Tech Digest – May 6, 2026
AI Capabilities Cross the Trust Line
GPT-5.5 Instant Cuts Hallucinations by Half — While a $29M Startup Fits 12 Million Tokens at a Fraction of the Cost
OpenAI’s GPT-5.5 Instant produces 52.5% fewer hallucinated claims than its predecessor on high-stakes prompts in medicine, law, and finance, and claimed the top spot on FrontierSWE, the hardest benchmark for ultra-long-horizon coding agents. Google’s multi-token prediction drafters added another efficiency lever, delivering 3x inference speedups for Gemma 4 with no quality loss.
Separately, Miami-based Subquadratic emerged from stealth with $29 million in seed funding and a 12-million-token context model — roughly 120 books in a single window — that demands approximately 1,000x less compute than standard transformer architectures. Its Sparse Attention mechanism scored 65.9% on MRCR v2, approaching frontier-model territory at a claimed fraction of the FLOPs. The full technical report has not yet been published, and prior subquadratic architectures have underperformed transformers at frontier scale — but if the claims hold, the cost structure of long-context AI shifts dramatically.
Note: The two biggest barriers to institutional adoption — “can I trust it?” and “can I afford to run it?” — are falling simultaneously. Hallucination rates halving on medical and legal prompts matters more to a procurement officer than any benchmark number.
Sources: OpenAI, Proximal (FrontierSWE), Google, SiliconAngle
Platform Intelligence & the Agent Stack
Apple Opens Intelligence to Third-Party Models — After Paying $250M for Overpromising the First Round
Apple’s iOS 27 will let users swap third-party AI models in and out of Apple Intelligence via the Settings app, treating intelligence itself like a default browser. Bloomberg reports the feature will give users control over which AI powers Siri, summarisation, and other system-level functions.
The move follows a $250 million settlement over the gap between Apple Intelligence’s marketing and its actual capabilities — a tangible reminder that AI hype now carries a price tag.
Note: Default browsers had antitrust battles. Default AI models will have bigger ones.
Sources: Bloomberg, New York Times
Meta and OpenAI Race to Own the Personal AI Layer
Meta is building a personal AI agent for its billions of users, according to the Financial Times, aiming to embed always-on assistance across its platforms. OpenAI is reportedly fast-tracking its first AI agent phone for mass production in the first half of 2027, according to supply chain analyst Ming-Chi Kuo. The two largest consumer AI platforms are converging on the same bet: the AI interface will be the primary interface, whether it lives in an app or in your hand.
Note: When the phone itself becomes an AI agent, every citizen-facing digital service — from tax portals to public transport apps — will need to interoperate with it.
Sources: Financial Times, Ming-Chi Kuo / X
Anthropic Ships 10 Finance Agents — From Pitchbooks to Month-End Close
Anthropic released ten pre-built finance agents covering pitchbook generation, credit memos, KYC screening, underwriting, month-end close, statement audits, and insurance claims processing. Each ships as a plugin for Claude Cowork and Claude Code, with full Microsoft 365 integration — a team can deploy Claude on real financial work in days rather than months. Data partnerships with Verisk, Dun & Bradstreet, Experian, S&P Capital IQ, Morningstar, and PitchBook provide the agents with institutional-grade source material.
Note: Ten agents, ten fewer RFPs. When back-office automation ships as a product rather than a project, the procurement conversation skips straight to integration.
The Silicon Supercycle
Samsung Crosses $1 Trillion as Global Chip Sales Hit $298.5 Billion in a Single Quarter
Samsung Electronics became the second Asian company to cross $1 trillion in market capitalisation, following TSMC, after shares more than quadrupled over the past year on AI-driven memory demand. The milestone pushed KOSPI above 7,000 for the first time. Global semiconductor sales reached $298.5 billion in Q1 2026, with March alone recording 79.2% year-over-year growth according to the Semiconductor Industry Association.
Micron began shipping its highest-capacity SSD, pushing past a $700 billion market cap and into the top ten US tech companies amid an AI-driven memory shortage. AMD’s Q2 forecast beat Wall Street expectations on sustained data-centre demand, sending shares up 12% in extended trading atop a 65% year-to-date run.
