Skip to content
← Skygena Signal
trends 6 min read LLM-drafted · human-edited

Nvidia CEO Under Investigation as OpenAI Hits $1.25T Valuation

Jensen Huang faces a board probe over China spy links, OpenAI launches GPT-6 Cyber and raises at $1.25T, and Anthropic's self-improving AI loop is now empirically documented.

by Skygena Editorial (LLM draft · human reviewed)

The week Nvidia’s CEO became the story, OpenAI crossed into trillion-dollar territory, and a safety report on Anthropic’s self-improving AI loop raised the kind of questions nobody wants to answer on a Friday afternoon.

The story of the week

Jensen Huang, the leather-jacketed face of the AI hardware boom, is now the subject of an independent board investigation after reports emerged that China’s Ministry of State Security recruited Megaspeed executive Alice Huang to gather intelligence on American AI chips — and, reportedly, to influence Nvidia’s posture on exporting advanced hardware to China. Huang sent a memo to employees acknowledging a personal relationship with Alice Huang. The board’s statement is careful and legalistic; the memo is not.

This landed in the same week as a separate story about US export controls: the Commerce Department apparently approved Nvidia H200 chip sales to a Chinese firm whose offices turned out to be an empty shell. Taken together, the two stories paint a picture of an export-control regime that is both politically fraught and operationally porous. European operators who depend on Nvidia silicon should watch the board investigation closely — not for the tabloid angle, but because any leadership disruption at Nvidia ripples through the entire supply chain. If you are negotiating GPU allocations or DGX contracts, factor in the possibility that Nvidia’s governance situation becomes a distraction that slows commercial decision-making for months.

New models & capabilities

OpenAI: GPT-6 Cyber and Images 2.6. OpenAI released GPT-6 Cyber, a model purpose-built for finding exploitable vulnerabilities and developing fixes, trained to stay within authorised scope. A companion post describes an autonomous pipeline that found and patched 948 privilege-escalation vulnerabilities in the Linux kernel — a striking demonstration of what “offensive-by-design, defensive-by-intent” looks like in practice. The model is also available in AWS GovCloud via Amazon Bedrock for US government customers. Separately, OpenAI launched Images 2.6 with faster generation, GIF creation, and new API-level editing controls.

OpenAI acquires Pierre Computer Company. The acquisition expands Codex’s tooling for code review and understanding — a signal that OpenAI sees the developer-tools layer as strategically important, not just a demo surface.

Google DeepMind: Gemini Omni 1.2 Flash. Google shipped Gemini Omni 1.2 Flash, bringing native 15-second video generation and editing, stronger visual consistency, and a 30 per cent price cut. For teams building multimodal applications on a budget, this is the model to benchmark against.

Google DeepMind: Antigravity. A quieter release: Google Antigravity with Cheminformatics Skills accelerates drug-discovery workflows, including compound search and novel macrocycle design. Pharma teams should pay attention.

Research worth knowing

Anthropic’s recursive improvement loop. The most consequential read this week is not a product launch. METR, an AI safety research organisation, published a report after an 11-week investigation into Anthropic’s use of AI to accelerate its own model development. The finding: Anthropic’s models are improving roughly twice as fast as a year ago, with “little indication of slowing down.” This is the self-improving loop that safety researchers have long warned about — except it is now empirically documented rather than hypothetical. Whether you find this thrilling or alarming depends on your priors, but it should sharpen any enterprise’s thinking about how quickly today’s model selection could become obsolete.

Reverse game theory for alignment. MIT Technology Review covered a mechanism-design approach — “reverse game theory” — that aims to control powerful AI agents by designing incentive structures that make rule-following the dominant strategy. Still academic, but relevant for anyone building multi-agent systems where you need agents to behave predictably.

Securing the agentic enterprise. A separate MIT Tech Review piece on data lineage in agentic workflows describes a familiar failure mode: agents pulling from restricted pricing systems and internal communications without human review, producing outputs that look right but aren’t authorised. If you are deploying agent chains in production, this is required reading.

CEO watch

OpenAI at $1.25 trillion. OpenAI raised another $11 billion at a $1.25 trillion valuation, up from $852 billion in March. Anthropic is reportedly nearing its IPO. These numbers have entered the realm where they say more about capital markets than about technology, but they matter operationally: vendors at this scale have every incentive to lock in enterprise customers early and tightly.

Apple pushes for agentic Siri. New CEO John Ternus is pressuring Craig Federighi to accelerate Siri’s agentic capabilities to compete with OpenAI’s Dots and Meta’s Muse. Apple’s installed base of devices makes any real agent capability instantly relevant to consumer-facing businesses in Europe.

Cybersecurity pressure. The Chainalysis breach — a “multifront cyberattack” involving AI agents that compromised credentials for every customer — underscores why OpenAI is investing so heavily in GPT-6 Cyber and partnering with Hacktron for vulnerability testing. The attack surface is growing faster than most security teams can staff for.

What it means for European operators

Nvidia governance risk is supply-chain risk. The Huang investigation may or may not result in leadership change, but the uncertainty alone should prompt European buyers to diversify GPU sourcing conversations. AMD MI400 and Intel Gaudi 3 deserve a second look, even if Nvidia remains the default.

Anthropic’s acceleration curve demands shorter planning horizons. If models are improving at double last year’s rate, a 12-month AI strategy is too long. European enterprises should plan in six-month cycles and structure vendor contracts with model-upgrade clauses.

Agentic security is not optional. Between the Chainalysis breach, the METR report, and MIT Technology Review’s data-lineage warning, the week delivered a clear message: deploying autonomous agents without lineage tracking and scope constraints is negligent, not innovative. European firms subject to the AI Act should treat data lineage as a compliance prerequisite, not a nice-to-have.

OpenAI’s cyber play has enterprise implications. GPT-6 Cyber is US-government-first, but its architecture — autonomous vulnerability discovery with authorised-scope guardrails — will shape what enterprise security vendors build next. European CISOs should start asking their vendors how they plan to integrate or compete with these capabilities, particularly under NIS2 obligations.

Heimstaden as a European case study. Buried in OpenAI’s content marketing this week: Heimstaden, a Nordic residential property manager, is using ChatGPT Enterprise and Codex across operations. It is one of the few named European enterprise deployments OpenAI has published. Worth studying as a reference architecture, even if your sector is different.

Sources

Thinking about AI in your business?

Skygena is a boutique European AI studio engineering autonomous agents and LLM products. If you're wrestling with where to start — or where to stop — we can help.

Book a 30-minute call