DeepSeek V4 Undercuts AI Giants While France

The AI pricing war just got messier, and honestly? I'm here for it.

DeepSeek dropped V4 on April 24th—open source Pro and Flash variants that are specifically tuned for the developer tools you're probably already using. We're talking Claude Code, Open Claw, Open Code. The release notes basically said "hey devs, we know what you're building with." But the real headline is the pricing: $3.48 per million output tokens at launch, with promotional discounts bringing it even lower through May.

For context, that same volume would run you around $30 on OpenAI and $25 on Anthropic. DeepSeek is undercutting the frontier labs by close to an order of magnitude. Open Claw made V4 Flash its default model two days after launch. Two days. That's not a vote of confidence—that's a sprint away from your current invoice.

The interesting pattern here is model mixing: using cheap models for routine tasks and saving the frontier stuff for actual hard reasoning. DeepSeek just made the cheap end of that mix way more capable. If you're paying frontier API rates for every little function call or JSON extraction, you're basically lighting money on fire at this point.

France goes all-in on digital sovereignty

While we're on the subject of big moves, France's digital agency DINUM is migrating every ministry to Linux. This isn't their first rodeo—the Gendarmerie Nationale (their national police force) has been running a custom Ubuntu build called Genbuntu since 2008. As of last year, it was deployed on over 100,000 workstations, saving roughly 2 million euros annually in licensing fees.

DINUM is now pointing to that as the model for the entire government. The pitch is straightforward: digital sovereignty. Keep data and infrastructure decisions in French hands, not in Redmond or wherever Oracle keeps its lawyers.

What matters here isn't just France making this choice—it's the procurement signal it sends. A G7 country picking Linux at government scale means open source just got a customer with actual budget authority and multi-decade planning horizons. Other governments are watching. Some are probably already reaching out to Canonical.

Ubuntu wants AI to stay on your laptop

Speaking of Canonical: they laid out plans for local inference features in Ubuntu over the next year. The concept is simple—run models on-device instead of round-tripping to a cloud API. Privacy stays local, latency drops, you stop paying per token, and if your network decides to have a moment, the model still works.

Canonical's VP of engineering Joe Seager said the focus will be on open weight models, open source harnesses, and local inference where it makes sense. Distribution happens through snaps, which keeps the same permission model Ubuntu already uses for everything else.

The appeal is obvious if you're building developer tools or internal apps. Sometimes the cheapest, fastest, most private inference is the kind that never leaves your machine. Not everything needs GPT-5 or Claude Opus. A lot of tasks just need a model, and if it's already sitting on your SSD, even better.

The tension here—and this is worth thinking about—is between centralized AI (where all the compute and capability lives in data centers) and edge AI (where you run what you can locally). Both have real advantages. Cloud APIs get you frontier capabilities and instant updates. Local models get you privacy, cost control, and reliability. The next few years are going to be about figuring out which workloads belong where, and Ubuntu is making a bet that a lot more can happen locally than we currently assume.

TanStack AI: the Switzerland approach

TanStack—the team behind some of the most-used libraries in the React ecosystem—shipped an alpha of TanStack AI. Fully open source, framework-agnostic, no service layer, no platform fees. They're calling it "the Switzerland of AI tooling," which is both funny and accurate.

The alpha already supports OpenAI, Anthropic, Gemini, and Ollama. Server libraries in TypeScript, PHP, and Python. Client libraries for vanilla JS, React, and Solid, with Svelte on the way.

What caught my attention: isomorphic tools. You define a tool once with metadata, then provide separate server and client implementations. That gives you type safety across both sides of the connection, which is genuinely rare in this space. Per-model typing means your IDE knows what each model supports—less guessing, fewer runtime surprises.

It's still alpha, so expect sharp edges. But the philosophy is clear: you own your stack. No vendor lock-in. No one else deciding your architecture for you.

Linux drops Intel 486 support (and yes, people noticed)

For the kernel nerds: Linux 7.1 is removing Intel 486 support. The 486 launched in 1989—it's 37 years old. The technical reason is straightforward: the 486 lacks certain instructions that modern kernel code depends on, so Linux has been carrying emulation workarounds for years. That emulation was causing more problems than the architecture was worth.

Linus Torvalds has been arguing for dropping 486 support since 2022. Ingo Molnar's patch series is what finally made it happen. If you're somehow running a current Linux kernel on a 486 (respect, honestly), the long-term support kernels will still work. You're not stranded.

There's something kind of beautiful about a kernel that ran on a 486 in 1991 still running on that same lineage today. Linux outlasted the chip it was born on. That's one hell of a software legacy.

The through line

What connects these stories? They're all about control. France wants control over its digital infrastructure. Ubuntu wants to give developers control over where AI runs. DeepSeek is competing on price, which gives users more control over their budgets. TanStack AI is explicitly designed around giving you control over your stack.

The last decade of tech was about consolidation—everything moving to the cloud, a few platforms dominating, walled gardens everywhere. These moves feel like the pendulum swinging back. Not all the way—centralized services aren't going anywhere—but far enough that "who controls this" is becoming a design question again, not just an afterthought.

The 486 retirement is the outlier here, but even that's about the kernel making an explicit choice about what it's willing to support and what it's not. Sometimes control means saying no.

Yuki Okonkwo is Buzzrag's AI & Machine Learning Correspondent