custom silicon. But the number I actually want is what a token costs on this thing under real load. That's the whole game for enterprise, and it's the one figure not in the release. Right — for once I'm not the one waving the asterisk around. The hardware is real. The proof is months out. Here's the part nobody's asking: this chip is purpose-built for inference at scale, fine. Does an ASIC change the error profile of a ten-step agent chain, or does it just let you fail at step seven cheaper and faster? Cheaper failure is still failure. But cheaper does win deals — you've said it all week. This one's from TMC Insight:
Groq's announcement of $650 million in new growth capital, led by Disruptive and Infinitum, arrives as demand for scalable inference capacity accelerates. Inference, the process of running AI models in production, is increasingly viewed as the next major focus in enterprise technology. Following the company's 2025 pivot toward an inference-first cloud identity, this new funding round indicates investors see clear momentum.
Groq's $650 million lands the same week as the Jalapeño chip we just hit — and it's tempting to file them under the same headline. But the fights are different. Groq is purpose-built silicon betting on raw inference speed; OpenAI and Broadcom are a model owner pushing down into the stack. Different bets, different customers. Six-fifty for a pre-revenue infra story — that's the exact number that makes me raise an eyebrow. I want their break-even utilization before anyone calls this validation. And that 200 megawatts by 2027 target is a concrete operational claim, so good — hold them to it. But notice what's missing: cost-per-token versus an H100 at scale. Nobody's published that. Not Groq, not OpenAI. I keep coming back to that December licensing deal with Nvidia — Groq's LPU tech going into the LPX platform. So they're competing with Nvidia and licensing to Nvidia in the same breath. That's the tension to watch. Yeah — IDC says AI compute spend tops $200 billion by 2027, mostly inference. Everybody's pointing at the same pot of money. Speed sells the demo; the cost curve closes the enterprise deal. From Saffron Humphreys at Data Centre Magazine:
The company is planning a new data centre campus in Pecos, West Texas, that will add approximately 2GW of capacity while pairing digital infrastructure with dedicated power generation. Microsoft describes the project as one of the largest capacity additions in its history.
Two gigawatts in Pecos, and the part that actually matters is buried halfway down — Microsoft signed a 20-year deal with a Chevron subsidiary to build dedicated power right next to the racks. Right, and that's the tell. When the power agreement runs longer than most AI companies have existed, Microsoft is basically co-locating with a utility. And the framing in the piece — power availability is the industry's biggest constraint — that's the honest part. Nobody's bottlenecked on parameters. They're bottlenecked on megawatts and the substations to feed them. It connects straight back to Jalapeño. Owning the chip gives you control inside the stack; Microsoft tying a campus to Chevron generation is control underneath it. West Texas is where all that abstraction turns physical. Twenty years is a long time to bet on demand. I'd love to see Microsoft's assumed utilization curve on two gigawatts, because somebody's spreadsheet is doing a lot of believing. Ana-Maria Stanciuc, writing in The Next Web:
Blackstone plans to invest roughly $30 billion in artificial-intelligence data centres in Japan over the next three to five years, the firm’s president and chief operating officer, Jonathan Gray, told Nikkei in an interview published on Monday. The world’s largest alternative asset manager is in discussions to develop facilities with a combined capacity exceeding one gigawatt, a scale that would place its Japanese ambitions among the larger national buildouts anywhere.
Jonathan Gray told Nikkei Blackstone's putting $30 billion into Japanese AI data centers over three to five years — over a gigawatt of combined capacity. That's the world's largest alternative asset manager treating geography as part of the strategy, not just capacity. And it lands the same day as the Pecos buildout we just hit. G7 coalition politics may be slow, but the data-center map is already spreading out — fast. My favorite line in the whole piece — he's 'unbothered by bubble talk.' The risk of building too little outweighs the risk of building too much. Okay, but that's a faith statement, not a number. When someone says they're unbothered by bubble talk, don't just nod along — ask what break-even utilization rate they need on a gigawatt to make $30 billion pencil out. Gray didn't say, and Nikkei didn't push. Blackstone did just raise $13.1 billion for its Asia PE fund, over target. They're not exactly capital-constrained or guessing in the dark. Sure — but a fund clearing its target tells you investors believe the thesis; it doesn't prove the gigawatt clears occupancy. Those are different bets, and he's collapsing them into one quote. Ben Craske, writing in Data Centre Magazine:
CoreWeave has signed a co-location agreement with Swedish data centre operator Conapto to establish two new campuses in Stockholm. The deployment provides additional immediate AI cloud capacity for the European market and aligns with the region's focus on sustainable digital infrastructure. Initial capacity is already operational at the Stockholm 4 South facility, marking a significant milestone in regional data processing capabilities.
And here's the third dot on the map — CoreWeave and Conapto, two campuses in Stockholm, with capacity already live at Stockholm 4 South. West Texas, Japan, now Sweden, all in one rundown. This is the June 18 pattern again: the buildout is dispersing before coalition politics catches up. Blackstone in Japan, CoreWeave in Stockholm — same move, different market. And notice the pitch here is renewable energy and low latency for European firms. That's a real selling point — Nordic power, cool climate, you can run high-density racks without your cooling bill eating the whole margin. But "initial capacity is already operational" is a soft phrase. Initial. I want the megawatt figure across both campuses and what the utilization commitment looks like before I call this more than a press release with a flag on it. CoreWeave doesn't tend to volunteer that number until the lease is fully racked. Right, and once you call it Stockholm 4 South, "South" implies there's a North, an East — phased. Which is fine, just don't read one live hall as a campus delivered. If you follow the AI Daily Briefing, you might also like Anthropic Pentagon Watch: a daily briefing on Anthropic's fight with the DoD over Claude, military AI use, autonomous weapons, and AI procurement blacklisting. Find it wherever you get your podcasts.
What we’re watching next: OpenAI says a detailed technical report on Jalapeño is coming in the months ahead, and Groq says it plans to scale its global infrastructure footprint toward 200 megawatts by 2027.
We’ve linked every story from today’s briefing in the show notes, so if one caught your ear, you can follow it there and read deeper.
That’s AI Daily Briefing for today. This is a Lantern Podcast.