Tech Digest – February 19, 2026
Agents Are Getting Longer Leashes — and Earning Paychecks
Claude Code Sessions Nearly Doubled in Length in Three Months — and One Agent Is Already Shipping Apps for Profit
Anthropic reports that the 99.9th percentile Claude Code session nearly doubled from 25 to 45 minutes between October and January — a steady climb toward longer autonomous missions without human intervention. Meanwhile, entrepreneur Austen Allred says his AI agent “Kelly” has shipped half a dozen apps to the App Store, earned thousands in contract revenue, and done it all without a human writing a single line of code. The gap between “tool you prompt” and “employee you brief” is closing month by month.
Note: Session length is an underappreciated metric. It’s a proxy for trust: how long can the system work before a human needs to check in? Doubling that window in one quarter means the range of tasks you can delegate is expanding on a steep curve.
Sources: Anthropic, Austen Allred (X)
GPT-5.3 Scores 72% on Smart Contract Exploitation — the Machines Can Already Audit the Money
OpenAI introduced EVMbench, a benchmark for evaluating AI on Ethereum smart contract security. GPT-5.3-Codex scored 72.2% on exploitation tasks, demonstrating that current models can identify and exploit most common vulnerabilities in the financial code that underpins decentralized finance. The benchmark covers real-world attack patterns, not toy examples.
Note: If AI can exploit 72% of smart contract vulnerabilities, it can audit them too. Any institution handling digital payments, procurement platforms, or automated financial processes should be asking: who is testing our code, and are they faster than the systems trying to break it?
Sources: OpenAI
The Capital Concentration Is Staggering
OpenAI Closing $100B Round at $830B — While xAI and Europe’s Largest Seed Round Stack Up Behind It
OpenAI is finalizing commitments for a $100 billion fundraise at an $830 billion valuation — a figure that would make it more valuable than all but a handful of public companies on Earth. In parallel, Saudi Arabia’s HUMAIN invested $3 billion in xAI’s Series E, and former DeepMind scientist David Silver is raising $1 billion for Ineffable Intelligence, which would be Europe’s largest seed round ever. The capital flowing into AI labs is now measured in the hundreds of billions per cycle.
Note: $100 billion for one company. $3 billion from a sovereign wealth fund. $1 billion for a European seed round. These aren’t venture bets — they’re infrastructure commitments at nation-state scale. The organizations receiving this capital will shape what tools are available, at what price, for the next decade.
Sources: The Information, HUMAIN, Economic Times
The Physical Layer Is Mutating
A Toilet Maker Gets 40% of Its Profits from AI Memory — and Activists Want More
Japanese toilet manufacturer Toto now derives 40% of its operating income from ceramics used in AI memory production, not bathrooms. The company’s stock has risen nearly 40% in the first two months of 2026, and activist investors are pushing it to reduce its plumbing business and double down on semiconductor materials. Toto’s pivot illustrates how deep AI demand reaches into supply chains that have nothing to do with software.
Note: When a toilet company becomes a semiconductor play, the AI supply chain has reached sectors that no procurement scan would flag. The materials bottleneck isn’t just chips — it’s the ceramics, substrates, and specialty inputs upstream of the fabs.
Sources: Tom’s Hardware
Microsoft Encodes 4.8 TB in Glass That Lasts 10,000 Years — Efficient Computer Targets a Trillion Ops Per Watt
Microsoft’s Project Silica, published in Nature, demonstrated encoding 4.8 terabytes of data across 301 layers in a glass substrate designed to last 10,000 years without degradation. Separately, Efficient Computer raised $60 million for a chip architecture targeting one trillion operations per watt — a 100x improvement over current efficiency benchmarks. Both developments address the same constraint from opposite directions: storing and processing ever-larger volumes of data within sustainable energy and physical footprints.
Note: For any institution managing long-term archives — cultural heritage, legal records, scientific data — glass-based storage changes the preservation calculus entirely. No migration cycles, no media degradation, no format obsolescence. The question shifts from “how do we keep refreshing this?” to “write once, forget for centuries.”
