Your dictated email circled the globe three times before landing in your document
Picture this. You open your mail client, hit a shortcut, and dictate a reply to a colleague. Three sentences, fifteen seconds of speech. Simple, fast, efficient.
But behind this apparent simplicity, a colossal infrastructure roared to life. Your voice left your workstation, crossed your router, rode your ISP's fiber optics, and reached a data center thousands of kilometers away—perhaps Virginia, perhaps Ireland, perhaps China. There, thousands of servers burned electricity to turn your voice into text. Then the result retraced the entire journey. Your dictated email circled the globe three times before landing in your document.
And that's just one email. Multiply by millions of daily dictations, meeting transcriptions, always-listening voice assistants, generative AI queries. The cloud is not immaterial. It carries a massive carbon footprint, an insatiable energy appetite, and a physical geography that turns every click into an intercontinental trip.
The cloud's hidden cost: the planet
Data centers account for roughly 2% of global electricity consumption—the equivalent of all of Japan. And that figure is skyrocketing: the generative AI explosion could double this consumption by 2027. Microsoft, Google, and Amazon are building mega-data centers at a frenzied pace, each one consuming the energy equivalent of a small city.
But electricity is only part of the problem. There's also water—billions of liters cooling these servers. There are rare earth materials—neodymium, cobalt, lithium—extracted under often precarious conditions to manufacture chips and hard drives. There's the short hardware lifecycle, replaced every 3 to 5 years. And above all, there's network traffic: every bit that travels consumes energy at every node it crosses.
The cloud's economic model rests on externalization. You don't see the data center, so you don't feel its impact. It's the inverse of "polluter pays": you pollute, someone else pays—the planet, downstream communities, future generations.
The political paradox of green tech
Here's the irony: the same governments signing ambitious climate agreements massively subsidize data center expansion. The European Union classified data centers as "critical infrastructure" during the pandemic, granting them energy exemptions. The United States offers colossal tax credits to cloud giants for installing server farms in rural areas desperate for jobs. China builds data centers in Xinjiang, coal-powered, to fuel its AI models.
Meanwhile, environmental regulations apply to traditional factories, not data centers. A cement producer must offset its emissions. A cloud operator can consume the equivalent of five nuclear plants with no equivalent obligation. Digital enjoys a cultural exemption: it's "clean," "immaterial," "the future." When in reality, it's one of the planet's most energy-hungry industries.
And there's the geopolitical question. Europe imports 80% of its cloud services from the United States. Every dictation, every AI query, crosses the Atlantic. It's disguised energy dependence: we pay not only in dollars, but in kilowatt-hours transported under the ocean. A digital sovereignty that would also be energy sovereignty remains to be built.
Local AI: digital sobriety
Against this excess, a radically different approach is emerging. Local artificial intelligence—models that run entirely on your workstation, with no internet connection, no remote server, no planetary round-trip.
The principle is simple: your computer is already on. You're already using it. Why not run AI where it's used, rather than sending your data on a journey of thousands of kilometers?
Tools like VoiceInk (macOS, open source, 100% offline), OpenWhispr (cross-platform, local Whisper), Whisper.cpp (the reference C/C++ engine used by most local tools), or Jan.ai (desktop interface for local LLMs) embody this philosophy. They prove that professional results are achievable without mobilizing a planetary infrastructure.
PerkySue also operates entirely offline. Whisper for speech recognition, llama.cpp for AI transformation—all on your workstation. No data leaves your computer. No remote server is solicited. No data center activated for your email.
The ecological gain is threefold:
Zero network traffic. Your voice doesn't circle the globe. It stays in your workstation's RAM, processed in milliseconds, then immediately erased. No TCP/IP packets, no intercontinental routing, no network infrastructure energy consumption.
Optimization of existing resources. Your computer consumes electricity whether it's running at 10% or 90% capacity. Running a local AI model uses resources that would otherwise be wasted on idle or background tasks. It's energy recycling.
Hardware longevity. A local model doesn't depend on external infrastructure that changes every two years. Your workstation, your models, your tools—they age together, without vendor-planned obsolescence.
