The Old 'Deutschland AG' Zombie Lives: Cartel of Telekom and SAP to Build Government App

Wildberger Commissions Cartel, Avoids Competition

Federal Digital Minister Karsten Wildberger (CDU) is bypassing the usual public tender process and directly commissioning SAP and Deutsche Telekom to develop an AI-based government administration app. The app is intended to become the central gateway to the state, allowing citizens to submit applications, book appointments, and interact with authorities in the future. Wildberger is using existing framework agreements the corporations have with the federal government to save months of procedural time. A tender is supposed to follow only after the prototype is developed, but the preliminary decision has already been made. In parallel, his ministry is actively supporting the planned merger between Aleph Alpha and the Canadian AI company Cohere, with the goal of creating an “AI champion” on the German infrastructure of Schwarz Digits. The federal government is even considering a future stake and plans to act as an anchor customer. → Handelsblatt KI-Briefing

Synthszr Take: Wildberger is building a German platform state, allowing its citizen interface to be monopolized by two corporations. The parallel to the Corona-Warn-App (€200 million, dormant since 2023) reveals the pattern: quick contracts awarded to familiar partners, high costs, limited impact. In Austria, the Corona app with more functionality cost only a fraction of that amount. The deal is reminiscent of China's WeChat strategy, but without the execution power: there, a private super-app became state infrastructure; here, state infrastructure is supposed to become a private super-app. SAP and Telekom are effectively being handed an oligopoly over the digital government interface—whoever defines the architecture shapes all future iterations. Meanwhile, Aleph Alpha, a technologically lagging company, is being kept alive through political deals while real AI innovation happens elsewhere. The old 'Deutschland AG' is experiencing a revival: the administration isn't being digitized, it's being cartelized.

Llama 4 Shatters All Parameter Limits

Meta is making parameter counts explode with Llama 4: Maverick with 400 billion parameters (128 experts, 17 billion active), Scout with 109 billion, and the yet-to-be-released Behemoth model with a whopping 2 trillion parameters in total. Scout promises a 10-million-token context, which shatters previous limits. The models are multimodal (text and image), and Scout is already ranked second on the LM Arena leaderboard, right behind Gemini 2.5 Pro. The catch: even with quantization, these Mixture-of-Experts giants don't fit on consumer GPUs. A Mac Studio with 256 GB of RAM can barely run the fp16 version of Scout at a modest 11.7 tokens per second. → Simon Willison's Newsletter

Synthszr Take: Meta is betting on the infrastructure of tomorrow. The models are so large that they can currently only run in specialized data centers; even a 64 GB Mac can only manage with 3-bit quantization. This is reminiscent of the early days of GPT-3, when OpenAI was the only company that could afford the hardware. Only this time, Meta is giving away the weights, betting that Moore's Law and specialized hardware (like Grok's LPUs) will drive down costs. The 10-million-token context length is the real game-changer: whoever can be the first to offer this capacity in production will define a new class of applications. Meta is building models for a world where compute becomes a commodity.

Perplexity Becomes a Finance API

Perplexity is connecting its computer agent to over 12,000 banks via Plaid integration, transforming the platform into a personal finance hub. Users can directly link accounts, credit cards, and loans, while the AI automatically creates budgets, develops debt repayment plans, and generates asset overviews. The integration follows Perplexity's US tax feature, which automatically fills out IRS forms. The pivot to an agent system has paid off: Perplexity's annual recurring revenue jumped to $450 million in March—a 50 percent increase in a single month. The computer function launched at the end of February and suddenly positions Perplexity against Mint, TurboTax, and established finance apps. → The Rundown AI

Synthszr Take: Perplexity is doing what the major tech corporations failed to do with their assistants: building AI directly where real user data resides. The Plaid integration isn't a feature; it's a declaration of war on the fragmented fintech landscape. While Google and Apple philosophize about privacy-compliant wallet solutions, Perplexity is creating facts on the ground with read-only banking access. This is reminiscent of WeChat's super-app strategy in China, except Perplexity comes from search instead of messaging. $450 million in annual recurring revenue after just one month of agent operation shows: users don't want more apps; they want an AI that orchestrates their existing data silos. Perplexity is achieving something Google never could.

