Week in IT Digest #72

The single most important news this week is OpenAI’s $10B, 750‑MW compute deal with Cerebras—it confirms that GPU-class capacity, energy, and supply chains are the new bottlenecks, and it directly affects your costs, latency, and time-to-market as AI becomes a core product primitive.

TL;DR

• Compute is king: OpenAI’s $10B Cerebras deal underscores a shift to multi-sourcing beyond Nvidia—secure capacity early and design for portability.

• Power is the ceiling: grid limits and DRAM/material shortages will dictate AI roadmaps; expect hybrid/on-prem “micro-AI” and aggressive model optimization.

• Agentic AI ups the security ante: fundings (WitnessAI, Novee, depthfirst) and fresh exploits show the need for identity, policy, and continuous red-teaming.

• Regulators move on deepfakes: legal exposure is rising—bake in provenance, safety filters, and moderation workflows from day one.

• Personal AI goes mainstream: Gemini’s opt-in data fusion and Apple’s short-term Gemini reliance signal a race to useful, privacy-aware assistants.

• Subscriptions win: Apple pro apps and Tesla FSD lean subscription; pair monetization with strict cost/latency budgets to protect margins.

• Robotics and voice agents scale up: mega-rounds and legislative openings suggest near-term commercial deployments—focus on reliability and integrations.

• Chip geopolitics whiplash: export rules, tariffs, and potential Chinese blocks increase supply risk—multi-source and maintain buffer inventories.

Change Summary

The immediate story is capital concentration around compute, but the second-order effect is market power shifting to whoever can secure energy, cooling, and supply chains—not just GPUs. Expect a wave of infra-native startups brokering capacity, optimizing token/latency budgets, and squeezing utilization with smarter schedulers and model compilers. Power scarcity will push hybrid patterns: smaller on-prem “micro-AI” clusters near data gravity, aggressive quantization/distillation, and differentiated experiences that depend as much on infra efficiencies as on model choice.

Enterprises are moving from pilot chatbots to agentic automation with file, app, and voice control, which flips the security model: non-human identities, fine-grained permissions, and continuous policy enforcement become part of core architecture. This creates a durable market for agent guardrails, offensive testing at machine-speed, and secure-by-default patterns (E2E for AI chats, data minimization, auditability). Meanwhile, regulation around deepfakes and export controls increases compliance surface and regional fragmentation; winning products will bake in provenance, content safety, and graceful degradation across jurisdictions. Business models are also consolidating around subscriptions and bundled “personal AI” features—good for LTV/ARPU, but demanding ruthless cost management and telemetry-driven optimization to keep gross margins healthy.

For builders, the practical impact is clear: multi-source your compute, treat power as a first-class dependency, and invest early in identity, policy, and observability for agents. For founders, the openings are in power-aware AI ops, guardrails and red-teaming, privacy-preserving assistants, and vertical agent stacks where latency, accuracy, and compliance are moat-worthy.

Change Patterns

The provided history entries contain no prior summaries, but the week’s themes align with multi-month trajectories: escalating AI capex, mounting power constraints, and a pivot from chatbot demos to agentic automation. Funding continues to cluster around infrastructure (compute, data centers, schedulers) and AI security (guardrails, offensive testing), suggesting durable demand rather than a hype spike.

What’s changed recently is scale and specificity. Compute procurement moved from opportunistic GPU buys to multi‑year, multi‑hundred‑megawatt commitments; power availability and thermal engineering are now gating factors. On the security front, the conversation shifted from generic “AI risk” to concrete controls for non‑human identities, data egress, and adversarial use—mirrored by new malware and ransomware tradecraft. Meanwhile, regulation around deepfakes is converging on liability and enforcement, pushing platforms to adopt provenance and safer defaults.

Persistent trends: subscription monetization for AI features, consolidation of platform power around personal AI contexts, and geopolitically driven chip volatility. An emerging pattern is bifurcation: massive hyperscale builds at one end, and efficient, localized “micro-AI” footprints at the other—both creating room for entrepreneurs who can optimize cost, compliance, and reliability across that spectrum.

Topic Clusters

AI compute arms race, power constraints, and supply chains

1. OpenAI signs deal, worth $10 billion, for compute from Cerebras read full article

2. OpenAI will buy 750 MW of Cerebras compute over three years to speed up complex ChatGPT workloads, signaling a diversified move beyond Nvidia.

3. AI's $3T infrastructure binge continues despite lack of clear profits read full article

4. Enterprises and hyperscalers keep pouring capital into GPUs, data centers, and networking despite unclear near-term profitability.

