Weekly Brief 25/2026

Week in a Nutshell

Week 25, 2026 delivered a collision of geopolitics, product launches, and capital markets that underscored just how interconnected the AI stack has become. The US government's abrupt shutdown of Anthropic's Mythos model — reportedly triggered by a jailbreak warning from Amazon — sent shockwaves through the industry and raised urgent questions about the coherence of America's AI export-control regime. Meanwhile, China continued its quiet industrialisation of AI: DeepSeek raised $7.4 billion at a $50 billion valuation, Z.ai's open-weights GLM-5.2 beat frontier US models at a fraction of the cost, and ByteDance explored domestic chip alternatives to Nvidia. On the product front, Microsoft's globally available Copilot Cowork redefined what enterprise AI automation looks like, even as a security vulnerability and a shareholder lawsuit complicated the narrative. Across hardware, robotics, and safety research, the week reinforced a single underlying truth: the race is no longer just about who has the best model, but who controls the infrastructure, the chips, the governance frameworks, and the geopolitical access that determine who gets to use AI at all.

---

Top Stories of the Week

1. US Government Shuts Down Anthropic's Mythos Model Over Jailbreak Fears

In an unprecedented move, the Trump administration issued an export control directive that effectively shelved Anthropic's most advanced models — Fable and Mythos — after Amazon's CEO reportedly alerted the White House to a jailbreak discovered in Fable 5. The order blocked access for many of America's closest allies, drew immediate condemnation from the AI community, and forced Anthropic to dispatch lobbyists to Washington in a bid to reverse the decision. By the end of the week, Trump had cleared Anthropic of being a security threat following compliance assurances, but the episode left the industry shaken.

The incident exposed the chaotic, improvised nature of US AI regulation in 2026. Rather than a structured review process with clear criteria, a single corporate alert triggered a sweeping government order that disrupted commercial contracts, diplomatic relationships, and the broader narrative of American AI leadership. Critics across the political spectrum argued it represented the worst possible model for oversight — opaque, reactive, and economically damaging.

The broader implications extend well beyond Anthropic. The episode has accelerated calls in Congress for a formal framework — including Rep. Josh Gottheimer's bill mandating government reviews of frontier models — and has intensified debate about whether export controls can meaningfully contain AI risks when open-weights models from China are simultaneously leaping ahead. Australia, notably, retained access to Mythos, highlighting the arbitrary geography of the new regulatory landscape.

2. Microsoft Launches Copilot Cowork Globally, Shifts to Usage-Based Pricing

Microsoft brought Copilot Cowork out of preview this week, making the agentic AI system — built in collaboration with Anthropic and powered by both Anthropic and OpenAI models — generally available to all Microsoft 365 Copilot customers worldwide. The launch followed a preview that included more than half of the Fortune 500, and was accompanied by deep integrations into Dynamics 365, GitHub, and the broader M365 ecosystem. The NHS England deal, covering 505,000 staff at a cost of £120 million, illustrated the scale of enterprise appetite for the platform.

The pricing shift to a usage-based model is arguably as significant as the product launch itself. Microsoft's admission that power users were driving compute costs unsustainably high marks a fundamental rethink of how enterprise AI is monetised, and analysts warned of incoming bill shock for organisations that have deployed agents without tracking consumption. The simultaneous exploration of DeepSeek V4 as a lower-cost model option for Cowork — despite the geopolitical sensitivity — signals that Microsoft is prioritising margin management over optics.

The week also brought complications: researchers disclosed the SearchLeak vulnerability, a prompt-injection attack against M365 Copilot that could exfiltrate 2FA codes and sensitive data, and a pension fund filed a proposed securities class action alleging Microsoft overhyped Copilot's capabilities. Together, these threads paint a picture of a product that is genuinely transformative but also maturing through a turbulent adolescence of security flaws, pricing upheaval, and investor scrutiny.

3. DeepSeek Raises $7.4B Series A at $50B+ Valuation as China's AI Ambitions Crystallise

DeepSeek, the Chinese AI lab that rattled global markets earlier in 2026 with its efficient open-source models, completed its first external funding round this week, raising approximately 51 billion yuan ($7.4 billion) at a valuation exceeding $59 billion. The deal reportedly grants founder Liang Wenfeng significant control over the company, and the scale of the raise — a Series A at this magnitude is essentially without precedent in AI — underscores the degree to which Chinese investors and the state are mobilising capital behind domestic AI champions.

