Callista AI Weekly (February 16 - February 22, 2026)
As of February 24, 2026, the AI landscape just witnessed one of its most consequential weeks of the year. A wave of major model releases, a landmark global summit in New Delhi, and a decisive pivot toward agentic AI in the enterprise converged to reshape priorities for boards and business leaders. At the center of this shift: AI systems that act, not just answer.
New AI Use Cases
This week’s announcements underscored a fundamental moment for agentic AI: enterprises are deploying autonomous agents to carry out multi-step tasks across compliance, operations, commerce, and customer service at scale.
Enterprise-grade agents in regulated sectors: Infosys and Anthropic revealed a strategic collaboration (February 17) to deploy agents that autonomously execute complex workflows such as claims processing, compliance reviews, and network operations across telecom, financial services, manufacturing, and software development. Infosys has already built a material AI services business (about $275 million per quarter), and the market signaled confidence with a 4.8% share price rise on the news.
Operationalizing agentic AI at scale: Cognizant and Google Cloud (February 16) expanded their partnership to take agentic AI from prototype to production through an Agent Development Lifecycle (ADLC) framework. The collaboration introduced pre-configured solutions targeting high-impact functions, including AI-powered contact centers and intelligent order management, designed to slot into existing enterprise operations.
Consumer commerce goes autonomous: Sea Ltd and Shopee introduced a forward-looking agentic shopping prototype capable of browsing products, managing orders, and completing purchases on behalf of users. Crucially, the partners also announced the Agent Payments Protocol (AP2), positioning it as foundational infrastructure for AI-mediated commerce in Southeast Asia. With Shopee holding 52% market share in the region’s e-commerce, AP2 could catalyze a broader market shift to agent-facilitated transactions.
“Talk with the network” in telecom: Ericsson launched Agentic rApp as a Service on AWS (February 17), enabling carriers to optimize networks via natural language commands. Field trials are already underway with Vivo Brazil. Ericsson’s AI footprint is significant. Its solutions process more than 100 million inferences daily across 11 million cells, providing immediate scale for agentic control in telecom operations.
Sovereign enterprise assistants for Swiss SMEs and regulated sectors: Swisscom introduced the “Swiss AI Assistant” (February 18), a generative AI chatbot for enterprise knowledge management that operates exclusively in Swiss data centers. Priced from CHF 149 per month for up to 10 employees, it targets SMEs and regulated organizations as a privacy-compliant alternative to foreign AI services, with planned integration of the national Apertus LLM. The move exemplifies how sovereign AI infrastructure can align advanced capability with stringent data residency and compliance requirements, which is key for Swiss businesses.
Brand-scale digital backbones: Unilever signed a five-year deal with Google Cloud to build an “AI-first digital backbone” across household brands like Dove, Vaseline, and Hellmann’s, embedding agentic workflows into a business with €50.5B in 2025 revenue. This is among the largest CPG-AI partnerships announced to date and effectively normalizes AI-mediated functions within large-scale, regulated, and brand-sensitive environments.
Taken together, these deployments confirm that autonomous AI agents are no longer a conceptual horizon. They are being built and integrated by the global systems integrators that large enterprises already rely on, and they are aligning with sectoral requirements around privacy, security, and control. For Swiss firms, especially in finance, healthcare, and telecom, this is a practical roadmap: leverage integrator-led frameworks, emphasize sovereign and compliant deployments, and prioritize multi-step workflows where agents can deliver fast operational ROI.
Major Vendor Updates
New model releases and hyperscale infrastructure deals reshaped competition on both capability and cost. The center of gravity is moving toward models that reason better, run cheaper, and support agentic architectures at enterprise scale.
Anthropic’s Claude Sonnet 4.6 (February 17): Anthropic made Sonnet 4.6 the default for Free and Pro users and pushed near-Opus performance for real-world office tasks at Sonnet pricing (3 dollars and 15 dollars per million input and output tokens). It improved coding and computer-use capabilities and offers a 1M-token context window in beta. The model set new records on OS World and SWE-Bench benchmarks. Anthropic followed with Claude Code Security (February 20), an AI-driven scanner that “reasons like a defender,” available for Enterprise and Team plans.
xAI’s Grok 4.20 in public beta: xAI introduced a four-agent parallel collaboration design in which specialized agents independently reason and synthesize outputs. Elon Musk reported a 65% reduction in hallucinations versus prior versions and unveiled a weekly update cadence via a “rapid learning” architecture. Pricing spans free with limits to 300 dollars per month for SuperGrok Heavy.
