AMD and Cerebras tag-team Nvidia's LPUs while Uncle Sam gets schooled on open AI
Curated by the Inblix editorial team
Two distinct fronts opened this week in the AI infrastructure wars. First, AMD and Cerebras announced a partnership that pairs AMD’s MI300X GPUs with Cerebras’s wafer-scale engines. The pitch is straightforward: give data centers a unified fabric that can handle both training and inference without being locked into Nvidia’s proprietary NVLink ecosystem. Cerebras CEO Andrew Feldman didn’t mince words — “The enemy of my enemy is my friend” — which tells you everything about who they’re gunning for. Spec-wise, AMD also dropped details on its Helios rack-scale system, claiming it beats Nvidia’s upcoming Vera Rubin on nearly every metric. On paper, anyway. Whether those numbers survive contact with real workloads is the multi-billion-dollar question.
Meanwhile, a coalition of industry leaders fired off a letter to the US government arguing for the strategic importance of open-weight AI models. The signatories — a who’s-who of AI research orgs and startups — want policymakers to understand that open models aren’t just charity for academics. They’re a competitive moat against proprietary systems controlled by a handful of companies. The letter’s framing is clever: national security, economic resilience, and innovation all hinge on keeping weights accessible. Conspicuously absent from the signatory list? The obvious suspects who profit most from closed, API-gated models.
On the security beat, things are getting weird. Russian operatives are now impersonating Signal’s support team to phish users — a reminder that encrypted comms are only as secure as the humans using them. Microsoft’s on-prem SharePoint is under active zero-day attack after patches failed to stick the landing. And in a grim milestone, Mikko Hyppönen noted it’s been exactly ten years since the first corporate ransomware attack, with no end in sight. His silver lining: infosec careers are recession-proof.
In open-source land, a quiet rebellion is brewing. Codeberg publicly booted “vibe-coded” AI-generated projects from its platform, drawing a hard line around human-authored FOSS. The message: if you didn’t write it, don’t commit it. GNOME’s Flashback desktop got a Windows-like taskbar that works without extensions, Cinnamon 6.8 teased optional Wayland support, and Debian shipped what may be its final x86-32 release. The old guard isn’t going quietly.
💡 Key Takeaways
- AMD's Helios rack specs beat Nvidia's Vera Rubin on paper, but paper specs don't run production workloads — the real test comes when these systems hit data centers.
- The open-weight AI letter to the US government is as much about market competition as it is about national security, and the companies that refused to sign it reveal the real battle lines.
- Codeberg's decision to ban AI-generated code is a signal that the open-source community is willing to fragment over the definition of authorship, not just licensing.
Keep reading: See related articles below for more coverage on this topic.
Get smarter about AI
The sharpest AI news, curated daily. Delivered free to your inbox.