Google and Intel Are Writing the Hardware Grammar of the AI Era

Google has expanded its TPU fleet 11.5-fold in just seven quarters. The energy consumption of these chips now surpasses Microsoft's entire AI compute stack, and the pace of expansion is accelerating. In the fourth quarter of 2025, Google alone added more computing power than xAI possesses in total. These figures show: the battle for AI dominance won't be decided by model architectures, but in silicon factories. Whoever controls the hardware defines the rules of the next economy. → The Business Engineer

Synthszr Take: Google is playing a different game than everyone else: while Microsoft and Meta are queuing up for Nvidia, Google is printing its own chips like a central bank prints money. This is reminiscent of Apple's vertical integration, only more radical: Apple controlled hardware for end-users; Google controls the means of production for the AI revolution itself. The TPU strategy is Google's Manhattan Project, only this time, everyone knows what's at stake. Microsoft can build as many Azure data centers as it wants; as long as they have to buy from Nvidia, they remain trapped in the value chain. Meanwhile, Google is rewriting the hardware grammar in which the next ten years of software will be written.

Alibaba (I): New AI Factory Buries Open Source

Alibaba is reorganizing its AI division with surgical precision. CEO Eddie Wu sent out two circulars within 23 days, established the Alibaba Token Hub (ATH) as a new business entity, and elevated the Tongyi Lab to a full-fledged business unit. The timing is no coincidence: between March and April, Alibaba released three new AI models (Qwen3.5-Omni, Wan2.7-Image, Qwen3.6-Plus), but for the first time, it held back the most commercially valuable capabilities behind APIs instead of releasing them as open weights. The appointment of Zhou Jingren as “Chief AI Architect” signals the shift from a research lab to a production factory. Li Feifei, a database specialist focused on reliability rather than algorithmic innovation, is taking over as the new CTO of Alibaba Cloud. The message is clear: tokens are becoming a tradable commodity like kilowatt-hours, and the best models are no longer free. → Hello China Tech

Synthszr Take: Alibaba is doing what Henry Ford did with the automotive industry: turning craftsmanship into assembly-line production. The analogy to kilowatt-hours is aptly chosen because, just like electricity in the early 20th century, AI compute has become a strategic resource that nations and corporations want to control. The retreat from open source marks a historic turning point: after years of sharing, the window is closing. Alibaba's three-step process (commercial logic through ATH, production logic through the Technical Committee, product logic through closed source) is reminiscent of China's Special Economic Zones of the 1980s: first experiment, then industrialize, and finally, monopolize. Open source was the bait to attract developers; now that critical mass has been reached, the trap is snapping shut.

Alibaba (II): Image Generation Becomes Production-Ready for Brands

Alibaba introduces Wan2.7-Image, an AI model for image generation that solves two persistent problems: the generic aesthetic of AI images and unpredictable color reproduction. The model allows for the reproduction of exact brand colors by inputting specific color codes and proportions—a feature that has been sorely lacking in professional design. With support for 3,000-token long text inputs, 12 languages, and the ability to render complex formulas and tables in print quality, Wan2.7 is positioned as a tool for serious applications. The “Click-to-Edit” feature allows for pixel-perfect editing of individual image areas, while the system can process up to nine reference images and generate twelve images simultaneously. In anonymized tests, the model beat leading competitors in visual quality and text rendering. → Hello China Tech

Synthszr Take: Alibaba is exploiting the weaknesses of Western image generators: while Midjourney and Dall-E compete on artistic vibes, China is building a tool for the real economy. The color code input provides control over the creative process—just as AutoCAD shifted architecture from artistic sketches to parametric precision. 3,000 tokens for text input sounds like overkill until you consider that Chinese e-commerce sites often contain novels' worth of product descriptions. The real innovation isn't in the technology, but in the understanding that professional users need to gain control over randomness. Alibaba isn't building toys for Silicon Valley VCs; it's building infrastructure for an economy where every Taobao merchant has to become a visual designer.

Gmail Gets End-to-End Encryption on Smartphones

Google is rolling out end-to-end encryption (E2EE) for Gmail on Android and iOS. Enterprise users with the appropriate licenses can now read and write encrypted emails directly in the app without additional tools. The messages arrive in the inbox as normal emails, as long as the recipients are using the Gmail app. Those using other email services can open encrypted messages in a browser. The feature is based on Client-Side Encryption (CSE), where companies control their own encryption keys—outside of Google's servers. Gmail CSE has been available in beta since December 2022 and reached general availability for Enterprise Plus and Education Plus customers in February 2023. → Techpresso

Synthszr Take: Google is turning a technical necessity into a selling point—and in doing so, reveals where the real battle for enterprise customers is being fought. While everyone is talking about AI features, Google is winning over large customers with the most boring promise of all: HIPAA compliance and data sovereignty. The genius of it is: the encryption works cross-platform, but only Gmail users get the full experience. It's the old Microsoft Exchange strategy, only more subtle: technically open, practically proprietary. Enterprise Plus licenses are becoming the ticket to a world where “Google can't read my data” is a premium feature—a remarkable reversal of Gmail's original business model.

