Adoption dailySeptember 28, 2026Web platform

Local compute is eating the cloud

WebGPU and AI agents show developers building for edge performance, while robotaxis cool as hardware struggles to keep pace.

The numbers: Passkeys is today's fastest-rising trend: installs at 1.33× the four-week pace. Momentum across 1 signal: 1.33×. 5 of 21 trends are climbing and 2 are cooling.

Backfilled on October 7, 2026 using only data published by September 28, 2026.

Passkeys: 887k npm installs a day on a typical day last week, against 667k the four weeks before

Bars are single days, the line is the 7-day average, the shaded band is the last 7 days and the dashed line is the four-week typical day (median).

Data through npm 09/26, Wikipedia 09/27, HN 09/27. Source: npm downloads API.

WebGPU installs reached 1.5M/day, up from 1.3M, while AI agent installs hit 653k/day. This combination signals a move toward local execution. Developers are no longer waiting for remote servers to process data. The browser now handles complex tasks directly on the user’s device, reducing latency and cost. This change alters how applications are architected, favoring lightweight, client-side logic over heavy backend dependencies.

The cooling of robotaxis, with Wikipedia interest dropping to 195/day from 251, contrasts with the rise of software-driven tools. Physical hardware adoption remains slower and more visible than digital tooling. While humanoid robots also cooled, the focus is shifting to software that runs on existing devices. The gap between hardware hype and software utility is widening, with developers prioritizing tools that work today over hardware that promises future capabilities.

OpenAI Codex saw a massive jump to 3.4M/day installs, up 117× from a year ago. This explosive growth in coding assistants parallels the rise of WebGPU. Both trends reflect a demand for speed and autonomy in the development process. As local compute capabilities improve, the need for powerful, autonomous coding tools increases. This creates a feedback loop where better hardware enables better software, which in turn drives further hardware adoption.

Watch next: Monitor if WebGPU adoption correlates with a drop in cloud API usage for real-time applications.

Column drafted by qwen/qwen3.8-27b from the numbers below; every figure checked against them.

The numbers

On a typical day last week, npm installs of @simplewebauthn/browser reached 887k, 1.33× the 667k of a typical day in the four weeks before.

Next on the board: Smart glasses at 1.27×, led by Wikipedia readers at 1.27× (285 on a typical day against 224).

Quiet adoption: Bun runtime installs run at 1.18× their four-week pace while attention sits at 0.86×. Developers are adding it faster than people are reading about it.

Cooling fastest: Humanoid robots at 0.66×, with Hacker News stories down to 0.4 on a typical day from 1.0. Part of that is a spike on Sep 16 in the weeks before.

The long view: OpenAI Codex installs over the last four weeks are 117× the same four weeks a year earlier (3.4M a day against 29k).

Trend pages: Passkeys Smart glasses Small modular reactors Humanoid robots

Quiet adoption Broad rise One-day spike

5 of 21 trends are climbing; Passkeys leads at 1.33×

Momentum is a typical day (median) of the last 7 divided by a typical day of the 28 before, averaged across every signal with enough data. 1× means no change. Rings mark each signal.

Data through npm 09/26, Wikipedia 09/27, HN 09/27. Source: npm downloads API, Wikimedia pageviews, Hacker News search.

3 of 7 developer tools are being installed faster than they are being read about

Across: attention momentum (Wikipedia readers and Hacker News stories). Up: install momentum (npm). Both on log scales; the 1× lines split the corners.

Data through npm 09/26, Wikipedia 09/27, HN 09/27. Source: npm downloads API, Wikimedia pageviews, Hacker News search.

AI agents draws the most Wikipedia readers against its own normal: 1.32× over the last week

Each cell is one day's readers divided by that trend's four-week typical day. Orange runs hot, dark runs cold.

Data through npm 09/26, Wikipedia 09/27, HN 09/27. Source: Wikimedia pageviews API, user agents only.

Every number on the board

All 22 trends, September 28, 2026. Per-day figures are a typical (median) day of the last week. * = base too small to count (under 1,000 installs, 100 readers or 1 story a day).

Passkeys	Web platform	1.33×	Climbing	887k	1.33×	90	1.12×*	0.6	1.33×*	0
Smart glasses	Devices	1.27×	Climbing	–	–	285	1.27×	1.4	3.1×*	0
Small modular reactors	Energy	1.23×	Climbing	–	–	539	1.24×	0.0	0×*	0
AI agents	AI agents	1.21×	Climbing	653k	1.19×	1,178	1.32×	12	1.13×	0
WebGPU	Web platform	1.15×	Climbing	1.5M	1.13×	141	1.17×	0.7	1.05×*	0
AI API SDKs	AI agents	1.13×	Steady	18M	1.13×	–	–	–	–	0
Vibe coding	AI coding	1.11×	Steady	–	–	1,676	1.22×	1.1	1×	0
SpaceX Starship	Space	1.08×	Steady	–	–	1,475	1.08×	0.0	0×*	0
Gemini CLI	AI coding	1.03×	Steady	64k	1.03×	–	–	0.0	n/a*	0
Model Context Protocol	AI agents	1.01×	Steady	10M	1.14×	1,620	1.14×	5.6	0.8×	0
Bun runtime	Web platform	1.01×	Steady	715k	1.18×	225	0.86×	0.7	0.74×*	0
Quantum computing	Devices	1×	Steady	–	–	1,919	1.1×	3.0	0.9×	1
Heat pumps	Energy	0.99×	Steady	–	–	473	0.99×	0.1	2×*	0
Starlink	Space	0.98×	Steady	–	–	1,989	0.98×	0.6	1.46×*	0
Local LLMs	AI agents	0.97×	Steady	164k	0.91×	706	1.03×	0.3	0.44×*	0
OpenAI Codex	AI coding	0.96×	Steady	3.5M	1.06×	39	1.11×*	3.7	0.87×	0
Solid-state batteries	Energy	0.94×	Steady	–	–	370	0.94×	0.0	0×*	0
eVTOL air taxis	Space	0.91×	Steady	–	–	190	0.91×	0.0	0×*	0
Claude Code	AI coding	0.89×	Steady	2.0M	0.8×	–	–	6.7	0.98×	0
Robotaxis	Robots	0.78×	Cooling	–	–	195	0.78×	0.0	0×*	0
Humanoid robots	Robots	0.66×	Cooling	–	–	372	1.01×	0.4	0.43×	0
Vehicle-to-grid	Energy	n/a	Not enough data	–	–	63	1.02×*	0.0	0×*	0

Method and sources

Every trend gets up to three free, daily signals. Installs: npm registry downloads for the packages named on each trend page (developer adoption; it counts CI and mirrors too, so it measures direction, not people). Attention: English Wikipedia readers per day, human agents only, and Hacker News stories with the term in the title. For each signal we divide a typical day (the median) of the last 7 complete days by a typical day of the 28 days before. Medians keep one launch-day spike from carrying a whole week; both windows are whole weeks, so weekend dips sit equally in both. Momentum is the geometric mean of the signals that clear a minimum base. npm reports some days as zero for every package at once; we treat those as missing.

The column is drafted by a free language model from the day's computed facts only, and is published only if every number in it matches those facts; otherwise the computed paragraphs stand alone. Theme labels come from Jev (or its Cloudflare Clef fallback, as noted) choosing from a fixed list of ten.

Stages: Breakout 1.5× and up, Climbing 1.15×, Steady 0.87× to 1.15×, Cooling below. Raw data: /api/trends/2026-09-28.json.