Tech Digest – March 4, 2026

AI Market & Trust

Anthropic Crosses $20 Billion Annualized Revenue — Doubled in Under Three Months

Anthropic is now approaching $20 billion in annualized revenue, more than doubling from $9 billion reported at year-end 2025. The growth comes despite a Pentagon supply-chain-risk designation and a cancelled $200 million Department of War contract — institutional friction that did not slow commercial adoption. The company’s trajectory puts it among the fastest-scaling enterprise software businesses ever recorded.

Note: The Pentagon friction is worth reading carefully: it signals that AI vendors are now subject to the same geopolitical supply-chain scrutiny as hardware manufacturers. Any EU institution evaluating AI procurement should expect similar classification questions to reach European contract processes within 12–18 months.

Sources: Bloomberg

ChatGPT Uninstalls Surge 295% After Pentagon Deal; Claude Climbs to No. 1 on the U.S. App Store

Following OpenAI’s announced Department of Defense contract, ChatGPT saw a 295% day-over-day spike in uninstalls and a 775% rise in 1-star reviews. Claude captured the No. 1 position on the U.S. App Store on a 51% download surge in the same window. The episode produced a measurable, quantified shift in market position within 48 hours of a single vendor decision.

Note: Trust is now a market variable with a measurable lag time measured in hours, not quarters. Vendor alignment with military or intelligence clients will increasingly shape institutional adoption decisions — especially in European public sector procurement, where defence entanglements directly affect permissible supplier lists.

Sources: TechCrunch

The Model Sprint

Three Major Models in One Week: GPT-5.3, Gemini Flash-Lite, and DeepSeek V4 Expected

OpenAI released GPT-5.3 Instant with hallucinations trimmed by roughly 30% and teased GPT-5.4 as “sooner than you think.” Google launched Gemini 3.1 Flash-Lite at $0.25 per million input tokens — with 2.5x faster time-to-first-answer and tunable reasoning depth — making capable AI effectively free at scale. DeepSeek is expected to release its trillion-parameter V4 this week, reportedly featuring natively multimodal 1-million-token context, timed to China’s Two Sessions parliamentary meetings.

Note: Capability, cost, and geopolitical timing moving simultaneously from three directions. The $0.25/million token price point means institutions can now run AI over large document sets — procedures, archives, citizen correspondence — for effectively no marginal cost per query. The procurement question is shifting from “can we afford this” to “what governance do we put around it.”

Sources: OpenAI, Google Blog, TechNode

Cursor’s AI Solves an Open Math Research Problem in Four Days — Beats the Official Human Answer

Cursor’s AI agent autonomously solved Problem Six of the First Proof challenge — a set of math research problems approximating the work of Stanford, MIT, and Berkeley academics — over four days, without hints, and produced a solution yielding stronger results than the official human-written answer. The task was completed inside a code editor, with no dedicated research infrastructure. Cursor CEO Michael Truell confirmed the result directly.

Note: Mathematical research is now a feature of a coding tool, not a vocation requiring years of specialization. The institutional implication isn’t about math — it’s about what category of knowledge work is next to become a plugin rather than a profession.

Sources: Michael Truell / Cursor (X)

Workforce & Market Structure

Zero Junior Hires Since 2024, 32 Million Roles Transformed Per Year, and Jamie Dimon Floating UBI

Gartner’s 2025 AI Job Impacts Analysis projects that AI will create more jobs than it eliminates starting in 2028–2029 — but that 32 million roles will be significantly transformed every year in the process. The ground-level picture is sharper: a CTO of a 50-person, $100M ARR startup reports zero junior hires since 2024, with senior employees now 3x more productive using AI. JPMorgan CEO Jamie Dimon warned publicly that mass automation will cause civil unrest if it moves too fast, noted that even his own firm will employ fewer people, and floated universal basic income as a potential release valve.

Note: The gap between Gartner’s 2028 net-positive forecast and the CTO’s 2024 hiring freeze is the policy window institutions need to act in. Workforce planning built around a 2028 equilibrium ignores what happens to entry-level pipelines between now and then — and who absorbs the transition cost.

