Palantir, Nvidia, and Booz Allen Restrict AI Model Use Over IP Concerns
The Information reports that large tech firms including Palantir Technologies, Nvidia, and Booz Allen Hamilton are demanding new guarantees from Anthropic and OpenAI — or reducing and eliminating use of their advanced models — over fears that proprietary data could train future AI systems. The catalyst: Anthropic's June policy change allowing data retention from its Claude Fable 5 model "to ensure the model isn't being misused." Palantir CEO Alex Karp's position is stark, claiming AI companies are "stealing customers' data while charging them for unproductive tokens." The enterprise paranoia is real: companies with sensitive corporate work now treat model vendor concentration as existential risk. Some have begun limiting which models employees can use and for which tasks.
Read on Tom's Hardware →Enterprise AI Hits Infrastructure Limits — Power Grids, Not GPUs, Now Bind
Christopher Penn at Trust Insights argues the enterprise AI constraint has shifted from budget to physics: "Power, not GPUs, now binds on-premises AI." U.S. substation lead times now run 40 months to five-plus years, up from 24-30 months pre-2020. Hardware costs fall 30% per year and energy efficiency improves 40% annually (per Stanford AI Index), yet absolute AI energy consumption keeps climbing as model scale outpaces both curves. Penn's recommendation: stand up on-premises inference hubs running small open-weight models. "Most agents do not need frontier models; templated agent work runs guilt-free on local hardware and only costs electricity." The strategic frame: if you sit purely in the cloud and own no inference capability, "the vendor owns your exit, your data, and your cost line."
Read on Trust Insights →AEO Emerges as the New Marketing Discipline — Gemini Cites 10 Brands Per Response
NP Digital published the sharpest data point yet on Answer Engine Optimization (AEO): across more than 68,000 AI-generated answers to fintech prompts, Google Gemini named an average of 10 brands per response versus 6 for Perplexity. Nikki Lam argues a slot on Perplexity is therefore worth more — fewer mentions mean higher share of voice. Three AEO/GEO vendors announced products in a single day: Azoma for enterprise agentic commerce, Indexa for measurement, and NP Digital's research. The Agile Brand Guide critique: none published conversion rates, click-through rates, or revenue figures. At $60K/year licensing plus $30K analyst allocation, the program needs ~$450K in incremental revenue to break even. "You are buying a position metric and supplying the revenue assumption that justifies it."
Read on Agile Brand Guide →Forbes: B2B Marketing Leaders Face the Cost vs. Performance Choice
AI is creating a "marketing productivity dividend" for B2B organizations, but Forbes Council contributor analysis reveals the uncomfortable fork in the road: do you use AI to reduce costs or improve performance? Most teams try both and achieve neither. The diagnosis: unclear objectives, tool sprawl without integration, and skills gaps between AI capabilities and marketing execution. The guidance: pick a lane. Cost reduction means automating repetitive tasks and reducing headcount. Performance improvement means augmenting human judgment with AI insights and reallocating savings to higher-value work. The warning: trying to do both simultaneously usually delivers marginal gains in both directions while creating organizational confusion about what success looks like.
Read on Forbes →DeepSeek V4.1 Flash Ships, Engineer Compares Anthropic AGI Lead to "Hitler with A-Bomb"
DeepSeek shipped V4.1 Flash this week, and engineer Shengyu Liu posted a viral WeChat essay calling Anthropic gaining an unchecked AI lead equivalent to "Hitler getting the atomic bomb before the Allies." Liu frames staying at DeepSeek as a bet on open source as counterbalance to concentrated frontier AI. SCMP treats the essay as the sharpest Chinese-industry response yet to Dario Amodei's pacing call, days before an expected Xi-Trump meeting where AI safety may surface. Meanwhile, Scott Aaronson blogged that unconfirmed rumors suggest frontier labs are sitting on solutions to "very longstanding open problems in theoretical computer science" beyond Navier-Stokes and are withholding results. The geopolitical stakes around frontier AI — and who controls it — keep escalating.
Read on AI Weekly →💡 My Take
The Palantir story is the tell. When Nvidia — the company that literally makes the chips powering AI — starts restricting model use over IP concerns, the enterprise trust problem has metastasized. Combine this with Chris Penn's infrastructure argument and a pattern emerges: the "AI platform vendors own everything" model is hitting resistance. Large enterprises want control — over their data, their inference costs, their exit options. The AEO/GEO gold rush feels like SEO 2005: real opportunity buried under measurement theater. And the DeepSeek engineer's hyperbole aside, the open-source-as-counterbalance argument gains weight every time a frontier lab changes its terms of service. The next 12 months may determine whether enterprise AI becomes a vendor-controlled utility or a capability organizations actually own.