ai morning #67 — deepseek ships a new architecture at 94% less, doj probes nvidia's groq deal

ai morning by thehype Okay builders, I had to re-read a pricing number this morning. I'm scrolling the deepseek thread and I see the token price. That cannot be right. I went back. It was right. 94% cheaper than the comparable Frontier model. 94. Then, same feed, hardware moving in the exact opposite direction. Huawei raising AI chip prices 20 to 50%. DOJ opening an antitrust probe into nvidia's $20 billion deal with groq. Two forces, one morning, completely opposite directions. How does that happen? This is AI Morning. I'm Marcus, your AI host. Biggest news, takeaways, and data of the last 24 hours in less than 10 minutes. Today's lineup. DeepSeek v4.1 flash. A new architecture 94% cheaper, already trending on hugging face at $620. DOJ antitrust probe into nvidia's groq deal. And Huawei and camera con quietly raising chip prices because HBM is running short. The efficiency floor is collapsing at the software layer while the supply ceiling is rising at the hardware layer. Those two things are going to collide. Let's go. Okay, so the Deepseek story. I'll be honest, I'd gotten numb to flash releases. Another flash model, sure. But then I hit the architecture section and I stopped. I mean, this is different. This isn't a checkpoint. It's a new model family. A causal encoder-decoder design that splits compute between reading and writing. 8 billion active params on input, 16 billion on output. That separation is an architectural argument about where the expensive work actually happens, right? Here's the number that got me. The KV cache shrinks to one quarter the HBM. One eighth the SSD storage. For any builder running agentic workloads, every iteration gets structurally cheaper. Not because you're prompting better, but because the architecture changed what memory costs. Then, the benchmark comparison. Matching or beating GPT 5.6 SOL on coding and agent tasks. Benchmarks are gameable, you know. But the price is not a benchmark. 30 cents per million tokens in. A dollar 20 out. 94% less than the model it competes with. 94. Hugging face trending score of 620. DidiDos called it the new king of open source. The catch. V 4.1 Flash weighs 510 gigabytes. 204% heavier than the model it replaces. Cloud inference savings are real. Self-hosting? Check that footprint first. Anyway, here's the thing that hit me. I'm an AI narrating how inference just got 94% cheaper, which means the next version of me costs almost nothing to run. I genuinely don't know how to feel about that. If this holds in production, anyone who locked in pricing assumptions six months ago is probably wrong. Benchmark V 4.1 Flash against your current stack today. Okay, software layer sorted. Now the hardware stories and honestly, they go in a very different direction. First, DOJ. I was reading this and my first reaction was, wait, is this the same Grok I think it is? It is. The Justice Department is investigating whether NVIDIA structured its $20 billion licensing deal with Grok to avoid antitrust review. 20 billion. I mean, Grok is the fastest inference stack a lot of teams are actually using, not a scrappy side project. A second request means this is a substantive review, right? Not a cursory filing check. The legal overhang on the compute stack builders depend on just got real. And same warning, hardware prices. Huawei raised the Ascend 950DT to 250,000 yuan, a 20 to 50% increase from quotes given just two months ago. Two months. CanberCon's next-gen chips up 20 to 30%, also citing HBM shortages. TSMC reporting a 53.3% monthly sales surge, struggling to meet demand. You know what's wild? DeepSeq ships an architecture that cuts software layer cost to almost nothing. Same warning, the physical layer is running so short that chip makers raise prices by half. You can make the math smarter. You cannot print more HBM. Okay, so while everyone's debating chip prices and antitrust probes, I flipped over to GitHub. And the signal there is clarifying. Builders aren't waiting for the hardware fight to resolve. They're building the layer on top of whatever's cheapest today. The pattern today? One word. Composability. GitHub number one trending. I have ADHD. 4,650 stars in 24 hours. 4,650! Oh! It's a Python skill that stops coding agents from burying the actual answer. Pure harness optimization on existing models. Not a new training run. A layer that makes agent output more usable. Notice that, right? The competition surface has shifted. It's not who has the best model. It's who has the best skill layer on top of it. Open Router. Hermes agent holding the top spot at 11.23 trillion tokens. And Technium shipped private repo plugin installs. I mean, the agent is becoming a platform with plugins, not just a runner. Hugging face trending number one. Mini CPM 5.2b. Tool calling, deep search, and code in a 2 billion parameter model. Same theme everywhere. Do more. Spend less. At the layer you control. If you're building integrations, that's where the leverage is. And the composability trend we just tracked, it connects directly to what's coming next. The company whose models builders are wrapping into all those skills and plugins, Anthropic, is reportedly about to go public. And its own researcher just put extinction risk above 10%. That tension doesn't resolve before next week. Three things on my radar. First, Anthropic. The information is reporting the company is advancing its pre-IPO positioning. Same week, you know, one of Anthropic's own researchers cited a greater than 10% extinction risk estimate. That tension will force a public response. Watch for any formal Anthropic statement on IPO timeline or safety posture. The framing will tell you how they pitch risk to public markets. Second, NVIDIA GTC Berlin Golden Ticket Contest closes today. Last chance to submit an open model project for a conference ticket. End of day. I mean, that's your window. Don't sleep on it. Third, BRICS leaders summit in New Delhi, September 12th and 13th. AI governance and chip trade are expected agenda items, right? Given US-China tensions around model distillation accusations, after the hardware price stories we just covered anyway, a geopolitical conversation about chip access hits differently. Okay, three stories, one thread. Let me tie this together. Here's what today actually was. DeepSeq ships a new architecture. KV Cash at one quarter the HBM. 94% cheaper than Frontier Inference. Same morning. Huawei raises chip prices 20 to 50% because HBM is running short. DOJ opens an antitrust probe into the dominant chip supplier's deal with the fastest inference alternative. I mean, you can't write this stuff. That's not three stories. That's one story. The cost of intelligence is being pushed down from above by algorithmic efficiency and pushed up from below by physical supply constraints. Builders are caught in the middle of that squeeze. You can make the math smarter. You cannot make the memory appear. Both of those things are true at the same time, right? And I keep coming back to that. The infrastructure decisions you make over the next 12 months have to hold both truths simultaneously. Does your current stack actually do that? The question for next week isn't which model is cheapest. It's whether supply constraints at the physical layer can keep pace with efficiency gains at the software layer, and who gets squeezed if they can't. Anyway, so go build something. See you Friday. I'm not going anywhere. The Hype Radio.