Tech Digest – April 1, 2026

Capital at Escape Velocity

OpenAI Closes at $852 Billion as Global AI Investment Hits $297 Billion in a Single Quarter

OpenAI completed its record-breaking funding round at an $852 billion post-money valuation, raising $122 billion in committed capital — anchored by Amazon ($50 billion), Nvidia ($30 billion), and SoftBank ($30 billion). For the first time, $3 billion came from retail investors. The company now generates $2 billion in monthly revenue, with enterprise accounting for over 40% and Codex serving more than 2 million weekly users, up fivefold in three months. The round landed in a quarter that saw global venture capital hit $297 billion — up 150% year-over-year — with AI startups capturing 81% of the total and just four companies absorbing 64%.

Note: Four companies captured nearly two-thirds of global venture capital in a single quarter. For anyone still modelling AI as a sector among sectors, the capital markets have moved on. The infrastructure budgets, procurement timelines, and talent pools available to everyone else are now shaped by this concentration — whether you’re investing in it or not.

Sources: CNBC, Bloomberg, Financial Times, OpenAI, Crunchbase

Oracle Cuts Thousands of Jobs to Fund AI Data Centre Buildout

Oracle is cutting thousands of positions as it accelerates spending on AI data centres, trading human headcount for compute capacity. The restructuring follows a pattern that has now repeated across multiple enterprise technology vendors in 2026: redirect the salary line into the server line. Oracle did not disclose exact figures but confirmed the cuts are directly tied to capital reallocation toward AI infrastructure.

Note: For institutions with Oracle contracts — and there are many in the public sector — the question is whether reduced headcount affects support quality and service continuity during the transition. This is the moment to review SLA terms and escalation paths.

Sources: CNBC

AI Security Under Stress

Anthropic’s Claude Code Leaks via npm Error — Supply Chain Attack Follows Within Hours

Anthropic accidentally published the full source code of Claude Code — a 512,000-line TypeScript codebase and one of the most closely guarded AI agent architectures — to the public npm registry. The cause: a developer omitted source map files from the packaging exclusion list. Third-party forensics revealed 44 hidden feature flags, an always-on background agent mode called KAIROS, an “undercover mode” that hides internal codenames, and a regex-based sentiment analyser to detect user frustration. Within hours, trojanized versions appeared on GitHub — users who installed or updated Claude Code via npm on March 31 between 00:21 and 03:29 UTC may have pulled a remote access trojan. Anthropic confirmed no customer data was exposed and attributed the incident to “plain developer error.”

Note: The leak itself is an embarrassment. The supply chain attack that followed is the institutional lesson. A single packaging error in one dependency turned a leading AI tool into a malware distribution vector within hours. Any organisation deploying AI tools via public package registries now has a live case study for why supply chain auditing cannot be deferred.

Sources: VentureBeat, The Hacker News, CNBC, The Register

The Compression Curve

A Full AI Model in 1.15 GB — Intelligence Density Jumps 10x

PrismML released Bonsai 8B, billed as the first commercially viable single-bit large language model. The 8-billion-parameter model requires just 1.15 GB of memory — a 14x smaller footprint than full-precision equivalents — while matching them on standard benchmarks, with 8x faster inference and 5x better energy efficiency. Released under Apache 2.0, it targets robotics, real-time agents, and edge computing. Meta researchers pushed the compression frontier further with TinyLoRA, training Qwen2.5 8B to 91% accuracy on a maths reasoning benchmark using just 13 parameters — 26 bytes total. Separately, Google introduced Veo 3.1 Lite, its most cost-effective video generation model, at less than half the cost of Veo 3.1 Fast with equivalent speed.

Note: An 8-billion-parameter model running in barely over a gigabyte of memory means capable AI runs on hardware institutions already own — laptops, tablets, edge devices. When the deployment barrier drops from “cloud GPU cluster” to “existing device,” the procurement question changes from “what infrastructure do we need?” to “what policy framework do we need for what’s already possible?”

