d-Matrix stacks DRAM on its chip and claims 100 TB/s

Hot Chips 2026 wrapped with a memory bet, and the funding calendar is unusually busy for a Monday.

d-Matrix says stacking memory directly on the compute die beats HBM by 20x on bandwidth density. The startup's Raptor accelerator puts a TSMC N4 logic die face-to-face on top of custom 3D DRAM and claims roughly 100 TB/s per card at 32 GB — against a practical ceiling of about 20 TB/s for today's HBM4 packages, which d-Matrix attributes to limited package "beachfront" and raw physics: driving 100 TB/s through HBM at 2.4 pJ/bit burns about 1.92 kW before any fabric traffic. The pitch numbers: 32.6 GB/s per square millimetre against roughly 1.5 for HBM parts, 13.5x better on power per GB/s, and a claimed 1,000 tokens per second per user while serving a 3-trillion-parameter-class model at 1M context — a 72-card scale-up that fits one rack can host a frontier model such as Kimi K3. That is the whole argument in one line: decode, not training, dominates wall-clock inference time, and decode is memory-bound, so the vendor that fixes the pipe wins the bill. We explained why HBM became the choke point of the AI buildout in AI 101 — What is HBM (high-bandwidth memory)?.

Buildots raised $130 million led by O.G. Venture Partners to sell construction schedules to the AI data-center boom. The Tel Aviv company films job sites with helmet-mounted cameras, turns the footage into a live digital representation of the project, and diffs it against the schedule and the 3D model to classify completed work and forecast delays. Total capital now sits at $297 million, up from a $45 million Series D roughly 16 months ago, with customers growing from about 50 construction companies to more than 100 large firms; Calcalist reports the round values Buildots near $1 billion. The demand story is the interesting part — the company says AI data-center construction is pulling its product, because a slipped substation or shell hands a delay straight to a cluster of GPUs that is already paid for. Eight years of real site footage is the moat it claims, versus general-purpose models trained on scraped internet data.

Apple's new Siri is built so its brain can be swapped for Claude or GPT, and reverse-engineered code shows how deep the hook goes. Sleuth "pdfu" found private frameworks in iOS 27 and macOS Golden Gate with two distinct paths: an extension protocol where "Ask Claude" routes a natural-language request to the third-party model while Siri keeps executing system actions, and a far more consequential inference provider inside "Model Manager Services" that can replace Apple's own server-side Siri model entirely — the outside model receives Apple's Siri planner prompt and tool definitions, requests system actions, gets personal data back, and answers through Siri's voice and interface. Apple has not opened the model-delegation entitlement to third parties and it isn't front-facing; today the "Ask…" path is limited to the ChatGPT extension in the macOS 27 release candidate. The Digital Markets Act, which the European Commission has said applies to Siri, is the obvious reason any of this plumbing exists in the first place.

If Apple is quietly engineered for a swappable assistant, should users pick their model per request — or is that exactly the support nightmare Apple refuses to own? Tell us in the comments.

Deep Dive — America's top open model still trails China's best

Meta released the weights for Muse Spark Open on Thursday and called it "the most capable US open-weight model to date." Independent measurement agrees with the first half of that sentence and complicates the second: on Artificial Analysis's Intelligence Index the model scores 39, first among American open weights and behind GLM-5.3 at 45, Kimi K3 at 44 and Alibaba's Qwen3.8 at 40. The qualifier in Meta's headline is doing the work. This is a real release, a real engineering achievement, and not

AI 101 — How do AI video generators work?

An AI video generator is a model that starts with a grid of random noise and repeatedly refines it into a moving picture that matches your prompt — the same trick image generators use, stretched across time. Nothing is filmed, and nothing is stitched together from clips the model has seen. The pixels in the finished video were invented, frame by frame, all at once. Why it matters right now Kling 4.0, Kuaishou's flagship video model, went fully live on October 15 with up to 30 seconds of vid

Houmo moves compute inside memory — and A20 enters production

Beijing chipmaker Houmo used its partner conference this week to lay out two compute-in-memory routes at once — and to say the first chip built on the newer one is already in production. Houmo's A20, its first chip built on a 3D compute-in-memory architecture, has entered production. The part supports up to 16 chips fully interconnected, claims a 50% improvement in compute utilization over the previous generation, and covers models from 3B to 30B parameters, according to co-founder and product

Kling 4.0 goes live with 30-second native video at 10-bit HDR

Kuaishou's Kling AI has spent two years becoming the video model people actually finish projects with. Its new flagship, Kling 4.0, went fully live on October 15 — and this round is less about prettier frames than about the two things that make a generated shot usable in a real edit: length and control. Kling 4.0 generates up to 30 seconds in a single native pass, outputs 10-bit HDR at 4K, 1080p and 21:9, and takes up to 15 reference assets in one task. That budget splits across ten images, fi

Read more

Deep Dive — America's top open model still trails China's best

AI 101 — How do AI video generators work?

Houmo moves compute inside memory — and A20 enters production

Kling 4.0 goes live with 30-second native video at 10-bit HDR