Tech Digest – June 25, 2026
AI Governance & Trust
Anthropic Accuses Alibaba of Harvesting 28.8 Million Claude Exchanges
Anthropic sent a letter to the US Senate Committee on Banking accusing Alibaba and its Qwen AI lab of the largest known distillation attack on its models. Between late April and early June, approximately 25,000 fraudulent accounts generated 28.8 million exchanges with Claude, systematically targeting capabilities in agentic reasoning, software engineering, and long-horizon tasks. Distillation — using a stronger model’s outputs to train a weaker one — is a known technique, but the scale and coordination described here is unprecedented.
Note: The commercial viability of frontier AI depends on the assumption that training costs can be recouped through usage. If a competitor can replicate key capabilities by harvesting outputs at industrial scale, the economics of model development change. For institutions evaluating AI vendors, the question becomes which provider can both build and defend its technology.
AI Chatbots Tested for Political Bias — ChatGPT Leans Left, Grok Surprises
A Washington Post investigation tested leading AI chatbots for political bias, finding ChatGPT mostly left-leaning in its responses, Gemini approximately neutral, and Grok — despite branding as a politically unfiltered alternative — citing left-leaning sources more than right-leaning ones. The findings arrive as frontier labs increasingly hire philosophers, not just engineers, to write the constitutions governing model behavior, according to The Economist.
Note: Every chatbot deployed for citizen services carries a political lean — whether by design, training data, or constitutional architecture. It’s the kind of subtle risk most procurement criteria are not built to surface.
Sources: Washington Post, The Economist
Cloudflare and Major Browsers Build Personhood Tokens — Bots Now Outnumber Humans Online
Cloudflare, Google Chrome, Mozilla Firefox, and Microsoft Edge announced PACT — Private Access Control Tokens — a protocol that lets trusted sites issue anonymous tokens verifying a real person is behind the traffic. The system replaces CAPTCHAs without revealing identity or enabling cross-site tracking. The urgency is concrete: Cloudflare reports bots now generate 58% of all HTTP traffic worldwide, a threshold CEO Matthew Prince had not expected until 2027.
Note: When more than half of web traffic is non-human, every digital service needs a position on how it authenticates visitors. PACT is being built by the browsers, not by governments — the standard will be set before most institutions know it exists.
Sources: Cloudflare, The Register
The Semiconductor Arms Race
Japan Unveils $2.3 Trillion Chip Plan as EU Joins US-Led Semiconductor Pact
Japanese Prime Minister Takaichi unveiled a 14-year investment plan worth ¥370 trillion ($2.3 trillion) through 2041, with ¥101.6 trillion earmarked for AI and semiconductors — aiming to quintuple Japan’s chip output. The announcement drove the Nikkei past 72,000.
Separately, the EU formally joined Pax Silica, a US-led initiative coordinating AI chip supply chains and export controls targeting China, committing to purchase at least $40 billion in US chips. The pact is not legally binding but politically significant. France objected, framing it as an attempt to undermine European tech sovereignty. Germany, Italy, and the Netherlands backed the agreement.
Note: The chip supply chain is splitting along alliance lines. For EU institutions running digital transformation programmes, this isn’t abstract geopolitics — it shapes which chips are available, at what price, and on what timeline. The tension between Pax Silica’s alignment with Washington and the EU’s own sovereignty ambitions is a contradiction that will show up in procurement.
Sources: Bloomberg, The Japan Times, Euractiv
Micron Posts Record $41.5 Billion Quarter and 84.9% Margins — Memory Is the Bottleneck
Micron reported quarterly revenue of $41.5 billion — up 346% year-over-year — with record gross margins of 84.9%. The company has secured $100 billion in minimum contracted customer agreements and guided Q4 revenue to $50 billion. Micron crossed a $1 trillion market capitalisation in May. On the earnings call, management noted that a single humanoid robot carries ten times more memory than the average autonomous vehicle, signalling what they called a “sustained multi-decade memory demand cycle.”
Note: When a memory chipmaker’s margins approach luxury-goods territory, it’s a supply-demand signal, not a product story. Every digital infrastructure project runs on memory — and the gap between what institutions budgeted and what memory will cost is widening.