Sources: PrismML, The Register, TinyLoRA (arXiv), Google

Cryptographic Countdown

Google Quantum AI: Breaking Crypto Needs 20x Fewer Qubits Than Previously Estimated

Google Quantum AI published a whitepaper demonstrating that breaking the elliptic curve cryptography protecting Bitcoin, Ethereum, and most major digital assets could require fewer than 500,000 physical qubits — a 20x reduction from prior estimates. The team compiled two quantum circuits implementing Shor’s algorithm for ECDLP-256: one using fewer than 1,200 logical qubits and 90 million Toffoli gates, another using fewer than 1,450 logical qubits with 70 million gates. Under their analysis, a superconducting quantum computer could crack a private key in approximately nine minutes once a transaction exposes the public key. Google urged the cryptocurrency community to begin transitioning to post-quantum cryptography.

Note: This is not only a cryptocurrency problem. Elliptic curve cryptography underpins TLS certificates, digital signatures, and authentication systems across every sector. A 20x reduction in estimated attack resources accelerates the timeline for “Q-Day” — when current encryption becomes unsafe. Three papers in three months have rewritten the quantum threat horizon. Any institution that has not yet begun evaluating post-quantum cryptographic standards is working against a shorter clock than it was in January.

Sources: Google Research, The Quantum Insider, CoinDesk, SiliconANGLE

Autonomous Systems Diversify

$1.75 Billion for Autonomous Warships, Humans Still Driving Tesla’s Robotaxis Below 10 mph

Saronic raised $1.75 billion at a $9.25 billion valuation to scale production of autonomous naval vessels for the US military, with plans to build more than 20 ships per year by 2027 from its expanding Louisiana shipyard. The round was led by Kleiner Perkins, with participation from Andreessen Horowitz, Advent International, and others. On the road, Tesla acknowledged that its robotaxis are sometimes operated by remote humans at speeds below 10 mph — an industry concession toward “centaur driving” where autonomy and human oversight share the controls. Meanwhile, Grab and WeRide launched Southeast Asia’s first driverless ride-hailing service in Singapore, adding another geography to the autonomous deployment map.

Note: The gap between autonomous claims and operational reality is where regulatory frameworks get built. Saronic’s $1.75 billion for unmanned warships and Tesla’s quiet admission that humans are still in the loop sit in the same policy space. For EU institutions watching the AI Act’s risk classifications unfold in practice, autonomous mobility — across sea, road, and air — is where deployment is outpacing the rules.

Sources: CNBC, Wired, Bloomberg

Nvidia Invests $2 Billion in Marvell to Build the Optical Wiring for Next-Generation AI Clusters

Nvidia invested $2 billion in Marvell Technology and announced a partnership to co-develop silicon photonics — the optical interconnect fabric for next-generation AI data centres. As AI clusters scale beyond the bandwidth limits of electrical copper wiring, silicon photonics replaces electrons with light, enabling higher throughput at lower power over longer distances. The investment signals that the AI infrastructure bottleneck is shifting from processors to the connections between them.

Note: The bottleneck moved. The next infrastructure cycle is not just “more GPUs” — it is a fundamentally different wiring architecture. Institutions planning cloud strategy or data centre procurement over a 3-5 year horizon should note that the physical layer underneath AI is changing, not just scaling.

Sources: Bloomberg

When Machines Prove Theorems

OpenAI Solves Three Erdős Conjectures — Mathematical Proof as Routine Deployment

OpenAI researcher Mehtaab Sawhney announced that an internal model solved three previously open problems posed by Paul Erdős, one of the most prolific mathematicians in history. Each proof was described as “short and elegant.” The achievement extends AI’s reach from empirical pattern-matching into formal mathematical reasoning — territory long considered among the last redoubts of uniquely human intellectual capability.

Sources: Mehtaab Sawhney (OpenAI)

The through-line today is compression — of model sizes, of timelines, of the distance between what’s possible and what’s planned for. AI fits in a gigabyte. Quantum threats arrive 20x sooner than estimated. A packaging typo turns a leading AI tool into a malware vector in hours.

The common thread for anyone running an institution: the assumptions you made last quarter about cost, security, and capability are already out of date. The question isn’t whether to act — it’s whether the planning cycle is fast enough to match the pace of what it’s planning for.