The funding news arrived alongside Z.ai's release of GLM-5.2, a 753-billion parameter open-weights model that reportedly outperforms GPT-5.5 on multiple long-horizon coding benchmarks at one-sixth the cost, sending Zhipu AI's Hong Kong-listed shares sharply higher. Taken together, the two stories crystallise a strategic pattern: China is not trying to win the frontier-model race on Western terms, but to industrialise AI deployment at scale, undercut US pricing, and capture users displaced by rising costs or export restrictions.

The week also saw ByteDance in talks with domestic chipmaker Iluvatar CoreX as an alternative to Nvidia hardware, while Sarvam in India closed a $234 million round to become the country's newest AI unicorn. The global diffusion of AI capability and capital is accelerating, and the US lead in frontier models is increasingly offset by a widening gap in cost-efficient deployment.

4. Nvidia Raises $25B in Bonds as AI Infrastructure Buildout Intensifies

Nvidia tapped debt markets for the first time since 2021, raising $25 billion through a bond sale that was met with enormous investor demand. The fundraise was accompanied by a striking data point: the company's data centre and AI revenue has grown 1,300-fold over the past 12 years. The capital is widely expected to fund expanded manufacturing capacity and R&D as the AI infrastructure buildout continues to consume staggering resources.

The broader hardware landscape this week reflected both the scale of Nvidia's dominance and the mounting pressure on it. Amazon's CEO Andy Jassy revealed that AWS's custom chip business already runs at a $20 billion annual revenue rate and is growing at triple-digit percentages, and confirmed the company is in talks to sell Trainium chips to third-party data centres — a direct challenge to Nvidia's core market. Qualcomm, meanwhile, was reported to be in advanced talks to acquire Tenstorrent, the AI chip design startup backed by Jim Keller, in a move that would significantly expand its AI silicon capabilities.

Intel and AMD added further texture by announcing ACE CPU extensions — a new matrix-multiplication instruction set designed to handle AI inference workloads more efficiently on conventional processors. The convergence of these stories points to a market in structural transition: Nvidia remains the dominant force, but its biggest customers are investing heavily in alternatives, and the chip architecture itself is diversifying to accommodate the heterogeneous demands of agentic AI at scale.

5. Noam Shazeer Leaves Google for OpenAI in Major Talent Shake-Up

Noam Shazeer, Google's vice president of engineering and co-lead of the Gemini model family, announced his departure to join OpenAI — one of the most significant individual talent moves in the AI industry in years. Shazeer is widely regarded as one of the architects of the transformer architecture and a foundational figure in modern deep learning. His departure deals a reputational and operational blow to Google DeepMind at a moment when it is competing on multiple fronts.

The move is particularly striking given that Shazeer had returned to Google from Character.AI in a reported $2.7 billion acqui-hire in 2024, and had been seen as a linchpin of Google's Gemini ambitions. OpenAI's ability to attract him signals that despite the enormous resources Google is deploying, its talent retention challenges at the very top of the research hierarchy remain acute.

The hire also comes as OpenAI announced the creation of a dedicated robotics division and continued to expand its footprint well beyond language models. For Google, the week brought additional pressure in the form of a $10M Google DeepMind-backed multi-agent AI safety research initiative — a sign that even as the company competes fiercely commercially, it is investing in the foundational safety research that Shazeer's departure may complicate.

---

By Topic

🧠 Large Language Models

The LLM landscape this week was dominated by two intertwined narratives: the US government's forced suppression of Anthropic's Mythos and Fable models, and China's continued open-weights offensive led by Z.ai's GLM-5.2. The cost and access dynamics are shifting dramatically — GLM-5.2 benchmarks ahead of GPT-5.5 at one-sixth the price, and its open-weights release is explicitly targeting users priced out of premium US models. Noam Shazeer's departure from Google to OpenAI added a dramatic personnel dimension to the competitive landscape. Yann LeCun's high-profile bet that LLMs are a 'dead end' and his dismissal of xAI as a failure generated debate, while Anthropic's own research paper on readable AI activations sparked nuanced discussion about the gap between interpretability and faithfulness. Clinical and biomedical LLM applications continued to mature, with multiple benchmarking frameworks and open medical AI initiatives published this week.

🤖 AI Agents & Automation

AI agents were the week's highest-volume topic, reflecting a market in full deployment mode — and beginning to grapple seriously with the consequences. OpenAI's launch of workspace agents in ChatGPT, powered by Codex and designed for complex enterprise workflows, was the headline product announcement, while Google and Microsoft's co-sponsorship of the Agentic Resource Discovery open spec signalled a move toward interoperability standards. Security emerged as a recurring and urgent theme: a tool-call attack capable of inflating AI costs by up to 658x, OWASP's argument that prompt injection may be a permanent architectural flaw, and multiple new funding rounds for agent identity and security platforms (Arcade's $60M, NewCore's $66M) all pointed to an industry racing to secure infrastructure it deployed before the threat model was fully understood. AWS's Summit announcements — including AgentCore web search and the AWS Continuum and Context services — and IBM's watsonx expansions illustrated how hyperscalers are embedding agentic capability deeper into enterprise stacks, while surveys showing that 85% of IT teams claim to have agents under control but only 42% know who owns them underscored the governance gap.

