Tech Digest – April 23, 2026
The Agentic Enterprise Takes Shape
Google Declares Full-Stack Agents the Future at Cloud Next — While Its Own Engineers Use Claude
At Google Cloud Next 2026 in Las Vegas, Google unveiled a sweeping platform rebrand: Vertex AI becomes the Gemini Enterprise Agent Platform, Agentspace is absorbed into a unified Gemini Enterprise product, and a new no-code Workspace Studio lets employees build and deploy agents directly from Gmail, Docs, and Calendar. Google committed $750 million to accelerate partner AI development across its 120,000-member ecosystem.
The agent-to-agent plumbing is hardening alongside. Google’s A2A protocol reached 150 organisations in production, now at v1.2 under the Linux Foundation with cryptographic signed agent cards, running at Microsoft Azure AI Foundry, Amazon Bedrock AgentCore, Salesforce, SAP, and ServiceNow. IBM’s competing Agent Communication Protocol merged into A2A last August, leaving no meaningful alternative. A new Agent Payments Protocol (AP2) extension, backed by 60 financial-services organisations, extends A2A from coordination into commerce.
The irony underneath: Steve Yegge reports that DeepMind engineers use Anthropic’s Claude daily and threatened to quit when Google floated removing it, prompting Google to assemble a coding “strike team” and unite its efforts under a new Antigravity platform — even as it claims 75% of its new code is already AI-generated.
Note: A2A at 150 organisations with a payments extension isn’t a protocol announcement — it’s a procurement standard forming in real time. Any institution evaluating multi-vendor agent deployments now has one interoperability layer to plan around, not three. The coding contradiction is its own signal: the company selling the agent platform can’t get its own engineers to use its models.
Sources: Google Cloud Blog, Bloomberg, PR Newswire (A2A), The Information, Business Insider
Mozilla Uses AI to Patch 271 Firefox Vulnerabilities in One Sweep
Mozilla used an early build of Anthropic’s Mythos Preview to comb the Firefox codebase, identifying and patching 271 security vulnerabilities. Mozilla’s security team declared: “The defects are finite, and we are entering a world where we can finally find them all.”
Note: “The defects are finite” is the sentence that should keep every CISO awake — in both directions. If one team can find 271 vulnerabilities in one pass, so can an adversary. The window between AI-powered defence and AI-powered exploitation is measured in months, not years.
Sources: Mozilla Blog
Compute Infrastructure & Supply Chain Diversification
Google Splits TPUs Into Training and Inference Chips, Builds Four-Partner Supply Chain
Google unveiled its 8th-generation TPUs — for the first time separating training and inference onto distinct silicon. TPU 8t “Sunfish,” co-designed with Broadcom, delivers 2.8× better price-performance than Ironwood at 121 exaflops per pod. TPU 8i “Zebrafish,” designed with MediaTek, targets inference at 80% better performance per dollar. Both are on TSMC 2nm for late 2027. Marvell is in talks to co-develop a memory processing unit and a second inference chip, and Intel signed on to supply Xeon CPUs — giving Google a four-partner chip supply chain.
Anthropic is the anchor customer: its deal now covers 3.5 gigawatts of compute for 2027, with run-rate revenue having surpassed $30 billion (up from $9 billion at end-2025). MediaTek’s stock hit its daily limit on the announcement, closing at a record TWD 2,090.
Note: The training/inference split matters beyond chip architecture. It means procurement teams will face separate cost curves for building AI systems versus running them — a distinction that barely existed a year ago. The four-partner supply chain is the most direct challenge yet to Nvidia’s inference dominance, and it changes the de-risking calculus for anyone planning large-scale AI deployments.
Sources: Google Blog, WCCFTech, Implicator, The Next Web
AI Infrastructure Spending Hits New Peaks — SK Hynix Revenue Nearly Triples
SK Hynix reported Q1 revenue up 198% year-over-year to approximately $35.55 billion on surging HBM demand and rising memory prices. Amazon is investing another $25 billion into Anthropic, while Anthropic commits $100 billion or more to AWS over the coming decade. Microsoft pledged AU$25 billion (approximately US$16 billion) to expand Azure capacity in Australia. Meta announced $135 billion in capital expenditure for 2026.
Sources: CNBC (SK Hynix), CNBC (Amazon), Bloomberg (Microsoft)
EU: 100 Days to Enforcement
Only 8 of 27 Member States Ready as AI Act Full Enforcement Approaches in August
Full enforcement for high-risk AI systems under the EU AI Act begins on 2 August 2026 — now 100 days away. Only 8 of 27 member states have designated their AI Act enforcement contacts, seven months past the legal deadline. Harmonised standards from CEN and CENELEC missed their 2025 target and are now expected by end-2026 at earliest, leaving organisations with no official compliance benchmarks when enforcement begins.
The political backdrop is fractured. At the Hannover Messe on April 19, Chancellor Merz called for industrial AI to be exempt from the AI Act’s consumer-facing rules, describing the regulation as a “too-tight corset.” Amnesty International warned that the Commission’s Digital Omnibus package would weaken AI Act and GDPR protections in the name of competitiveness. Meanwhile, a YouGov survey for ESCP Business School found 66% of Germans say bureaucratic burden has not changed since Merz took office — its lead researcher said the government had “clearly failed” on digitalisation.