Sources: Nature, Efficient Computer
OpenAI Anchors a 1-GW Data Center in India — Meta Spends $65M on Politicians to Clear the Permits
OpenAI will anchor Tata’s new HyperVault data center in India with capacity up to 1 gigawatt — enough to power a mid-sized city. Meanwhile, Meta reportedly spent $65 million on AI-friendly political candidates in the last election cycle, specifically to ease permitting and zoning for data center construction. The infrastructure buildout is now large enough that companies are shaping the political environment to remove physical constraints.
Note: A 1-GW data center in India and $65 million in political spending to clear permits in the U.S. — these are not technology decisions. They’re infrastructure and real estate plays at industrial scale. The geography of compute capacity is being redrawn right now, and institutions that depend on cloud services will feel the effects in latency, pricing, and data sovereignty.
Sources: OpenAI, The New York Times
The Workforce Is Being Restructured in Real Time
Accenture Ties Promotions to AI Usage — While a Newsroom Redesigns Roles Around It
Accenture is tying employee promotions to AI adoption and tracking weekly login frequency to AI tools. At Cleveland.com, the newsroom took a different approach: it assigned reporter writing to an “AI rewrite specialist,” freeing up an extra workday per reporter for street journalism. The result — reporters are returning with more story ideas than the newsroom can handle. Two models, one direction: organizations that treat AI as a productivity layer are restructuring roles now, not later.
Note: Accenture’s approach is the stick — use it or stall. Cleveland.com’s is the redesign — change who does what so humans do more of what they’re best at. Both assume the same thing: AI proficiency is no longer optional for knowledge workers. The institutions that figure out which model fits their culture will adapt faster.
Sources: Financial Times, Cleveland.com
Andrew Yang Warns of Millions Displaced in 12-18 Months — UK Tribunals Already See a 33% Surge in AI Grievances
Andrew Yang published “The End of the Office,” warning that AI automation will displace millions of white-collar workers within 12 to 18 months and estimating the U.S. could lose 20-50% of its 70 million office jobs over the next several years. The friction is already visible: UK employment tribunals report a 33% increase in AI-generated “slop grievances” — cases where AI-drafted complaints flood the system. The displacement signal and the institutional friction signal are arriving simultaneously.
Note: Yang’s timeline may be aggressive, but the direction matches what Suleyman said yesterday. More telling is the UK tribunal data: a 33% surge in AI-generated complaints means the legal and administrative systems that process workforce disputes are themselves being overwhelmed by AI output. The tools creating displacement are also creating the paperwork around it.
Sources: Andrew Yang, Financial Times
Measured Impact
Survey of 12,000 EU Firms: AI Lifts Productivity 4%
A CEPR/VoxEU study surveying over 12,000 firms across the European Union found that AI adoption is associated with a 4% productivity increase. The data covers firms across sectors and sizes, providing the first large-scale EU-specific measurement of AI’s productivity effect in practice — not in lab conditions or vendor case studies, but in real operating environments.
Note: Four percent is modest enough to be credible and large enough to matter at scale. For an institution processing thousands of citizen requests, 4% means measurably fewer backlogs, shorter wait times, or reduced overtime — without changing headcount. This is the kind of baseline data that belongs in a business case.
Sources: CEPR/VoxEU
AI Matches Expert Diagnosis Across 2,919 Rare Diseases at 95.4% Agreement
DeepRare, published in Nature, achieved 95.4% agreement with human experts across 2,919 rare disease diagnoses. Rare diseases affect over 300 million people worldwide, and patients typically endure years-long diagnostic odysseys. The system doesn’t replace clinicians — it shortens the path to the right specialist and the right diagnosis by orders of magnitude.
Note: Rare disease diagnosis is one of the clearest cases where AI doesn’t threaten jobs — it fills a gap that humans can’t cover at scale. For health and social services, this kind of diagnostic support means earlier intervention, lower long-term costs, and fewer patients falling through the cracks.
Sources: Nature