Alibaba’s Qwen 3.5: Alibaba released a 397B-parameter mixture-of-experts model with just 17B active parameters per pass. It supports 201 languages and a one-million-token context window under an Apache 2.0 open license. Alibaba claims performance parity with leading US proprietary models at dramatically lower cost, alongside decoding speeds up to 19 times faster than earlier versions. Google also shipped Gemini 3.1 Pro, reportedly doubling reasoning performance on ARC-AGI-2 over its predecessor.
Chinese market momentum: Zhipu AI surged around 43% and MiniMax rose roughly 14% on February 20, both surpassing HK$300 billion in market cap. The moves followed pre-holiday launches: Zhipu’s GLM-5 (744B parameters, fully open-source) and MiniMax’s M2.5, which matched Claude Sonnet-class performance at one twentieth the cost. DeepSeek V4 did not launch but its anticipation alone moved markets.
Hyperscale infrastructure realignment: Meta and NVIDIA announced a multi-year, multi-generational strategic partnership (February 17) valued by analysts at 35 to 67 billion dollars. Meta will deploy millions of NVIDIA Blackwell and Rubin GPUs, standalone Grace CPUs at unprecedented scale, and Spectrum-X networking. NVIDIA Confidential Computing will underpin AI features inside encrypted WhatsApp messaging, which is crucial for Meta’s 3+ billion users. Meta also adopted Vera CPUs for potential 2027 deployment, making this among the largest AI infrastructure agreements on record.
Spatial intelligence bets: World Labs raised 1 billion dollars (February 18) from AMD, Autodesk, NVIDIA, and others, backing “spatial intelligence,” that is AI that reasons over 3D environments. The investment signals an industry-wide push into world models as the next frontier beyond text-centric LLMs.
Global infrastructure commitments from New Delhi: The India AI Impact Summit (February 16–21) served as the stage for substantial pledges. Microsoft committed 50 billion dollars toward AI infrastructure for the Global South by decade’s end. Adani committed 100 billion dollars for AI data centers with renewable energy in India by 2035. OpenAI became the first customer of TCS’s data center business. India earmarked 1.1 billion dollars for a state-backed AI venture fund and targeted 200 billion dollars in AI investment over two years.
The competitive playbook is shifting. New model families emphasize reasoning, context length, and agentic coordination, while open-licensed, cost-efficient architectures like Qwen 3.5 are narrowing the performance-price gap with top proprietary systems. At the same time, hyperscale infrastructure alliances are setting the tempo for how and where enterprise AI can scale securely, with encrypted-by-design functionality now part of frontline user experiences.
AI Governance Developments
This week’s governance arc was defined by breadth over depth and by visible divergence among major powers on the path forward. A first-of-its-kind global declaration landed in New Delhi, punctuated by sharply different national strategies and the swift expansion of state-level rules.
The New Delhi Declaration on AI Impact (February 22): Endorsed by 89 signatories, including both the US and China, the declaration is organized around seven pillars such as democratizing AI resources, economic growth, trusted AI, science, social empowerment, human capital, and resilient AI. Notably, it omits the word “safety” entirely, favoring “security” and voluntary, industry-led approaches. Safety advocates criticized the gap between discussions and the final text.
US position: The White House OSTP Director announced, “We totally reject global governance of AI,” and introduced the American AI Export Program, National Champions Initiative, and Tech Corps. The stance underlined a preference for national competitiveness and voluntary measures over binding international rules.
EU initiatives: Executive Vice-President Henna Virkkunen launched the Frontier AI Grand Challenge, focused on building sovereign European AI models, and opened a European Legal Gateway Office in India, signaling ongoing efforts to strengthen regional capacity while promoting European governance approaches abroad.
UK as early mover on AI regulation: Prime Minister Keir Starmer announced (February 16) that AI chatbots will be brought under the Online Safety Act’s illegal content duties. This is a first major extension of online safety regulation directly to AI systems, with compliance backed by enforcement.
Rapid evolution of US state rules: As of February 20, 78 chatbot bills were active across 27 states. Oregon’s SB 1546 passed the Senate 26 to 1, and Washington advanced six AI-related bills past a significant deadline. Meanwhile, the White House pressed Utah to drop its own AI child safety bill, illustrating escalating federal-state tensions.
