Microsoft tells employees to stop tokenmaxxing, sets division-level AI budgets
THE SO WHAT
When Microsoft tells teams to stop “tokenmaxxing” and moves to division-level AI budgets, it’s admitting that unconstrained LLM usage can quietly blow up cloud bills. If you’re scaling AI internally, you need metering, cost dashboards, and usage guardrails now — not after finance calls.
READ THE SOURCE
MORE FROM THE WIRE
Applied AIOpen-weight AI models are catching up to the frontier. The safety gap remains.
Open-weight GLM-5.2 approaching frontier capability without comparable safety mitigations means powerful models are no longer gated by a few labs’ policies. If you ship on-prem or customer-controlled AI, you now have to assume clients can — and some will — swap in far less-governed models, so build your own guardrails and monitoring accordingly.
Applied AICracken opens self-serve access to its AI-powered offensive cybersecurity platform
Offensive security is going product-led — Cracken’s self-serve AI red-teaming puts real attack-path simulation in the hands of individual practitioners. CISOs should expect more internal teams to run their own tests and need governance on who can launch what against production-like environments.
Applied AIAnthropic signs $10 billion deal with AI cloud startup Volta
Anthropic locking in a reported $10B deal with Volta shows that AI-native cloud providers can win massive, multi-year commitments — compute is fragmenting beyond the big three hyperscalers. If you’re planning large-scale training or inference, your vendor map should now include specialized AI clouds with different economics and hardware mixes.
Applied AIAmazon admits it accidentally shelled out $1.8 million for Claude to finish its menial coding tasks
Blowing an 860% budget on Claude Sonnet for “menial” coding tasks is a clean example of AI spend running ahead of cost controls and workload design. If you’re rolling out coding assistants, you need hard usage caps, per-team cost dashboards, and clear rules on when to use premium models versus cheaper tiers.