Sources: CNBC, Investing.com
Qualcomm and OpenAI Both Move to Break Nvidia’s Inference Grip
Qualcomm announced a $3.92 billion acquisition of Modular, whose MAX inference engine and Mojo programming language let developers write AI code once and run it across any processor — a direct challenge to Nvidia’s CUDA lock-in. Alongside the deal, Qualcomm unveiled the Dragonfly C1000, a 250+ core data centre CPU with Meta as its first customer, marking a return to the server market after years of absence.
Separately, OpenAI and Broadcom unveiled Jalapeño, OpenAI’s first custom inference chip, designed and taped out in nine months — possibly the fastest ASIC development cycle in high-performance semiconductors. Early testing shows substantially better performance per watt than current hardware, with deployment planned for late 2026.
Note: CUDA’s dominance has been the single largest vendor lock-in in AI infrastructure. Two independent attacks on it in one day — one from a chip company buying a software stack, one from an AI lab designing its own silicon — suggest the next round of institutional AI procurement may carry genuine hardware optionality for the first time.
Sources: Reuters, CNBC, OpenAI
Agents & Discovery
Google Bakes Computer Use Into Gemini 3.5 Flash — Agents Now Operate Any Screen
Google released computer use as a built-in capability of Gemini 3.5 Flash, enabling developers to build agents that see, reason, and act across browser, mobile, and desktop environments. The model scores 78.4 on OSWorld-Verified, matching Anthropic’s Sonnet 4.6 while significantly outperforming the previous Gemini Flash generation. Google also shipped enterprise safeguards: mandatory user confirmation for irreversible actions and automatic shutdown if prompt injection is detected.
Note: Computer use in a lightweight, fast model means screen-operating agents aren’t a premium experiment — they’re approaching commodity infrastructure. The enterprise safeguards are telling: Google expects these to run in production environments, not demos.
Sources: Google AI Blog
Ed Witten Credits Claude Opus 4.8 With Generalising a Physics Calculation
In a paper on corrections to entanglement wedge reconstruction, Edward Witten — widely regarded as the most influential living theoretical physicist — credited Anthropic’s Claude Opus 4.8 with “generalising this calculation beyond some simple special cases.” It is believed to be the first time Witten has credited an AI model as a contributor to his research. The work extends the CCKLP construction in AdS/CFT holography, demonstrating that reconstruction errors are exponentially small.
Note: When the most influential living theoretical physicist credits a language model, the question isn’t whether AI can do research. It’s how quickly funding bodies, journals, and university governance catch up.
Sources: arXiv
Energy for Compute
Home Batteries and EVs Become 16 Gigawatts of Data Centre Power
Sunrun, Tesla, and Renew Home announced a partnership to aggregate home batteries and electric vehicles into 16 gigawatts of flexible power capacity, with data centres and large-load facilities as the primary customers. The model treats distributed residential energy storage as a virtual power plant, dispatching stored energy to high-demand commercial users during peak loads.
Note: The data centre energy problem is producing novel grid architectures. Using thousands of home batteries as dispatchable capacity turns residential infrastructure into industrial supply — a model that merges the EU’s twin transition into a single wire.
Sources: GlobeNewsWire
Humanoid Robots Leave the Lab
Agility Robotics Goes Public at $2.5 Billion as 1X Opens Its Humanoid Platform
Agility Robotics is going public through a SPAC merger with Churchill Capital XI at a $2.5 billion valuation — the largest capital raise in humanoid robotics to date. The company expects over $620 million in proceeds, backed by Foxconn, and has deployed nearly 100 of its Digit robots in commercial logistics settings.
On the same day, 1X CEO Bernt Bornich announced the company will accelerate delivery of its NEO humanoid to developers, calling a Western humanoid development platform “too important for the future of humanity to gatekeep.” The open-platform approach positions NEO as shared infrastructure rather than a proprietary product.
Note: A public market valuation and an open platform in the same week suggest humanoid robotics is crossing from prototype to deployment phase. The EU’s manufacturing and logistics sectors face the same labour shortages driving US adoption — when the hardware is ready, the demand will already be waiting.
Sources: WSJ, TechCrunch, Bernt Bornich (1X CEO)
The infrastructure of the AI era is being poured simultaneously — in chip fabs, browser protocols, home batteries, and stock exchange listings. Japan commits $2.3 trillion over 14 years. The EU signs a geopolitical chip pact while arguing internally about sovereignty. Cloudflare builds plumbing to sort humans from bots because the bots already outnumber us online. When Ed Witten credits a language model and a humanoid robot company files for a public listing, the question for institutions isn’t whether this is happening. It’s whether their plans reflect the speed.