China’s enforcement approach: Authorities penalized more than 13,000 accounts and removed hundreds of thousands of posts for unlabeled AI-generated content in response to the viral reach of ByteDance’s Seedance 2.0. The moves showcase China’s emphasis on compliance and content control.
India’s first binding AI content rules: Updated IT Amendment Rules took effect on February 20, defining “synthetically generated content” and imposing due diligence obligations, which marks the country’s first binding regulation specific to AI content.
Swiss Focus: Governance, Sovereign AI, and Deep-Tech Momentum
Swiss leadership on the world stage: President Guy Parmelin attended the India AI Summit (February 19–20) and announced that Switzerland will host the 2027 AI Summit in Geneva, continuing the Bletchley, Seoul, Paris, Delhi sequence. Switzerland is championing the Council of Europe’s AI Framework Convention, the first binding international AI treaty, and is pursuing a sector-specific approach to domestic AI regulation rather than a comprehensive “Swiss AI Act.” Draft legislation to implement the Council of Europe convention is expected by end of 2026.
Sovereign AI for Swiss enterprises: Swisscom’s launch of the Swiss AI Assistant reflects a practical, market-driven answer to privacy, compliance, and data residency requirements. With pricing from CHF 149 per month for up to 10 employees and operations wholly in Swiss data centers, the solution targets SMEs and regulated industries with a home-sovereign alternative to foreign tools and plans to integrate the national Apertus LLM.
Investment and research depth: Zurich drew fresh international capital as DIC Corporation and Emerald Technology Ventures unveiled a 62 million dollar Physical AI investment platform, with DIC establishing a Zurich subsidiary in spring 2026. Swiss research institutions also delivered globally relevant advances this week:
ETH Zurich and USI: “Zero Knowledge (About) Encryption” documented 25 successful attacks against major password managers Bitwarden, LastPass, Dashlane, and 1Password, which serve more than 60 million users collectively.
ETH Zurich and MIT: Self-Distillation Fine-Tuning (SDFT) enables LLMs to learn new skills without catastrophic forgetting. This is directly relevant to maintaining performance as enterprises extend agent capabilities.
EPFL: Dynami-CAL GraphNet (February 20) models complex dynamical processes while respecting Newton’s laws, published in Nature Communications.
Startup ecosystem priorities: The Swiss Startup Association released the “Startup-Agenda Schweiz” with 20 reforms supported by 18 partners. Switzerland leads Europe in AI talent density at 4.8% of core AI professionals, yet struggles to retain scale-ups. In 2025, 31 of 42 startup exits went to international buyers. The agenda points to structural reforms needed to translate world-class talent into long-term domestic champions.
Breakthrough Research
Leading-edge research advanced autonomous reasoning in the physical world, accelerated drug discovery, and revealed new materials, all with potential implications for industrial innovation, life sciences, and energy technology.
Robotics world models: NVIDIA’s DreamZero (February 17) is a 14-billion-parameter World Action Model that jointly predicts future video frames and robot actions. It doubles generalization performance over state-of-the-art Vision-Language-Action models when adapting to new tasks and transfers across entirely new robot embodiments with only 30 minutes of play data. Jim Fan described it as the “GPT-2 moment” for robotics. If sustained, this shift could enable more reliable and sample-efficient training of factory and field robots.
Drug discovery leap: Isomorphic Labs’ IsoDDE, detailed in Nature on February 19, more than doubles AlphaFold 3’s accuracy on challenging protein-ligand benchmarks. On antibody–antigen structure prediction, the system achieves 2.3 times improvement over AlphaFold 3 and 19.8 times over Boltz-2. Columbia’s Mohammed AlQuraishi called it “a major advance, on the scale of an AlphaFold 4.” While the fully proprietary approach drew open-science criticism, the performance gains underscore how rapidly AI is advancing in therapeutics design.
Biomanufacturing optimization: MIT researchers showed on February 16 that an LLM can optimize the genetic code of industrial yeast used in biopharmaceutical manufacturing. The result points to potential reductions in the 15 to 20% of drug commercialization costs tied to manufacturing development, suggesting AI could meaningfully lower the expense and time-to-market of complex biologics.
