Good morning. Welcome to the first newsletter of 2026.
Usually we would focus on the news from the last two weeks (and yes, there was some) but as this is the first 2026’s edition, let’s use it to rewind on 2025 and try to guess where we might be going in 2026.
2025 wasn't just another year of "AI hype", it was the year the narrative fractured. It was the year a Chinese lab crashed the US stock market with a $6 million budget, the year "chatbots" died to make way for "agents," and the year Big Tech spent nearly $50 billion acquiring start-up to make it all work.
But before we talk about the technology, we have to talk about the money. Because 2025 will be remembered as the year the numbers stopped making sense to anyone.
What we should remember of 2025
The Financial Stratosphere: $5 Trillion Caps and the "Capex War"
If you blinked in October, you might have missed history. Nvidia briefly became the first company to shatter the $5 trillion market cap ceiling. To put that in perspective: one chipmaker became worth more than the entire GDP of Germany.
The "Capex" Cold War: The story of 2025 was the reckless, essential spending of the "Hyperscalers." Amazon, Google, Meta, Microsoft, and Oracle collectively poured an estimated $400B+ into AI infrastructure. Oracle’s stock gyrated wildly as it announced a $300 billion data center plan, raising fears of a debt-fueled bubble.
The Bubble Debate: By Q4, the industry was locked in a fierce debate. On one side, bears pointed to the "circular economy"—startups buying Nvidia GPUs with money invested by Nvidia. On the other, bulls pointed to Nvidia’s $130.5 billion in revenue (up 114%) as proof that the demand was real, not speculative. The market ended the year high, but anxious.
The "DeepSeek Shock" and the Efficiency War
It feels like a lifetime ago, but it was only last January that DeepSeek, a Chinese lab funded by a quant hedge fund, broke the "Scaling Laws." They released DeepSeek-R1 and published the "Manifold-Constrained Hyper-Connections" paper, proving they could match US frontier models at a training cost of just ~$6 million—a fraction of the $100M+ spent by US rivals.
The Market Reaction: The realization that "intelligence" might be cheap and efficient rather than expensive and hardware-dependent panic-sold Nvidia stock, causing a 17% drop and wiping out nearly $600 billion in value in a single day on January 27.
The Legacy: While Nvidia recovered, the "DeepSeek moment" ended the era of lazy scaling. Efficiency is now the primary metric.
The Model Wars: Personality Crises and Viral Bananas
2025 taught us that users care as much about "vibes" as they do about IQ.
The OpenAI Stumble: The August release of GPT-5 was a PR disaster. Users revolted against its "cold," clinical personality, forcing CEO Sam Altman to apologize and re-tune the model's "warmth."
Google's Redemption: Google finally found its footing with Gemini 3, which dominated benchmarks in late 2025. But the real surprise was "Nano Banana" (Gemini 2.5 Flash Image), a viral hit that solved character consistency in images and became a massive creative tool.
The Copyright Wall: Anthropic agreed to a landmark $1.5 billion settlement with authors in September, effectively ending the "Wild West" of scraping copyrighted books for training data.
The M&A Supercycle: Buying the "Agentic" Infrastructure
While consumers played with chatbots, enterprises bought the plumbing. 2025 saw massive consolidation as tech giants raced to build "Agentic" ecosystems—systems where AI doesn't just talk, but acts.
Google acquired Wiz ($32B): The largest deal of the year, securing cloud infrastructure for an AI-native world.
Salesforce acquired Informatica ($8B): A move to secure the data layer necessary for its "Agentforce" strategy.
More recently, Meta acquired Manus ($2B): Mark Zuckerberg bought the Chinese-founded agent startup in December to turn WhatsApp and Meta AI into a "global concierge" that can book travel and handle tasks autonomously.
5. Science over Syntax: Physical AI
The most consequential breakthroughs happened in the lab. Google DeepMind’s released their incredible World model Genie 3 getting closer and closer to a model capable of “real world” understanding. This marks the transition of AI from generating text to generating physical solutions and tackling “real world” problems.
What to expect for 2026
The Year of "Lazy Thinking" (and the backlash)
Gartner predicts that in 2026, we will hit a cultural wall: the "atrophy of critical thinking." As reliance on AI copilots grows, expect a backlash where 50% of organizations implement "AI-free" skills assessments to ensure employees can still think without an algorithm.
Strategy Implication: We may need to audit our own teams not just for AI literacy, but for retained human competency.
From "Chatbots" to "Agentlakes"
The buzzword for 2026 is Multiagent Systems. We are moving away from single bots to "Agentlakes"—ecosystems where agents from different vendors (Salesforce, Microsoft, external legal bots) negotiate and execute complex workflows together.
The Shift: You won't ask an AI to "write an email." You “might” ask an agent to "manage the renewal process," and it will coordinate with legal, finance, and CRM agents to finish the job.
Apple's "Siri LLM" Moment (FINALLY?)
After sitting out the wildest parts of 2025, Apple is expected to launch its revamped Siri (powered by LLMs, likely integrating Gemini) in Spring 2026. This will be the mass-market test for whether "Agentic AI" can work on a consumer device.
Physical AI gets Real
With the rise of world model and improvements in vision models, 2026 will be the year Physical AI leaves the demo stage. Expect Tesla’s Optimus Gen 3 and Figure 03 to move from viral videos to actual pilot deployments in factories. It might be the right moment to start looking at robotic start-up, who will pave the way forward.
The "Chip Tax" and Geopolitics
The Trump administration's proposal to allow high-end chip sales to China in exchange for a 25% revenue tariff signals a shift from "containment" to "extraction." Expect 2026 to be defined by transactional tech diplomacy, complicating supply chains for global companies like Zurich.
In a nutshell: 2025 was the year the machines woke up to capitalism. 2026 is the year they get to work. Now will we finally get a model so powerful that we would call it AGI? Perhaps with Gemini 3 or Opus 4.5 we already have? Only time will tell! But what is certain is that things ought to get more and more interesting - for the better or for the worse.
We will return to our standard format next week. Happy New Year!