🛡️ AI Safety & Alignment

AI safety discourse this week oscillated between acute policy drama and longer-term existential debate. The Anthropic Mythos shutdown — and its subsequent partial reversal — dominated the policy conversation, with critics arguing that reactive export controls are a poor substitute for structured risk assessment frameworks. Anthropic itself urged a global pause in AI development, citing self-improvement risks, while Google DeepMind committed $10M to multi-agent safety research and a prominent safety researcher publicly shifted toward supporting model releases — reflecting the internal tensions within the field. On the more alarming end of the spectrum, Stuart Russell warned of AI being 'faster and more effective than Hitler,' Roman Yampolskiy reiterated his 99.9% extinction probability estimate, and a CTCP report documented rising anti-technology violence including the firebombing of Sam Altman's home. Interpretability research continued to advance, with Anthropic's readable activations paper generating both excitement and methodological critique, and Rep. Gottheimer's proposed mandatory government review bill adding legislative momentum to a fractious regulatory environment.

🛠️ AI Tools & Products

The AI tools landscape this week was almost entirely defined by Microsoft and its Copilot ecosystem. Copilot Cowork's global general availability, the shift to usage-based pricing, NHS England's £120M deployment, the Dynamics 365 integration, and GitHub's agent finder all landed within the same week — a coordinated product offensive that nonetheless attracted legal, security, and commercial scrutiny in equal measure. The SearchLeak vulnerability, which allowed one-click exfiltration of 2FA codes through M365 Copilot, and a shareholder class action alleging Copilot was overhyped, complicated the launch narrative. GitHub's separate disclosure that it had to turn to Amazon Web Services because its own infrastructure could not keep up with AI coding demand was a rare moment of vulnerability for the Microsoft ecosystem. Beyond Microsoft, IBM watsonx, HPE GreenLake, and a wave of vertical-specific AI tools continued to proliferate, while the Twindo offline AI copilot for field technicians illustrated how the frontier of AI deployment is extending into environments far removed from cloud-connected enterprise offices.

🎨 Image & Video Generation

Video generation reached a new competitive intensity this week with xAI's Grok Imagine Video 1.5 reaching general availability, topping the AI video leaderboard and priced 86% below Sora 2 Pro — a pricing signal that mirrors the broader cost compression happening across the AI stack. Google integrated Gemini Omni into Flow and YouTube Shorts, bringing conversational video editing to consumer scale, while Midjourney made a surprising pivot into medical spa body scanning, marking one of the more unexpected strategic pivots of the year. Microsoft Research's Mirage video world model introduced persistent spatial memory in latent space, pointing toward a future where AI video generation understands scene continuity rather than just frame-to-frame coherence. The EU's ongoing difficulty in defining deepfakes — with major retailers lobbying for exemptions from transparency rules — highlighted the regulatory lag that continues to trail creative AI capabilities.

🦾 Robotics & Embodied AI

Robotics saw a burst of significant activity this week anchored by capital, acquisitions, and new form-factor launches. Neura Robotics raised $1.4B backed by Nvidia and Amazon, Genesis AI (backed by Eric Schmidt) unveiled its first non-humanoid general-purpose industrial robot in partnership with LG, and Meta quietly acquired Assured Robot Intelligence to accelerate its humanoid programme. OpenAI's announcement of a dedicated robotics division raised immediate questions about competitive dynamics with Tesla, while Samsung was reported to be weighing an investment in Boston Dynamics. Alibaba launched a new suite of physical AI models designed for robotic task execution, and Shenzhen-based IO-AI Tech's teleoperation workforce — humans controlling humanoids via VR rigs — offered a vivid snapshot of how the gap between robot capability and full autonomy is currently being bridged in China's hardware capital. The $6B in Q1 2026 funding for embodied AI world models underscored investor conviction, even as analysts warned the LLM scaling analogy may not hold for physical AI.

🔬 AI Research

AI research highlights this week demonstrated the technology's expanding reach into high-stakes scientific and medical domains. A UC Davis team reported that an ALS patient is working a full-time job using a brain-computer interface powered by machine learning to translate neural activity into speech, marking a meaningful quality-of-life milestone for BCI technology. On the drug discovery front, a James Collins-led team published findings in Science Translational Medicine showing that a deep learning pipeline identified two novel compounds active against antibiotic-resistant gonorrhea — a concrete example of AI accelerating the antibiotic discovery pipeline. Swedish researchers also reported a machine-learning framework that embeds physical laws directly into neural networks, with potential implications for quantum computing simulation. Separately, reinforcement learning research showed broad alignment gains across benchmarks, adding empirical weight to the argument that RL-based training may be a path toward more reliably aligned models.

