Tech Digest – February 26, 2026
AI Agents Hit Production
Claude Agents Ship a Feature Over a Weekend — Karpathy Says Programming Changed More in Two Months Than in Years
An Anthropic engineer wrote a spec, pointed Claude at an Asana board, and left for the weekend. Claude decomposed the spec into tickets, spawned agents for each one, and by Monday the feature had shipped. Separately, Andrej Karpathy — former Tesla AI lead and OpenAI co-founder — said that programming has changed more in the past two months than in years of incremental progress, specifically crediting the step-change to coding agents that “basically didn’t work before December.”
Note: The shift here isn’t theoretical. When a single spec produces working software over a weekend, the bottleneck in digital projects moves from development capacity to decision quality. The institutions that define requirements clearly will be the ones that benefit first.
Sources: Vivek Ravisankar (Anthropic), Andrej Karpathy
AI-Generated Vulnerability Reports Overwhelm the National Vulnerability Database by 100–200×
AI-powered security tools have flooded the U.S. National Vulnerability Database (NVD) with an estimated 30,000 CVEs, overwhelming the system by a factor of 100 to 200 compared to historical volumes. The NVD — the central registry used globally for tracking software vulnerabilities — is now buried under reports generated faster than human reviewers can triage them.
Note: The same tools that find vulnerabilities at machine speed will soon be the only tools that can triage them at machine speed. Any institution running custom or legacy software should expect its exposure surface to expand sharply — not because attackers got smarter, but because the scanners did.
Sources: The Register
Agent Platforms Consolidate: Anthropic Acquires Vercept, Perplexity Launches Multi-Model Orchestration
Anthropic acquired Vercept to accelerate Claude’s computer-use capabilities — the ability for AI agents to operate software interfaces the way a human would. Separately, Perplexity launched “Perplexity Computer,” a system that orchestrates 19 models in parallel, with Opus routing each task to the best-suited model. Google’s Gemini also began automating multi-step tasks on Android, starting with actions like ordering a ride. The agent layer is thickening across all major platforms simultaneously.
Note: Computer-use agents don’t need APIs or integrations — they use existing interfaces. That changes the integration calculus for any institution running web-based tools. The question shifts from “does the vendor offer an API?” to “can an agent use the screen?”
Sources: Anthropic, Perplexity, TechCrunch
Infrastructure, Capital & Energy
Nvidia Data Center Revenue Hits $63.3B as Compute Bottleneck Grows Daily
Nvidia reported $63.3 billion in data center revenue for Q4, up 75% year over year. The company has extended $3.5 billion in guarantees to firms leasing land, power, and facilities — four times the prior quarter — effectively underwriting the physical buildout its chips require. Google’s Logan Kilpatrick warned that the compute supply-demand gap is growing by single-digit percentage points every day.
Note: When Nvidia starts guaranteeing real estate deals, the AI infrastructure build is no longer a software story. It’s a construction and power story — and that’s where it intersects with every institution’s energy costs and data center availability.
Sources: CNBC, The Information, Logan Kilpatrick (Google)
Amazon Ties $35B of Its $50B OpenAI Investment to an IPO or “Reaching AGI”
Amazon has reportedly structured $35 billion of its $50 billion OpenAI investment as contingent on either an IPO or the company demonstrating it has reached AGI — artificial general intelligence. The remaining $15 billion is committed regardless. The deal structure suggests Amazon’s bet is less about current capabilities and more about positioning for a threshold event.
Note: When the world’s largest cloud provider writes a contract with “reaching AGI” as a trigger clause, that’s no longer a research milestone — it’s a procurement condition. The definition of AGI is now a $35 billion question with contractual consequences.