The numbers speak
A typical cloud transcription request consumes about 0.3 to 0.5 grams of CO₂. That seems negligible—until you multiply it by millions of daily dictations. A company of 100 employees, each dictating 20 emails per day, generates roughly 200 kg of CO₂ per year from cloud transcription alone. The equivalent of a round-trip flight from Paris to London.
The same company with a local solution like PerkySue? Zero grams of CO₂ from network traffic. The electricity consumed is that of the already-powered workstation, with a marginal footprint near zero compared to the cloud.
And this calculation doesn't even account for LLM queries. A call to a cloud model generates 10 to 100 times more emissions than a modest local model. Giants with 100+ billion parameters consume dozens of watts per query. A local 7-billion-parameter model, optimized, uses less than an LED bulb.
Pareto's law applied to ecology
Skeptics will object that local models are less powerful. True, in absolute terms. But eighty percent of our daily uses—email drafting, text reformulation, translation, note-taking—are perfectly handled by "modest" models.
Why mobilize a planetary infrastructure for a task your workstation can accomplish in silence? It's like taking a plane to buy bread at the corner bakery. Technically possible. Ecologically absurd.
The hybrid approach—local by default, cloud by exception—drastically reduces carbon footprint without sacrificing productivity. Reserve external tools for cases where they bring real, irreducible value. Use the power at your fingertips for everything else.
Tools like Note67 (transcription + meeting summaries, all local on Mac), Scriberr (self-hosted file transcription via Docker), or Screenpipe (continuous local capture with search) show this approach scales. Each in their own way, they prove local isn't a technical niche—it's a credible alternative.
Toward collective technological responsibility
The stakes extend beyond the individual. When companies adopt local tools, they collectively reduce pressure on cloud infrastructure. A city whose administrations use free local software saves tons of CO₂. A country whose businesses master their AI tools reduces its energy dependence on foreign data centers.
Open source is the foundation of this ecological resilience. Accessible, modifiable, auditable source code guarantees that the tool will outlive its creator without requiring new infrastructure. That a community of developers can optimize it, make it lighter, faster, more sober. That no company can unilaterally impose an update doubling energy consumption.
PerkySue, under Apache 2.0 license, embodies this philosophy. Every line of code is visible. Every user can verify what the tool consumes—and see that it consumes almost nothing compared to a cloud alternative. A competent developer can modify the code, reduce the footprint, create an even lighter version. The tool belongs to its community, not to a shareholder maximizing energy profit.
The choice of ecological intentionality
Adopting local AI is not a rejection of technology. It's a choice of ecological intentionality. It's consciously deciding which resources deserve to be mobilized, which emissions we're willing to generate, which infrastructural dependencies we're willing to create.
It's also a question of real cost. A $15/month cloud subscription seems modest. But it hides an ecological cost that nobody bills: the CO₂ emitted, the water consumed, the materials extracted. Local AI offers an alternative where the ecological cost is transparent, measurable, and above all—minimal.
Tools like VoiceInk (macOS, open source, 100% offline), OpenWhispr (cross-platform, local Whisper), Whisper.cpp (the reference C/C++ engine), Jan.ai (desktop interface for local LLMs), Scriberr (self-hosted via Docker), or PerkySue (Windows, Apache 2.0, 16 languages, integrated TTS) embody this philosophy. Each in their own way, they prove digital sobriety isn't a utopia—it's a technical choice available today. Which tool you choose depends on your operating system, your specific needs, your appetite for configuration. What matters isn't which one you adopt, but that you adopt local.
Next time you hit a shortcut to dictate text, ask yourself: where does my voice go? What distance does it travel? What footprint does it leave? If the answers trouble you, perhaps it's time to choose sobriety.
About the Author
Jérôme Corbiau is the creator of PerkySue, a local voice dictation tool with AI that works entirely offline, with no remote server or data transmitted. He is also co-founder and software architect of My App Zone SRL (Brussels), and creator of the Cloud Neareo platform — an award-winning CMS notably by Microsoft and the Public Service of Wallonia, deployed in museums and heritage sites. His work aims at a constant objective: putting technology at the service of the user, rather than the reverse.
P.S. — If local voice dictation interests you, I've open-sourced what I use daily: github.com/PerkySue/PerkySue. No account, no cloud, just a hotkey. Windows only for now — and I know that's a limitation.
Top comments (0)