Gemini x Notebooks: The Karpathy Wiki for Everyone

Google is integrating NotebookLM directly into Gemini, creating personal knowledge bases that are synchronized between the two apps. Users can organize chats by topic, save custom instructions, and add relevant documents as context. Depending on the subscription tier, a varying number of sources can be integrated, enabling more complex projects. The seamless synchronization allows users to leverage specific features of both apps—such as Video Overviews in NotebookLM or the web search in Gemini. The rollout begins this week for Google AI Ultra, Pro, and Plus subscribers on the web. → TAAFT - There's An AI For That

Synthszr Take: Google is turning Gemini into an integrated development environment for thoughts. What Visual Studio Code is for programmers, Notebooks is intended to be for knowledge workers: a central workspace with persistent context, modular components, and cross-tool integration. The subscription tiers with source limits show where Google sees the real value—not in the model itself, but in the manageable context window. NotebookLM was the trial balloon; now comes the industrialization. The next logical step: team notebooks with shared knowledge bases that learn from individual research.

GRID Makes VC Decisions Transparent: Pitch Analysis as an Operating System for European Startups

The Berlin-based startup GRID (grid.nma.vc) is transforming the black box of fundraising into a structured process. The platform analyzes pitch decks in 60 seconds, evaluates them based on 10 VC criteria, and specifically points out which slides are weak. The centerpiece is DIALECTIC: a multi-agent system that simulates an internal VC debate—with evidence-based pro and con arguments—and delivers an INVEST/PASS/UNCERTAIN verdict with a confidence score. Additionally, GRID matches startups against a database of 6,296 European and Israeli VC firms and 21,248 investment professionals, filtered by sector, stage, geography, and ticket size. The fourth component is a Growth Agent that identifies the first customers. → Nico Lumma from Five Things

Synthszr Take: GRID is industrializing VC due diligence from the inside out. The business model is reminiscent of the early days of search engine optimization: whoever knows and caters to the evaluation criteria wins. For years, VCs have hidden their decision-making patterns behind the myth of 'pattern recognition'—now, these patterns are becoming a commodity. The Dialectic feature is particularly clever: it not only shows what VCs might be thinking but also forces founders to sharpen their own story against structured objections. If all startups optimize for the same ten criteria, the VC decision becomes purely a matter of execution.

Techno-Capitalism vs. Innovation: Can This Economic System Last Another Decade?

Michael Kratsios, director of the White House Office of Science and Technology Policy, along with Reid Hoffman and 500 other CEOs at the Semafor World Economy Summit, poses the central systemic question: Can the American model of capitalism sustain another 50 years of technological leadership? The golden era of post-war innovation was based on massive government investment (Cold War, the internet backbone, patent reforms) and an expanding middle class. Today, private capital dominates AI infrastructure with sums exceeding the Apollo program, while the top 10% of income earners account for almost half of consumption. At the same time, AI models like Claude are threatening the very existence of data processing giants: Snowflake and Databricks are losing value on the stock market, while ThoughtSpot investors warn that Anthropic and OpenAI could take over their analysis services themselves within two years. Paul Smith from Anthropic promises cooperation (“If your data is well-structured in Snowflake, you'll make progress faster”), but the threat is unmistakable. → Semafor Technology

Synthszr Take: The American tech economy is experiencing its decadent phase: the old order still works, but the foundations are already crumbling. What was once supported by government research funding and broad consumption now depends on the drip feed of private mega-investments and a narrow luxury consumer class. The parallel to the late Roman economy is compelling: ever-larger latifundia (tech monopolies) are displacing small farmers (startups), while the legions (talent) increasingly consist of mercenaries (H-1B visas). That data infrastructure providers could be the first victims of the AI revolution follows a perverse logic: they have structured the data so well that the AI models no longer need them to. Techno-capitalism eats its most successful children first.