💼 AI Business & Funding

Funding activity this week was headlined by DeepSeek's landmark $7.4B Series A at a $59B valuation — a round whose scale and structure have no real precedent in AI startup history and which reflects the degree to which Chinese capital is concentrating behind domestic AI champions. Beyond DeepSeek, the week saw Baseten reportedly closing in on a $1.5B raise at a $13B valuation just five months after its last mega-round, world-models startup Odyssey raising $310M at a $1.45B valuation, and Genspark extending its Series B to $485M at a $2.6B post-money valuation. India's Sarvam achieved unicorn status with a $234M round led by HCLTech, underscoring the geographic broadening of the AI funding wave. Databricks' acquisition of Panther Labs in a cybersecurity push and Qualcomm's reported talks to buy chip designer Tenstorrent reflected how AI-adjacent M&A is reshaping the competitive landscape beyond pure model development.

⚡ Hardware & Infrastructure

Hardware infrastructure was one of the week's most consequential topic areas, with Nvidia's $25B bond sale setting the tone for a market still in aggressive expansion mode despite growing competitive pressure. Amazon's revelation that its custom chip business already generates $20B annually — and its move to sell Trainium to third-party data centres — represents the most credible near-term challenge to Nvidia's data centre dominance. Qualcomm's reported talks to acquire Tenstorrent would add a formidable design capability to its AI silicon portfolio, while Intel and AMD's joint ACE CPU extensions signal that conventional processors are being retooled to absorb AI inference workloads more efficiently. On the geopolitical dimension, the US chip export regime suffered a symbolic defeat as China showed little interest in the Nvidia H20-class chips the Trump administration had approved, while ByteDance's talks with domestic chipmaker Iluvatar CoreX illustrated how Chinese hyperscalers are actively building supply chains independent of US silicon. The £1.7B UK infrastructure buildout announced by Nebius Group with Nvidia, including a dedicated robotics lab, rounded out a week that confirmed AI infrastructure investment remains structurally unconstrained.

💻 Tech Industry

Three stories this week captured the broader tectonic forces reshaping the tech industry. Amazon's role in triggering the Anthropic Mythos shutdown — a major investor effectively alerting the government about the safety of its own portfolio company's product — illustrated the deeply entangled interests and conflicts that now define the frontier AI ecosystem. Noam Shazeer's move from Google to OpenAI confirmed that talent competition at the very top of the research hierarchy remains fierce and consequential, with individual departures capable of shifting the perceived trajectory of entire model families. Microsoft's quiet success selling OpenAI models to Chinese enterprises despite US-China tech rivalry demonstrated that commercial pragmatism is consistently outpacing geopolitical posturing, and that the real competitive map of AI does not align neatly with national borders.

---

Emerging Trends

The dominant cross-topic pattern of Week 25 is the collision between AI's accelerating commercial deployment and the governance, security, and geopolitical frameworks struggling to keep pace. The US government's shutdown of Anthropic's Mythos model, the SearchLeak vulnerability in M365 Copilot, the 658x cost-inflation agent attack, and OWASP's argument that prompt injection is an architectural flaw rather than a patchable bug all point to an industry that has deployed capability faster than it has built the defensive infrastructure to protect it. A second clear theme is cost compression and China's open-weights strategy: GLM-5.2 beating GPT-5.5 at one-sixth the cost, DeepSeek's $7.4B raise, and Microsoft's exploration of DeepSeek V4 for Copilot Cowork together suggest that the era of premium frontier-model pricing is under serious structural pressure, with Chinese open-weights releases functioning as a persistent deflationary force. Talent concentration at OpenAI — crystallised by Noam Shazeer's departure from Google — and the simultaneous launch of dedicated robotics divisions at both OpenAI and Meta signal that the next competitive battleground is shifting from language models toward physical AI and embodied systems. Finally, the sheer volume of enterprise agentic AI deployments, governance platforms, and identity security funding rounds confirms that agentic AI has crossed the threshold from experimental to operational — and that the industry is now urgently building the control planes, audit frameworks, and identity infrastructure it should have designed before deployment began.

---

By the Numbers

• Total articles: 209
• Most active topic: AI Tools & Products
• Top sources: reuters.com, techcrunch.com, siliconangle.com
• Topics covered: 10
• Average importance: 3.5/5