Tech Digest – February 20, 2026

AI Capabilities & Cost

Gemini 3.1 Pro Takes the Lead — at Half the Cost of Its Competitors

Google launched Gemini 3.1 Pro with state-of-the-art scores: 44.4% on Humanity’s Last Exam (without tools), 77.1% on ARC-AGI-2, and 94.3% on GPQA Diamond. Artificial Analysis ranked it the new Intelligence Index leader. The more striking detail: running their evaluations on Gemini 3.1 Pro cost less than half what the same tests cost on Claude Opus 4.6 or GPT-5.2.

Note: The performance gap at the top keeps narrowing while the price gap widens downward. For any institution evaluating AI tools for internal use, the procurement question is shifting from “which model is best” to “which model is best per euro.”

Sources: Google AI Blog, Artificial Analysis

Anthropic Growing 10x Per Year — May Overtake OpenAI Revenue by Mid-2026

Epoch AI analysis shows Anthropic has grown at roughly 10x per year since hitting $1 billion in annualized revenue, compared to OpenAI’s 3.4x over the same period. If the trend holds, Anthropic could surpass OpenAI’s revenue by mid-2026. Separately, Nvidia is reportedly close to finalizing a $30 billion investment in OpenAI, deepening the vertical integration between chipmakers and foundation model companies.

Note: Two years ago, the AI vendor market looked like a one-horse race. Now there are at least three frontier providers competing on price, performance, and ecosystem. That’s good news for buyers — but the $30B Nvidia-OpenAI link is a reminder that “competition” doesn’t mean these companies are independent of each other.

Sources: Epoch AI, Financial Times

Infrastructure & Capital Flows

US Now Imports More From Taiwan Than China — First Time Since 1992

New US trade data shows that American imports from Taiwan have surpassed those from China for the first time since 1992. The shift is driven by semiconductor purchases feeding AI infrastructure buildout. The trade war reduced direct US-China trade volumes, but the AI boom accelerated Taiwan’s rise as the dominant import partner.

Note: Every digital project in Europe runs on chips that overwhelmingly flow through one island. This isn’t a trivia item — it’s the single biggest supply chain dependency in institutional IT, and it just became the defining line in US trade.

Sources: Joey Politano (US trade data analysis)

AMD GPUs Used as Loan Collateral for the First Time — Chips Become Financial Instruments

AMD is backstopping a $300 million loan to AI infrastructure company Crusoe, using AI chips as collateral. This is the first known instance of AMD GPUs functioning as a financial instrument — following a pattern Nvidia pioneered. The deal signals that compute hardware now carries asset-class characteristics similar to real estate or commodities.

Note: When banks accept GPUs as collateral, compute isn’t just expensive — it’s scarce enough to underwrite debt. Institutions planning large infrastructure purchases should expect hardware availability and pricing to behave more like capital markets than retail procurement.

Sources: The Information

Amazon Overtakes Walmart as World’s Largest Company by Revenue

Amazon has dethroned Walmart as the world’s largest company by revenue, a shift powered primarily by its cloud and data center business rather than retail growth. The milestone reflects how infrastructure spending — not consumer commerce — is now the dominant revenue driver for the world’s biggest company.

Sources: Bloomberg

Energy & the Compute Buildout

Data Centers Build a Shadow Power Grid — SoftBank Plans 9.2 GW Gas Plant

AI data centers are spawning a parallel energy infrastructure of off-grid natural gas plants, bypassing the public grid entirely because solar and wind variability cannot reliably support continuous compute loads. SoftBank is forming a consortium to build a $33 billion gas-fired power plant on the Ohio-Kentucky border. At 9.2 GW, it would be the largest power plant in the United States. Separately, Type One Energy submitted a licensing application for its Infinity One stellarator fusion reactor in Tennessee, targeting a working prototype by 2029.

Note: The energy story has split in two. Short-term: gas plants being built at national-infrastructure scale to feed AI demand. Long-term: fusion timelines are now measured in licensing applications, not theoretical papers. For EU energy and sustainability planning, the tension between “twin transition” goals and the raw power appetite of AI infrastructure is no longer abstract.

Sources: Washington Post, Nikkei Asia, Popular Mechanics

Workforce & Synthetic Labor

Freelancers Replaced at 33:1 — One Dollar of Human Work Now Costs Three Cents in AI

New research finds that businesses are replacing freelancers on platforms like Upwork and Fiverr with AI at a ratio of $1 of human labor to $0.03 in AI spending. The study documents a structural shift in how organizations source knowledge work — not a marginal efficiency gain, but an order-of-magnitude cost collapse. The cultural echo is already registering: BAFTA has introduced “human achievement” as a guiding principle for its awards, the first major cinematic institution to formally distinguish between human and machine-generated creative work.

Note: Three cents on the dollar. That ratio doesn’t leave room for “upskilling” to compete on price. The question for any institution outsourcing knowledge work is no longer whether AI will be cheaper — it’s whether the remaining human premium buys quality, accountability, or just habit.

Sources: arXiv, Financial Times (BAFTA)

500,000 People Sign Up to Be Hired by AI Agents

RentAHuman.ai, a platform where AI agents can search, book, and pay humans for physical tasks via API, reports over 500,000 signups since its February launch. The platform provides MCP integration so autonomous agents can hire humans with a single function call. Actual task fulfillment remains limited — most gigs are low-value and the crypto-only payment model filters for novelty over serious work. But the signup volume, even if inflated, signals willingness to participate in a labor market where machines are the employers.

Note: The platform is more proof-of-concept than labor market. But half a million people raising their hand to work for bots is a data point that workforce planners shouldn’t dismiss. The inversion — AI as buyer of human labor — was theoretical six months ago.

Sources: RentAHuman.ai, Interesting Engineering, Futurism

Why Europe Can’t Restructure: 62 Months of Salary to Let One Employee Go in Spain

A Works in Progress analysis of European labor costs finds that corporate restructuring in Spain costs the equivalent of 62 months of salary per employee — compared to 7 months in the United States. France and Germany come in at 38 and 31 months respectively, with Italy at 52. The analysis argues these costs explain why Europe has failed to build companies that can take repeated, large-scale innovation bets: the downside of failure is structurally higher, suppressing both venture investment and strategic risk-taking.

Note: When the cost of course-correcting is measured in years of salary per person, institutions and companies alike default to inaction. In a period where AI is compressing product and service cycles from years to months, Europe’s restructuring costs are not just a labor policy issue — they’re a speed-of-adaptation issue.

Sources: Works in Progress