Sources: Gartner (X), Alex Turnbull (X), UBI Works / Dimon remarks (X)

SaaSpocalypse: Per-Seat Pricing Collapsing, No Venture-Backed SaaS IPOs on the Horizon

The term “SaaSpocalypse” is now being used to describe a structural breakdown in the per-seat SaaS pricing model, with AI alternatives undercutting subscription software across multiple categories. No venture-backed SaaS companies have filed for IPO. Analysts describe the investment climate as defined by “FOBO” — Fear Of Becoming Obsolete — with capital flowing away from established SaaS incumbents and toward AI-native alternatives.

Note: Institutions sitting on multi-year SaaS contracts should be asking whether the vendor will exist in its current form when renewal comes. “Upgrade path” is no longer a vendor feature — it’s a procurement risk factor.

Sources: TechCrunch

Quinn Emanuel Deploys AI Platform That Surfaces Impeachment Evidence While a Witness Is Still Speaking

An attorney at Quinn Emanuel, one of the world’s largest litigation firms, described a proprietary in-house AI platform that processes discovery documents in real time during trial — surfacing impeachment evidence and contradictions while a witness is still on the stand. The platform draws from all documents produced in discovery and operates as a live adversarial tool. Quinn Emanuel has been among the most aggressive adopters of AI litigation tools, including TrialView (used in a £335M fraud case) and Syllo AI for rapid evidence review.

Note: When AI becomes a standard component of litigation, institutions that haven’t documented their processes, decisions, and data handling are at an asymmetric disadvantage in any dispute. Audit trails and records governance aren’t just compliance requirements — they’re legal exposure management.

Sources: TBPN / Quinn Emanuel attorney (X), TrialView

Hardware, Power & Infrastructure

Apple’s M5 Pro and M5 Max Are Designed Around One Assumption: The Primary Workload Is Running LLMs Locally

Apple announced the M5 Pro and M5 Max, bonding two third-generation 3nm dies into a Fusion Architecture SoC with Neural Accelerators embedded inside each GPU core, delivering 4x the AI performance of M4. Commentary from chip analysts frames the architectural decision plainly: Apple has reorganized its silicon around the assumption that local LLM inference is the defining workload of a professional laptop in 2026. No cloud required.

Note: Local inference at this performance level changes the data sovereignty calculus. Sensitive institutional workflows that can’t go to the cloud have a hardware path now — one that doesn’t require a dedicated on-premise server room.

Sources: Apple Newsroom, Aakash Gupta (X)

Peking University Achieves 1nm Ferroelectric Transistors — Ten Times More Energy-Efficient Than Previous Records

A team at Peking University demonstrated ferroelectric transistors at a 1-nanometer gate length, switching at just 0.6V — an order of magnitude more energy-efficient than previously recorded results. The breakthrough advances the physics frontier on chip density and power consumption simultaneously, with direct implications for the cost and environmental footprint of AI inference at scale.

Note: Energy consumption is currently the primary physical constraint on AI deployment. A 10x efficiency improvement at the transistor level, if it reaches production, compresses the timeline on affordable, sustainable AI infrastructure — including for institutions that can’t build next to a nuclear plant.

Sources: TrendForce

Power Is the New Constraint: $876M Bet on Fuel Cells, Iowa County Enacts Strict Data Center Zoning Near Nuclear Plant

Leopold Aschenbrenner’s $5.52 billion Situational Awareness portfolio is led by an $876 million stake in Bloom Energy — a fuel-cell company that generates electricity on-site from natural gas — not in Nvidia or any AI lab. Meanwhile, Linn County, Iowa unanimously enacted one of America’s most comprehensive data center zoning ordinances, requiring mandatory water studies, noise limits, and community impact agreements, prompted by Google’s plans for a six-building campus co-located with a restarting nuclear plant under a 25-year power purchase agreement.

Note: When the sharpest AI infrastructure analyst bets on power generation over chips, and a rural Iowa county is ahead of most European municipalities in data center governance, the planning signal is the same from both ends: energy and water are the binding variables in the next phase of digital infrastructure. European public administrations approving or hosting digital infrastructure projects should be adding power and water impact assessments to their evaluation criteria now.

Sources: Limitless / Aschenbrenner 13F (X), Inside Climate News, Corridor Business