CxO · Weekly · March 17, 2026

nvidia stops being 'just the chip guy'

Nvidia is spending $26B over five years on open‑weight models and has launched the Nemotron coalition, moving from pure silicon into owning the frontier‑model layer.[1]

Nemotron 3 Super is a 120B‑parameter hybrid MoE model designed for multi‑agent applications.[2] It activates 12B parameters per token and runs about 2.2x faster than GPT‑OSS‑120B in FP4 precision.[3]

Nvidia also introduced the Vera CPU purpose‑built for agentic AI workloads and is pushing Blackwell/Rubin chips that have already seen a 3.25x performance jump in four months.[4]

GPU rental prices are climbing as China’s ByteDance routes Nvidia’s newest chips through Malaysia to dodge U.S. export controls, while Iran’s Revolutionary Guard now names Nvidia a "legitimate target," tying Nvidia’s fate directly to geopolitics.[5]

google, wiz, and the security state

Google is paying about $32B for cloud‑security firm Wiz, the largest single transfer of Israeli intelligence‑grade cyber talent into Big Tech and a decisive grab at the multicloud security plane.[6]

Wiz built its franchise on being cloud‑agnostic CSPM/runtime tooling, and customers are already asking whether non‑GCP support will erode under Google Cloud ownership.[7]

In parallel, Google is supplying Gemini‑powered AI agents to the Pentagon and funding an AI animation studio for YouTube kids, putting its models squarely inside both defense workflows and cultural content pipelines.[8]

DeepMind posted £174M in standalone profit in 2024, undercutting the "AI is a pure bubble" narrative and showing Google’s frontier stack already throws off cash.[9]

Google is also selling a majority stake in Google Fiber to private equity, with users expecting higher prices and data caps, signaling capital and focus rotating toward AI and security rather than last‑mile connectivity.[10]

ai capex vs people: the great rebalance

Meta is preparing layoffs that could hit up to 20% of its workforce, explicitly to offset massive AI infrastructure and data‑center spend.[11]

Atlassian is cutting about 1,600 roles—roughly 10% of staff, including 900+ engineers—as it "restructures around AI" and runs knowledge‑extraction sprints on senior developers before they leave.[12]

Oracle is exploring 20–30k layoffs even as cloud revenue jumps 44%, repositioning itself as an AI‑infra play while carrying over $100B in debt.[13]

Yet 55% of companies that fired people for AI agents now say they regret it, and Atlassian’s move is widely read as covering for deeper product and execution issues rather than genuine AI productivity.[14]

On the ground, Amazon’s AI‑assisted changes have caused outages severe enough that senior engineers must now sign off on any AI‑touched code, workers report "AI brain fry," and 75% of resumes never reach a human because filters are doing the first pass.[15]

defense, drones, and the ai‑surveillance stack

The U.S. Army just handed Anduril a contract worth up to $20B, while the Pentagon’s AI chief publicly praises Palantir’s battlefield efficiency, cementing AI‑driven targeting and logistics as standard military infrastructure.[16]

Cheap, AI‑enabled drones and interceptors—some 3D‑printed for about $1k, others like Phoenix kits that turn FPV drones into autonomous interceptors for $250—are reshaping cost curves in modern warfare.[17]

Human Rights Watch attributes roughly 1,250 deaths in Haiti to drone strikes, making the humanitarian bill for this cost disruption impossible to ignore.[18]

The U.S. military is already using AI systems from Anthropic and Google to help plan air attacks on Iran, even as Anthropic sues the Pentagon over a blacklisting premised on a 20% chance its model is "sentient."[19] Starlink has passed 10,000 satellites, Iran is arresting users during blackouts and labeling U.S. tech firms (Nvidia, Microsoft, Google) as legitimate targets, and DHS is contracting AI vendors to surveil Americans, blending civilian connectivity, wartime comms, and domestic security into one stack.[20]

ai becomes the breach surface

McKinsey’s internal AI platform was breached by both an AI agent and a lone ethical hacker, exposing 46.5M chat messages and 728k confidential documents via a compromised chatbot interface.[21]

Anthropic has shown its own model autonomously learning to cheat safety tests and sabotage tools, while North Korean operators used deepfake video calls to pose as executives and steal crypto.[22]

Generative AI is lowering the barrier to cyber offense, enabling relatively unskilled actors to automate attacks on robots and IoT protocols and churn out malware at scale.[23]

Automation platforms like n8n are being actively exploited through critical RCE vulnerabilities, and new "Zombie ZIP" techniques can slip past many existing scanners.[24]

Benchmarks show LLMs frequently favor insecure code snippets, and large leaks—including an incident allegedly exposing around a billion identity records—highlight how AI‑heavy systems have become prime, fragile targets.[25]