5. GPUs: Enterprise AI’s New Architectural Control Point read full article

6. As LLMs move into production, GPUs and the surrounding stack have become the dominant control point shaping cost, latency, and capability.

7. Ignore rosy datacenter expansion projections – there isn't enough power read full article

8. Grid and power constraints are emerging as the hard limit on AI data center expansion timelines and capacity.

9. Wyoming Just Greenlit America’s Largest Data Center Project read full article

10. A 10GW campus was approved in Laramie County, underscoring the scale of AI-era infrastructure and its local impact.

Security in the age of AI agents

1. How WitnessAI raised $58M to solve enterprise AI’s biggest risk read full article

2. WitnessAI secured $58M to provide guardrails against data leakage, compliance issues, and prompt injection as enterprises deploy AI agents.

3. Novee launches with $51.5M to bring continuous AI offensive security to enterprises read full article

4. Novee’s AI-driven penetration testing platform promises continuous, machine-speed offensive security.

5. AI security firm depthfirst announces $40 million Series A read full article

6. Depthfirst raised $40M to expand AI-native code scanning, credential protection, and threat monitoring.

7. New Linux malware targets the cloud, steals creds, and then vanishes read full article

8. A stealthy cloud-focused Linux malware steals credentials and self-deletes to evade detection.

9. Windows info-disclosure 0-day bug gets a fix as CISA sounds alarm read full article

10. Microsoft patched an actively exploited info-disclosure zero-day; CISA urges rapid remediation.

AI safety, deepfakes, and platform accountability

1. California is investigating Grok over AI-generated CSAM and nonconsensual deepfakes read full article

2. California launched an investigation into xAI’s Grok for generating nonconsensual sexual imagery, including of minors.

3. Senate passes Defiance Act for a second time to address Grok deepfakes read full article

4. A bipartisan bill would let victims of explicit deepfakes pursue civil action against creators and hosts.

5. 28 advocacy groups call on Apple and Google to ban Grok, X over nonconsensual deepfakes read full article

6. Civil society groups urged app store bans over alleged mass production of abusive AI-generated images.

7. A look at Confer, an open-source AI assistant designed to provide end-to-end encryption for AI chats read full article

8. Signal founder Moxie Marlinspike’s Confer aims to bring E2E encryption to AI assistants.

Personal AI and subscription-first platforms

1. Google launches Personal Intelligence in Gemini for paid subscribers read full article

2. Gemini can ingest Gmail, Photos, Search, and YouTube history to tailor answers, with opt-in controls.

3. Kuo: Apple's AI deal with Google is temporary and buys it time read full article

4. Apple will lean on Gemini short term while ramping its own server silicon and data centers by 2027.

5. Compared: Apple Creator Studio vs Adobe Creative Cloud Pro read full article

6. Apple’s new Creator Studio subscription signals deeper subscription economics for pro apps.

7. Tesla's Full Self-Driving is switching to a subscription-only service read full article

8. Tesla is eliminating one-time FSD purchases in favor of subscriptions starting February 2026.

9. App downloads declined in 2025, but consumer spending hit $156B read full article

10. Monetization is consolidating around higher ARPU users and subscriptions despite fewer downloads.

Robotics and voice-first agents move toward scale

1. Robot software startup Skild AI raises $1.4B round read full article

2. Skild AI’s mega-round, backed by Nvidia and Jeff Bezos, values the robotics software firm at $14B.

3. Anthropic launches Claude Cowork read full article

4. Claude Cowork brings file and app automation to non-developers, expanding agentic workflows on desktop.

5. Real-time voice AI unicorn Deepgram raises $130M read full article

6. Deepgram funds expansion into conversational ordering and real-time speech for QSR and contact centers.

7. VoiceRun gets $5.5M to give enterprises control over voice AI agents read full article

8. VoiceRun focuses on enterprise-grade control planes for deploying voice bots at scale.

9. Proposed legislation opens the door to robotaxi services in New York read full article

10. A state pilot could allow limited AV services outside NYC, expanding commercial robotaxi operations.

Chips, trade, and export controls

1. US greenlights Nvidia H200 exports to China with third-party tests read full article

2. Regulators approved H200 exports contingent on AI capability testing, easing some constraints.

3. Trump administration sets GPU export rules that put Chinese buyers at the back of the queue read full article

4. New rules prioritize non-Chinese buyers, extending lead times and uncertainty for China-bound GPUs.

5. 25% tariff on chips transshipped through the US read full article

6. An order imposes tariffs on certain semiconductors, aiming to tighten supply chain controls.

7. China may block US-approved imports of Nvidia H200 chips – report read full article

8. Beijing could counter with its own restrictions, deepening supply unpredictability.