New magnetic materials: An AI system at the University of New Hampshire mined scientific literature to assemble a 67,573-compound database of magnetic materials, identifying 25 previously unrecognized compounds that remain magnetic at high temperatures. Published February 19 in Nature Communications, this offers a path toward rare-earth alternatives for electric vehicles and clean energy systems with implications for supply chain resilience and costs in energy transition technologies.
Generative video capabilities: ByteDance’s Seedance 2.0 was described by CNN as “the most sophisticated of its kind to date,” and its rapid spread across social platforms provoked a Motion Picture Association backlash while catalyzing China’s crackdown on unlabeled AI content. Beyond media implications, the episode underscores rising policy salience for synthetic media governance and labeling, which is increasingly relevant to brand safety, customer trust, and legal exposure across markets.
Conclusion
Three shifts now define the state of AI coming out of February 16–22:
Agentic AI crossed from prototype to production. Agents no longer simply answer, they act. They are shopping on Shopee, optimizing telecom networks at Ericsson, and processing insurance claims for Infosys clients. The emergence of implementation frameworks and infrastructure, such as Cognizant’s ADLC, Shopee’s AP2, and telecom-focused rApps, signals that integration standards for agent interoperability are forming quickly. For Swiss enterprises, the path to practical adoption is clearer. Identify multi-step workflows with measurable ROI and implement via trusted integrators and sovereign environments where required.
The global governance consensus fractured. The New Delhi Declaration earned unprecedented breadth, with 89 signatories including both the US and China, by deliberately avoiding the contested topic of “safety” in favor of “security” and voluntary approaches. The US explicitly rejected binding global governance while the EU focused on building sovereign model capacity. The UK moved first on regulatory extension. US states are legislating fastest, with federal–state tensions rising. China is enforcing strict content controls. India implemented its first binding AI content rules. Switzerland’s selection as host of the 2027 AI Summit positions Geneva as a plausible bridge across regulatory divides, with domestic policy oriented toward sector-specific regulation and alignment with the Council of Europe AI Framework Convention.
The cost curve collapsed further. Alibaba’s Qwen 3.5 claims parity with leading US proprietary systems at a fraction of the cost under an open license. MiniMax’s M2.5 reportedly rivals Claude at one twentieth the price. Anthropic pushed Opus-class performance to Sonnet pricing. The moat around frontier models is narrowing, and competitive advantage is shifting from the model alone to the ecosystem that surrounds it. Integrators that deliver outcomes, agent infrastructure, sovereign operations, and compliant data processing are becoming decisive. Swisscom’s Swiss AI Assistant for SMEs and regulated industries is a strong illustration of this shift.
Ready to explore how Agentic AI can transform your organization? Visit us at https://www.callista.ch/agentic-ai to discover how we can guide your journey into this exciting new era of AI-powered productivity.
Sources
TechCrunch — "Anthropic releases Sonnet 4.6" — February 17, 2026
CNBC — "Meta expands Nvidia deal to use millions of AI chips in data center build-out" — February 17, 2026
TechCrunch — "All the important news from the ongoing India AI Impact Summit" — February 22, 2026
Fortune — "India's AI Impact Summit closes with the New Delhi Declaration and a $200 billion boost" — February 23, 2026
TechCrunch — "Infosys partners with Anthropic to build enterprise-grade AI agents" — February 17, 2026
CNN — "Seedance 2.0: China's latest AI is so good it's spooked Hollywood" — February 20, 2026
CNN — "Microsoft pledges $50 billion to tackle AI inequality" — February 18, 2026
Cognizant Press Release — "Cognizant Expands Strategic Partnership with Google Cloud to Operationalize Agentic AI" — February 16, 2026
CNBC — "AI chatbot firms face stricter regulation in online safety laws protecting children in the UK" — February 16, 2026
Bloomberg — "AI Pioneer Fei-Fei Li's Startup World Labs Raises $1 Billion" — February 18, 2026
Ericsson Press Release — "Ericsson launches Agentic rApp as a Service on AWS" — February 17, 2026
Nature — "'An AlphaFold 4' — scientists marvel at DeepMind drug spin-off's exclusive new AI" — February 19, 2026
Swisscom — "Swiss AI Assistant: Swisscom's AI chatbot for in-house knowledge management" — February 18, 2026
SWI swissinfo.ch — "Switzerland to host 2027 world summit on artificial intelligence" — February 22, 2026
ScienceDaily — "AI breakthrough could replace rare earth magnets in electric vehicles" — February 18, 2026
