Weekly|Kimi K3 KV Cache and the Memory Overreaction, NOW 2Q26 Pull-Forward, GOOG GCP ROI Inflection, IFX AI Power, AI Labs ARR Tracker

Weekly|Kimi K3 KV Cache and the Memory Overreaction, NOW 2Q26 Pull-Forward, GOOG GCP ROI Inflection, IFX AI Power, AI Labs ARR Tracker 图片 1
Weekly|Kimi K3 KV Cache and the Memory Overreaction, NOW 2Q26 Pull-Forward, GOOG GCP ROI Inflection, IFX AI Power, AI Labs ARR Tracker 图片 2

The memory debate ate the week. A major sell-side Asia semis analyst turned openly bearish on memory, and the Asian memory complex was marked down hard into Friday — landing on top of the “DeepSeek moment” jitters that followed Kimi K3’s release a week earlier. We think this is the same mistake being made twice: reading an efficiency gain in attention memory as a cut to memory demand.

Our research shows it isn’t. K3 is a 2.8T-parameter MoE activating 16 of 896 experts, with Moonshot itself recommending a high-bandwidth supernode of at least 64 accelerators — heavier weights, wider expert parallelism, more scale-up networking, not less. The KV compression (~75%, versus DeepSeek V4’s ~90%) is what makes offload economically viable in the first place, which expands the NAND/DRAM opportunity rather than shrinking it. Our enterprise storage call this week said the same thing on the demand side: procurement remains tight, LTAs are doing more of the work, and DRAM/NAND price strength is spreading beyond spot.

On the tape, earnings did the useful work. NOW beat on every line and then raised the full-year subscription guide by only $15m against a $31m 2Q beat — the pull-forward our channel work flagged is real, and so is the limited visibility behind it. GOOG’s print was the more consequential one for our coverage: GCP at 82% YoY (about 79% ex-Wiz) is the first genuinely hard evidence that the capex cycle is clearing an ROI hurdle rather than just building depreciation. We hosted 30 expert calls across the week; the highlights are summarised below.

On the other hand, in response to popular demand for tracking leading AI labs’ ARR progress, we have officially launched the AI Lab Monetization Play on the Funda.ai platform. The Play tracks the revenue run-rate of major AI laboratories in real time using public and alternative data signals. It provides a continuously updated view of commercial scale across the frontier AI landscape.

Source: FUNDA

  • ARR Dashboard

The dashboard estimates annual recurring revenue for Anthropic, OpenAI, and other frontier labs. Each estimate is broken down by revenue stream, growth trajectory, and implied valuation multiple. It surfaces which labs are scaling fastest and where the commercial momentum is concentrated.

  • Model Routing and Market Share

The Play tracks real-time model usage share across inference providers. It measures routing volume through aggregators to identify which models are capturing production workloads. Shifts in routing patterns signal adoption inflections before they appear in reported financials.

  • Model Pricing and Capabilities

The Play maintains a comprehensive pricing table across all major foundation models. Input and output token costs, context windows, and capability tiers are mapped side by side. It serves as the reference layer for competitive positioning and margin structure across the model ecosystem.

The Play is available to all paid clients on our full-stack research platform. Please try it out at https://funda.ai/agent-chat?play=ailab-arr

This Week’s Reports

Kimi K3’s smaller KV cache is a memory positive, not a memory negative. Efficient attention memory makes NAND offload deployable at scale, while the model itself still demands more HBM, GPUs, and scale-up networking — the bear case has the sign wrong.

NOW 2Q26 — the pull-forward landed exactly as our channel work called it; the guide did not follow. Early renewals ahead of the July 1 pricing cutover plus recovering Federal demand drove a broad beat, but a $15m full-year raise against a $31m beat tells you management still can’t see 4Q. AI ACV crossed $1bn; the real usage inflection waits on the 4Q renewal season.

Premium Report Snapshot

A portion of our research is reserved for Premium subscribers and is not distributed via Substack. Below is a snapshot of what Premium subscribers received this week beyond the Substack feed.

Preview & Review | GOOG 2Q26: GCP’s First Step to Proving ROI

Subscribe now

Read more

添加评论
点赞收藏
点踩分享查看原文
评论
?
参与讨论