Note: A $298.5 billion quarter means the physical layer of the AI transition is scaling faster than any digital transformation plan written in 2024 anticipated. These aren’t stock movements — they’re infrastructure signals.
Sources: CNBC (Samsung), SIA, CNBC (Micron), Reuters (AMD)
China Targets 70% Domestic Wafers This Year — Apple Explores Intel and Samsung as US Fabs
China is targeting 70% domestic silicon wafer usage by year’s end, according to Nikkei, accelerating its drive for semiconductor self-sufficiency. Separately, Apple is exploring Intel and Samsung as US-based fabrication partners beyond TSMC, diversifying its chip supply chain away from a single manufacturer. The news drove Intel up 13% to an all-time high, capping a 114% monthly rally that has rewritten the chip-stock landscape.
Note: Two parallel supply chains are being built — not by accident, but by policy. Any institution planning IT procurement on a 3-5 year horizon is buying into one of these ecosystems, whether it realises it or not.
Sources: Nikkei Asia, Bloomberg, CNBC (Intel)
Compute Infrastructure at Every Scale
OpenAI Plans $50 Billion in Compute This Year — Anthropic Commits $200 Billion to Google Over Five
OpenAI will spend $50 billion on compute in 2026 alone, co-founder Greg Brockman stated during court testimony. Anthropic is committing $200 billion to Google Cloud over five years, a single contract now representing over 40% of Google’s disclosed cloud revenue backlog, according to The Information.
At the residential end of the spectrum, Span is deploying XFRA mini data centres inside PulteGroup neighbourhoods, embedding Nvidia GPUs into spare grid capacity and turning suburban homes into inference nodes. Compute infrastructure is expanding from cul-de-sacs to quarter-trillion-dollar contracts simultaneously.
Note: The SEC is proposing to replace quarterly 10-Q earnings reports with semiannual 10-S filings — even regulators are conceding that the capital cycles driving these investments no longer fit a 90-day reporting window.
Sources: Bloomberg (OpenAI), The Information, CNBC (Span/Nvidia/Pulte), CNBC (SEC)
AI Crosses Into Health and Safety
Meta Runs Bone Scans on User Photos — Pennsylvania Sues Character.AI Over AI Doctors
Meta has begun running AI bone-structure analysis on user photos to identify under-13 accounts on Instagram and Facebook — performing radiology without radiation and turning ordinary selfies into clinical signal. Separately, Pennsylvania’s governor sued Character.AI in the first lawsuit of its kind by a US state executive, alleging the platform’s chatbots impersonated licensed medical professionals in conversations with minors.
Note: One company turned selfies into clinical data. Another’s chatbot passed, inadvertently, a bedside Turing test. Both are now explaining themselves to regulators. The EU AI Act’s high-risk classification framework was built for exactly this territory.
Workforce Rewritten
Coinbase Cuts 14% of Staff — DeepMind UK Workers Unionise Over Military Deal
Coinbase is laying off 14% of its workforce. CEO Brian Armstrong explained that engineers now ship in days what teams used to ship in weeks, with even non-technical employees pushing production code. At Google DeepMind’s UK operation, workers voted to unionise — not over pay or conditions, but over a partnership with the US military, in one of the first cases of AI-sector collective action driven by ethical objections to deployment.
Note: Two workforce stories pulling in opposite directions. In one, productivity gains erase roles. In the other, workers organise not to resist automation but to control where it’s pointed. Both threads will shape how EU AI labour policy develops.
Sources: Brian Armstrong / X, The Guardian
Today’s signals point in the same direction at every layer. Chip demand surges 79% year-over-year. Hallucination rates halve. Platforms commoditise AI like they once commoditised browsers. Two companies commit a quarter-trillion dollars to compute infrastructure while a startup claims to fit 12 million tokens into a fraction of the cost. The gap is no longer between those who use AI and those who don’t — it’s between those building the next five years on current assumptions and those who’ve noticed the assumptions changed.
