Kog is going deeper to squeeze more inference out of GPUs
THE SO WHAT
If Kog is right that GPUs can handle agentic workflows with smarter scheduling and kernel-level tricks, the near-term constraint shifts from hardware scarcity to software efficiency. Before you budget for more GPUs, pressure vendors on how they handle long-context, multi-step agents on today’s cards.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIAI Companies Work for Better Data, Not Better Models
If Anthropic is reportedly in talks to buy a data startup like Decart for $6B, the center of gravity is shifting from model architecture to proprietary, high-signal data streams. For most teams, the edge won’t be “our fine-tuned model” but “our rights, pipelines, and feedback loops around domain-specific data.”
Applied AITech Visionary Says the Big AI Labs Don’t Get What People Want
Tim O’Reilly’s critique—paired with his backing of open source—captures a growing gap between centralized assistant platforms and what developers actually want: control, transparency, and local leverage. If you’re building on AI, hedge your stack with at least one open model path so you’re not fully exposed to a single lab’s product thesis.
Applied AIArm’s co-founder says AI will create more value than any revolution before it, and be a rollercoaster
Hauser’s point isn’t that AI is big — it’s that the value curve will be violent, with Europe at risk of becoming a ‘technology colony’ again if it doesn’t own more of the stack. European operators should be treating AI infra, IP, and talent retention as sovereignty questions, not just innovation theater.
Applied AIFive of Europe’s biggest companies just bought compute Mistral has not built
Mistral selling future ‘compute units’ to ASML and other giants is pre-selling capacity like a power plant — locking in demand and financing data center buildout before the racks exist. If you’re a large buyer, expect more vendors to push you toward long-term capacity contracts rather than pay-as-you-go API spend.