Sources: The Information
Seven Tech Giants Head to White House for Energy Self-Supply — Bavaria Backs Proxima Fusion with €400M for Europe’s First Stellarator
Seven major technology companies are expected at the White House in March to sign agreements for building their own electricity supply to sustain AI infrastructure growth. Meanwhile, in Europe, the Free State of Bavaria committed €400 million toward Proxima Fusion’s €2 billion “Alpha” demonstration stellarator in Garching, near Munich. The MoU — signed with RWE and the Max Planck Institute for Plasma Physics — targets net energy gain by the early 2030s and lays the path toward “Stellaris,” a commercial fusion plant at the former Gundremmingen nuclear site. An industrial consortium of over 30 European companies, including Siemens Energy, Air Liquide, and Eni, has formed around the project.
Note: The US side of this story is tech companies routing around the grid. The European side is different: public-private co-financing, state-backed research infrastructure, industrial consortium. Both are responses to the same pressure — AI compute needs power that doesn’t exist yet — but the European model creates procurement and workforce opportunities that the self-supply model does not.
Sources: Fox News, Proxima Fusion, TechFundingNews
Physical AI Crosses the Deployment Line
Nvidia Trains Humanoid from 20,000 Hours of Human Video — No Robot Needed in the Loop
Nvidia demonstrated a humanoid robot with dexterous hands that learned to assemble model cars, operate syringes, sort cards, and fold shirts — trained primarily from 20,000+ hours of egocentric human video, with no robot involved in the training process. The researchers reported a near-perfect scaling law: after pretraining, a single teleop demonstration was sufficient to learn a new task. Analysts expect this result to trigger a field-wide shift toward using human video data for robotic generalization. Separately, Alphabet merged its robotics subsidiary Intrinsic back into Google to combine DeepMind’s AI with real-world hardware deployment.
Note: The bottleneck for useful robots was never hardware — it was teaching them. Training from human video instead of robot-specific data means every hour of recorded human work becomes training data. The scaling curve just changed slope.
Sources: Jim Fan (Nvidia), Intrinsic / Google
36 Cleaning Robots Cover 2.7 Million Square Meters in Shenzhen — Government-Procured, Cost-Benchmarked
In Shenzhen’s Shijing subdistrict, 36 autonomous cleaning robots built by CowaRobot are covering 2.7 million square meters of streets and public spaces. The machines are procured by local government with explicit performance targets: each robot is expected to replace three to five human workers, and the city benchmarks cost-effectiveness against the average sanitation worker salary of roughly ¥70,000 (~€9,000) per year. A single worker-and-robot team now handles 7–8 km of sidewalk daily, up from 1 km per worker alone.
Note: This is not a pilot or a demo. It’s a city government procuring robots against a defined cost benchmark, with performance evaluations in place. The deployment model — government buyer, quantified ROI threshold, phased human-robot teaming — is directly transferable to European municipal operations.
Sources: Yicai Global
Economic & Policy Signals
Fed Governor Waller: “Never Seen the Economy Grow Like This Without Jobs”
Federal Reserve Governor Christopher Waller told the National Association for Business Economics that U.S. payroll employment likely fell in 2025 — only the third non-recession year that has happened since 1945 — while the economy continued to grow. Revised BLS data show job creation averaged just 15,000 per month, and Waller expects further downward revisions. He noted that firms remain reluctant to hire and that AI is contributing to sustained weakness in labor demand, while productivity gains have kept output expanding.
Note: An economy that grows while shedding jobs is an economy where technology is doing more of the work. For Europe, this data point underscores a practical reality: the institutions and regions that adopt productivity-enhancing tools will capture the growth — and those that wait may find themselves competing for a shrinking pool of workers who can deliver the same output manually. The strongest case for digital transformation was never ideology. It’s arithmetic.
Sources: Federal Reserve, Fortune
$265 Million Amassed for AI Regulation Lobbying War
Groups on both sides of AI regulation — those pushing for tighter controls and those seeking to keep rules permissive — have collectively amassed at least $265 million in financial resources for lobbying efforts, according to the Financial Times. The battle over who steers AI governance is now a well-funded, multi-front campaign with direct implications for how quickly AI tools can be deployed in regulated sectors.
Note: This is the number that tells you the policy window is closing. Institutions waiting for “clear regulatory guidance” before acting should note: the guidance is being shaped right now, by groups spending nine figures to influence it.
Sources: Financial Times
