Huawei ships its first logic-folded chip — and bonding is the gate
Huawei has started mass production on a phone chip that engineers its way around the lithography wall instead of through it, and in doing so it has quietly moved the competitive bottleneck from transistor density to how tightly two silicon wafers can be glued together.
Huawei's Kirin 9050 Pro is the first mass-produced logic-folding chip, stacking multiple dies and wiring them with more than 50 million hybrid bonds. The design puts circuits that would normally route sideways on a flat plane into two layers joined vertically, shortening the paths that set a chip's clock speed. Huawei claims transistor density up 53.5% over the previous generation, an NPU that is 140% faster while drawing 66% less power, GPU performance up 40% with power down 58%, and CPU multi-core throughput up about 23% on the Mate 90 Pro Max. Independent teardowns have confirmed the stacked layout and the roughly 1.5-micron hybrid-bonding pitch — the number that actually decides how much logic can participate in the fold, because more connections per unit area means more modules can be split across layers.
That is the part worth watching. Hybrid-bonding pitch is a manufacturing metric, not a design one, so Huawei's roadmap — toward 1 micron and beyond, with more than 100 million interconnects — is really a bet on its packaging supply chain rather than its architects. The company says next-generation Kirin 2027 silicon is already in hand and testing to expectations. Tom's Hardware also reports Qualcomm plans to license patents behind the architecture, which would be the clearest sign yet that vertical stacking is becoming standard practice rather than a workaround for one company's fab constraints. Huawei spent RMB 192.3 billion on research and development in 2025, 21.8% of revenue, and this is what that buys: a phone SoC whose headline claim is that it runs a 30-billion-parameter mixture-of-experts model locally while adding under 2.5 GB of memory pressure.
Bloomberg is opening its ASKB conversational AI to every Bloomberg Terminal user, and letting firms splice their own documents into the answer. The Journal's CIO Journal first reported the expansion; WIRED's April interview put the installed base at about 375,000 Terminal seats, with ASKB in the hands of roughly a third of them after a limited rollout early this year. Firms can now connect proprietary research and internal chat content, or upload their own PDFs and Word files for analysis, all cited back with Bloomberg Query Language references. Bloomberg CTO Shawn Edwards has said the conversational layer will become the primary way people interface with the Terminal — which, if true, turns a $32,000-a-year data terminal into an enterprise AI surface with a captive audience. No ASKB-specific pricing has been disclosed; it rides on existing Terminal entitlements.
Figma released a Markdown spec layer so that AI agents — its own and everyone else's — design inside a brand's rules. Figma Specs is a document that defines a company's design language and can pull in canvas elements, FigJam diagrams, code layers, design libraries and written guidelines, giving both human teams and third-party agents one place to read the constraints from. Announced at Figma's first Config event in India, it is the platform's answer to the non-determinism problem: agents can build fast, but they need a spec to build the right thing. The company also moved code layers from private beta into open beta and expanded its MCP catalog by 20 partners, including Grok and Indian model maker Sarvam. Figma's chief design officer Loredana Crisan told TechCrunch that teams now produce 17 times more lines of code than before while releases are up only 1.3 times — most of that code exists to discover a product, not to ship it.
What to watch: whether Qualcomm licensing the folding patents turns cross-layer stacking into a licensing business, and how fast hybrid-bonding capacity can scale outside Asia.
If a phone can run a 30B model locally, what exactly are we still paying cloud inference for? Tell us in the comments.
