What happened
Perplexity CEO Aravind Srinivas told CNBC on September 4 that banks like Morgan Stanley and JPMorgan want "air gapped implementation" with AI running on their own premises.
Why it matters
He argues data centers alone can't support billions of AI agents needing a terawatt of power, pointing to NVIDIA's DGX Spark for local inference.
What to watch
Whether an on-prem shift compresses hyperscaler capex behind NVIDIA's Q3 FY27 revenue guidance of $108B, or just routes it differently.
WHO IT HITSEnterprise IT leaders and infrastructure planners at financial institutions, along with cloud providers and NVIDIA's data center business, face potential shifts in where AI compute is deployed as banks demand air-gapped local systems.
Ask the AI about this article →
Summaries like this, in your inbox every morning.
Srinivas's argument is that a future with a billion always-on AI agents would need a terawatt of power, which data centers alone cannot provide. He sees untapped compute in the RAM and power of offices and homes, a view that challenges the narrative centered on massive cloud infrastructure.
NVIDIA is positioned on both sides of this equation. Its Data Center revenue reached $89.02B (+117% YoY), driven by hyperscaler capex, but CFO Kress also highlighted on-prem strength and non-hyperscaler categories making up half of the business. The company is mobilizing over $500B with partners for centralized AI infrastructure while also selling DGX Spark for desks.
The key question is whether the bank demand for air-gapped systems becomes a significant enough trend to route spending away from hyperscalers, or whether it simply becomes a different line item on NVIDIA's invoice. With Q3 FY27 guidance at $108B, NVIDIA's bet is that it sells the silicon regardless of where the workloads run.
For example, today's edition would include:
AI-summarized, only the topics you pick — one digest a day via Email, Slack, or Discord.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. Q&As are published on this page for other readers too.
Meta will no longer factor AI dashboards or token counters into engineer performance reviews, executives Maher…
Pampa Energía plans a facility of up to 500 megawatts in Neuquén province, seeking investors with first contra…
SoftBank introduced an AI system for video interview screening in May 2020, cutting screening time by about 70…
OpenAI's Astra model is now available, beating the ARC-AGI-3 benchmark and scoring the highest on Zapier's Aut…
Mistral AI raised €3 billion (about $3.58 billion) at a post-money valuation of more than €21 billion (about $…
A new OpenAI blog post announces that its GPT‑6 Astra model can now act as a "semi-autonomous agent" to handle…
