Google's TurboQuant Paper Triggers Storage Stock Drops, Community Implements It in 48 Hours

On March 24, Google Research released the ICLR 2026 paper "TurboQuant", which proposes a mathematical method to compress the key‑value (KV) cache used during large‑model inference. The technique combines PolarQuant with a Walsh‑Hadamard rotation, achieving 3.8× to 6.4× compression while adding only 0.23 % perplexity (PPL) over the q8_0 baseline.

The KV cache is a major memory bottleneck for on‑device LLM inference because its size grows linearly with context length. Reducing its footprint could lower demand for high‑bandwidth memory (HBM) and DRAM, a factor that has driven storage‑chip price increases over the past two years. After the paper’s release, analysts noted a rapid market reaction: SanDisk shares fell 11 %, Micron 7 %, Samsung 5 %, and DDR5 kit prices dropped $100 overnight.

Within 48 hours, the open‑source community produced at least three independent implementations. The most complete, TurboQuant+, was authored by @TheTom, who added a Metal GPU kernel to llama.cpp supporting Apple Silicon and CUDA and evaluated seven model families. Another contributor, @no_stp_on_snek, documented the end‑to‑end process from the paper to Metal and CUDA ports, reporting a prototype built in 25 minutes with AI‑assisted code generation.

Benchmark results on an M5 Max show three compression levels:

turbo4 : 3.8× compression, +0.23 % PPL, quality closest to q8_0.

turbo3 : 4.6–5.1× compression, +1.06 % PPL, preferred for maximum reduction.

turbo2 : 6.4× compression, +6.48 % PPL, useful when VRAM is extremely scarce.

An additional optimization called Sparse V skips de‑quantizing V‑vectors whose attention weights are near zero in long contexts, yielding a 22.8 % decode‑speed boost with no PPL change.

At the same time, Anthropic announced on March 30 that Claude Code now supports “Computer Use”, enabling the model to control mouse and keyboard, open applications, fill forms, and test code directly from the CLI via the /mcp command. This moves the capability from a separate API into the everyday developer workflow.

OpenAI responded by releasing a Codex plugin for Claude Code, authored by Dominik Kundel. After installing the plugin from the Claude Code marketplace, users can invoke three commands: /codex:review – standard read‑only code review. /codex:adversarial-review – challenges implementation assumptions, suited for high‑risk changes. /codex:rescue – hands the stalled task to Codex for completion.

Installation requires a ChatGPT subscription (Free tier or higher) or an OpenAI API key and Node.js 18.18+.

Putting these developments together, TurboQuant demonstrates that the efficiency ceiling for local inference is still far from reached, and the rapid community implementation shows that the toolchain is mature enough for individual developers to bridge the gap to research. Claude Code’s Computer Use feature and the OpenAI Codex plugin turn coding agents from mere code generators into full‑cycle assistants that can run, test, and review code autonomously. As models become more efficient and toolchains more self‑contained, on‑device AI will handle increasingly complex engineering tasks, while storage‑chip demand growth may begin to lag behind efficiency gains.