Meta's in-house silicon just got weirder in a way I actually respect. The new MTIA 400 does two jobs that rarely share a chip: it trains the AI models running inside Meta, and then it helps decide which ad lands in your feed. Same silicon, two very different brains.
Early coverage from The Register suggests it's faster than Blackwell in some training workloads. Before anyone starts sharpening pitchforks — no, it's not replacing AMD or Nvidia for anyone outside Menlo Park. Not yet. That's the honest version of the story, and honestly it's still the interesting version. A chip that trains the model and then serves the ads that quietly fund that same model is basically Meta's entire business model stamped onto a piece of hardware. You don't see that kind of vertical thinking often, and when you do, it usually means the company is serious about owning the stack end to end.
Meanwhile, the GPU king just pulled a move that's more awkward than hostile. Nvidia paused the financing program it launched in July — the one where it gave AI cloud companies credit support in exchange for a slice of their revenue. WSJ reports internal worries about antitrust scrutiny and about how much control Nvidia could dictate over how customers run their businesses. In the early weeks it apparently irked partners by demanding they only rent chips to approved customers, and signaling it preferred capacity spread across many small AI firms rather than one big tenant. My take, for what it's worth: Nvidia doesn't need to be everyone's bank. Selling the shovels at a fat margin is already a phenomenal business, and trying to own the whole food chain is how you get regulators breathing down your neck. The program might get revamped, but the pause itself tells you the era of the chip company quietly running your cloud is not arriving quietly.
Over in the open-source corner, Deep Cogito banked a $43M Series A led by TQ Ventures, with Benchmark, Nexus, Atreides, South Park Commons, and Zscaler chipping in — pushing total outside funding past $56M. Two ex-Google founders, Drishan Arora and Dhruv Malrana, are chasing the self-improving model dream. Their open Cogito line just hit v2.1 671B, which they claim was ahead of any other US open model at launch while using less tokens than comparable reasoning models. That last bit matters more than it sounds. Token cost is the quiet tax every developer pays, and if you've ever watched a reasoning model burn through context on a trivial question, you know exactly why process supervision is their secret sauce — grading each step a model takes instead of just the final answer. I've had agents spiral into five paragraphs of deliberation for a two-line fix; a technique that trims the unnecessary steps is the kind of thing you feel in your monthly bill before you feel it in the output.
On the enterprise side, adoption is getting boring in the best way. Wipro is rolling out Gemini Enterprise internally and plans to train more than 10,000 specialists, nudging AI past the toy phase and into daily business workflows. And Joget launched an Agent Lab where businesses submit real workflow problems and get a free, expert-built agent in return. It's a marketing play dressed as a community program, sure, but the smart part is the signal it generates — real companies describing real bottlenecks, out in the open.
One more from the fringes that I keep chewing on: Bilibili, the Chinese video platform, says it's targeting 40-45% gross margin and roughly 15-20% operating margin, with ad revenue up 28% and a global launch for its Lumi Master AI thing scheduled for September 17. A social video site talking like a margin-obsessed software company is a good reminder that the AI gold rush isn't just in chips and models — it's in every recommendation engine quietly getting better at keeping you on the page.
To be fair, none of this week is going to change your life by Friday. The MTIA 400 still has a distribution problem, Nvidia's pause is corporate caution more than strategy, and open-source self-improving models remain a research bet with a real chance of fizzling. But the direction is consistent: everyone is consolidating — chips, clouds, enterprise tools — and the ones doing it without screaming are the ones to watch. If I'm honest, I'm most curious about whether the split-personality chip actually ships at scale, because if it does, the ad-serving side of AI just got a whole lot cheaper to run.
For anyone digging into similar territory, a Manual Assistant I've been using for day-to-day AI workflow notes has been handy for keeping track of which agent runs are actually costing you tokens — small thing, but it stops the drift.
Top comments (0)