AMD Threadripper Halo Station: 96-Core AI Workstation with MI350P
At IFA 2026, AMD announced the Threadripper Halo Station, a workstation packing a 96-core Threadripper Pro 9995WX and dual liquid-cooled MI350P accelerators, designed to run trillion-parameter models. AMD claims it's "the most powerful workstation in the world," though no OEM partners have been announced yet.
Core specs
- CPU: Threadripper Pro 9995WX — 96 Zen 5 cores, 192 threads, boost up to 5.4 GHz, 384 MB L3 cache, 350W TDP.
- Accelerators: Dual liquid-cooled Instinct MI350P, with a path to four.
- Memory: 2 TB DDR5 system RAM.
- HBM: 288 GB HBM3E on the GPUs, with up to 576 GB supported (presumably with four accelerators).
The machine essentially reconfigures a server tray into a tower, swapping an EPYC host for the Threadripper. AMD has not detailed the full system design or pricing, but street price for just the CPU, GPUs, and 2 TB of RAM is over $100,000. A fully configured unit—with storage, power, and cooling—could easily exceed $150,000.
AMD hasn't announced which OEMs will build and ship the system. Given the price and target workload (running trillion-parameter models locally), this is squarely aimed at AI research labs and enterprise teams that need on-prem inference without cloud round-trips.
📖 Read the full source: HN AI Agents
👀 See Also

Agentic AI Failure Modes and Developmental Scaffolding
Agentic AI systems fail in production through alignment drift, context loss across handoffs, boundary violations, and coordination collapse. The source proposes a 'developmental scaffolding' approach with five components: coherence monitoring, coordination repair, consent and boundary awareness, relational continuity, and adaptive governance.

Startups Report Spending More on AI Compute Than Human Salaries
AI startups like Swan AI report monthly AI compute bills exceeding $113k, with CEOs describing this as 'tokenmaxxing' where AI spending replaces traditional headcount budgets.
Claude Code v2.1.224 Adds Self-Hosted Runners, Cross-Session Messaging
Anthropic's Claude Code v2.1.224 adds self-hosted environments via `claude self-hosted-runner`, cross-session messaging, archive plugin source, and fixes a bunch of sandbox and remote control issues.

Cerebras releases Step-3.5-Flash-REAP models with 40% memory reduction
Cerebras has released Step-3.5-Flash-REAP models that use REAP (Router-weighted Expert Activation Pruning) to compress 196B parameter models to 121B while maintaining near-identical performance. The models work with vanilla vLLM and are optimized for resource-constrained environments.