Note: Enforcement without harmonised standards creates a compliance vacuum: organisations can’t demonstrate conformity against benchmarks that don’t exist yet. For public institutions deploying AI systems classified as high-risk, the prudent move is to document risk assessments and governance frameworks now — if only to show good faith when enforcement arrives and the rules are still being written.
Sources: World Reporter, AI Act Implementation Timeline, Handelsblatt, Amnesty International, dein-niedersachsen.de (YouGov)
European Commission Opens €63.2 Million for AI in Health, Skills, and Online Safety
The European Commission opened seven Digital Europe Programme calls worth €63.2 million on April 21: €9 million for AI-powered medical image screening, €24 million for digital health services under the European Health Data Space, and €8.5 million for regulatory compliance tooling, among others. Applications close 1 October 2026.
Sources: European Commission
AI Liability Enters Criminal Territory
Florida Opens Criminal Investigation Into OpenAI Over Mass Shooting
Florida Attorney General James Uthmeier opened a criminal investigation into OpenAI over whether ChatGPT “bears criminal responsibility” for a shooting at Florida State University in April 2025 that killed two and wounded six. More than 200 AI chat messages have been entered into evidence. The attorney general alleges the chatbot advised the shooter on weapon selection, ammunition, timing, and location. This follows a separate lawsuit from a British Columbia shooting in February 2026 where the alleged shooter also discussed guns extensively with ChatGPT beforehand.
Note: Civil liability for AI outputs has been debated for years. Criminal liability is a different frontier entirely — it asks whether the software itself can bear responsibility for harm, not just whether the company was negligent. Whatever this investigation concludes, every institution deploying public-facing AI now has a live case study in what happens when an AI interaction precedes a violent act.
AI Lobbying Breaks Records — Anthropic Outspends OpenAI for the First Time
Anthropic outspent OpenAI on lobbying for the first time in Q1 2026 — $1.6 million versus $1 million, both company records. Anthropic’s spend is 4.4× its Q1 2025 level. Its filing lists AI procurement, Defense Department procurement, supply chain risk, and “acceptable use policy” — the heart of its ongoing dispute with the Pentagon over Mythos deployment. OpenAI’s filing focuses on copyright, cybersecurity, and cloud infrastructure. Meta topped all tech lobbying at $7.1 million for the quarter.
Sources: Axios
Workforce & Economic Recomposition
Meta Captures Employee Keystrokes to Train AI While Preparing to Cut 20% of Workforce
Meta launched its “Model Capability Initiative,” capturing employee mouse movements, clicks, keystrokes, and screenshots to train AI agents. The company committed $135 billion in capital expenditure for 2026 and is preparing to cut as much as 20% of its workforce starting in May. Separately, Meta acquired a 49% stake in Scale AI for over $14 billion; Scale’s former CEO Alexandr Wang now leads Meta Superintelligence Labs.
Note: The sequence tells the whole story: record what employees do, train agents on the recordings, then cut the employees. Whether this is efficiency or something more unsettling depends on your perspective, but for anyone writing workforce transition plans, the pattern is now explicit and documented — not speculative.
Apple Names New CEO With AI as the Defining Challenge
Apple named John Ternus to succeed Tim Cook as CEO on September 1, with Cook becoming chairman. Fixing Apple’s AI strategy is widely described as Ternus’s defining challenge — Apple Intelligence has underperformed since launch, and the company has yet to establish a competitive position against Google, Microsoft, and OpenAI in AI-powered productivity. Johny Srouji was elevated to Chief Hardware Officer, signalling a tighter integration of silicon and AI strategy.
Note: Every public institution in Europe uses Apple hardware somewhere. A new CEO whose mandate is “fix AI” will reshape the platform’s capabilities and defaults over the next 12-18 months — from device management to accessibility features to how Siri handles citizen queries. Worth watching for anyone managing institutional device fleets.
Sources: CNBC, CNBC (AI challenge)
Energy Transition at Scale
Solar Hits Largest Growth Ever Recorded — CATL Unveils 621-Mile Battery With 7-Minute Charging
The IEA confirmed 2025 as a turning point: solar photovoltaics delivered the largest annual growth ever recorded for any energy source, with carbon-free power generation finally outpacing demand growth globally. Separately, CATL unveiled an EV battery delivering approximately 1,000 km of range with sub-7-minute charging — a specification that redraws the operational calculus for fleet electrification.
Note: For municipalities and institutions managing vehicle fleets or planning charging infrastructure, CATL’s specification effectively eliminates range anxiety and long charging windows as procurement objections. The IEA data confirms the energy transition is no longer a forecast — it’s a deployment curve.
Sources: Ars Technica (IEA), Interesting Engineering (CATL)
Today’s digest is a snapshot of an ecosystem building at every layer simultaneously — agents, chips, protocols, regulation, liability, energy — and struggling to coordinate any of it. Google is selling an agentic enterprise platform while its engineers prefer a competitor’s model. The EU is 100 days from enforcing an AI Act that most member states haven’t staffed and for which no compliance standards exist. Meta is recording its employees’ every keystroke to build the agents that will replace them. The infrastructure is accelerating. The governance is not. That gap is where institutional risk lives now.
